diff --git a/.gitattributes b/.gitattributes index a6344aac8c09253b3b630fb776ae94478aa0275b..56e30a68897ed412d35a6bd63af5c62ac615175a 100644 --- a/.gitattributes +++ b/.gitattributes @@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text *.zip filter=lfs diff=lfs merge=lfs -text *.zst filter=lfs diff=lfs merge=lfs -text *tfevents* filter=lfs diff=lfs merge=lfs -text +assets/13.gif filter=lfs diff=lfs merge=lfs -text diff --git a/LICENSE b/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..3ec467e1d7fb12cd1a0aab9711ecb5c127fd6349 --- /dev/null +++ b/LICENSE @@ -0,0 +1,201 @@ + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright Tencent + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. \ No newline at end of file diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..0698f3ae9db5aceb3ffe77e48c6779f1a75d378e --- /dev/null +++ b/README.md @@ -0,0 +1,87 @@ +## ___***ToonCrafter_with_SketchGuidance***___ +This repository is an implementation that recreates the SketchGuidance feature of "ToonCrafter". + +- https://github.com/ToonCrafter/ToonCrafter +- https://arxiv.org/pdf/2405.17933 + +https://github.com/user-attachments/assets/f72f287d-f848-4982-8f91-43c49d037007 + +# ToonCrafter with Sketch Guidance + +## Overview +This repository contains the ToonCrafter model with additional sketch guidance functionality. The model is hosted on GitHub and can be accessed via the following link: + +- **Repository:** [https://github.com/svjack/ToonCrafter_with_SketchGuidance](https://github.com/svjack/ToonCrafter_with_SketchGuidance) + +## Installation + +### Clone the Repository +To clone the repository and set up the environment, follow these steps: + +```bash +git clone https://github.com/svjack/ToonCrafter_with_SketchGuidance && cd ToonCrafter_with_SketchGuidance +``` + +### List Control Models +After cloning the repository, you can list the control models available: + +```bash +ls control_models +``` + +### Download Sketch Encoder +Download the `sketch_encoder.ckpt` file from the Hugging Face repository: + +```bash +wget https://huggingface.co/Doubiiu/ToonCrafter/resolve/main/sketch_encoder.ckpt +``` + +### Copy Sketch Encoder to Control Models +Move the downloaded `sketch_encoder.ckpt` file to the `control_models` directory: + +```bash +cp sketch_encoder.ckpt control_models +``` + +## Usage + +### Run the Gradio App +Once the repository is cloned and the necessary files are in place, you can run the Gradio app using the following command: + +```bash +python gradio_app.py +``` + +This will start the Gradio interface, allowing you to interact with the ToonCrafter model with sketch guidance. + +## 🧰 Models + +|Model|Resolution|GPU Mem. & Inference Time (A100, ddim 50steps)|Checkpoint| +|:---------|:---------|:--------|:--------| +|ToonCrafter_512|320x512| TBD (`perframe_ae=True`)|[Hugging Face](https://huggingface.co/Doubiiu/ToonCrafter/blob/main/model.ckpt)| +|SketchEncoder|TBD| TBD |[Hugging Face](https://huggingface.co/Doubiiu/ToonCrafter/blob/main/sketch_encoder.ckpt)| + + +Currently, ToonCrafter can support generating videos of up to 16 frames with a resolution of 512x320. The inference time can be reduced by using fewer DDIM steps. + + + +## βš™οΈ Setup + +### Install Environment via Anaconda (Recommended) +```bash +conda create -n tooncrafter python=3.8.5 +conda activate tooncrafter +pip install -r requirements.txt +``` + + +## πŸ’« Inference + +### 1. Local Gradio demo +1. Download pretrained ToonCrafter_512 and put the model.ckpt in checkpoints/tooncrafter_512_interp_v1/model.ckpt. +2. Download pretrained SketchEncoder and put the model.ckpt in control_models/sketch_encoder.ckpt. + +```bash + python gradio_app.py +``` diff --git a/assets/00.gif b/assets/00.gif new file mode 100644 index 0000000000000000000000000000000000000000..569c93b7020cf54297481a59775658dca88919a1 Binary files /dev/null and b/assets/00.gif differ diff --git a/assets/01.gif b/assets/01.gif new file mode 100644 index 0000000000000000000000000000000000000000..d739d4716b7cc9b5331c665b3dcac7e183d0d53f Binary files /dev/null and b/assets/01.gif differ diff --git a/assets/02.gif b/assets/02.gif new file mode 100644 index 0000000000000000000000000000000000000000..8f3821c128cd3330ed0ddfab187fbd215a8221c0 Binary files /dev/null and b/assets/02.gif differ diff --git a/assets/03.gif b/assets/03.gif new file mode 100644 index 0000000000000000000000000000000000000000..103daaf59a89d611468dbaae4c0543b556b07b56 Binary files /dev/null and b/assets/03.gif differ diff --git a/assets/04.gif b/assets/04.gif new file mode 100644 index 0000000000000000000000000000000000000000..65b0ed24e2d7455423d9634934a4180edab5e938 Binary files /dev/null and b/assets/04.gif differ diff --git a/assets/05.gif b/assets/05.gif new file mode 100644 index 0000000000000000000000000000000000000000..e9308b2be69aea8a851dd7a430ffc13a33b448d6 Binary files /dev/null and b/assets/05.gif differ diff --git a/assets/06.gif b/assets/06.gif new file mode 100644 index 0000000000000000000000000000000000000000..2d548b6335bb67b5918e869b2c4f91cb8994bc8f Binary files /dev/null and b/assets/06.gif differ diff --git a/assets/07.gif b/assets/07.gif new file mode 100644 index 0000000000000000000000000000000000000000..6d64d1be55419e56c417199c2072c2d215f10f1c Binary files /dev/null and b/assets/07.gif differ diff --git a/assets/08.gif b/assets/08.gif new file mode 100644 index 0000000000000000000000000000000000000000..e6ea4eb0a558493ec5ad87107ebd9a886268d8dd Binary files /dev/null and b/assets/08.gif differ diff --git a/assets/09.gif b/assets/09.gif new file mode 100644 index 0000000000000000000000000000000000000000..859bd4313902de439ae8120325394ab095e9b2da Binary files /dev/null and b/assets/09.gif differ diff --git a/assets/10.gif b/assets/10.gif new file mode 100644 index 0000000000000000000000000000000000000000..47d9f0aebb69ac0e4b282fded5072d34705c7f6e Binary files /dev/null and b/assets/10.gif differ diff --git a/assets/11.gif b/assets/11.gif new file mode 100644 index 0000000000000000000000000000000000000000..2b4fec5f03f83b245cec772020b0841dfaff310a Binary files /dev/null and b/assets/11.gif differ diff --git a/assets/12.gif b/assets/12.gif new file mode 100644 index 0000000000000000000000000000000000000000..33ad8663174e0a7be8cd15e53e115385587fe3ef Binary files /dev/null and b/assets/12.gif differ diff --git a/assets/13.gif b/assets/13.gif new file mode 100644 index 0000000000000000000000000000000000000000..aa2543371f82545863136a1eaf63e7de512dd134 --- /dev/null +++ b/assets/13.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:179af7d265d8790c0ca31a5898f870961b0a738b02d9fd0c991a3a75651cbb56 +size 1030647 diff --git a/assets/72105_388.mp4_00-00.png b/assets/72105_388.mp4_00-00.png new file mode 100644 index 0000000000000000000000000000000000000000..8c9f797d81531d00138ba594995a346174904cea Binary files /dev/null and b/assets/72105_388.mp4_00-00.png differ diff --git a/assets/72105_388.mp4_00-01.png b/assets/72105_388.mp4_00-01.png new file mode 100644 index 0000000000000000000000000000000000000000..e094f5f7ddc3d74ba5f161cefc4b1fb0359bc9c0 Binary files /dev/null and b/assets/72105_388.mp4_00-01.png differ diff --git a/assets/72109_125.mp4_00-00.png b/assets/72109_125.mp4_00-00.png new file mode 100644 index 0000000000000000000000000000000000000000..8e3458152124e93c0f79b76727205a18aa36803e Binary files /dev/null and b/assets/72109_125.mp4_00-00.png differ diff --git a/assets/72109_125.mp4_00-01.png b/assets/72109_125.mp4_00-01.png new file mode 100644 index 0000000000000000000000000000000000000000..e8abd1c0ff0fda455fc66dd8f64005a91ec879d3 Binary files /dev/null and b/assets/72109_125.mp4_00-01.png differ diff --git a/assets/72110_255.mp4_00-00.png b/assets/72110_255.mp4_00-00.png new file mode 100644 index 0000000000000000000000000000000000000000..7350bf4f9cd080d6b83eb7263cc681208262a8d4 Binary files /dev/null and b/assets/72110_255.mp4_00-00.png differ diff --git a/assets/72110_255.mp4_00-01.png b/assets/72110_255.mp4_00-01.png new file mode 100644 index 0000000000000000000000000000000000000000..83b4f318eb3138567a844c0cc924bbbcaf00f451 Binary files /dev/null and b/assets/72110_255.mp4_00-01.png differ diff --git a/assets/74302_1349_frame1.png b/assets/74302_1349_frame1.png new file mode 100644 index 0000000000000000000000000000000000000000..6b1d1ecdcf8ee5693a1076797e812eeefbff1084 Binary files /dev/null and b/assets/74302_1349_frame1.png differ diff --git a/assets/74302_1349_frame3.png b/assets/74302_1349_frame3.png new file mode 100644 index 0000000000000000000000000000000000000000..64837a531f7c8fe5c08c1300956d2385971a32a7 Binary files /dev/null and b/assets/74302_1349_frame3.png differ diff --git a/assets/Japan_v2_1_070321_s3_frame1.png b/assets/Japan_v2_1_070321_s3_frame1.png new file mode 100644 index 0000000000000000000000000000000000000000..b875686659b151a87eff55f84a9e4502bedeab1d Binary files /dev/null and b/assets/Japan_v2_1_070321_s3_frame1.png differ diff --git a/assets/Japan_v2_1_070321_s3_frame3.png b/assets/Japan_v2_1_070321_s3_frame3.png new file mode 100644 index 0000000000000000000000000000000000000000..7564a86c2ee8b48613452bfd0cbff79a4d39bc04 Binary files /dev/null and b/assets/Japan_v2_1_070321_s3_frame3.png differ diff --git a/assets/Japan_v2_2_062266_s2_frame1.png b/assets/Japan_v2_2_062266_s2_frame1.png new file mode 100644 index 0000000000000000000000000000000000000000..4cc9b95450dda7e9ccf5992fd9f4c77f88fb476b Binary files /dev/null and b/assets/Japan_v2_2_062266_s2_frame1.png differ diff --git a/assets/Japan_v2_2_062266_s2_frame3.png b/assets/Japan_v2_2_062266_s2_frame3.png new file mode 100644 index 0000000000000000000000000000000000000000..e78eeccb4008280eecbdf9b633197d7fe757c2ad Binary files /dev/null and b/assets/Japan_v2_2_062266_s2_frame3.png differ diff --git a/assets/frame0001_05.png b/assets/frame0001_05.png new file mode 100644 index 0000000000000000000000000000000000000000..4bb2b39f18e08181f5650bd71267f403d18c40fe Binary files /dev/null and b/assets/frame0001_05.png differ diff --git a/assets/frame0001_09.png b/assets/frame0001_09.png new file mode 100644 index 0000000000000000000000000000000000000000..5a25c5c05a3ba6fae88d79c7b1655195e373ef0c Binary files /dev/null and b/assets/frame0001_09.png differ diff --git a/assets/frame0001_10.png b/assets/frame0001_10.png new file mode 100644 index 0000000000000000000000000000000000000000..e77c14492fae5dc63889c3c2181718851c937876 Binary files /dev/null and b/assets/frame0001_10.png differ diff --git a/assets/frame0001_11.png b/assets/frame0001_11.png new file mode 100644 index 0000000000000000000000000000000000000000..06fd5a9c0767d8dc938af1520ddce1bb3d4a8f74 Binary files /dev/null and b/assets/frame0001_11.png differ diff --git a/assets/frame0016_10.png b/assets/frame0016_10.png new file mode 100644 index 0000000000000000000000000000000000000000..ab4fc0e66712f87ff1039f4cc6c8eebd67e8e353 Binary files /dev/null and b/assets/frame0016_10.png differ diff --git a/assets/frame0016_11.png b/assets/frame0016_11.png new file mode 100644 index 0000000000000000000000000000000000000000..1275a8f0f50dd422e58f95922f17894e0011e6a8 Binary files /dev/null and b/assets/frame0016_11.png differ diff --git a/assets/sketch_sample/frame_1.png b/assets/sketch_sample/frame_1.png new file mode 100644 index 0000000000000000000000000000000000000000..e8779f40b231cec340b974f13272791bf8af8e11 Binary files /dev/null and b/assets/sketch_sample/frame_1.png differ diff --git a/assets/sketch_sample/frame_2.png b/assets/sketch_sample/frame_2.png new file mode 100644 index 0000000000000000000000000000000000000000..fd69ae305472d5469293f3a4e464478a93c6d6b8 Binary files /dev/null and b/assets/sketch_sample/frame_2.png differ diff --git a/assets/sketch_sample/sample.mov b/assets/sketch_sample/sample.mov new file mode 100644 index 0000000000000000000000000000000000000000..9d0f2ea8ab0c7850751738c05bfe713488614fdd Binary files /dev/null and b/assets/sketch_sample/sample.mov differ diff --git a/cldm/cldm.py b/cldm/cldm.py new file mode 100644 index 0000000000000000000000000000000000000000..06839d1869f142bea19f2512a684c8cf47022640 --- /dev/null +++ b/cldm/cldm.py @@ -0,0 +1,478 @@ +import einops +import torch +import torch as th +import torch.nn as nn + +from ldm.modules.diffusionmodules.util import ( + conv_nd, + linear, + zero_module, + timestep_embedding, +) + +from einops import rearrange, repeat +from torchvision.utils import make_grid +from ldm.modules.attention import SpatialTransformer +from ldm.modules.diffusionmodules.openaimodel import TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock +from lvdm.modules.networks.openaimodel3d import UNetModel +from ldm.models.diffusion.ddpm import LatentDiffusion +from ldm.util import log_txt_as_img, exists, instantiate_from_config +from ldm.models.diffusion.ddim import DDIMSampler + + +class ControlledUnetModel(UNetModel): + def forward(self, x, timesteps, context=None, features_adapter=None, fs=None, control = None, **kwargs): + b,_,t,_,_ = x.shape + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).type(x.dtype) + emb = self.time_embed(t_emb) + ## repeat t times for context [(b t) 77 768] & time embedding + ## check if we use per-frame image conditioning + _, l_context, _ = context.shape + if l_context == 77 + t*16: ## !!! HARD CODE here + context_text, context_img = context[:,:77,:], context[:,77:,:] + context_text = context_text.repeat_interleave(repeats=t, dim=0) + context_img = rearrange(context_img, 'b (t l) c -> (b t) l c', t=t) + context = torch.cat([context_text, context_img], dim=1) + else: + context = context.repeat_interleave(repeats=t, dim=0) + emb = emb.repeat_interleave(repeats=t, dim=0) + + ## always in shape (b t) c h w, except for temporal layer + x = rearrange(x, 'b c t h w -> (b t) c h w') + + ## combine emb + if self.fs_condition: + if fs is None: + fs = torch.tensor( + [self.default_fs] * b, dtype=torch.long, device=x.device) + fs_emb = timestep_embedding(fs, self.model_channels, repeat_only=False).type(x.dtype) + + fs_embed = self.fps_embedding(fs_emb) + fs_embed = fs_embed.repeat_interleave(repeats=t, dim=0) + emb = emb + fs_embed + + h = x.type(self.dtype) + adapter_idx = 0 + hs = [] + with torch.no_grad(): + for id, module in enumerate(self.input_blocks): + h = module(h, emb, context=context, batch_size=b) + if id ==0 and self.addition_attention: + h = self.init_attn(h, emb, context=context, batch_size=b) + ## plug-in adapter features + if ((id+1)%3 == 0) and features_adapter is not None: + h = h + features_adapter[adapter_idx] + adapter_idx += 1 + hs.append(h) + if features_adapter is not None: + assert len(features_adapter)==adapter_idx, 'Wrong features_adapter' + + h = self.middle_block(h, emb, context=context, batch_size=b) + + if control is not None: + h += control.pop() + + for module in self.output_blocks: + if control is None: + h = torch.cat([h, hs.pop()], dim=1) + else: + h = torch.cat([h, hs.pop() + control.pop()], dim=1) + h = module(h, emb, context=context, batch_size=b) + + h = h.type(x.dtype) + y = self.out(h) + + # reshape back to (b c t h w) + y = rearrange(y, '(b t) c h w -> b c t h w', b=b) + return y + + +class ControlNet(nn.Module): + def __init__( + self, + image_size, + in_channels, + model_channels, + hint_channels, + num_res_blocks, + attention_resolutions, + dropout=0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + use_checkpoint=False, + use_fp16=False, + num_heads=-1, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False, + resblock_updown=False, + use_new_attention_order=False, + use_spatial_transformer=False, # custom transformer support + transformer_depth=1, # custom transformer support + context_dim=None, # custom transformer support + n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model + legacy=True, + disable_self_attentions=None, + num_attention_blocks=None, + disable_middle_self_attn=False, + use_linear_in_transformer=False, + ): + super().__init__() + if use_spatial_transformer: + assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' + + if context_dim is not None: + assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' + from omegaconf.listconfig import ListConfig + if type(context_dim) == ListConfig: + context_dim = list(context_dim) + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' + + if num_head_channels == -1: + assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' + + self.dims = dims + self.image_size = image_size + self.in_channels = in_channels + self.model_channels = model_channels + if isinstance(num_res_blocks, int): + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + else: + if len(num_res_blocks) != len(channel_mult): + raise ValueError("provide num_res_blocks either as an int (globally constant) or " + "as a list/tuple (per-level) with the same length as channel_mult") + self.num_res_blocks = num_res_blocks + if disable_self_attentions is not None: + # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not + assert len(disable_self_attentions) == len(channel_mult) + if num_attention_blocks is not None: + assert len(num_attention_blocks) == len(self.num_res_blocks) + assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) + print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " + f"This option has LESS priority than attention_resolutions {attention_resolutions}, " + f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " + f"attention will still not be set.") + + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.use_checkpoint = use_checkpoint + self.dtype = th.float16 if use_fp16 else th.float32 + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.predict_codebook_ids = n_embed is not None + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels)]) + + self.input_hint_block = TimestepEmbedSequential( + conv_nd(dims, hint_channels, 16, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 16, 16, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 16, 32, 3, padding=1, stride=2), + nn.SiLU(), + conv_nd(dims, 32, 32, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 32, 96, 3, padding=1, stride=2), + nn.SiLU(), + conv_nd(dims, 96, 96, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 96, 256, 3, padding=1, stride=2), + nn.SiLU(), + zero_module(conv_nd(dims, 256, model_channels, 3, padding=1)) + ) + + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + # num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self.zero_convs.append(self.make_zero_conv(ch)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + self.zero_convs.append(self.make_zero_conv(ch)) + ds *= 2 + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + # num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self.middle_block_out = self.make_zero_conv(ch) + self._feature_size += ch + + def make_zero_conv(self, channels): + return TimestepEmbedSequential(zero_module(conv_nd(self.dims, channels, channels, 1, padding=0))) + + def forward(self, x, hint, timesteps, context, **kwargs): + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + guided_hint = self.input_hint_block(hint, emb, context) + + outs = [] + + h = x.type(self.dtype) + + for module, zero_conv in zip(self.input_blocks, self.zero_convs): + if guided_hint is not None: + h = module(h, emb, context) + h += guided_hint + guided_hint = None + else: + h = module(h, emb, context) + outs.append(zero_conv(h, emb, context, True)) + + h = self.middle_block(h, emb, context) + outs.append(self.middle_block_out(h, emb, context)) + + return outs + + +class ControlLDM(LatentDiffusion): + + def __init__(self, control_stage_config, control_key, only_mid_control, *args, **kwargs): + super().__init__(*args, **kwargs) + self.control_model = instantiate_from_config(control_stage_config) + self.control_key = control_key + self.only_mid_control = only_mid_control + self.control_scales = [1.0] * 13 + + @torch.no_grad() + def get_input(self, batch, k, bs=None, *args, **kwargs): + x, c = super().get_input(batch, self.first_stage_key, *args, **kwargs) + control = batch[self.control_key] + if bs is not None: + control = control[:bs] + control = control.to(self.device) + control = einops.rearrange(control, 'b h w c -> b c h w') + control = control.to(memory_format=torch.contiguous_format).float() + return x, dict(c_crossattn=[c], c_concat=[control]) + + def apply_model(self, x_noisy, t, cond, *args, **kwargs): + assert isinstance(cond, dict) + diffusion_model = self.model.diffusion_model + + cond_txt = torch.cat(cond['c_crossattn'], 1) + + if cond['c_concat'] is None: + eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=None, only_mid_control=self.only_mid_control) + else: + control = self.control_model(x=x_noisy, hint=torch.cat(cond['c_concat'], 1), timesteps=t, context=cond_txt) + control = [c * scale for c, scale in zip(control, self.control_scales)] + eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=control, only_mid_control=self.only_mid_control) + + return eps + + @torch.no_grad() + def get_unconditional_conditioning(self, N): + return self.get_learned_conditioning([""] * N) + + @torch.no_grad() + def log_images(self, batch, N=4, n_row=2, sample=False, ddim_steps=50, ddim_eta=0.0, return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=False, unconditional_guidance_scale=9.0, unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + use_ddim = ddim_steps is not None + + log = dict() + z, c = self.get_input(batch, self.first_stage_key, bs=N) + c_cat, c = c["c_concat"][0][:N], c["c_crossattn"][0][:N] + N = min(z.shape[0], N) + n_row = min(z.shape[0], n_row) + log["reconstruction"] = self.decode_first_stage(z) + log["control"] = c_cat * 2.0 - 1.0 + log["conditioning"] = log_txt_as_img((512, 512), batch[self.cond_stage_key], size=16) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_cross = self.get_unconditional_conditioning(N) + uc_cat = c_cat # torch.zeros_like(c_cat) + uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]} + samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc_full, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + return log + + @torch.no_grad() + def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs): + ddim_sampler = DDIMSampler(self) + b, c, h, w = cond["c_concat"][0].shape + shape = (self.channels, h // 8, w // 8) + samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, shape, cond, verbose=False, **kwargs) + return samples, intermediates + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.control_model.parameters()) + if not self.sd_locked: + params += list(self.model.diffusion_model.output_blocks.parameters()) + params += list(self.model.diffusion_model.out.parameters()) + opt = torch.optim.AdamW(params, lr=lr) + return opt + + def low_vram_shift(self, is_diffusing): + if is_diffusing: + self.model = self.model.cuda() + self.control_model = self.control_model.cuda() + self.first_stage_model = self.first_stage_model.cpu() + self.cond_stage_model = self.cond_stage_model.cpu() + else: + self.model = self.model.cpu() + self.control_model = self.control_model.cpu() + self.first_stage_model = self.first_stage_model.cuda() + self.cond_stage_model = self.cond_stage_model.cuda() diff --git a/cldm/ddim_hacked.py b/cldm/ddim_hacked.py new file mode 100644 index 0000000000000000000000000000000000000000..25b1bc947272ad14d7f7e5e4d1809005253b63d0 --- /dev/null +++ b/cldm/ddim_hacked.py @@ -0,0 +1,317 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor + + +class DDIMSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + dynamic_threshold=None, + ucg_schedule=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): ctmp = ctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + elif isinstance(conditioning, list): + for ctmp in conditioning: + if ctmp.shape[0] != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold, + ucg_schedule=ucg_schedule + ) + return samples, intermediates + + @torch.no_grad() + def ddim_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None, + ucg_schedule=None): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + if ucg_schedule is not None: + assert len(ucg_schedule) == len(time_range) + unconditional_guidance_scale = ucg_schedule[i] + + outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold) + img, pred_x0 = outs + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + dynamic_threshold=None): + b, *_, device = *x.shape, x.device + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + model_output = self.model.apply_model(x, t, c) + else: + model_t = self.model.apply_model(x, t, c) + model_uncond = self.model.apply_model(x, t, unconditional_conditioning) + model_output = model_uncond + unconditional_guidance_scale * (model_t - model_uncond) + + if self.model.parameterization == "v": + e_t = self.model.predict_eps_from_z_and_v(x, t, model_output) + else: + e_t = model_output + + if score_corrector is not None: + assert self.model.parameterization == "eps", 'not implemented' + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + if self.model.parameterization != "v": + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + else: + pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output) + + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + + if dynamic_threshold is not None: + raise NotImplementedError() + + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + @torch.no_grad() + def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None, + unconditional_guidance_scale=1.0, unconditional_conditioning=None, callback=None): + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + num_reference_steps = timesteps.shape[0] + + assert t_enc <= num_reference_steps + num_steps = t_enc + + if use_original_steps: + alphas_next = self.alphas_cumprod[:num_steps] + alphas = self.alphas_cumprod_prev[:num_steps] + else: + alphas_next = self.ddim_alphas[:num_steps] + alphas = torch.tensor(self.ddim_alphas_prev[:num_steps]) + + x_next = x0 + intermediates = [] + inter_steps = [] + for i in tqdm(range(num_steps), desc='Encoding Image'): + t = torch.full((x0.shape[0],), timesteps[i], device=self.model.device, dtype=torch.long) + if unconditional_guidance_scale == 1.: + noise_pred = self.model.apply_model(x_next, t, c) + else: + assert unconditional_conditioning is not None + e_t_uncond, noise_pred = torch.chunk( + self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)), + torch.cat((unconditional_conditioning, c))), 2) + noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond) + + xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next + weighted_noise_pred = alphas_next[i].sqrt() * ( + (1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred + x_next = xt_weighted + weighted_noise_pred + if return_intermediates and i % ( + num_steps // return_intermediates) == 0 and i < num_steps - 1: + intermediates.append(x_next) + inter_steps.append(i) + elif return_intermediates and i >= num_steps - 2: + intermediates.append(x_next) + inter_steps.append(i) + if callback: callback(i) + + out = {'x_encoded': x_next, 'intermediate_steps': inter_steps} + if return_intermediates: + out.update({'intermediates': intermediates}) + return x_next, out + + @torch.no_grad() + def stochastic_encode(self, x0, t, use_original_steps=False, noise=None): + # fast, but does not allow for exact reconstruction + # t serves as an index to gather the correct alphas + if use_original_steps: + sqrt_alphas_cumprod = self.sqrt_alphas_cumprod + sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod + else: + sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas) + sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas + + if noise is None: + noise = torch.randn_like(x0) + return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 + + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise) + + @torch.no_grad() + def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + use_original_steps=False, callback=None): + + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + timesteps = timesteps[:t_start] + + time_range = np.flip(timesteps) + total_steps = timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='Decoding image', total=total_steps) + x_dec = x_latent + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long) + x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning) + if callback: callback(i) + return x_dec diff --git a/cldm/hack.py b/cldm/hack.py new file mode 100644 index 0000000000000000000000000000000000000000..454361e9d036cd1a6a79122c2fd16b489e4767b1 --- /dev/null +++ b/cldm/hack.py @@ -0,0 +1,111 @@ +import torch +import einops + +import ldm.modules.encoders.modules +import ldm.modules.attention + +from transformers import logging +from ldm.modules.attention import default + + +def disable_verbosity(): + logging.set_verbosity_error() + print('logging improved.') + return + + +def enable_sliced_attention(): + ldm.modules.attention.CrossAttention.forward = _hacked_sliced_attentin_forward + print('Enabled sliced_attention.') + return + + +def hack_everything(clip_skip=0): + disable_verbosity() + ldm.modules.encoders.modules.FrozenCLIPEmbedder.forward = _hacked_clip_forward + ldm.modules.encoders.modules.FrozenCLIPEmbedder.clip_skip = clip_skip + print('Enabled clip hacks.') + return + + +# Written by Lvmin +def _hacked_clip_forward(self, text): + PAD = self.tokenizer.pad_token_id + EOS = self.tokenizer.eos_token_id + BOS = self.tokenizer.bos_token_id + + def tokenize(t): + return self.tokenizer(t, truncation=False, add_special_tokens=False)["input_ids"] + + def transformer_encode(t): + if self.clip_skip > 1: + rt = self.transformer(input_ids=t, output_hidden_states=True) + return self.transformer.text_model.final_layer_norm(rt.hidden_states[-self.clip_skip]) + else: + return self.transformer(input_ids=t, output_hidden_states=False).last_hidden_state + + def split(x): + return x[75 * 0: 75 * 1], x[75 * 1: 75 * 2], x[75 * 2: 75 * 3] + + def pad(x, p, i): + return x[:i] if len(x) >= i else x + [p] * (i - len(x)) + + raw_tokens_list = tokenize(text) + tokens_list = [] + + for raw_tokens in raw_tokens_list: + raw_tokens_123 = split(raw_tokens) + raw_tokens_123 = [[BOS] + raw_tokens_i + [EOS] for raw_tokens_i in raw_tokens_123] + raw_tokens_123 = [pad(raw_tokens_i, PAD, 77) for raw_tokens_i in raw_tokens_123] + tokens_list.append(raw_tokens_123) + + tokens_list = torch.IntTensor(tokens_list).to(self.device) + + feed = einops.rearrange(tokens_list, 'b f i -> (b f) i') + y = transformer_encode(feed) + z = einops.rearrange(y, '(b f) i c -> b (f i) c', f=3) + + return z + + +# Stolen from https://github.com/basujindal/stable-diffusion/blob/main/optimizedSD/splitAttention.py +def _hacked_sliced_attentin_forward(self, x, context=None, mask=None): + h = self.heads + + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + del context, x + + q, k, v = map(lambda t: einops.rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + limit = k.shape[0] + att_step = 1 + q_chunks = list(torch.tensor_split(q, limit // att_step, dim=0)) + k_chunks = list(torch.tensor_split(k, limit // att_step, dim=0)) + v_chunks = list(torch.tensor_split(v, limit // att_step, dim=0)) + + q_chunks.reverse() + k_chunks.reverse() + v_chunks.reverse() + sim = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device) + del k, q, v + for i in range(0, limit, att_step): + q_buffer = q_chunks.pop() + k_buffer = k_chunks.pop() + v_buffer = v_chunks.pop() + sim_buffer = torch.einsum('b i d, b j d -> b i j', q_buffer, k_buffer) * self.scale + + del k_buffer, q_buffer + # attention, what we cannot get enough of, by chunks + + sim_buffer = sim_buffer.softmax(dim=-1) + + sim_buffer = torch.einsum('b i j, b j d -> b i d', sim_buffer, v_buffer) + del v_buffer + sim[i:i + att_step, :, :] = sim_buffer + + del sim_buffer + sim = einops.rearrange(sim, '(b h) n d -> b n (h d)', h=h) + return self.to_out(sim) diff --git a/cldm/logger.py b/cldm/logger.py new file mode 100644 index 0000000000000000000000000000000000000000..6a8803846f2a8979f87f3cf9ea5b12869439e62f --- /dev/null +++ b/cldm/logger.py @@ -0,0 +1,76 @@ +import os + +import numpy as np +import torch +import torchvision +from PIL import Image +from pytorch_lightning.callbacks import Callback +from pytorch_lightning.utilities.distributed import rank_zero_only + + +class ImageLogger(Callback): + def __init__(self, batch_frequency=2000, max_images=4, clamp=True, increase_log_steps=True, + rescale=True, disabled=False, log_on_batch_idx=False, log_first_step=False, + log_images_kwargs=None): + super().__init__() + self.rescale = rescale + self.batch_freq = batch_frequency + self.max_images = max_images + if not increase_log_steps: + self.log_steps = [self.batch_freq] + self.clamp = clamp + self.disabled = disabled + self.log_on_batch_idx = log_on_batch_idx + self.log_images_kwargs = log_images_kwargs if log_images_kwargs else {} + self.log_first_step = log_first_step + + @rank_zero_only + def log_local(self, save_dir, split, images, global_step, current_epoch, batch_idx): + root = os.path.join(save_dir, "image_log", split) + for k in images: + grid = torchvision.utils.make_grid(images[k], nrow=4) + if self.rescale: + grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w + grid = grid.transpose(0, 1).transpose(1, 2).squeeze(-1) + grid = grid.numpy() + grid = (grid * 255).astype(np.uint8) + filename = "{}_gs-{:06}_e-{:06}_b-{:06}.png".format(k, global_step, current_epoch, batch_idx) + path = os.path.join(root, filename) + os.makedirs(os.path.split(path)[0], exist_ok=True) + Image.fromarray(grid).save(path) + + def log_img(self, pl_module, batch, batch_idx, split="train"): + check_idx = batch_idx # if self.log_on_batch_idx else pl_module.global_step + if (self.check_frequency(check_idx) and # batch_idx % self.batch_freq == 0 + hasattr(pl_module, "log_images") and + callable(pl_module.log_images) and + self.max_images > 0): + logger = type(pl_module.logger) + + is_train = pl_module.training + if is_train: + pl_module.eval() + + with torch.no_grad(): + images = pl_module.log_images(batch, split=split, **self.log_images_kwargs) + + for k in images: + N = min(images[k].shape[0], self.max_images) + images[k] = images[k][:N] + if isinstance(images[k], torch.Tensor): + images[k] = images[k].detach().cpu() + if self.clamp: + images[k] = torch.clamp(images[k], -1., 1.) + + self.log_local(pl_module.logger.save_dir, split, images, + pl_module.global_step, pl_module.current_epoch, batch_idx) + + if is_train: + pl_module.train() + + def check_frequency(self, check_idx): + return check_idx % self.batch_freq == 0 + + def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx): + if not self.disabled: + self.log_img(pl_module, batch, batch_idx, split="train") diff --git a/cldm/model.py b/cldm/model.py new file mode 100644 index 0000000000000000000000000000000000000000..fed3c31ac145b78907c7f771d1d8db6fb32d92ed --- /dev/null +++ b/cldm/model.py @@ -0,0 +1,28 @@ +import os +import torch + +from omegaconf import OmegaConf +from ldm.util import instantiate_from_config + + +def get_state_dict(d): + return d.get('state_dict', d) + + +def load_state_dict(ckpt_path, location='cpu'): + _, extension = os.path.splitext(ckpt_path) + if extension.lower() == ".safetensors": + import safetensors.torch + state_dict = safetensors.torch.load_file(ckpt_path, device=location) + else: + state_dict = get_state_dict(torch.load(ckpt_path, map_location=torch.device(location))) + state_dict = get_state_dict(state_dict) + print(f'Loaded state_dict from [{ckpt_path}]') + return state_dict + + +def create_model(config_path): + config = OmegaConf.load(config_path) + model = instantiate_from_config(config.model).cpu() + print(f'Loaded model config from [{config_path}]') + return model diff --git a/configs/cldm_v21.yaml b/configs/cldm_v21.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4f92d882d7642384318c74d9042165413187b072 --- /dev/null +++ b/configs/cldm_v21.yaml @@ -0,0 +1,17 @@ +control_stage_config: + target: cldm.cldm.ControlNet + params: + use_checkpoint: True + image_size: 32 # unused + in_channels: 4 + hint_channels: 1 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False diff --git a/configs/inference_512_v1.0.yaml b/configs/inference_512_v1.0.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1e526a2d58766fe9abe267512f4e67d98d399a97 --- /dev/null +++ b/configs/inference_512_v1.0.yaml @@ -0,0 +1,103 @@ +model: + target: lvdm.models.ddpm3d.LatentVisualDiffusion + params: + rescale_betas_zero_snr: True + parameterization: "v" + linear_start: 0.00085 + linear_end: 0.012 + num_timesteps_cond: 1 + timesteps: 1000 + first_stage_key: video + cond_stage_key: caption + cond_stage_trainable: False + conditioning_key: hybrid + image_size: [40, 64] + channels: 4 + scale_by_std: False + scale_factor: 0.18215 + use_ema: False + uncond_type: 'empty_seq' + use_dynamic_rescale: true + base_scale: 0.7 + fps_condition_type: 'fps' + perframe_ae: True + loop_video: true + unet_config: + target: cldm.cldm.ControlledUnetModel + params: + in_channels: 8 + out_channels: 4 + model_channels: 320 + attention_resolutions: + - 4 + - 2 + - 1 + num_res_blocks: 2 + channel_mult: + - 1 + - 2 + - 4 + - 4 + dropout: 0.1 + num_head_channels: 64 + transformer_depth: 1 + context_dim: 1024 + use_linear: true + use_checkpoint: True + temporal_conv: True + temporal_attention: True + temporal_selfatt_only: true + use_relative_position: false + use_causal_attention: False + temporal_length: 16 + addition_attention: true + image_cross_attention: true + default_fs: 24 + fs_condition: true + + first_stage_config: + target: lvdm.models.autoencoder.AutoencoderKL_Dualref + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder + params: + freeze: true + layer: "penultimate" + + img_cond_stage_config: + target: lvdm.modules.encoders.condition.FrozenOpenCLIPImageEmbedderV2 + params: + freeze: true + + image_proj_stage_config: + target: lvdm.modules.encoders.resampler.Resampler + params: + dim: 1024 + depth: 4 + dim_head: 64 + heads: 12 + num_queries: 16 + embedding_dim: 1280 + output_dim: 1024 + ff_mult: 4 + video_length: 16 diff --git a/configs/training_1024_v1.0/config.yaml b/configs/training_1024_v1.0/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2fa67398d15d66253c1db76261f6ea0a24c40a66 --- /dev/null +++ b/configs/training_1024_v1.0/config.yaml @@ -0,0 +1,166 @@ +model: + pretrained_checkpoint: checkpoints/dynamicrafter_1024_v1/model.ckpt + base_learning_rate: 1.0e-05 + scale_lr: False + target: lvdm.models.ddpm3d.LatentVisualDiffusion + params: + rescale_betas_zero_snr: True + parameterization: "v" + linear_start: 0.00085 + linear_end: 0.012 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: video + cond_stage_key: caption + cond_stage_trainable: False + image_proj_model_trainable: True + conditioning_key: hybrid + image_size: [72, 128] + channels: 4 + scale_by_std: False + scale_factor: 0.18215 + use_ema: False + uncond_prob: 0.05 + uncond_type: 'empty_seq' + rand_cond_frame: true + use_dynamic_rescale: true + base_scale: 0.3 + fps_condition_type: 'fps' + perframe_ae: True + + unet_config: + target: lvdm.modules.networks.openaimodel3d.UNetModel + params: + in_channels: 8 + out_channels: 4 + model_channels: 320 + attention_resolutions: + - 4 + - 2 + - 1 + num_res_blocks: 2 + channel_mult: + - 1 + - 2 + - 4 + - 4 + dropout: 0.1 + num_head_channels: 64 + transformer_depth: 1 + context_dim: 1024 + use_linear: true + use_checkpoint: True + temporal_conv: True + temporal_attention: True + temporal_selfatt_only: true + use_relative_position: false + use_causal_attention: False + temporal_length: 16 + addition_attention: true + image_cross_attention: true + default_fs: 10 + fs_condition: true + + first_stage_config: + target: lvdm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder + params: + freeze: true + layer: "penultimate" + + img_cond_stage_config: + target: lvdm.modules.encoders.condition.FrozenOpenCLIPImageEmbedderV2 + params: + freeze: true + + image_proj_stage_config: + target: lvdm.modules.encoders.resampler.Resampler + params: + dim: 1024 + depth: 4 + dim_head: 64 + heads: 12 + num_queries: 16 + embedding_dim: 1280 + output_dim: 1024 + ff_mult: 4 + video_length: 16 + +data: + target: utils_data.DataModuleFromConfig + params: + batch_size: 1 + num_workers: 12 + wrap: false + train: + target: lvdm.data.webvid.WebVid + params: + data_dir: + meta_path: <.csv FILE> + video_length: 16 + frame_stride: 6 + load_raw_resolution: true + resolution: [576, 1024] + spatial_transform: resize_center_crop + random_fs: true ## if true, we uniformly sample fs with max_fs=frame_stride (above) + +lightning: + precision: 16 + # strategy: deepspeed_stage_2 + trainer: + benchmark: True + accumulate_grad_batches: 2 + max_steps: 100000 + # logger + log_every_n_steps: 50 + # val + val_check_interval: 0.5 + gradient_clip_algorithm: 'norm' + gradient_clip_val: 0.5 + callbacks: + model_checkpoint: + target: pytorch_lightning.callbacks.ModelCheckpoint + params: + every_n_train_steps: 9000 #1000 + filename: "{epoch}-{step}" + save_weights_only: True + metrics_over_trainsteps_checkpoint: + target: pytorch_lightning.callbacks.ModelCheckpoint + params: + filename: '{epoch}-{step}' + save_weights_only: True + every_n_train_steps: 10000 #20000 # 3s/step*2w= + batch_logger: + target: callbacks.ImageLogger + params: + batch_frequency: 500 + to_local: False + max_images: 8 + log_images_kwargs: + ddim_steps: 50 + unconditional_guidance_scale: 7.5 + timestep_spacing: uniform_trailing + guidance_rescale: 0.7 \ No newline at end of file diff --git a/configs/training_1024_v1.0/run.sh b/configs/training_1024_v1.0/run.sh new file mode 100644 index 0000000000000000000000000000000000000000..c1f895d85a2307b57f13033c9b17d1f1c639f732 --- /dev/null +++ b/configs/training_1024_v1.0/run.sh @@ -0,0 +1,37 @@ +# NCCL configuration +# export NCCL_DEBUG=INFO +# export NCCL_IB_DISABLE=0 +# export NCCL_IB_GID_INDEX=3 +# export NCCL_NET_GDR_LEVEL=3 +# export NCCL_TOPO_FILE=/tmp/topo.txt + +# args +name="training_1024_v1.0" +config_file=configs/${name}/config.yaml + +# save root dir for logs, checkpoints, tensorboard record, etc. +save_root="" + +mkdir -p $save_root/$name + +## run +CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m torch.distributed.launch \ +--nproc_per_node=$HOST_GPU_NUM --nnodes=1 --master_addr=127.0.0.1 --master_port=12352 --node_rank=0 \ +./main/trainer.py \ +--base $config_file \ +--train \ +--name $name \ +--logdir $save_root \ +--devices $HOST_GPU_NUM \ +lightning.trainer.num_nodes=1 + +## debugging +# CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch \ +# --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12352 --node_rank=0 \ +# ./main/trainer.py \ +# --base $config_file \ +# --train \ +# --name $name \ +# --logdir $save_root \ +# --devices 4 \ +# lightning.trainer.num_nodes=1 \ No newline at end of file diff --git a/configs/training_512_v1.0/config.yaml b/configs/training_512_v1.0/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6cd69c2a39a9e7d9931645724a0496f25ef9ab29 --- /dev/null +++ b/configs/training_512_v1.0/config.yaml @@ -0,0 +1,166 @@ +model: + pretrained_checkpoint: checkpoints/dynamicrafter_512_v1/model.ckpt + base_learning_rate: 1.0e-05 + scale_lr: False + target: lvdm.models.ddpm3d.LatentVisualDiffusion + params: + rescale_betas_zero_snr: True + parameterization: "v" + linear_start: 0.00085 + linear_end: 0.012 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: video + cond_stage_key: caption + cond_stage_trainable: False + image_proj_model_trainable: True + conditioning_key: hybrid + image_size: [40, 64] + channels: 4 + scale_by_std: False + scale_factor: 0.18215 + use_ema: False + uncond_prob: 0.05 + uncond_type: 'empty_seq' + rand_cond_frame: true + use_dynamic_rescale: true + base_scale: 0.7 + fps_condition_type: 'fps' + perframe_ae: True + + unet_config: + target: lvdm.modules.networks.openaimodel3d.UNetModel + params: + in_channels: 8 + out_channels: 4 + model_channels: 320 + attention_resolutions: + - 4 + - 2 + - 1 + num_res_blocks: 2 + channel_mult: + - 1 + - 2 + - 4 + - 4 + dropout: 0.1 + num_head_channels: 64 + transformer_depth: 1 + context_dim: 1024 + use_linear: true + use_checkpoint: True + temporal_conv: True + temporal_attention: True + temporal_selfatt_only: true + use_relative_position: false + use_causal_attention: False + temporal_length: 16 + addition_attention: true + image_cross_attention: true + default_fs: 10 + fs_condition: true + + first_stage_config: + target: lvdm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder + params: + freeze: true + layer: "penultimate" + + img_cond_stage_config: + target: lvdm.modules.encoders.condition.FrozenOpenCLIPImageEmbedderV2 + params: + freeze: true + + image_proj_stage_config: + target: lvdm.modules.encoders.resampler.Resampler + params: + dim: 1024 + depth: 4 + dim_head: 64 + heads: 12 + num_queries: 16 + embedding_dim: 1280 + output_dim: 1024 + ff_mult: 4 + video_length: 16 + +data: + target: utils_data.DataModuleFromConfig + params: + batch_size: 2 + num_workers: 12 + wrap: false + train: + target: lvdm.data.webvid.WebVid + params: + data_dir: + meta_path: <.csv FILE> + video_length: 16 + frame_stride: 6 + load_raw_resolution: true + resolution: [320, 512] + spatial_transform: resize_center_crop + random_fs: true ## if true, we uniformly sample fs with max_fs=frame_stride (above) + +lightning: + precision: 16 + # strategy: deepspeed_stage_2 + trainer: + benchmark: True + accumulate_grad_batches: 2 + max_steps: 100000 + # logger + log_every_n_steps: 50 + # val + val_check_interval: 0.5 + gradient_clip_algorithm: 'norm' + gradient_clip_val: 0.5 + callbacks: + model_checkpoint: + target: pytorch_lightning.callbacks.ModelCheckpoint + params: + every_n_train_steps: 9000 #1000 + filename: "{epoch}-{step}" + save_weights_only: True + metrics_over_trainsteps_checkpoint: + target: pytorch_lightning.callbacks.ModelCheckpoint + params: + filename: '{epoch}-{step}' + save_weights_only: True + every_n_train_steps: 10000 #20000 # 3s/step*2w= + batch_logger: + target: callbacks.ImageLogger + params: + batch_frequency: 500 + to_local: False + max_images: 8 + log_images_kwargs: + ddim_steps: 50 + unconditional_guidance_scale: 7.5 + timestep_spacing: uniform_trailing + guidance_rescale: 0.7 \ No newline at end of file diff --git a/configs/training_512_v1.0/run.sh b/configs/training_512_v1.0/run.sh new file mode 100644 index 0000000000000000000000000000000000000000..46320b4a94fb3859e814be903e05ed3d9f1d5cd2 --- /dev/null +++ b/configs/training_512_v1.0/run.sh @@ -0,0 +1,37 @@ +# NCCL configuration +# export NCCL_DEBUG=INFO +# export NCCL_IB_DISABLE=0 +# export NCCL_IB_GID_INDEX=3 +# export NCCL_NET_GDR_LEVEL=3 +# export NCCL_TOPO_FILE=/tmp/topo.txt + +# args +name="training_512_v1.0" +config_file=configs/${name}/config.yaml + +# save root dir for logs, checkpoints, tensorboard record, etc. +save_root="" + +mkdir -p $save_root/$name + +## run +CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m torch.distributed.launch \ +--nproc_per_node=$HOST_GPU_NUM --nnodes=1 --master_addr=127.0.0.1 --master_port=12352 --node_rank=0 \ +./main/trainer.py \ +--base $config_file \ +--train \ +--name $name \ +--logdir $save_root \ +--devices $HOST_GPU_NUM \ +lightning.trainer.num_nodes=1 + +## debugging +# CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch \ +# --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12352 --node_rank=0 \ +# ./main/trainer.py \ +# --base $config_file \ +# --train \ +# --name $name \ +# --logdir $save_root \ +# --devices 4 \ +# lightning.trainer.num_nodes=1 \ No newline at end of file diff --git a/control_models/_put_your_control_models_.txt b/control_models/_put_your_control_models_.txt new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/gradio_app.py b/gradio_app.py new file mode 100644 index 0000000000000000000000000000000000000000..9b8d77851e669ce66c2eccc465c04068e020ad36 --- /dev/null +++ b/gradio_app.py @@ -0,0 +1,84 @@ +import os, argparse +import sys +import gradio as gr +from scripts.gradio.i2v_test_application import Image2Video +sys.path.insert(1, os.path.join(sys.path[0], 'lvdm')) + + +i2v_examples_interp_512 = [ + ['prompts/512_interp/74906_1462_frame1.png', 'walking man', 50, 7.5, 1.0, 10, 123, 'prompts/512_interp/74906_1462_frame3.png'], + ['prompts/512_interp/Japan_v2_2_062266_s2_frame1.png', 'an anime scene', 50, 7.5, 1.0, 10, 789, 'prompts/512_interp/Japan_v2_2_062266_s2_frame3.png'], + ['prompts/512_interp/Japan_v2_3_119235_s2_frame1.png', 'an anime scene', 50, 7.5, 1.0, 10, 123, 'prompts/512_interp/Japan_v2_3_119235_s2_frame3.png'], +] + + + + +def dynamicrafter_demo(result_dir='./tmp/', res=512): + if res == 1024: + resolution = '576_1024' + css = """#input_img {max-width: 1024px !important} #output_vid {max-width: 1024px; max-height:576px}""" + elif res == 512: + resolution = '320_512' + css = """#input_img {max-width: 512px !important} #output_vid {max-width: 512px; max-height: 320px} #input_img2 {max-width: 512px !important} #output_vid {max-width: 512px; max-height: 320px}""" + elif res == 256: + resolution = '256_256' + css = """#input_img {max-width: 256px !important} #output_vid {max-width: 256px; max-height: 256px}""" + else: + raise NotImplementedError(f"Unsupported resolution: {res}") + image2video = Image2Video(result_dir, resolution=resolution) + with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface: + + + + with gr.Tab(label='ToonCrafter_320x512'): + with gr.Column(): + with gr.Row(): + with gr.Column(): + with gr.Row(): + i2v_input_image = gr.Image(label="Input Image1",elem_id="input_img") + frame_guides = gr.Video(label="Input Guidance",elem_id="input_guidance", autoplay=True,show_share_button=True) + with gr.Row(): + i2v_input_text = gr.Text(label='Prompts') + with gr.Row(): + i2v_seed = gr.Slider(label='Random Seed', minimum=0, maximum=50000, step=1, value=123) + i2v_eta = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, label='ETA', value=1.0, elem_id="i2v_eta") + i2v_cfg_scale = gr.Slider(minimum=1.0, maximum=15.0, step=0.5, label='CFG Scale', value=7.5, elem_id="i2v_cfg_scale") + with gr.Row(): + i2v_steps = gr.Slider(minimum=1, maximum=60, step=1, elem_id="i2v_steps", label="Sampling steps", value=50) + i2v_motion = gr.Slider(minimum=5, maximum=30, step=1, elem_id="i2v_motion", label="FPS", value=10) + control_scale = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, elem_id="i2v_ctrl_scale", label="control_scale", value=0.6) + i2v_end_btn = gr.Button("Generate") + with gr.Column(): + with gr.Row(): + i2v_input_image2 = gr.Image(label="Input Image2",elem_id="input_img2") + with gr.Row(): + i2v_output_video = gr.Video(label="Generated Video",elem_id="output_vid",autoplay=True,show_share_button=True) + + gr.Examples(examples=i2v_examples_interp_512, + inputs=[i2v_input_image, i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed, i2v_input_image2, control_scale], + outputs=[i2v_output_video], + fn = image2video.get_image, + cache_examples=False, + ) + i2v_end_btn.click(inputs=[i2v_input_image, i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed, i2v_input_image2, frame_guides, control_scale], + outputs=[i2v_output_video], + fn = image2video.get_image + ) + + + return dynamicrafter_iface + +def get_parser(): + parser = argparse.ArgumentParser() + return parser + +if __name__ == "__main__": + parser = get_parser() + args = parser.parse_args() + + result_dir = os.path.join('./', 'results') + dynamicrafter_iface = dynamicrafter_demo(result_dir) + dynamicrafter_iface.queue(max_size=12) + dynamicrafter_iface.launch(max_threads=1, share=True) + #dynamicrafter_iface.launch(server_name='0.0.0.0', server_port=80, max_threads=1) \ No newline at end of file diff --git a/ldm/data/__init__.py b/ldm/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/data/util.py b/ldm/data/util.py new file mode 100644 index 0000000000000000000000000000000000000000..5b60ceb2349e3bd7900ff325740e2022d2903b1c --- /dev/null +++ b/ldm/data/util.py @@ -0,0 +1,24 @@ +import torch + +from ldm.modules.midas.api import load_midas_transform + + +class AddMiDaS(object): + def __init__(self, model_type): + super().__init__() + self.transform = load_midas_transform(model_type) + + def pt2np(self, x): + x = ((x + 1.0) * .5).detach().cpu().numpy() + return x + + def np2pt(self, x): + x = torch.from_numpy(x) * 2 - 1. + return x + + def __call__(self, sample): + # sample['jpg'] is tensor hwc in [-1, 1] at this point + x = self.pt2np(sample['jpg']) + x = self.transform({"image": x})["image"] + sample['midas_in'] = x + return sample \ No newline at end of file diff --git a/ldm/models/autoencoder.py b/ldm/models/autoencoder.py new file mode 100644 index 0000000000000000000000000000000000000000..d122549995ce2cd64092c81a58419ed4a15a02fd --- /dev/null +++ b/ldm/models/autoencoder.py @@ -0,0 +1,219 @@ +import torch +import pytorch_lightning as pl +import torch.nn.functional as F +from contextlib import contextmanager + +from ldm.modules.diffusionmodules.model import Encoder, Decoder +from ldm.modules.distributions.distributions import DiagonalGaussianDistribution + +from ldm.util import instantiate_from_config +from ldm.modules.ema import LitEma + + +class AutoencoderKL(pl.LightningModule): + def __init__(self, + ddconfig, + lossconfig, + embed_dim, + ckpt_path=None, + ignore_keys=[], + image_key="image", + colorize_nlabels=None, + monitor=None, + ema_decay=None, + learn_logvar=False + ): + super().__init__() + self.learn_logvar = learn_logvar + self.image_key = image_key + self.encoder = Encoder(**ddconfig) + self.decoder = Decoder(**ddconfig) + self.loss = instantiate_from_config(lossconfig) + assert ddconfig["double_z"] + self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1) + self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) + self.embed_dim = embed_dim + if colorize_nlabels is not None: + assert type(colorize_nlabels)==int + self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1)) + if monitor is not None: + self.monitor = monitor + + self.use_ema = ema_decay is not None + if self.use_ema: + self.ema_decay = ema_decay + assert 0. < ema_decay < 1. + self.model_ema = LitEma(self, decay=ema_decay) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) + + def init_from_ckpt(self, path, ignore_keys=list()): + sd = torch.load(path, map_location="cpu")["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + self.load_state_dict(sd, strict=False) + print(f"Restored from {path}") + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.parameters()) + self.model_ema.copy_to(self) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self) + + def encode(self, x): + h = self.encoder(x) + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + return posterior + + def decode(self, z): + z = self.post_quant_conv(z) + dec = self.decoder(z) + return dec + + def forward(self, input, sample_posterior=True): + posterior = self.encode(input) + if sample_posterior: + z = posterior.sample() + else: + z = posterior.mode() + dec = self.decode(z) + return dec, posterior + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float() + return x + + def training_step(self, batch, batch_idx, optimizer_idx): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + + if optimizer_idx == 0: + # train encoder+decoder+logvar + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return aeloss + + if optimizer_idx == 1: + # train the discriminator + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + + self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return discloss + + def validation_step(self, batch, batch_idx): + log_dict = self._validation_step(batch, batch_idx) + with self.ema_scope(): + log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema") + return log_dict + + def _validation_step(self, batch, batch_idx, postfix=""): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step, + last_layer=self.get_last_layer(), split="val"+postfix) + + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step, + last_layer=self.get_last_layer(), split="val"+postfix) + + self.log(f"val{postfix}/rec_loss", log_dict_ae[f"val{postfix}/rec_loss"]) + self.log_dict(log_dict_ae) + self.log_dict(log_dict_disc) + return self.log_dict + + def configure_optimizers(self): + lr = self.learning_rate + ae_params_list = list(self.encoder.parameters()) + list(self.decoder.parameters()) + list( + self.quant_conv.parameters()) + list(self.post_quant_conv.parameters()) + if self.learn_logvar: + print(f"{self.__class__.__name__}: Learning logvar") + ae_params_list.append(self.loss.logvar) + opt_ae = torch.optim.Adam(ae_params_list, + lr=lr, betas=(0.5, 0.9)) + opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(), + lr=lr, betas=(0.5, 0.9)) + return [opt_ae, opt_disc], [] + + def get_last_layer(self): + return self.decoder.conv_out.weight + + @torch.no_grad() + def log_images(self, batch, only_inputs=False, log_ema=False, **kwargs): + log = dict() + x = self.get_input(batch, self.image_key) + x = x.to(self.device) + if not only_inputs: + xrec, posterior = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec.shape[1] > 3 + x = self.to_rgb(x) + xrec = self.to_rgb(xrec) + log["samples"] = self.decode(torch.randn_like(posterior.sample())) + log["reconstructions"] = xrec + if log_ema or self.use_ema: + with self.ema_scope(): + xrec_ema, posterior_ema = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec_ema.shape[1] > 3 + xrec_ema = self.to_rgb(xrec_ema) + log["samples_ema"] = self.decode(torch.randn_like(posterior_ema.sample())) + log["reconstructions_ema"] = xrec_ema + log["inputs"] = x + return log + + def to_rgb(self, x): + assert self.image_key == "segmentation" + if not hasattr(self, "colorize"): + self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x)) + x = F.conv2d(x, weight=self.colorize) + x = 2.*(x-x.min())/(x.max()-x.min()) - 1. + return x + + +class IdentityFirstStage(torch.nn.Module): + def __init__(self, *args, vq_interface=False, **kwargs): + self.vq_interface = vq_interface + super().__init__() + + def encode(self, x, *args, **kwargs): + return x + + def decode(self, x, *args, **kwargs): + return x + + def quantize(self, x, *args, **kwargs): + if self.vq_interface: + return x, None, [None, None, None] + return x + + def forward(self, x, *args, **kwargs): + return x + diff --git a/ldm/models/diffusion/__init__.py b/ldm/models/diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/models/diffusion/ddim.py b/ldm/models/diffusion/ddim.py new file mode 100644 index 0000000000000000000000000000000000000000..27ead0ea914c64c747b64e690662899fb3801144 --- /dev/null +++ b/ldm/models/diffusion/ddim.py @@ -0,0 +1,336 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor + + +class DDIMSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + dynamic_threshold=None, + ucg_schedule=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): ctmp = ctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + elif isinstance(conditioning, list): + for ctmp in conditioning: + if ctmp.shape[0] != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold, + ucg_schedule=ucg_schedule + ) + return samples, intermediates + + @torch.no_grad() + def ddim_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None, + ucg_schedule=None): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + if ucg_schedule is not None: + assert len(ucg_schedule) == len(time_range) + unconditional_guidance_scale = ucg_schedule[i] + + outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold) + img, pred_x0 = outs + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + dynamic_threshold=None): + b, *_, device = *x.shape, x.device + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + model_output = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + if isinstance(c, dict): + assert isinstance(unconditional_conditioning, dict) + c_in = dict() + for k in c: + if isinstance(c[k], list): + c_in[k] = [torch.cat([ + unconditional_conditioning[k][i], + c[k][i]]) for i in range(len(c[k]))] + else: + c_in[k] = torch.cat([ + unconditional_conditioning[k], + c[k]]) + elif isinstance(c, list): + c_in = list() + assert isinstance(unconditional_conditioning, list) + for i in range(len(c)): + c_in.append(torch.cat([unconditional_conditioning[i], c[i]])) + else: + c_in = torch.cat([unconditional_conditioning, c]) + model_uncond, model_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + model_output = model_uncond + unconditional_guidance_scale * (model_t - model_uncond) + + if self.model.parameterization == "v": + e_t = self.model.predict_eps_from_z_and_v(x, t, model_output) + else: + e_t = model_output + + if score_corrector is not None: + assert self.model.parameterization == "eps", 'not implemented' + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + if self.model.parameterization != "v": + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + else: + pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output) + + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + + if dynamic_threshold is not None: + raise NotImplementedError() + + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + @torch.no_grad() + def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None, + unconditional_guidance_scale=1.0, unconditional_conditioning=None, callback=None): + num_reference_steps = self.ddpm_num_timesteps if use_original_steps else self.ddim_timesteps.shape[0] + + assert t_enc <= num_reference_steps + num_steps = t_enc + + if use_original_steps: + alphas_next = self.alphas_cumprod[:num_steps] + alphas = self.alphas_cumprod_prev[:num_steps] + else: + alphas_next = self.ddim_alphas[:num_steps] + alphas = torch.tensor(self.ddim_alphas_prev[:num_steps]) + + x_next = x0 + intermediates = [] + inter_steps = [] + for i in tqdm(range(num_steps), desc='Encoding Image'): + t = torch.full((x0.shape[0],), i, device=self.model.device, dtype=torch.long) + if unconditional_guidance_scale == 1.: + noise_pred = self.model.apply_model(x_next, t, c) + else: + assert unconditional_conditioning is not None + e_t_uncond, noise_pred = torch.chunk( + self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)), + torch.cat((unconditional_conditioning, c))), 2) + noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond) + + xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next + weighted_noise_pred = alphas_next[i].sqrt() * ( + (1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred + x_next = xt_weighted + weighted_noise_pred + if return_intermediates and i % ( + num_steps // return_intermediates) == 0 and i < num_steps - 1: + intermediates.append(x_next) + inter_steps.append(i) + elif return_intermediates and i >= num_steps - 2: + intermediates.append(x_next) + inter_steps.append(i) + if callback: callback(i) + + out = {'x_encoded': x_next, 'intermediate_steps': inter_steps} + if return_intermediates: + out.update({'intermediates': intermediates}) + return x_next, out + + @torch.no_grad() + def stochastic_encode(self, x0, t, use_original_steps=False, noise=None): + # fast, but does not allow for exact reconstruction + # t serves as an index to gather the correct alphas + if use_original_steps: + sqrt_alphas_cumprod = self.sqrt_alphas_cumprod + sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod + else: + sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas) + sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas + + if noise is None: + noise = torch.randn_like(x0) + return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 + + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise) + + @torch.no_grad() + def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + use_original_steps=False, callback=None): + + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + timesteps = timesteps[:t_start] + + time_range = np.flip(timesteps) + total_steps = timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='Decoding image', total=total_steps) + x_dec = x_latent + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long) + x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning) + if callback: callback(i) + return x_dec \ No newline at end of file diff --git a/ldm/models/diffusion/ddpm.py b/ldm/models/diffusion/ddpm.py new file mode 100644 index 0000000000000000000000000000000000000000..f71a44af48c8cba8e97849b7e6813b3e6f9fe83c --- /dev/null +++ b/ldm/models/diffusion/ddpm.py @@ -0,0 +1,1797 @@ +""" +wild mixture of +https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py +https://github.com/CompVis/taming-transformers +-- merci +""" + +import torch +import torch.nn as nn +import numpy as np +import pytorch_lightning as pl +from torch.optim.lr_scheduler import LambdaLR +from einops import rearrange, repeat +from contextlib import contextmanager, nullcontext +from functools import partial +import itertools +from tqdm import tqdm +from torchvision.utils import make_grid +from pytorch_lightning.utilities.distributed import rank_zero_only +from omegaconf import ListConfig + +from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config +from ldm.modules.ema import LitEma +from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution +from ldm.models.autoencoder import IdentityFirstStage, AutoencoderKL +from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like +from ldm.models.diffusion.ddim import DDIMSampler + + +__conditioning_keys__ = {'concat': 'c_concat', + 'crossattn': 'c_crossattn', + 'adm': 'y'} + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +def uniform_on_device(r1, r2, shape, device): + return (r1 - r2) * torch.rand(*shape, device=device) + r2 + + +class DDPM(pl.LightningModule): + # classic DDPM with Gaussian diffusion, in image space + def __init__(self, + unet_config, + timesteps=1000, + beta_schedule="linear", + loss_type="l2", + ckpt_path=None, + ignore_keys=[], + load_only_unet=False, + monitor="val/loss", + use_ema=True, + first_stage_key="image", + image_size=256, + channels=3, + log_every_t=100, + clip_denoised=True, + linear_start=1e-4, + linear_end=2e-2, + cosine_s=8e-3, + given_betas=None, + original_elbo_weight=0., + v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta + l_simple_weight=1., + conditioning_key=None, + parameterization="eps", # all assuming fixed variance schedules + scheduler_config=None, + use_positional_encodings=False, + learn_logvar=False, + logvar_init=0., + make_it_fit=False, + ucg_training=None, + reset_ema=False, + reset_num_ema_updates=False, + ): + super().__init__() + assert parameterization in ["eps", "x0", "v"], 'currently only supporting "eps" and "x0" and "v"' + self.parameterization = parameterization + print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode") + self.cond_stage_model = None + self.clip_denoised = clip_denoised + self.log_every_t = log_every_t + self.first_stage_key = first_stage_key + self.image_size = image_size # try conv? + self.channels = channels + self.use_positional_encodings = use_positional_encodings + self.model = DiffusionWrapper(unet_config, conditioning_key) + count_params(self.model, verbose=True) + self.use_ema = use_ema + if self.use_ema: + self.model_ema = LitEma(self.model) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + self.use_scheduler = scheduler_config is not None + if self.use_scheduler: + self.scheduler_config = scheduler_config + + self.v_posterior = v_posterior + self.original_elbo_weight = original_elbo_weight + self.l_simple_weight = l_simple_weight + + if monitor is not None: + self.monitor = monitor + self.make_it_fit = make_it_fit + if reset_ema: assert exists(ckpt_path) + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet) + if reset_ema: + assert self.use_ema + print(f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.") + self.model_ema = LitEma(self.model) + if reset_num_ema_updates: + print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ") + assert self.use_ema + self.model_ema.reset_num_updates() + + self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps, + linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s) + + self.loss_type = loss_type + + self.learn_logvar = learn_logvar + logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,)) + if self.learn_logvar: + self.logvar = nn.Parameter(self.logvar, requires_grad=True) + else: + self.register_buffer('logvar', logvar) + + self.ucg_training = ucg_training or dict() + if self.ucg_training: + self.ucg_prng = np.random.RandomState() + + def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if exists(given_betas): + betas = given_betas + else: + betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + + # calculations for posterior q(x_{t-1} | x_t, x_0) + posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / ( + 1. - alphas_cumprod) + self.v_posterior * betas + # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) + self.register_buffer('posterior_variance', to_torch(posterior_variance)) + # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain + self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20)))) + self.register_buffer('posterior_mean_coef1', to_torch( + betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))) + self.register_buffer('posterior_mean_coef2', to_torch( + (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod))) + + if self.parameterization == "eps": + lvlb_weights = self.betas ** 2 / ( + 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)) + elif self.parameterization == "x0": + lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod)) + elif self.parameterization == "v": + lvlb_weights = torch.ones_like(self.betas ** 2 / ( + 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))) + else: + raise NotImplementedError("mu not supported") + lvlb_weights[0] = lvlb_weights[1] + self.register_buffer('lvlb_weights', lvlb_weights, persistent=False) + assert not torch.isnan(self.lvlb_weights).all() + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.model.parameters()) + self.model_ema.copy_to(self.model) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.model.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + @torch.no_grad() + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + if self.make_it_fit: + n_params = len([name for name, _ in + itertools.chain(self.named_parameters(), + self.named_buffers())]) + for name, param in tqdm( + itertools.chain(self.named_parameters(), + self.named_buffers()), + desc="Fitting old weights to new weights", + total=n_params + ): + if not name in sd: + continue + old_shape = sd[name].shape + new_shape = param.shape + assert len(old_shape) == len(new_shape) + if len(new_shape) > 2: + # we only modify first two axes + assert new_shape[2:] == old_shape[2:] + # assumes first axis corresponds to output dim + if not new_shape == old_shape: + new_param = param.clone() + old_param = sd[name] + if len(new_shape) == 1: + for i in range(new_param.shape[0]): + new_param[i] = old_param[i % old_shape[0]] + elif len(new_shape) >= 2: + for i in range(new_param.shape[0]): + for j in range(new_param.shape[1]): + new_param[i, j] = old_param[i % old_shape[0], j % old_shape[1]] + + n_used_old = torch.ones(old_shape[1]) + for j in range(new_param.shape[1]): + n_used_old[j % old_shape[1]] += 1 + n_used_new = torch.zeros(new_shape[1]) + for j in range(new_param.shape[1]): + n_used_new[j] = n_used_old[j % old_shape[1]] + + n_used_new = n_used_new[None, :] + while len(n_used_new.shape) < len(new_shape): + n_used_new = n_used_new.unsqueeze(-1) + new_param /= n_used_new + + sd[name] = new_param + + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys:\n {missing}") + if len(unexpected) > 0: + print(f"\nUnexpected Keys:\n {unexpected}") + + def q_mean_variance(self, x_start, t): + """ + Get the distribution q(x_t | x_0). + :param x_start: the [N x C x ...] tensor of noiseless inputs. + :param t: the number of diffusion steps (minus 1). Here, 0 means one step. + :return: A tuple (mean, variance, log_variance), all of x_start's shape. + """ + mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start) + variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape) + log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape) + return mean, variance, log_variance + + def predict_start_from_noise(self, x_t, t, noise): + return ( + extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - + extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise + ) + + def predict_start_from_z_and_v(self, x_t, t, v): + # self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + # self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t - + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v + ) + + def predict_eps_from_z_and_v(self, x_t, t, v): + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * v + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * x_t + ) + + def q_posterior(self, x_start, x_t, t): + posterior_mean = ( + extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start + + extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t + ) + posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape) + posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape) + return posterior_mean, posterior_variance, posterior_log_variance_clipped + + def p_mean_variance(self, x, t, clip_denoised: bool): + model_out = self.model(x, t) + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + if clip_denoised: + x_recon.clamp_(-1., 1.) + + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, t, clip_denoised=True, repeat_noise=False): + b, *_, device = *x.shape, x.device + model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised) + noise = noise_like(x.shape, device, repeat_noise) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def p_sample_loop(self, shape, return_intermediates=False): + device = self.betas.device + b = shape[0] + img = torch.randn(shape, device=device) + intermediates = [img] + for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps): + img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long), + clip_denoised=self.clip_denoised) + if i % self.log_every_t == 0 or i == self.num_timesteps - 1: + intermediates.append(img) + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, batch_size=16, return_intermediates=False): + image_size = self.image_size + channels = self.channels + return self.p_sample_loop((batch_size, channels, image_size, image_size), + return_intermediates=return_intermediates) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def get_v(self, x, noise, t): + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise - + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x + ) + + def get_loss(self, pred, target, mean=True): + if self.loss_type == 'l1': + loss = (target - pred).abs() + if mean: + loss = loss.mean() + elif self.loss_type == 'l2': + if mean: + loss = torch.nn.functional.mse_loss(target, pred) + else: + loss = torch.nn.functional.mse_loss(target, pred, reduction='none') + else: + raise NotImplementedError("unknown loss type '{loss_type}'") + + return loss + + def p_losses(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + model_out = self.model(x_noisy, t) + + loss_dict = {} + if self.parameterization == "eps": + target = noise + elif self.parameterization == "x0": + target = x_start + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError(f"Parameterization {self.parameterization} not yet supported") + + loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3]) + + log_prefix = 'train' if self.training else 'val' + + loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()}) + loss_simple = loss.mean() * self.l_simple_weight + + loss_vlb = (self.lvlb_weights[t] * loss).mean() + loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb}) + + loss = loss_simple + self.original_elbo_weight * loss_vlb + + loss_dict.update({f'{log_prefix}/loss': loss}) + + return loss, loss_dict + + def forward(self, x, *args, **kwargs): + # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size + # assert h == img_size and w == img_size, f'height and width of image must be {img_size}' + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + return self.p_losses(x, t, *args, **kwargs) + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = rearrange(x, 'b h w c -> b c h w') + x = x.to(memory_format=torch.contiguous_format).float() + return x + + def shared_step(self, batch): + x = self.get_input(batch, self.first_stage_key) + loss, loss_dict = self(x) + return loss, loss_dict + + def training_step(self, batch, batch_idx): + for k in self.ucg_training: + p = self.ucg_training[k]["p"] + val = self.ucg_training[k]["val"] + if val is None: + val = "" + for i in range(len(batch[k])): + if self.ucg_prng.choice(2, p=[1 - p, p]): + batch[k][i] = val + + loss, loss_dict = self.shared_step(batch) + + self.log_dict(loss_dict, prog_bar=True, + logger=True, on_step=True, on_epoch=True) + + self.log("global_step", self.global_step, + prog_bar=True, logger=True, on_step=True, on_epoch=False) + + if self.use_scheduler: + lr = self.optimizers().param_groups[0]['lr'] + self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False) + + return loss + + @torch.no_grad() + def validation_step(self, batch, batch_idx): + _, loss_dict_no_ema = self.shared_step(batch) + with self.ema_scope(): + _, loss_dict_ema = self.shared_step(batch) + loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema} + self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True) + self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True) + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self.model) + + def _get_rows_from_list(self, samples): + n_imgs_per_row = len(samples) + denoise_grid = rearrange(samples, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs): + log = dict() + x = self.get_input(batch, self.first_stage_key) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + x = x.to(self.device)[:N] + log["inputs"] = x + + # get diffusion row + diffusion_row = list() + x_start = x[:n_row] + + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(x_start) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + diffusion_row.append(x_noisy) + + log["diffusion_row"] = self._get_rows_from_list(diffusion_row) + + if sample: + # get denoise row + with self.ema_scope("Plotting"): + samples, denoise_row = self.sample(batch_size=N, return_intermediates=True) + + log["samples"] = samples + log["denoise_row"] = self._get_rows_from_list(denoise_row) + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + if self.learn_logvar: + params = params + [self.logvar] + opt = torch.optim.AdamW(params, lr=lr) + return opt + + +class LatentDiffusion(DDPM): + """main class""" + + def __init__(self, + first_stage_config, + cond_stage_config, + num_timesteps_cond=None, + cond_stage_key="image", + cond_stage_trainable=False, + concat_mode=True, + cond_stage_forward=None, + conditioning_key=None, + scale_factor=1.0, + scale_by_std=False, + force_null_conditioning=False, + *args, **kwargs): + self.force_null_conditioning = force_null_conditioning + self.num_timesteps_cond = default(num_timesteps_cond, 1) + self.scale_by_std = scale_by_std + assert self.num_timesteps_cond <= kwargs['timesteps'] + # for backwards compatibility after implementation of DiffusionWrapper + if conditioning_key is None: + conditioning_key = 'concat' if concat_mode else 'crossattn' + if cond_stage_config == '__is_unconditional__' and not self.force_null_conditioning: + conditioning_key = None + ckpt_path = kwargs.pop("ckpt_path", None) + reset_ema = kwargs.pop("reset_ema", False) + reset_num_ema_updates = kwargs.pop("reset_num_ema_updates", False) + ignore_keys = kwargs.pop("ignore_keys", []) + super().__init__(conditioning_key=conditioning_key, *args, **kwargs) + self.concat_mode = concat_mode + self.cond_stage_trainable = cond_stage_trainable + self.cond_stage_key = cond_stage_key + try: + self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1 + except: + self.num_downs = 0 + if not scale_by_std: + self.scale_factor = scale_factor + else: + self.register_buffer('scale_factor', torch.tensor(scale_factor)) + self.instantiate_first_stage(first_stage_config) + self.instantiate_cond_stage(cond_stage_config) + self.cond_stage_forward = cond_stage_forward + self.clip_denoised = False + self.bbox_tokenizer = None + + self.restarted_from_ckpt = False + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys) + self.restarted_from_ckpt = True + if reset_ema: + assert self.use_ema + print( + f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.") + self.model_ema = LitEma(self.model) + if reset_num_ema_updates: + print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ") + assert self.use_ema + self.model_ema.reset_num_updates() + + def make_cond_schedule(self, ): + self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long) + ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long() + self.cond_ids[:self.num_timesteps_cond] = ids + + @rank_zero_only + @torch.no_grad() + def on_train_batch_start(self, batch, batch_idx, dataloader_idx): + # only for very first batch + if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt: + assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously' + # set rescale weight to 1./std of encodings + print("### USING STD-RESCALING ###") + x = super().get_input(batch, self.first_stage_key) + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + del self.scale_factor + self.register_buffer('scale_factor', 1. / z.flatten().std()) + print(f"setting self.scale_factor to {self.scale_factor}") + print("### USING STD-RESCALING ###") + + def register_schedule(self, + given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s) + + self.shorten_cond_schedule = self.num_timesteps_cond > 1 + if self.shorten_cond_schedule: + self.make_cond_schedule() + + def instantiate_first_stage(self, config): + model = instantiate_from_config(config) + self.first_stage_model = model.eval() + self.first_stage_model.train = disabled_train + for param in self.first_stage_model.parameters(): + param.requires_grad = False + + def instantiate_cond_stage(self, config): + if not self.cond_stage_trainable: + if config == "__is_first_stage__": + print("Using first stage also as cond stage.") + self.cond_stage_model = self.first_stage_model + elif config == "__is_unconditional__": + print(f"Training {self.__class__.__name__} as an unconditional model.") + self.cond_stage_model = None + # self.be_unconditional = True + else: + model = instantiate_from_config(config) + self.cond_stage_model = model.eval() + self.cond_stage_model.train = disabled_train + for param in self.cond_stage_model.parameters(): + param.requires_grad = False + else: + assert config != '__is_first_stage__' + assert config != '__is_unconditional__' + model = instantiate_from_config(config) + self.cond_stage_model = model + + def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False): + denoise_row = [] + for zd in tqdm(samples, desc=desc): + denoise_row.append(self.decode_first_stage(zd.to(self.device), + force_not_quantize=force_no_decoder_quantization)) + n_imgs_per_row = len(denoise_row) + denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W + denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + def get_first_stage_encoding(self, encoder_posterior): + if isinstance(encoder_posterior, DiagonalGaussianDistribution): + z = encoder_posterior.sample() + elif isinstance(encoder_posterior, torch.Tensor): + z = encoder_posterior + else: + raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented") + return self.scale_factor * z + + def get_learned_conditioning(self, c): + if self.cond_stage_forward is None: + if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode): + c = self.cond_stage_model.encode(c) + if isinstance(c, DiagonalGaussianDistribution): + c = c.mode() + else: + c = self.cond_stage_model(c) + else: + assert hasattr(self.cond_stage_model, self.cond_stage_forward) + c = getattr(self.cond_stage_model, self.cond_stage_forward)(c) + return c + + def meshgrid(self, h, w): + y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1) + x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1) + + arr = torch.cat([y, x], dim=-1) + return arr + + def delta_border(self, h, w): + """ + :param h: height + :param w: width + :return: normalized distance to image border, + wtith min distance = 0 at border and max dist = 0.5 at image center + """ + lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2) + arr = self.meshgrid(h, w) / lower_right_corner + dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0] + dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0] + edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0] + return edge_dist + + def get_weighting(self, h, w, Ly, Lx, device): + weighting = self.delta_border(h, w) + weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"], + self.split_input_params["clip_max_weight"], ) + weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device) + + if self.split_input_params["tie_braker"]: + L_weighting = self.delta_border(Ly, Lx) + L_weighting = torch.clip(L_weighting, + self.split_input_params["clip_min_tie_weight"], + self.split_input_params["clip_max_tie_weight"]) + + L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device) + weighting = weighting * L_weighting + return weighting + + def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code + """ + :param x: img of size (bs, c, h, w) + :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1]) + """ + bs, nc, h, w = x.shape + + # number of crops in image + Ly = (h - kernel_size[0]) // stride[0] + 1 + Lx = (w - kernel_size[1]) // stride[1] + 1 + + if uf == 1 and df == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params) + + weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx)) + + elif uf > 1 and df == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf), + dilation=1, padding=0, + stride=(stride[0] * uf, stride[1] * uf)) + fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2) + + weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx)) + + elif df > 1 and uf == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df), + dilation=1, padding=0, + stride=(stride[0] // df, stride[1] // df)) + fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2) + + weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx)) + + else: + raise NotImplementedError + + return fold, unfold, normalization, weighting + + @torch.no_grad() + def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False, + cond_key=None, return_original_cond=False, bs=None, return_x=False): + x = super().get_input(batch, k) + if bs is not None: + x = x[:bs] + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + + if self.model.conditioning_key is not None and not self.force_null_conditioning: + if cond_key is None: + cond_key = self.cond_stage_key + if cond_key != self.first_stage_key: + if cond_key in ['caption', 'coordinates_bbox', "txt"]: + xc = batch[cond_key] + elif cond_key in ['class_label', 'cls']: + xc = batch + else: + xc = super().get_input(batch, cond_key).to(self.device) + else: + xc = x + if not self.cond_stage_trainable or force_c_encode: + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + c = self.get_learned_conditioning(xc.to(self.device)) + else: + c = xc + if bs is not None: + c = c[:bs] + + if self.use_positional_encodings: + pos_x, pos_y = self.compute_latent_shifts(batch) + ckey = __conditioning_keys__[self.model.conditioning_key] + c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y} + + else: + c = None + xc = None + if self.use_positional_encodings: + pos_x, pos_y = self.compute_latent_shifts(batch) + c = {'pos_x': pos_x, 'pos_y': pos_y} + out = [z, c] + if return_first_stage_outputs: + xrec = self.decode_first_stage(z) + out.extend([x, xrec]) + if return_x: + out.extend([x]) + if return_original_cond: + out.append(xc) + return out + + @torch.no_grad() + def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False): + if predict_cids: + if z.dim() == 4: + z = torch.argmax(z.exp(), dim=1).long() + z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None) + z = rearrange(z, 'b h w c -> b c h w').contiguous() + + z = 1. / self.scale_factor * z + return self.first_stage_model.decode(z) + + @torch.no_grad() + def encode_first_stage(self, x): + return self.first_stage_model.encode(x) + + def shared_step(self, batch, **kwargs): + x, c = self.get_input(batch, self.first_stage_key) + loss = self(x, c) + return loss + + def forward(self, x, c, *args, **kwargs): + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + if self.model.conditioning_key is not None: + assert c is not None + if self.cond_stage_trainable: + c = self.get_learned_conditioning(c) + if self.shorten_cond_schedule: # TODO: drop this option + tc = self.cond_ids[t].to(self.device) + c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float())) + return self.p_losses(x, c, t, *args, **kwargs) + + def apply_model(self, x_noisy, t, cond, return_ids=False): + if isinstance(cond, dict): + # hybrid case, cond is expected to be a dict + pass + else: + if not isinstance(cond, list): + cond = [cond] + key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn' + cond = {key: cond} + + x_recon = self.model(x_noisy, t, **cond) + + if isinstance(x_recon, tuple) and not return_ids: + return x_recon[0] + else: + return x_recon + + def _predict_eps_from_xstart(self, x_t, t, pred_xstart): + return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \ + extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) + + def _prior_bpd(self, x_start): + """ + Get the prior KL term for the variational lower-bound, measured in + bits-per-dim. + This term can't be optimized, as it only depends on the encoder. + :param x_start: the [N x C x ...] tensor of inputs. + :return: a batch of [N] KL values (in bits), one per batch element. + """ + batch_size = x_start.shape[0] + t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device) + qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t) + kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0) + return mean_flat(kl_prior) / np.log(2.0) + + def p_losses(self, x_start, cond, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + model_output = self.apply_model(x_noisy, t, cond) + + loss_dict = {} + prefix = 'train' if self.training else 'val' + + if self.parameterization == "x0": + target = x_start + elif self.parameterization == "eps": + target = noise + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError() + + loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3]) + loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()}) + + logvar_t = self.logvar[t].to(self.device) + loss = loss_simple / torch.exp(logvar_t) + logvar_t + # loss = loss_simple / torch.exp(self.logvar) + self.logvar + if self.learn_logvar: + loss_dict.update({f'{prefix}/loss_gamma': loss.mean()}) + loss_dict.update({'logvar': self.logvar.data.mean()}) + + loss = self.l_simple_weight * loss.mean() + + loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3)) + loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean() + loss_dict.update({f'{prefix}/loss_vlb': loss_vlb}) + loss += (self.original_elbo_weight * loss_vlb) + loss_dict.update({f'{prefix}/loss': loss}) + + return loss, loss_dict + + def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False, + return_x0=False, score_corrector=None, corrector_kwargs=None): + t_in = t + model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids) + + if score_corrector is not None: + assert self.parameterization == "eps" + model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs) + + if return_codebook_ids: + model_out, logits = model_out + + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + else: + raise NotImplementedError() + + if clip_denoised: + x_recon.clamp_(-1., 1.) + if quantize_denoised: + x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon) + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + if return_codebook_ids: + return model_mean, posterior_variance, posterior_log_variance, logits + elif return_x0: + return model_mean, posterior_variance, posterior_log_variance, x_recon + else: + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, + return_codebook_ids=False, quantize_denoised=False, return_x0=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None): + b, *_, device = *x.shape, x.device + outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, + return_codebook_ids=return_codebook_ids, + quantize_denoised=quantize_denoised, + return_x0=return_x0, + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) + if return_codebook_ids: + raise DeprecationWarning("Support dropped.") + model_mean, _, model_log_variance, logits = outputs + elif return_x0: + model_mean, _, model_log_variance, x0 = outputs + else: + model_mean, _, model_log_variance = outputs + + noise = noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + + if return_codebook_ids: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1) + if return_x0: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0 + else: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False, + img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0., + score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None, + log_every_t=None): + if not log_every_t: + log_every_t = self.log_every_t + timesteps = self.num_timesteps + if batch_size is not None: + b = batch_size if batch_size is not None else shape[0] + shape = [batch_size] + list(shape) + else: + b = batch_size = shape[0] + if x_T is None: + img = torch.randn(shape, device=self.device) + else: + img = x_T + intermediates = [] + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + + if start_T is not None: + timesteps = min(timesteps, start_T) + iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation', + total=timesteps) if verbose else reversed( + range(0, timesteps)) + if type(temperature) == float: + temperature = [temperature] * timesteps + + for i in iterator: + ts = torch.full((b,), i, device=self.device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img, x0_partial = self.p_sample(img, cond, ts, + clip_denoised=self.clip_denoised, + quantize_denoised=quantize_denoised, return_x0=True, + temperature=temperature[i], noise_dropout=noise_dropout, + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) + if mask is not None: + assert x0 is not None + img_orig = self.q_sample(x0, ts) + img = img_orig * mask + (1. - mask) * img + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(x0_partial) + if callback: callback(i) + if img_callback: img_callback(img, i) + return img, intermediates + + @torch.no_grad() + def p_sample_loop(self, cond, shape, return_intermediates=False, + x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, start_T=None, + log_every_t=None): + + if not log_every_t: + log_every_t = self.log_every_t + device = self.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + intermediates = [img] + if timesteps is None: + timesteps = self.num_timesteps + + if start_T is not None: + timesteps = min(timesteps, start_T) + iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed( + range(0, timesteps)) + + if mask is not None: + assert x0 is not None + assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match + + for i in iterator: + ts = torch.full((b,), i, device=device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img = self.p_sample(img, cond, ts, + clip_denoised=self.clip_denoised, + quantize_denoised=quantize_denoised) + if mask is not None: + img_orig = self.q_sample(x0, ts) + img = img_orig * mask + (1. - mask) * img + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(img) + if callback: callback(i) + if img_callback: img_callback(img, i) + + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None, + verbose=True, timesteps=None, quantize_denoised=False, + mask=None, x0=None, shape=None, **kwargs): + if shape is None: + shape = (batch_size, self.channels, self.image_size, self.image_size) + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + return self.p_sample_loop(cond, + shape, + return_intermediates=return_intermediates, x_T=x_T, + verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised, + mask=mask, x0=x0) + + @torch.no_grad() + def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs): + if ddim: + ddim_sampler = DDIMSampler(self) + shape = (self.channels, self.image_size, self.image_size) + samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, + shape, cond, verbose=False, **kwargs) + + else: + samples, intermediates = self.sample(cond=cond, batch_size=batch_size, + return_intermediates=True, **kwargs) + + return samples, intermediates + + @torch.no_grad() + def get_unconditional_conditioning(self, batch_size, null_label=None): + if null_label is not None: + xc = null_label + if isinstance(xc, ListConfig): + xc = list(xc) + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + if hasattr(xc, "to"): + xc = xc.to(self.device) + c = self.get_learned_conditioning(xc) + else: + if self.cond_stage_key in ["class_label", "cls"]: + xc = self.cond_stage_model.get_unconditional_conditioning(batch_size, device=self.device) + return self.get_learned_conditioning(xc) + else: + raise NotImplementedError("todo") + if isinstance(c, list): # in case the encoder gives us a list + for i in range(len(c)): + c[i] = repeat(c[i], '1 ... -> b ...', b=batch_size).to(self.device) + else: + c = repeat(c, '1 ... -> b ...', b=batch_size).to(self.device) + return c + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=50, ddim_eta=0., return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, + return_first_stage_outputs=True, + force_c_encode=True, + return_original_cond=True, + bs=N) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', "cls"]: + try: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + except KeyError: + # probably no "human_label" in batch + pass + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance( + self.first_stage_model, IdentityFirstStage): + # also display when quantizing x0 while sampling + with ema_scope("Plotting Quantized Denoised"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + quantize_denoised=True) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True, + # quantize_denoised=True) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_x0_quantized"] = x_samples + + if unconditional_guidance_scale > 1.0: + uc = self.get_unconditional_conditioning(N, unconditional_guidance_label) + if self.model.conditioning_key == "crossattn-adm": + uc = {"c_crossattn": [uc], "c_adm": c["c_adm"]} + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + if inpaint: + # make a simple center square + b, h, w = z.shape[0], z.shape[2], z.shape[3] + mask = torch.ones(N, h, w).to(self.device) + # zeros will be filled in + mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0. + mask = mask[:, None, ...] + with ema_scope("Plotting Inpaint"): + samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta, + ddim_steps=ddim_steps, x0=z[:N], mask=mask) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_inpainting"] = x_samples + log["mask"] = mask + + # outpaint + mask = 1. - mask + with ema_scope("Plotting Outpaint"): + samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta, + ddim_steps=ddim_steps, x0=z[:N], mask=mask) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_outpainting"] = x_samples + + if plot_progressive_rows: + with ema_scope("Plotting Progressives"): + img, progressives = self.progressive_denoising(c, + shape=(self.channels, self.image_size, self.image_size), + batch_size=N) + prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") + log["progressive_row"] = prog_row + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + if self.cond_stage_trainable: + print(f"{self.__class__.__name__}: Also optimizing conditioner params!") + params = params + list(self.cond_stage_model.parameters()) + if self.learn_logvar: + print('Diffusion model optimizing logvar') + params.append(self.logvar) + opt = torch.optim.AdamW(params, lr=lr) + if self.use_scheduler: + assert 'target' in self.scheduler_config + scheduler = instantiate_from_config(self.scheduler_config) + + print("Setting up LambdaLR scheduler...") + scheduler = [ + { + 'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule), + 'interval': 'step', + 'frequency': 1 + }] + return [opt], scheduler + return opt + + @torch.no_grad() + def to_rgb(self, x): + x = x.float() + if not hasattr(self, "colorize"): + self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x) + x = nn.functional.conv2d(x, weight=self.colorize) + x = 2. * (x - x.min()) / (x.max() - x.min()) - 1. + return x + + +class DiffusionWrapper(pl.LightningModule): + def __init__(self, diff_model_config, conditioning_key): + super().__init__() + self.sequential_cross_attn = diff_model_config.pop("sequential_crossattn", False) + self.diffusion_model = instantiate_from_config(diff_model_config) + self.conditioning_key = conditioning_key + assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm', 'hybrid-adm', 'crossattn-adm'] + + def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, c_adm=None): + if self.conditioning_key is None: + out = self.diffusion_model(x, t) + elif self.conditioning_key == 'concat': + xc = torch.cat([x] + c_concat, dim=1) + out = self.diffusion_model(xc, t) + elif self.conditioning_key == 'crossattn': + if not self.sequential_cross_attn: + cc = torch.cat(c_crossattn, 1) + else: + cc = c_crossattn + out = self.diffusion_model(x, t, context=cc) + elif self.conditioning_key == 'hybrid': + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc) + elif self.conditioning_key == 'hybrid-adm': + assert c_adm is not None + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc, y=c_adm) + elif self.conditioning_key == 'crossattn-adm': + assert c_adm is not None + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(x, t, context=cc, y=c_adm) + elif self.conditioning_key == 'adm': + cc = c_crossattn[0] + out = self.diffusion_model(x, t, y=cc) + else: + raise NotImplementedError() + + return out + + +class LatentUpscaleDiffusion(LatentDiffusion): + def __init__(self, *args, low_scale_config, low_scale_key="LR", noise_level_key=None, **kwargs): + super().__init__(*args, **kwargs) + # assumes that neither the cond_stage nor the low_scale_model contain trainable params + assert not self.cond_stage_trainable + self.instantiate_low_stage(low_scale_config) + self.low_scale_key = low_scale_key + self.noise_level_key = noise_level_key + + def instantiate_low_stage(self, config): + model = instantiate_from_config(config) + self.low_scale_model = model.eval() + self.low_scale_model.train = disabled_train + for param in self.low_scale_model.parameters(): + param.requires_grad = False + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False): + if not log_mode: + z, c = super().get_input(batch, k, force_c_encode=True, bs=bs) + else: + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + x_low = batch[self.low_scale_key][:bs] + x_low = rearrange(x_low, 'b h w c -> b c h w') + x_low = x_low.to(memory_format=torch.contiguous_format).float() + zx, noise_level = self.low_scale_model(x_low) + if self.noise_level_key is not None: + # get noise level from batch instead, e.g. when extracting a custom noise level for bsr + raise NotImplementedError('TODO') + + all_conds = {"c_concat": [zx], "c_crossattn": [c], "c_adm": noise_level} + if log_mode: + # TODO: maybe disable if too expensive + x_low_rec = self.low_scale_model.decode(zx) + return z, all_conds, x, xrec, xc, x_low, x_low_rec, noise_level + return z, all_conds + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None, + plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True, + unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc, x_low, x_low_rec, noise_level = self.get_input(batch, self.first_stage_key, bs=N, + log_mode=True) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + log["x_lr"] = x_low + log[f"x_lr_rec_@noise_levels{'-'.join(map(lambda x: str(x), list(noise_level.cpu().numpy())))}"] = x_low_rec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', 'cls']: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label) + # TODO explore better "unconditional" choices for the other keys + # maybe guide away from empty text label and highest noise level and maximally degraded zx? + uc = dict() + for k in c: + if k == "c_crossattn": + assert isinstance(c[k], list) and len(c[k]) == 1 + uc[k] = [uc_tmp] + elif k == "c_adm": # todo: only run with text-based guidance? + assert isinstance(c[k], torch.Tensor) + #uc[k] = torch.ones_like(c[k]) * self.low_scale_model.max_noise_level + uc[k] = c[k] + elif isinstance(c[k], list): + uc[k] = [c[k][i] for i in range(len(c[k]))] + else: + uc[k] = c[k] + + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + if plot_progressive_rows: + with ema_scope("Plotting Progressives"): + img, progressives = self.progressive_denoising(c, + shape=(self.channels, self.image_size, self.image_size), + batch_size=N) + prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") + log["progressive_row"] = prog_row + + return log + + +class LatentFinetuneDiffusion(LatentDiffusion): + """ + Basis for different finetunas, such as inpainting or depth2image + To disable finetuning mode, set finetune_keys to None + """ + + def __init__(self, + concat_keys: tuple, + finetune_keys=("model.diffusion_model.input_blocks.0.0.weight", + "model_ema.diffusion_modelinput_blocks00weight" + ), + keep_finetune_dims=4, + # if model was trained without concat mode before and we would like to keep these channels + c_concat_log_start=None, # to log reconstruction of c_concat codes + c_concat_log_end=None, + *args, **kwargs + ): + ckpt_path = kwargs.pop("ckpt_path", None) + ignore_keys = kwargs.pop("ignore_keys", list()) + super().__init__(*args, **kwargs) + self.finetune_keys = finetune_keys + self.concat_keys = concat_keys + self.keep_dims = keep_finetune_dims + self.c_concat_log_start = c_concat_log_start + self.c_concat_log_end = c_concat_log_end + if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint' + if exists(ckpt_path): + self.init_from_ckpt(ckpt_path, ignore_keys) + + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + + # make it explicit, finetune by including extra input channels + if exists(self.finetune_keys) and k in self.finetune_keys: + new_entry = None + for name, param in self.named_parameters(): + if name in self.finetune_keys: + print( + f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only") + new_entry = torch.zeros_like(param) # zero init + assert exists(new_entry), 'did not find matching parameter to modify' + new_entry[:, :self.keep_dims, ...] = sd[k] + sd[k] = new_entry + + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys: {missing}") + if len(unexpected) > 0: + print(f"Unexpected Keys: {unexpected}") + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True) + c_cat, c = c["c_concat"][0], c["c_crossattn"][0] + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', 'cls']: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if not (self.c_concat_log_start is None and self.c_concat_log_end is None): + log["c_concat_decoded"] = self.decode_first_stage(c_cat[:, self.c_concat_log_start:self.c_concat_log_end]) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label) + uc_cat = c_cat + uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]} + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc_full, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + return log + + +class LatentInpaintDiffusion(LatentFinetuneDiffusion): + """ + can either run as pure inpainting model (only concat mode) or with mixed conditionings, + e.g. mask as concat and text via cross-attn. + To disable finetuning mode, set finetune_keys to None + """ + + def __init__(self, + concat_keys=("mask", "masked_image"), + masked_image_key="masked_image", + *args, **kwargs + ): + super().__init__(concat_keys, *args, **kwargs) + self.masked_image_key = masked_image_key + assert self.masked_image_key in concat_keys + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpainting' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + c_cat = list() + for ck in self.concat_keys: + cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float() + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + bchw = z.shape + if ck != self.masked_image_key: + cc = torch.nn.functional.interpolate(cc, size=bchw[-2:]) + else: + cc = self.get_first_stage_encoding(self.encode_first_stage(cc)) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super(LatentInpaintDiffusion, self).log_images(*args, **kwargs) + log["masked_image"] = rearrange(args[0]["masked_image"], + 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float() + return log + + +class LatentDepth2ImageDiffusion(LatentFinetuneDiffusion): + """ + condition on monocular depth estimation + """ + + def __init__(self, depth_stage_config, concat_keys=("midas_in",), *args, **kwargs): + super().__init__(concat_keys=concat_keys, *args, **kwargs) + self.depth_model = instantiate_from_config(depth_stage_config) + self.depth_stage_key = concat_keys[0] + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for depth2img' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + assert len(self.concat_keys) == 1 + c_cat = list() + for ck in self.concat_keys: + cc = batch[ck] + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + cc = self.depth_model(cc) + cc = torch.nn.functional.interpolate( + cc, + size=z.shape[2:], + mode="bicubic", + align_corners=False, + ) + + depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3], + keepdim=True) + cc = 2. * (cc - depth_min) / (depth_max - depth_min + 0.001) - 1. + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super().log_images(*args, **kwargs) + depth = self.depth_model(args[0][self.depth_stage_key]) + depth_min, depth_max = torch.amin(depth, dim=[1, 2, 3], keepdim=True), \ + torch.amax(depth, dim=[1, 2, 3], keepdim=True) + log["depth"] = 2. * (depth - depth_min) / (depth_max - depth_min) - 1. + return log + + +class LatentUpscaleFinetuneDiffusion(LatentFinetuneDiffusion): + """ + condition on low-res image (and optionally on some spatial noise augmentation) + """ + def __init__(self, concat_keys=("lr",), reshuffle_patch_size=None, + low_scale_config=None, low_scale_key=None, *args, **kwargs): + super().__init__(concat_keys=concat_keys, *args, **kwargs) + self.reshuffle_patch_size = reshuffle_patch_size + self.low_scale_model = None + if low_scale_config is not None: + print("Initializing a low-scale model") + assert exists(low_scale_key) + self.instantiate_low_stage(low_scale_config) + self.low_scale_key = low_scale_key + + def instantiate_low_stage(self, config): + model = instantiate_from_config(config) + self.low_scale_model = model.eval() + self.low_scale_model.train = disabled_train + for param in self.low_scale_model.parameters(): + param.requires_grad = False + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for upscaling-ft' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + assert len(self.concat_keys) == 1 + # optionally make spatial noise_level here + c_cat = list() + noise_level = None + for ck in self.concat_keys: + cc = batch[ck] + cc = rearrange(cc, 'b h w c -> b c h w') + if exists(self.reshuffle_patch_size): + assert isinstance(self.reshuffle_patch_size, int) + cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w', + p1=self.reshuffle_patch_size, p2=self.reshuffle_patch_size) + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + if exists(self.low_scale_model) and ck == self.low_scale_key: + cc, noise_level = self.low_scale_model(cc) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + if exists(noise_level): + all_conds = {"c_concat": [c_cat], "c_crossattn": [c], "c_adm": noise_level} + else: + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super().log_images(*args, **kwargs) + log["lr"] = rearrange(args[0]["lr"], 'b h w c -> b c h w') + return log diff --git a/ldm/models/diffusion/dpm_solver/__init__.py b/ldm/models/diffusion/dpm_solver/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7427f38c07530afbab79154ea8aaf88c4bf70a08 --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/__init__.py @@ -0,0 +1 @@ +from .sampler import DPMSolverSampler \ No newline at end of file diff --git a/ldm/models/diffusion/dpm_solver/dpm_solver.py b/ldm/models/diffusion/dpm_solver/dpm_solver.py new file mode 100644 index 0000000000000000000000000000000000000000..095e5ba3ce0b1aa7f4b3f1e2e5d8fff7cfe6dc8c --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/dpm_solver.py @@ -0,0 +1,1154 @@ +import torch +import torch.nn.functional as F +import math +from tqdm import tqdm + + +class NoiseScheduleVP: + def __init__( + self, + schedule='discrete', + betas=None, + alphas_cumprod=None, + continuous_beta_0=0.1, + continuous_beta_1=20., + ): + """Create a wrapper class for the forward SDE (VP type). + *** + Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t. + We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images. + *** + The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ). + We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper). + Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have: + log_alpha_t = self.marginal_log_mean_coeff(t) + sigma_t = self.marginal_std(t) + lambda_t = self.marginal_lambda(t) + Moreover, as lambda(t) is an invertible function, we also support its inverse function: + t = self.inverse_lambda(lambda_t) + =============================================================== + We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]). + 1. For discrete-time DPMs: + For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by: + t_i = (i + 1) / N + e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1. + We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3. + Args: + betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details) + alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details) + Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`. + **Important**: Please pay special attention for the args for `alphas_cumprod`: + The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that + q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ). + Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have + alpha_{t_n} = \sqrt{\hat{alpha_n}}, + and + log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}). + 2. For continuous-time DPMs: + We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise + schedule are the default settings in DDPM and improved-DDPM: + Args: + beta_min: A `float` number. The smallest beta for the linear schedule. + beta_max: A `float` number. The largest beta for the linear schedule. + cosine_s: A `float` number. The hyperparameter in the cosine schedule. + cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule. + T: A `float` number. The ending time of the forward process. + =============================================================== + Args: + schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs, + 'linear' or 'cosine' for continuous-time DPMs. + Returns: + A wrapper object of the forward SDE (VP type). + + =============================================================== + Example: + # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', betas=betas) + # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod) + # For continuous-time DPMs (VPSDE), linear schedule: + >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.) + """ + + if schedule not in ['discrete', 'linear', 'cosine']: + raise ValueError( + "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format( + schedule)) + + self.schedule = schedule + if schedule == 'discrete': + if betas is not None: + log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0) + else: + assert alphas_cumprod is not None + log_alphas = 0.5 * torch.log(alphas_cumprod) + self.total_N = len(log_alphas) + self.T = 1. + self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1)) + self.log_alpha_array = log_alphas.reshape((1, -1,)) + else: + self.total_N = 1000 + self.beta_0 = continuous_beta_0 + self.beta_1 = continuous_beta_1 + self.cosine_s = 0.008 + self.cosine_beta_max = 999. + self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * ( + 1. + self.cosine_s) / math.pi - self.cosine_s + self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.)) + self.schedule = schedule + if schedule == 'cosine': + # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T. + # Note that T = 0.9946 may be not the optimal setting. However, we find it works well. + self.T = 0.9946 + else: + self.T = 1. + + def marginal_log_mean_coeff(self, t): + """ + Compute log(alpha_t) of a given continuous-time label t in [0, T]. + """ + if self.schedule == 'discrete': + return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device), + self.log_alpha_array.to(t.device)).reshape((-1)) + elif self.schedule == 'linear': + return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0 + elif self.schedule == 'cosine': + log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.)) + log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0 + return log_alpha_t + + def marginal_alpha(self, t): + """ + Compute alpha_t of a given continuous-time label t in [0, T]. + """ + return torch.exp(self.marginal_log_mean_coeff(t)) + + def marginal_std(self, t): + """ + Compute sigma_t of a given continuous-time label t in [0, T]. + """ + return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t))) + + def marginal_lambda(self, t): + """ + Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T]. + """ + log_mean_coeff = self.marginal_log_mean_coeff(t) + log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff)) + return log_mean_coeff - log_std + + def inverse_lambda(self, lamb): + """ + Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t. + """ + if self.schedule == 'linear': + tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb)) + Delta = self.beta_0 ** 2 + tmp + return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0) + elif self.schedule == 'discrete': + log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb) + t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]), + torch.flip(self.t_array.to(lamb.device), [1])) + return t.reshape((-1,)) + else: + log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb)) + t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * ( + 1. + self.cosine_s) / math.pi - self.cosine_s + t = t_fn(log_alpha) + return t + + +def model_wrapper( + model, + noise_schedule, + model_type="noise", + model_kwargs={}, + guidance_type="uncond", + condition=None, + unconditional_condition=None, + guidance_scale=1., + classifier_fn=None, + classifier_kwargs={}, +): + """Create a wrapper function for the noise prediction model. + DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to + firstly wrap the model function to a noise prediction model that accepts the continuous time as the input. + We support four types of the diffusion model by setting `model_type`: + 1. "noise": noise prediction model. (Trained by predicting noise). + 2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0). + 3. "v": velocity prediction model. (Trained by predicting the velocity). + The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2]. + [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models." + arXiv preprint arXiv:2202.00512 (2022). + [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models." + arXiv preprint arXiv:2210.02303 (2022). + + 4. "score": marginal score function. (Trained by denoising score matching). + Note that the score function and the noise prediction model follows a simple relationship: + ``` + noise(x_t, t) = -sigma_t * score(x_t, t) + ``` + We support three types of guided sampling by DPMs by setting `guidance_type`: + 1. "uncond": unconditional sampling by DPMs. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + 2. "classifier": classifier guidance sampling [3] by DPMs and another classifier. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + The input `classifier_fn` has the following format: + `` + classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond) + `` + [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis," + in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794. + 3. "classifier-free": classifier-free guidance sampling by conditional DPMs. + The input `model` has the following format: + `` + model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score + `` + And if cond == `unconditional_condition`, the model output is the unconditional DPM output. + [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance." + arXiv preprint arXiv:2207.12598 (2022). + + The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999) + or continuous-time labels (i.e. epsilon to T). + We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise: + `` + def model_fn(x, t_continuous) -> noise: + t_input = get_model_input_time(t_continuous) + return noise_pred(model, x, t_input, **model_kwargs) + `` + where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver. + =============================================================== + Args: + model: A diffusion model with the corresponding format described above. + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + model_type: A `str`. The parameterization type of the diffusion model. + "noise" or "x_start" or "v" or "score". + model_kwargs: A `dict`. A dict for the other inputs of the model function. + guidance_type: A `str`. The type of the guidance for sampling. + "uncond" or "classifier" or "classifier-free". + condition: A pytorch tensor. The condition for the guided sampling. + Only used for "classifier" or "classifier-free" guidance type. + unconditional_condition: A pytorch tensor. The condition for the unconditional sampling. + Only used for "classifier-free" guidance type. + guidance_scale: A `float`. The scale for the guided sampling. + classifier_fn: A classifier function. Only used for the classifier guidance. + classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function. + Returns: + A noise prediction model that accepts the noised data and the continuous time as the inputs. + """ + + def get_model_input_time(t_continuous): + """ + Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time. + For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N]. + For continuous-time DPMs, we just use `t_continuous`. + """ + if noise_schedule.schedule == 'discrete': + return (t_continuous - 1. / noise_schedule.total_N) * 1000. + else: + return t_continuous + + def noise_pred_fn(x, t_continuous, cond=None): + if t_continuous.reshape((-1,)).shape[0] == 1: + t_continuous = t_continuous.expand((x.shape[0])) + t_input = get_model_input_time(t_continuous) + if cond is None: + output = model(x, t_input, **model_kwargs) + else: + output = model(x, t_input, cond, **model_kwargs) + if model_type == "noise": + return output + elif model_type == "x_start": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims) + elif model_type == "v": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x + elif model_type == "score": + sigma_t = noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return -expand_dims(sigma_t, dims) * output + + def cond_grad_fn(x, t_input): + """ + Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t). + """ + with torch.enable_grad(): + x_in = x.detach().requires_grad_(True) + log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs) + return torch.autograd.grad(log_prob.sum(), x_in)[0] + + def model_fn(x, t_continuous): + """ + The noise predicition model function that is used for DPM-Solver. + """ + if t_continuous.reshape((-1,)).shape[0] == 1: + t_continuous = t_continuous.expand((x.shape[0])) + if guidance_type == "uncond": + return noise_pred_fn(x, t_continuous) + elif guidance_type == "classifier": + assert classifier_fn is not None + t_input = get_model_input_time(t_continuous) + cond_grad = cond_grad_fn(x, t_input) + sigma_t = noise_schedule.marginal_std(t_continuous) + noise = noise_pred_fn(x, t_continuous) + return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad + elif guidance_type == "classifier-free": + if guidance_scale == 1. or unconditional_condition is None: + return noise_pred_fn(x, t_continuous, cond=condition) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t_continuous] * 2) + c_in = torch.cat([unconditional_condition, condition]) + noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2) + return noise_uncond + guidance_scale * (noise - noise_uncond) + + assert model_type in ["noise", "x_start", "v"] + assert guidance_type in ["uncond", "classifier", "classifier-free"] + return model_fn + + +class DPM_Solver: + def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.): + """Construct a DPM-Solver. + We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0"). + If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver). + If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++). + In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True. + The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales. + Args: + model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]): + `` + def model_fn(x, t_continuous): + return noise + `` + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model. + thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1]. + max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding. + + [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b. + """ + self.model = model_fn + self.noise_schedule = noise_schedule + self.predict_x0 = predict_x0 + self.thresholding = thresholding + self.max_val = max_val + + def noise_prediction_fn(self, x, t): + """ + Return the noise prediction model. + """ + return self.model(x, t) + + def data_prediction_fn(self, x, t): + """ + Return the data prediction model (with thresholding). + """ + noise = self.noise_prediction_fn(x, t) + dims = x.dim() + alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t) + x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims) + if self.thresholding: + p = 0.995 # A hyperparameter in the paper of "Imagen" [1]. + s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1) + s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims) + x0 = torch.clamp(x0, -s, s) / s + return x0 + + def model_fn(self, x, t): + """ + Convert the model to the noise prediction model or the data prediction model. + """ + if self.predict_x0: + return self.data_prediction_fn(x, t) + else: + return self.noise_prediction_fn(x, t) + + def get_time_steps(self, skip_type, t_T, t_0, N, device): + """Compute the intermediate time steps for sampling. + Args: + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + N: A `int`. The total number of the spacing of the time steps. + device: A torch device. + Returns: + A pytorch tensor of the time steps, with the shape (N + 1,). + """ + if skip_type == 'logSNR': + lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device)) + lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device)) + logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device) + return self.noise_schedule.inverse_lambda(logSNR_steps) + elif skip_type == 'time_uniform': + return torch.linspace(t_T, t_0, N + 1).to(device) + elif skip_type == 'time_quadratic': + t_order = 2 + t = torch.linspace(t_T ** (1. / t_order), t_0 ** (1. / t_order), N + 1).pow(t_order).to(device) + return t + else: + raise ValueError( + "Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type)) + + def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device): + """ + Get the order of each step for sampling by the singlestep DPM-Solver. + We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast". + Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is: + - If order == 1: + We take `steps` of DPM-Solver-1 (i.e. DDIM). + - If order == 2: + - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of DPM-Solver-2. + - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If order == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2. + ============================================ + Args: + order: A `int`. The max order for the solver (2 or 3). + steps: A `int`. The total number of function evaluations (NFE). + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + device: A torch device. + Returns: + orders: A list of the solver order of each step. + """ + if order == 3: + K = steps // 3 + 1 + if steps % 3 == 0: + orders = [3, ] * (K - 2) + [2, 1] + elif steps % 3 == 1: + orders = [3, ] * (K - 1) + [1] + else: + orders = [3, ] * (K - 1) + [2] + elif order == 2: + if steps % 2 == 0: + K = steps // 2 + orders = [2, ] * K + else: + K = steps // 2 + 1 + orders = [2, ] * (K - 1) + [1] + elif order == 1: + K = 1 + orders = [1, ] * steps + else: + raise ValueError("'order' must be '1' or '2' or '3'.") + if skip_type == 'logSNR': + # To reproduce the results in DPM-Solver paper + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device) + else: + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[ + torch.cumsum(torch.tensor([0, ] + orders)).to(device)] + return timesteps_outer, orders + + def denoise_to_zero_fn(self, x, s): + """ + Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization. + """ + return self.data_prediction_fn(x, s) + + def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False): + """ + DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + if self.predict_x0: + phi_1 = torch.expm1(-h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + ) + if return_intermediate: + return x_t, {'model_s': model_s} + else: + return x_t + else: + phi_1 = torch.expm1(h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + ) + if return_intermediate: + return x_t, {'model_s': model_s} + else: + return x_t + + def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False, + solver_type='dpm_solver'): + """ + Singlestep solver DPM-Solver-2 from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + r1: A `float`. The hyperparameter of the second-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + if r1 is None: + r1 = 0.5 + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + s1 = ns.inverse_lambda(lambda_s1) + log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff( + s1), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t) + alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t) + + if self.predict_x0: + phi_11 = torch.expm1(-r1 * h) + phi_1 = torch.expm1(-h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = ( + expand_dims(sigma_s1 / sigma_s, dims) * x + - expand_dims(alpha_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + - (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s) + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * ( + model_s1 - model_s) + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_1 = torch.expm1(h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = ( + expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x + - expand_dims(sigma_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s) + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s) + ) + if return_intermediate: + return x_t, {'model_s': model_s, 'model_s1': model_s1} + else: + return x_t + + def singlestep_dpm_solver_third_update(self, x, s, t, r1=1. / 3., r2=2. / 3., model_s=None, model_s1=None, + return_intermediate=False, solver_type='dpm_solver'): + """ + Singlestep solver DPM-Solver-3 from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + r1: A `float`. The hyperparameter of the third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`). + If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + if r1 is None: + r1 = 1. / 3. + if r2 is None: + r2 = 2. / 3. + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + lambda_s2 = lambda_s + r2 * h + s1 = ns.inverse_lambda(lambda_s1) + s2 = ns.inverse_lambda(lambda_s2) + log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff( + s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std( + s2), ns.marginal_std(t) + alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t) + + if self.predict_x0: + phi_11 = torch.expm1(-r1 * h) + phi_12 = torch.expm1(-r2 * h) + phi_1 = torch.expm1(-h) + phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1. + phi_2 = phi_1 / h + 1. + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = ( + expand_dims(sigma_s1 / sigma_s, dims) * x + - expand_dims(alpha_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + expand_dims(sigma_s2 / sigma_s, dims) * x + - expand_dims(alpha_s2 * phi_12, dims) * model_s + + r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s) + ) + elif solver_type == 'taylor': + D1_0 = (1. / r1) * (model_s1 - model_s) + D1_1 = (1. / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2. * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + expand_dims(alpha_t * phi_2, dims) * D1 + - expand_dims(alpha_t * phi_3, dims) * D2 + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_12 = torch.expm1(r2 * h) + phi_1 = torch.expm1(h) + phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1. + phi_2 = phi_1 / h - 1. + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = ( + expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x + - expand_dims(sigma_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x + - expand_dims(sigma_s2 * phi_12, dims) * model_s + - r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s) + ) + elif solver_type == 'taylor': + D1_0 = (1. / r1) * (model_s1 - model_s) + D1_1 = (1. / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2. * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - expand_dims(sigma_t * phi_2, dims) * D1 + - expand_dims(sigma_t * phi_3, dims) * D2 + ) + + if return_intermediate: + return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2} + else: + return x_t + + def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"): + """ + Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + ns = self.noise_schedule + dims = x.dim() + model_prev_1, model_prev_0 = model_prev_list + t_prev_1, t_prev_0 = t_prev_list + lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda( + t_prev_0), ns.marginal_lambda(t) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0 = h_0 / h + D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1) + if self.predict_x0: + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + - 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0 + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0 + ) + else: + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0 + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0 + ) + return x_t + + def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'): + """ + Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + dims = x.dim() + model_prev_2, model_prev_1, model_prev_0 = model_prev_list + t_prev_2, t_prev_1, t_prev_0 = t_prev_list + lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda( + t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_1 = lambda_prev_1 - lambda_prev_2 + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0, r1 = h_0 / h, h_1 / h + D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1) + D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2) + D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1) + D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1) + if self.predict_x0: + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1 + - expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h ** 2 - 0.5), dims) * D2 + ) + else: + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1 + - expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h ** 2 - 0.5), dims) * D2 + ) + return x_t + + def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None, + r2=None): + """ + Singlestep DPM-Solver with the order `order` from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + r1: A `float`. The hyperparameter of the second-order or third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate) + elif order == 2: + return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate, + solver_type=solver_type, r1=r1) + elif order == 3: + return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate, + solver_type=solver_type, r1=r1, r2=r2) + else: + raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order)) + + def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'): + """ + Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1]) + elif order == 2: + return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + elif order == 3: + return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + else: + raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order)) + + def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5, + solver_type='dpm_solver'): + """ + The adaptive step size solver based on singlestep DPM-Solver. + Args: + x: A pytorch tensor. The initial value at time `t_T`. + order: A `int`. The (higher) order of the solver. We only support order == 2 or 3. + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + h_init: A `float`. The initial step size (for logSNR). + atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1]. + rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05. + theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1]. + t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the + current time and `t_0` is less than `t_err`. The default setting is 1e-5. + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_0: A pytorch tensor. The approximated solution at time `t_0`. + [1] A. Jolicoeur-Martineau, K. Li, R. PichΓ©-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021. + """ + ns = self.noise_schedule + s = t_T * torch.ones((x.shape[0],)).to(x) + lambda_s = ns.marginal_lambda(s) + lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x)) + h = h_init * torch.ones_like(s).to(x) + x_prev = x + nfe = 0 + if order == 2: + r1 = 0.5 + lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, + solver_type=solver_type, + **kwargs) + elif order == 3: + r1, r2 = 1. / 3., 2. / 3. + lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, + return_intermediate=True, + solver_type=solver_type) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2, + solver_type=solver_type, + **kwargs) + else: + raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order)) + while torch.abs((s - t_0)).mean() > t_err: + t = ns.inverse_lambda(lambda_s + h) + x_lower, lower_noise_kwargs = lower_update(x, s, t) + x_higher = higher_update(x, s, t, **lower_noise_kwargs) + delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev))) + norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)) + E = norm_fn((x_higher - x_lower) / delta).max() + if torch.all(E <= 1.): + x = x_higher + s = t + x_prev = x_lower + lambda_s = ns.marginal_lambda(s) + h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s) + nfe += order + print('adaptive solver nfe', nfe) + return x + + def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform', + method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver', + atol=0.0078, rtol=0.05, + ): + """ + Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`. + ===================================================== + We support the following algorithms for both noise prediction model and data prediction model: + - 'singlestep': + Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver. + We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps). + The total number of function evaluations (NFE) == `steps`. + Given a fixed NFE == `steps`, the sampling procedure is: + - If `order` == 1: + - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2. + - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If `order` == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2. + - 'multistep': + Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`. + We initialize the first `order` values by lower order multistep solvers. + Given a fixed NFE == `steps`, the sampling procedure is: + Denote K = steps. + - If `order` == 1: + - We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2. + - If `order` == 3: + - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3. + - 'singlestep_fixed': + Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3). + We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE. + - 'adaptive': + Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper). + We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`. + You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs + (NFE) and the sample quality. + - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2. + - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3. + ===================================================== + Some advices for choosing the algorithm: + - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs: + Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False) + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3, + skip_type='time_uniform', method='singlestep') + - For **guided sampling with large guidance scale** by DPMs: + Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True) + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2, + skip_type='time_uniform', method='multistep') + We support three types of `skip_type`: + - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images** + - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**. + - 'time_quadratic': quadratic time for the time steps. + ===================================================== + Args: + x: A pytorch tensor. The initial value at time `t_start` + e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution. + steps: A `int`. The total number of function evaluations (NFE). + t_start: A `float`. The starting time of the sampling. + If `T` is None, we use self.noise_schedule.T (default is 1.0). + t_end: A `float`. The ending time of the sampling. + If `t_end` is None, we use 1. / self.noise_schedule.total_N. + e.g. if total_N == 1000, we have `t_end` == 1e-3. + For discrete-time DPMs: + - We recommend `t_end` == 1. / self.noise_schedule.total_N. + For continuous-time DPMs: + - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15. + order: A `int`. The order of DPM-Solver. + skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'. + method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'. + denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step. + Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1). + This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and + score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID + for diffusion models sampling by diffusion SDEs for low-resolutional images + (such as CIFAR-10). However, we observed that such trick does not matter for + high-resolutional images. As it needs an additional NFE, we do not recommend + it for high-resolutional images. + lower_order_final: A `bool`. Whether to use lower order solvers at the final steps. + Only valid for `method=multistep` and `steps < 15`. We empirically find that + this trick is a key to stabilizing the sampling by DPM-Solver with very few steps + (especially for steps <= 10). So we recommend to set it to be `True`. + solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`. + atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + Returns: + x_end: A pytorch tensor. The approximated solution at time `t_end`. + """ + t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end + t_T = self.noise_schedule.T if t_start is None else t_start + device = x.device + if method == 'adaptive': + with torch.no_grad(): + x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol, + solver_type=solver_type) + elif method == 'multistep': + assert steps >= order + timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device) + assert timesteps.shape[0] - 1 == steps + with torch.no_grad(): + vec_t = timesteps[0].expand((x.shape[0])) + model_prev_list = [self.model_fn(x, vec_t)] + t_prev_list = [vec_t] + # Init the first `order` values by lower order multistep DPM-Solver. + for init_order in tqdm(range(1, order), desc="DPM init order"): + vec_t = timesteps[init_order].expand(x.shape[0]) + x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order, + solver_type=solver_type) + model_prev_list.append(self.model_fn(x, vec_t)) + t_prev_list.append(vec_t) + # Compute the remaining values by `order`-th order multistep DPM-Solver. + for step in tqdm(range(order, steps + 1), desc="DPM multistep"): + vec_t = timesteps[step].expand(x.shape[0]) + if lower_order_final and steps < 15: + step_order = min(order, steps + 1 - step) + else: + step_order = order + x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, step_order, + solver_type=solver_type) + for i in range(order - 1): + t_prev_list[i] = t_prev_list[i + 1] + model_prev_list[i] = model_prev_list[i + 1] + t_prev_list[-1] = vec_t + # We do not need to evaluate the final model value. + if step < steps: + model_prev_list[-1] = self.model_fn(x, vec_t) + elif method in ['singlestep', 'singlestep_fixed']: + if method == 'singlestep': + timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order, + skip_type=skip_type, + t_T=t_T, t_0=t_0, + device=device) + elif method == 'singlestep_fixed': + K = steps // order + orders = [order, ] * K + timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device) + for i, order in enumerate(orders): + t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1] + timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(), + N=order, device=device) + lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner) + vec_s, vec_t = t_T_inner.tile(x.shape[0]), t_0_inner.tile(x.shape[0]) + h = lambda_inner[-1] - lambda_inner[0] + r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h + r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h + x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2) + if denoise_to_zero: + x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0) + return x + + +############################################################# +# other utility functions +############################################################# + +def interpolate_fn(x, xp, yp): + """ + A piecewise linear function y = f(x), using xp and yp as keypoints. + We implement f(x) in a differentiable way (i.e. applicable for autograd). + The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.) + Args: + x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver). + xp: PyTorch tensor with shape [C, K], where K is the number of keypoints. + yp: PyTorch tensor with shape [C, K]. + Returns: + The function values f(x), with shape [N, C]. + """ + N, K = x.shape[0], xp.shape[1] + all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2) + sorted_all_x, x_indices = torch.sort(all_x, dim=2) + x_idx = torch.argmin(x_indices, dim=2) + cand_start_idx = x_idx - 1 + start_idx = torch.where( + torch.eq(x_idx, 0), + torch.tensor(1, device=x.device), + torch.where( + torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx, + ), + ) + end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1) + start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2) + end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2) + start_idx2 = torch.where( + torch.eq(x_idx, 0), + torch.tensor(0, device=x.device), + torch.where( + torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx, + ), + ) + y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1) + start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2) + end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2) + cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x) + return cand + + +def expand_dims(v, dims): + """ + Expand the tensor `v` to the dim `dims`. + Args: + `v`: a PyTorch tensor with shape [N]. + `dim`: a `int`. + Returns: + a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`. + """ + return v[(...,) + (None,) * (dims - 1)] \ No newline at end of file diff --git a/ldm/models/diffusion/dpm_solver/sampler.py b/ldm/models/diffusion/dpm_solver/sampler.py new file mode 100644 index 0000000000000000000000000000000000000000..7d137b8cf36718c1c58faa09f9dd919e5fb2977b --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/sampler.py @@ -0,0 +1,87 @@ +"""SAMPLING ONLY.""" +import torch + +from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver + + +MODEL_TYPES = { + "eps": "noise", + "v": "v" +} + + +class DPMSolverSampler(object): + def __init__(self, model, **kwargs): + super().__init__() + self.model = model + to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device) + self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod)) + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + + print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}') + + device = self.model.betas.device + if x_T is None: + img = torch.randn(size, device=device) + else: + img = x_T + + ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod) + + model_fn = model_wrapper( + lambda x, t, c: self.model.apply_model(x, t, c), + ns, + model_type=MODEL_TYPES[self.model.parameterization], + guidance_type="classifier-free", + condition=conditioning, + unconditional_condition=unconditional_conditioning, + guidance_scale=unconditional_guidance_scale, + ) + + dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False) + x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, lower_order_final=True) + + return x.to(device), None \ No newline at end of file diff --git a/ldm/models/diffusion/plms.py b/ldm/models/diffusion/plms.py new file mode 100644 index 0000000000000000000000000000000000000000..7002a365d27168ced0a04e9a4d83e088f8284eae --- /dev/null +++ b/ldm/models/diffusion/plms.py @@ -0,0 +1,244 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm +from functools import partial + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like +from ldm.models.diffusion.sampling_util import norm_thresholding + + +class PLMSSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + if ddim_eta != 0: + raise ValueError('ddim_eta must be 0 for PLMS') + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + dynamic_threshold=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for PLMS sampling is {size}') + + samples, intermediates = self.plms_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold, + ) + return samples, intermediates + + @torch.no_grad() + def plms_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + dynamic_threshold=None): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running PLMS Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps) + old_eps = [] + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + old_eps=old_eps, t_next=ts_next, + dynamic_threshold=dynamic_threshold) + img, pred_x0, e_t = outs + old_eps.append(e_t) + if len(old_eps) >= 4: + old_eps.pop(0) + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None, + dynamic_threshold=None): + b, *_, device = *x.shape, x.device + + def get_model_output(x, t): + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + e_t = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + c_in = torch.cat([unconditional_conditioning, c]) + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + return e_t + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + + def get_x_prev_and_pred_x0(e_t, index): + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + if dynamic_threshold is not None: + pred_x0 = norm_thresholding(pred_x0, dynamic_threshold) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + e_t = get_model_output(x, t) + if len(old_eps) == 0: + # Pseudo Improved Euler (2nd order) + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index) + e_t_next = get_model_output(x_prev, t_next) + e_t_prime = (e_t + e_t_next) / 2 + elif len(old_eps) == 1: + # 2nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (3 * e_t - old_eps[-1]) / 2 + elif len(old_eps) == 2: + # 3nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12 + elif len(old_eps) >= 3: + # 4nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24 + + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index) + + return x_prev, pred_x0, e_t diff --git a/ldm/models/diffusion/sampling_util.py b/ldm/models/diffusion/sampling_util.py new file mode 100644 index 0000000000000000000000000000000000000000..7eff02be6d7c54d43ee6680636ac0698dd3b3f33 --- /dev/null +++ b/ldm/models/diffusion/sampling_util.py @@ -0,0 +1,22 @@ +import torch +import numpy as np + + +def append_dims(x, target_dims): + """Appends dimensions to the end of a tensor until it has target_dims dimensions. + From https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/utils.py""" + dims_to_append = target_dims - x.ndim + if dims_to_append < 0: + raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less') + return x[(...,) + (None,) * dims_to_append] + + +def norm_thresholding(x0, value): + s = append_dims(x0.pow(2).flatten(1).mean(1).sqrt().clamp(min=value), x0.ndim) + return x0 * (value / s) + + +def spatial_norm_thresholding(x0, value): + # b c h w + s = x0.pow(2).mean(1, keepdim=True).sqrt().clamp(min=value) + return x0 * (value / s) \ No newline at end of file diff --git a/ldm/modules/attention.py b/ldm/modules/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..509cd873768f0dd75a75ab3fcdd652822b12b59f --- /dev/null +++ b/ldm/modules/attention.py @@ -0,0 +1,341 @@ +from inspect import isfunction +import math +import torch +import torch.nn.functional as F +from torch import nn, einsum +from einops import rearrange, repeat +from typing import Optional, Any + +from ldm.modules.diffusionmodules.util import checkpoint + + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + +# CrossAttn precision handling +import os +_ATTN_PRECISION = os.environ.get("ATTN_PRECISION", "fp32") + +def exists(val): + return val is not None + + +def uniq(arr): + return{el: True for el in arr}.keys() + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def max_neg_value(t): + return -torch.finfo(t.dtype).max + + +def init_(tensor): + dim = tensor.shape[-1] + std = 1 / math.sqrt(dim) + tensor.uniform_(-std, std) + return tensor + + +# feedforward +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + nn.Linear(dim, inner_dim), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def Normalize(in_channels): + return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = rearrange(q, 'b c h w -> b (h w) c') + k = rearrange(k, 'b c h w -> b c (h w)') + w_ = torch.einsum('bij,bjk->bik', q, k) + + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, 'b c h w -> b c (h w)') + w_ = rearrange(w_, 'b i j -> b j i') + h_ = torch.einsum('bij,bjk->bik', v, w_) + h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h) + h_ = self.proj_out(h_) + + return x+h_ + + +class CrossAttention(nn.Module): + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.scale = dim_head ** -0.5 + self.heads = heads + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), + nn.Dropout(dropout) + ) + + def forward(self, x, context=None, mask=None): + h = self.heads + + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + # force cast to fp32 to avoid overflowing + if _ATTN_PRECISION =="fp32": + with torch.autocast(enabled=False, device_type = 'cuda'): + q, k = q.float(), k.float() + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + else: + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + + del q, k + + if exists(mask): + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + sim = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', sim, v) + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + return self.to_out(out) + + +class MemoryEfficientCrossAttention(nn.Module): + # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0): + super().__init__() + print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using " + f"{heads} heads.") + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.heads = heads + self.dim_head = dim_head + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)) + self.attention_op: Optional[Any] = None + + def forward(self, x, context=None, mask=None): + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + + # actually compute the attention, what we cannot get enough of + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + return self.to_out(out) + + +class BasicTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention + } + def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, + disable_self_attn=False): + super().__init__() + attn_mode = "softmax-xformers" if XFORMERS_IS_AVAILBLE else "softmax" + assert attn_mode in self.ATTENTION_MODES + attn_cls = self.ATTENTION_MODES[attn_mode] + self.disable_self_attn = disable_self_attn + self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, + context_dim=context_dim if self.disable_self_attn else None) # is a self-attention if not self.disable_self_attn + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, + heads=n_heads, dim_head=d_head, dropout=dropout) # is self-attn if context is none + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.norm3 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + + def forward(self, x, context=None): + return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint) + + def _forward(self, x, context=None): + x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x + x = self.attn2(self.norm2(x), context=context) + x + x = self.ff(self.norm3(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data. + First, project the input (aka embedding) + and reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + NEW: use_linear for more efficiency instead of the 1x1 convs + """ + def __init__(self, in_channels, n_heads, d_head, + depth=1, dropout=0., context_dim=None, + disable_self_attn=False, use_linear=False, + use_checkpoint=True): + super().__init__() + if exists(context_dim) and not isinstance(context_dim, list): + context_dim = [context_dim] + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = Normalize(in_channels) + if not use_linear: + self.proj_in = nn.Conv2d(in_channels, + inner_dim, + kernel_size=1, + stride=1, + padding=0) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + self.transformer_blocks = nn.ModuleList( + [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d], + disable_self_attn=disable_self_attn, checkpoint=use_checkpoint) + for d in range(depth)] + ) + if not use_linear: + self.proj_out = zero_module(nn.Conv2d(inner_dim, + in_channels, + kernel_size=1, + stride=1, + padding=0)) + else: + self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) + self.use_linear = use_linear + + def forward(self, x, context=None): + # note: if no context is given, cross-attention defaults to self-attention + if not isinstance(context, list): + context = [context] + b, c, h, w = x.shape + x_in = x + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, 'b c h w -> b (h w) c').contiguous() + if self.use_linear: + x = self.proj_in(x) + for i, block in enumerate(self.transformer_blocks): + x = block(x, context=context[i]) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous() + if not self.use_linear: + x = self.proj_out(x) + return x + x_in + diff --git a/ldm/modules/diffusionmodules/__init__.py b/ldm/modules/diffusionmodules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/diffusionmodules/model.py b/ldm/modules/diffusionmodules/model.py new file mode 100644 index 0000000000000000000000000000000000000000..b089eebbe1676d8249005bb9def002ff5180715b --- /dev/null +++ b/ldm/modules/diffusionmodules/model.py @@ -0,0 +1,852 @@ +# pytorch_diffusion + derived encoder decoder +import math +import torch +import torch.nn as nn +import numpy as np +from einops import rearrange +from typing import Optional, Any + +from ldm.modules.attention import MemoryEfficientCrossAttention + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + print("No module 'xformers'. Proceeding without it.") + + +def get_timestep_embedding(timesteps, embedding_dim): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: + From Fairseq. + Build sinusoidal embeddings. + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + assert len(timesteps.shape) == 1 + + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) + emb = emb.to(device=timesteps.device) + emb = timesteps.float()[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0,1,0,0)) + return emb + + +def nonlinearity(x): + # swish + return x*torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32): + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + return x + + +class Downsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=2, + padding=0) + + def forward(self, x): + if self.with_conv: + pad = (0,1,0,1) + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + return x + + +class ResnetBlock(nn.Module): + def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, + dropout, temb_channels=512): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels) + self.conv1 = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if temb_channels > 0: + self.temb_proj = torch.nn.Linear(temb_channels, + out_channels) + self.norm2 = Normalize(out_channels) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d(out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + else: + self.nin_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x+h + + +class AttnBlock(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = q.reshape(b,c,h*w) + q = q.permute(0,2,1) # b,hw,c + k = k.reshape(b,c,h*w) # b,c,hw + w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b,c,h*w) + w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b,c,h,w) + + h_ = self.proj_out(h_) + + return x+h_ + +class MemoryEfficientAttnBlock(nn.Module): + """ + Uses xformers efficient implementation, + see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + Note: this is a single-head self-attention operation + """ + # + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.attention_op: Optional[Any] = None + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + B, C, H, W = q.shape + q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v)) + + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(B, t.shape[1], 1, C) + .permute(0, 2, 1, 3) + .reshape(B * 1, t.shape[1], C) + .contiguous(), + (q, k, v), + ) + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + out = ( + out.unsqueeze(0) + .reshape(B, 1, out.shape[1], C) + .permute(0, 2, 1, 3) + .reshape(B, out.shape[1], C) + ) + out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C) + out = self.proj_out(out) + return x+out + + +class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention): + def forward(self, x, context=None, mask=None): + b, c, h, w = x.shape + x = rearrange(x, 'b c h w -> b (h w) c') + out = super().forward(x, context=context, mask=mask) + out = rearrange(out, 'b (h w) c -> b c h w', h=h, w=w, c=c) + return x + out + + +def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None): + assert attn_type in ["vanilla", "vanilla-xformers", "memory-efficient-cross-attn", "linear", "none"], f'attn_type {attn_type} unknown' + if XFORMERS_IS_AVAILBLE and attn_type == "vanilla": + attn_type = "vanilla-xformers" + print(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + assert attn_kwargs is None + return AttnBlock(in_channels) + elif attn_type == "vanilla-xformers": + print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...") + return MemoryEfficientAttnBlock(in_channels) + elif type == "memory-efficient-cross-attn": + attn_kwargs["query_dim"] = in_channels + return MemoryEfficientCrossAttentionWrapper(**attn_kwargs) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + raise NotImplementedError() + + +class Model(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = self.ch*4 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + self.use_timestep = use_timestep + if self.use_timestep: + # timestep embedding + self.temb = nn.Module() + self.temb.dense = nn.ModuleList([ + torch.nn.Linear(self.ch, + self.temb_ch), + torch.nn.Linear(self.temb_ch, + self.temb_ch), + ]) + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + skip_in = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + if i_block == self.num_res_blocks: + skip_in = ch*in_ch_mult[i_level] + block.append(ResnetBlock(in_channels=block_in+skip_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x, t=None, context=None): + #assert x.shape[2] == x.shape[3] == self.resolution + if context is not None: + # assume aligned context, cat along channel axis + x = torch.cat((x, context), dim=1) + if self.use_timestep: + # timestep embedding + assert t is not None + temb = get_timestep_embedding(t, self.ch) + temb = self.temb.dense[0](temb) + temb = nonlinearity(temb) + temb = self.temb.dense[1](temb) + else: + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block]( + torch.cat([h, hs.pop()], dim=1), temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + def get_last_layer(self): + return self.conv_out.weight + + +class Encoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla", + **ignore_kwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + 2*z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # timestep embedding + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class Decoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, + attn_type="vanilla", **ignorekwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,)+tuple(ch_mult) + block_in = ch*ch_mult[self.num_resolutions-1] + curr_res = resolution // 2**(self.num_resolutions-1) + self.z_shape = (1,z_channels,curr_res,curr_res) + print("Working with z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape))) + + # z to block_in + self.conv_in = torch.nn.Conv2d(z_channels, + block_in, + kernel_size=3, + stride=1, + padding=1) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, z): + #assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + + # middle + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block](h, temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + if self.tanh_out: + h = torch.tanh(h) + return h + + +class SimpleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, *args, **kwargs): + super().__init__() + self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1), + ResnetBlock(in_channels=in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=2 * in_channels, + out_channels=4 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=4 * in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + nn.Conv2d(2*in_channels, in_channels, 1), + Upsample(in_channels, with_conv=True)]) + # end + self.norm_out = Normalize(in_channels) + self.conv_out = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + for i, layer in enumerate(self.model): + if i in [1,2,3]: + x = layer(x, None) + else: + x = layer(x) + + h = self.norm_out(x) + h = nonlinearity(h) + x = self.conv_out(h) + return x + + +class UpsampleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, + ch_mult=(2,2), dropout=0.0): + super().__init__() + # upsampling + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + block_in = in_channels + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.res_blocks = nn.ModuleList() + self.upsample_blocks = nn.ModuleList() + for i_level in range(self.num_resolutions): + res_block = [] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + res_block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + self.res_blocks.append(nn.ModuleList(res_block)) + if i_level != self.num_resolutions - 1: + self.upsample_blocks.append(Upsample(block_in, True)) + curr_res = curr_res * 2 + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # upsampling + h = x + for k, i_level in enumerate(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.res_blocks[i_level][i_block](h, None) + if i_level != self.num_resolutions - 1: + h = self.upsample_blocks[k](h) + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class LatentRescaler(nn.Module): + def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2): + super().__init__() + # residual block, interpolate, residual block + self.factor = factor + self.conv_in = nn.Conv2d(in_channels, + mid_channels, + kernel_size=3, + stride=1, + padding=1) + self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + self.attn = AttnBlock(mid_channels) + self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + + self.conv_out = nn.Conv2d(mid_channels, + out_channels, + kernel_size=1, + ) + + def forward(self, x): + x = self.conv_in(x) + for block in self.res_block1: + x = block(x, None) + x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor)))) + x = self.attn(x) + for block in self.res_block2: + x = block(x, None) + x = self.conv_out(x) + return x + + +class MergedRescaleEncoder(nn.Module): + def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, + ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + intermediate_chn = ch * ch_mult[-1] + self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult, + z_channels=intermediate_chn, double_z=False, resolution=resolution, + attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv, + out_ch=None) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn, + mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth) + + def forward(self, x): + x = self.encoder(x) + x = self.rescaler(x) + return x + + +class MergedRescaleDecoder(nn.Module): + def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8), + dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + tmp_chn = z_channels*ch_mult[-1] + self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout, + resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks, + ch_mult=ch_mult, resolution=resolution, ch=ch) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn, + out_channels=tmp_chn, depth=rescale_module_depth) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Upsampler(nn.Module): + def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2): + super().__init__() + assert out_size >= in_size + num_blocks = int(np.log2(out_size//in_size))+1 + factor_up = 1.+ (out_size % in_size) + print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}") + self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels, + out_channels=in_channels) + self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2, + attn_resolutions=[], in_channels=None, ch=in_channels, + ch_mult=[ch_mult for _ in range(num_blocks)]) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Resize(nn.Module): + def __init__(self, in_channels=None, learned=False, mode="bilinear"): + super().__init__() + self.with_conv = learned + self.mode = mode + if self.with_conv: + print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode") + raise NotImplementedError() + assert in_channels is not None + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=4, + stride=2, + padding=1) + + def forward(self, x, scale_factor=1.0): + if scale_factor==1.0: + return x + else: + x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor) + return x diff --git a/ldm/modules/diffusionmodules/openaimodel.py b/ldm/modules/diffusionmodules/openaimodel.py new file mode 100644 index 0000000000000000000000000000000000000000..9633c38707900b50830d24ca20cb33834be1de17 --- /dev/null +++ b/ldm/modules/diffusionmodules/openaimodel.py @@ -0,0 +1,786 @@ +from abc import abstractmethod +import math + +import numpy as np +import torch as th +import torch.nn as nn +import torch.nn.functional as F + +from ldm.modules.diffusionmodules.util import ( + checkpoint, + conv_nd, + linear, + avg_pool_nd, + zero_module, + normalization, + timestep_embedding, +) +from ldm.modules.attention import SpatialTransformer +from ldm.util import exists + + +# dummy replace +def convert_module_to_f16(x): + pass + +def convert_module_to_f32(x): + pass + + +## go +class AttentionPool2d(nn.Module): + """ + Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py + """ + + def __init__( + self, + spacial_dim: int, + embed_dim: int, + num_heads_channels: int, + output_dim: int = None, + ): + super().__init__() + self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5) + self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1) + self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1) + self.num_heads = embed_dim // num_heads_channels + self.attention = QKVAttention(self.num_heads) + + def forward(self, x): + b, c, *_spatial = x.shape + x = x.reshape(b, c, -1) # NC(HW) + x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1) + x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1) + x = self.qkv_proj(x) + x = self.attention(x) + x = self.c_proj(x) + return x[:, :, 0] + + +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + + @abstractmethod + def forward(self, x, emb): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + """ + A sequential module that passes timestep embeddings to the children that + support it as an extra input. + """ + + def forward(self, x, emb, context=None, check=False): + for layer in self: + if isinstance(layer, TimestepBlock): + x = layer(x, emb) + elif isinstance(layer, SpatialTransformer): + x = layer(x, context) + else: + x = layer(x) + return x + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + if use_conv: + self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding) + + def forward(self, x): + assert x.shape[1] == self.channels + if self.dims == 3: + x = F.interpolate( + x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" + ) + else: + x = F.interpolate(x, scale_factor=2, mode="nearest") + if self.use_conv: + x = self.conv(x) + return x + +class TransposedUpsample(nn.Module): + 'Learned 2x upsampling without padding' + def __init__(self, channels, out_channels=None, ks=5): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + + self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2) + + def forward(self,x): + return self.up(x) + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd( + dims, self.channels, self.out_channels, 3, stride=stride, padding=padding + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x, emb): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + return checkpoint( + self._forward, (x, emb), self.parameters(), self.use_checkpoint + ) + + + def _forward(self, x, emb): + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class AttentionBlock(nn.Module): + """ + An attention block that allows spatial positions to attend to each other. + Originally ported from here, but adapted to the N-d case. + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. + """ + + def __init__( + self, + channels, + num_heads=1, + num_head_channels=-1, + use_checkpoint=False, + use_new_attention_order=False, + ): + super().__init__() + self.channels = channels + if num_head_channels == -1: + self.num_heads = num_heads + else: + assert ( + channels % num_head_channels == 0 + ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" + self.num_heads = channels // num_head_channels + self.use_checkpoint = use_checkpoint + self.norm = normalization(channels) + self.qkv = conv_nd(1, channels, channels * 3, 1) + if use_new_attention_order: + # split qkv before split heads + self.attention = QKVAttention(self.num_heads) + else: + # split heads before split qkv + self.attention = QKVAttentionLegacy(self.num_heads) + + self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) + + def forward(self, x): + return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! + #return pt_checkpoint(self._forward, x) # pytorch + + def _forward(self, x): + b, c, *spatial = x.shape + x = x.reshape(b, c, -1) + qkv = self.qkv(self.norm(x)) + h = self.attention(qkv) + h = self.proj_out(h) + return (x + h).reshape(b, c, *spatial) + + +def count_flops_attn(model, _x, y): + """ + A counter for the `thop` package to count the operations in an + attention operation. + Meant to be used like: + macs, params = thop.profile( + model, + inputs=(inputs, timestamps), + custom_ops={QKVAttention: QKVAttention.count_flops}, + ) + """ + b, c, *spatial = y[0].shape + num_spatial = int(np.prod(spatial)) + # We perform two matmuls with the same number of ops. + # The first computes the weight matrix, the second computes + # the combination of the value vectors. + matmul_ops = 2 * b * (num_spatial ** 2) * c + model.total_ops += th.DoubleTensor([matmul_ops]) + + +class QKVAttentionLegacy(nn.Module): + """ + A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", q * scale, k * scale + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class QKVAttention(nn.Module): + """ + A module which performs QKV attention and splits in a different order. + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.chunk(3, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", + (q * scale).view(bs * self.n_heads, ch, length), + (k * scale).view(bs * self.n_heads, ch, length), + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + :param in_channels: channels in the input Tensor. + :param model_channels: base channel count for the model. + :param out_channels: channels in the output Tensor. + :param num_res_blocks: number of residual blocks per downsample. + :param attention_resolutions: a collection of downsample rates at which + attention will take place. May be a set, list, or tuple. + For example, if this contains 4, then at 4x downsampling, attention + will be used. + :param dropout: the dropout probability. + :param channel_mult: channel multiplier for each level of the UNet. + :param conv_resample: if True, use learned convolutions for upsampling and + downsampling. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param num_classes: if specified (as an int), then this model will be + class-conditional with `num_classes` classes. + :param use_checkpoint: use gradient checkpointing to reduce memory usage. + :param num_heads: the number of attention heads in each attention layer. + :param num_heads_channels: if specified, ignore num_heads and instead use + a fixed channel width per attention head. + :param num_heads_upsample: works with num_heads to set a different number + of heads for upsampling. Deprecated. + :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. + :param resblock_updown: use residual blocks for up/downsampling. + :param use_new_attention_order: use a different attention pattern for potentially + increased efficiency. + """ + + def __init__( + self, + image_size, + in_channels, + model_channels, + out_channels, + num_res_blocks, + attention_resolutions, + dropout=0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + num_classes=None, + use_checkpoint=False, + use_fp16=False, + num_heads=-1, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False, + resblock_updown=False, + use_new_attention_order=False, + use_spatial_transformer=False, # custom transformer support + transformer_depth=1, # custom transformer support + context_dim=None, # custom transformer support + n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model + legacy=True, + disable_self_attentions=None, + num_attention_blocks=None, + disable_middle_self_attn=False, + use_linear_in_transformer=False, + ): + super().__init__() + if use_spatial_transformer: + assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' + + if context_dim is not None: + assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' + from omegaconf.listconfig import ListConfig + if type(context_dim) == ListConfig: + context_dim = list(context_dim) + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' + + if num_head_channels == -1: + assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' + + self.image_size = image_size + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + if isinstance(num_res_blocks, int): + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + else: + if len(num_res_blocks) != len(channel_mult): + raise ValueError("provide num_res_blocks either as an int (globally constant) or " + "as a list/tuple (per-level) with the same length as channel_mult") + self.num_res_blocks = num_res_blocks + if disable_self_attentions is not None: + # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not + assert len(disable_self_attentions) == len(channel_mult) + if num_attention_blocks is not None: + assert len(num_attention_blocks) == len(self.num_res_blocks) + assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) + print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " + f"This option has LESS priority than attention_resolutions {attention_resolutions}, " + f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " + f"attention will still not be set.") + + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.num_classes = num_classes + self.use_checkpoint = use_checkpoint + self.dtype = th.float16 if use_fp16 else th.float32 + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.predict_codebook_ids = n_embed is not None + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.num_classes is not None: + if isinstance(self.num_classes, int): + self.label_emb = nn.Embedding(num_classes, time_embed_dim) + elif self.num_classes == "continuous": + print("setting up linear c_adm embedding layer") + self.label_emb = nn.Linear(1, time_embed_dim) + else: + raise ValueError() + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(self.num_res_blocks[level] + 1): + ich = input_block_chans.pop() + layers = [ + ResBlock( + ch + ich, + time_embed_dim, + dropout, + out_channels=model_channels * mult, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or i < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads_upsample, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + if level and i == self.num_res_blocks[level]: + out_ch = ch + layers.append( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + if self.predict_codebook_ids: + self.id_predictor = nn.Sequential( + normalization(ch), + conv_nd(dims, model_channels, n_embed, 1), + #nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits + ) + + def convert_to_fp16(self): + """ + Convert the torso of the model to float16. + """ + self.input_blocks.apply(convert_module_to_f16) + self.middle_block.apply(convert_module_to_f16) + self.output_blocks.apply(convert_module_to_f16) + + def convert_to_fp32(self): + """ + Convert the torso of the model to float32. + """ + self.input_blocks.apply(convert_module_to_f32) + self.middle_block.apply(convert_module_to_f32) + self.output_blocks.apply(convert_module_to_f32) + + def forward(self, x, timesteps=None, context=None, y=None,**kwargs): + """ + Apply the model to an input batch. + :param x: an [N x C x ...] Tensor of inputs. + :param timesteps: a 1-D batch of timesteps. + :param context: conditioning plugged in via crossattn + :param y: an [N] Tensor of labels, if class-conditional. + :return: an [N x C x ...] Tensor of outputs. + """ + assert (y is not None) == ( + self.num_classes is not None + ), "must specify y if and only if the model is class-conditional" + hs = [] + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + if self.num_classes is not None: + assert y.shape[0] == x.shape[0] + emb = emb + self.label_emb(y) + + h = x.type(self.dtype) + for module in self.input_blocks: + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb, context) + h = h.type(x.dtype) + if self.predict_codebook_ids: + return self.id_predictor(h) + else: + return self.out(h) diff --git a/ldm/modules/diffusionmodules/upscaling.py b/ldm/modules/diffusionmodules/upscaling.py new file mode 100644 index 0000000000000000000000000000000000000000..03816662098ce1ffac79bd939b892e867ab91988 --- /dev/null +++ b/ldm/modules/diffusionmodules/upscaling.py @@ -0,0 +1,81 @@ +import torch +import torch.nn as nn +import numpy as np +from functools import partial + +from ldm.modules.diffusionmodules.util import extract_into_tensor, make_beta_schedule +from ldm.util import default + + +class AbstractLowScaleModel(nn.Module): + # for concatenating a downsampled image to the latent representation + def __init__(self, noise_schedule_config=None): + super(AbstractLowScaleModel, self).__init__() + if noise_schedule_config is not None: + self.register_schedule(**noise_schedule_config) + + def register_schedule(self, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def forward(self, x): + return x, None + + def decode(self, x): + return x + + +class SimpleImageConcat(AbstractLowScaleModel): + # no noise level conditioning + def __init__(self): + super(SimpleImageConcat, self).__init__(noise_schedule_config=None) + self.max_noise_level = 0 + + def forward(self, x): + # fix to constant noise level + return x, torch.zeros(x.shape[0], device=x.device).long() + + +class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel): + def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False): + super().__init__(noise_schedule_config=noise_schedule_config) + self.max_noise_level = max_noise_level + + def forward(self, x, noise_level=None): + if noise_level is None: + noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long() + else: + assert isinstance(noise_level, torch.Tensor) + z = self.q_sample(x, noise_level) + return z, noise_level + + + diff --git a/ldm/modules/diffusionmodules/util.py b/ldm/modules/diffusionmodules/util.py new file mode 100644 index 0000000000000000000000000000000000000000..637363dfe34799e70cfdbcd11445212df9d9ca1f --- /dev/null +++ b/ldm/modules/diffusionmodules/util.py @@ -0,0 +1,270 @@ +# adopted from +# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# and +# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +# and +# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py +# +# thanks! + + +import os +import math +import torch +import torch.nn as nn +import numpy as np +from einops import repeat + +from ldm.util import instantiate_from_config + + +def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if schedule == "linear": + betas = ( + torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2 + ) + + elif schedule == "cosine": + timesteps = ( + torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s + ) + alphas = timesteps / (1 + cosine_s) * np.pi / 2 + alphas = torch.cos(alphas).pow(2) + alphas = alphas / alphas[0] + betas = 1 - alphas[1:] / alphas[:-1] + betas = np.clip(betas, a_min=0, a_max=0.999) + + elif schedule == "sqrt_linear": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) + elif schedule == "sqrt": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5 + else: + raise ValueError(f"schedule '{schedule}' unknown.") + return betas.numpy() + + +def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True): + if ddim_discr_method == 'uniform': + c = num_ddpm_timesteps // num_ddim_timesteps + ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c))) + elif ddim_discr_method == 'quad': + ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int) + else: + raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"') + + # assert ddim_timesteps.shape[0] == num_ddim_timesteps + # add one to get the final alpha values right (the ones from first scale to data during sampling) + steps_out = ddim_timesteps + 1 + if verbose: + print(f'Selected timesteps for ddim sampler: {steps_out}') + return steps_out + + +def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True): + # select alphas for computing the variance schedule + alphas = alphacums[ddim_timesteps] + alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist()) + + # according the the formula provided in https://arxiv.org/abs/2010.02502 + sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)) + if verbose: + print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}') + print(f'For the chosen value of eta, which is {eta}, ' + f'this results in the following sigma_t schedule for ddim sampler {sigmas}') + return sigmas, alphas, alphas_prev + + +def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, + which defines the cumulative product of (1-beta) over time from t = [0,1]. + :param num_diffusion_timesteps: the number of betas to produce. + :param alpha_bar: a lambda that takes an argument t from 0 to 1 and + produces the cumulative product of (1-beta) up to that + part of the diffusion process. + :param max_beta: the maximum beta to use; use values lower than 1 to + prevent singularities. + """ + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return np.array(betas) + + +def extract_into_tensor(a, t, x_shape): + b, *_ = t.shape + out = a.gather(-1, t) + return out.reshape(b, *((1,) * (len(x_shape) - 1))) + + +def checkpoint(func, inputs, params, flag): + """ + Evaluate a function without caching intermediate activations, allowing for + reduced memory at the expense of extra compute in the backward pass. + :param func: the function to evaluate. + :param inputs: the argument sequence to pass to `func`. + :param params: a sequence of parameters `func` depends on but does not + explicitly take as arguments. + :param flag: if False, disable gradient checkpointing. + """ + if flag: + args = tuple(inputs) + tuple(params) + return CheckpointFunction.apply(func, len(inputs), *args) + else: + return func(*inputs) + + +class CheckpointFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, run_function, length, *args): + ctx.run_function = run_function + ctx.input_tensors = list(args[:length]) + ctx.input_params = list(args[length:]) + ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(), + "dtype": torch.get_autocast_gpu_dtype(), + "cache_enabled": torch.is_autocast_cache_enabled()} + with torch.no_grad(): + output_tensors = ctx.run_function(*ctx.input_tensors) + return output_tensors + + @staticmethod + def backward(ctx, *output_grads): + ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors] + with torch.enable_grad(), \ + torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs): + # Fixes a bug where the first op in run_function modifies the + # Tensor storage in place, which is not allowed for detach()'d + # Tensors. + shallow_copies = [x.view_as(x) for x in ctx.input_tensors] + output_tensors = ctx.run_function(*shallow_copies) + input_grads = torch.autograd.grad( + output_tensors, + ctx.input_tensors + ctx.input_params, + output_grads, + allow_unused=True, + ) + del ctx.input_tensors + del ctx.input_params + del output_tensors + return (None, None) + input_grads + + +def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False): + """ + Create sinusoidal timestep embeddings. + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an [N x dim] Tensor of positional embeddings. + """ + if not repeat_only: + half = dim // 2 + freqs = torch.exp( + -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half + ).to(device=timesteps.device) + args = timesteps[:, None].float() * freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) + else: + embedding = repeat(timesteps, 'b -> b d', d=dim) + return embedding + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def scale_module(module, scale): + """ + Scale the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().mul_(scale) + return module + + +def mean_flat(tensor): + """ + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def normalization(channels): + """ + Make a standard normalization layer. + :param channels: number of input channels. + :return: an nn.Module for normalization. + """ + return GroupNorm32(32, channels) + + +# PyTorch 1.7 has SiLU, but we support PyTorch 1.5. +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + + +class GroupNorm32(nn.GroupNorm): + def forward(self, x): + return super().forward(x.float()).type(x.dtype) + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def linear(*args, **kwargs): + """ + Create a linear module. + """ + return nn.Linear(*args, **kwargs) + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +class HybridConditioner(nn.Module): + + def __init__(self, c_concat_config, c_crossattn_config): + super().__init__() + self.concat_conditioner = instantiate_from_config(c_concat_config) + self.crossattn_conditioner = instantiate_from_config(c_crossattn_config) + + def forward(self, c_concat, c_crossattn): + c_concat = self.concat_conditioner(c_concat) + c_crossattn = self.crossattn_conditioner(c_crossattn) + return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]} + + +def noise_like(shape, device, repeat=False): + repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1))) + noise = lambda: torch.randn(shape, device=device) + return repeat_noise() if repeat else noise() \ No newline at end of file diff --git a/ldm/modules/distributions/__init__.py b/ldm/modules/distributions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/distributions/distributions.py b/ldm/modules/distributions/distributions.py new file mode 100644 index 0000000000000000000000000000000000000000..f2b8ef901130efc171aa69742ca0244d94d3f2e9 --- /dev/null +++ b/ldm/modules/distributions/distributions.py @@ -0,0 +1,92 @@ +import torch +import numpy as np + + +class AbstractDistribution: + def sample(self): + raise NotImplementedError() + + def mode(self): + raise NotImplementedError() + + +class DiracDistribution(AbstractDistribution): + def __init__(self, value): + self.value = value + + def sample(self): + return self.value + + def mode(self): + return self.value + + +class DiagonalGaussianDistribution(object): + def __init__(self, parameters, deterministic=False): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device) + + def sample(self): + x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device) + return x + + def kl(self, other=None): + if self.deterministic: + return torch.Tensor([0.]) + else: + if other is None: + return 0.5 * torch.sum(torch.pow(self.mean, 2) + + self.var - 1.0 - self.logvar, + dim=[1, 2, 3]) + else: + return 0.5 * torch.sum( + torch.pow(self.mean - other.mean, 2) / other.var + + self.var / other.var - 1.0 - self.logvar + other.logvar, + dim=[1, 2, 3]) + + def nll(self, sample, dims=[1,2,3]): + if self.deterministic: + return torch.Tensor([0.]) + logtwopi = np.log(2.0 * np.pi) + return 0.5 * torch.sum( + logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, + dim=dims) + + def mode(self): + return self.mean + + +def normal_kl(mean1, logvar1, mean2, logvar2): + """ + source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12 + Compute the KL divergence between two gaussians. + Shapes are automatically broadcasted, so batches can be compared to + scalars, among other use cases. + """ + tensor = None + for obj in (mean1, logvar1, mean2, logvar2): + if isinstance(obj, torch.Tensor): + tensor = obj + break + assert tensor is not None, "at least one argument must be a Tensor" + + # Force variances to be Tensors. Broadcasting helps convert scalars to + # Tensors, but it does not work for torch.exp(). + logvar1, logvar2 = [ + x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor) + for x in (logvar1, logvar2) + ] + + return 0.5 * ( + -1.0 + + logvar2 + - logvar1 + + torch.exp(logvar1 - logvar2) + + ((mean1 - mean2) ** 2) * torch.exp(-logvar2) + ) diff --git a/ldm/modules/ema.py b/ldm/modules/ema.py new file mode 100644 index 0000000000000000000000000000000000000000..bded25019b9bcbcd0260f0b8185f8c7859ca58c4 --- /dev/null +++ b/ldm/modules/ema.py @@ -0,0 +1,80 @@ +import torch +from torch import nn + + +class LitEma(nn.Module): + def __init__(self, model, decay=0.9999, use_num_upates=True): + super().__init__() + if decay < 0.0 or decay > 1.0: + raise ValueError('Decay must be between 0 and 1') + + self.m_name2s_name = {} + self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32)) + self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates + else torch.tensor(-1, dtype=torch.int)) + + for name, p in model.named_parameters(): + if p.requires_grad: + # remove as '.'-character is not allowed in buffers + s_name = name.replace('.', '') + self.m_name2s_name.update({name: s_name}) + self.register_buffer(s_name, p.clone().detach().data) + + self.collected_params = [] + + def reset_num_updates(self): + del self.num_updates + self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int)) + + def forward(self, model): + decay = self.decay + + if self.num_updates >= 0: + self.num_updates += 1 + decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates)) + + one_minus_decay = 1.0 - decay + + with torch.no_grad(): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + + for key in m_param: + if m_param[key].requires_grad: + sname = self.m_name2s_name[key] + shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) + shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) + else: + assert not key in self.m_name2s_name + + def copy_to(self, model): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + for key in m_param: + if m_param[key].requires_grad: + m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) + else: + assert not key in self.m_name2s_name + + def store(self, parameters): + """ + Save the current parameters for restoring later. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + temporarily stored. + """ + self.collected_params = [param.clone() for param in parameters] + + def restore(self, parameters): + """ + Restore the parameters stored with the `store` method. + Useful to validate the model with EMA parameters without affecting the + original optimization process. Store the parameters before the + `copy_to` method. After validation (or model saving), use this to + restore the former parameters. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. + """ + for c_param, param in zip(self.collected_params, parameters): + param.data.copy_(c_param.data) diff --git a/ldm/modules/encoders/__init__.py b/ldm/modules/encoders/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/encoders/modules.py b/ldm/modules/encoders/modules.py new file mode 100644 index 0000000000000000000000000000000000000000..4edd5496b9e668ea72a5be39db9cca94b6a42f9b --- /dev/null +++ b/ldm/modules/encoders/modules.py @@ -0,0 +1,213 @@ +import torch +import torch.nn as nn +from torch.utils.checkpoint import checkpoint + +from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel + +import open_clip +from ldm.util import default, count_params + + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + + +class IdentityEncoder(AbstractEncoder): + + def encode(self, x): + return x + + +class ClassEmbedder(nn.Module): + def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1): + super().__init__() + self.key = key + self.embedding = nn.Embedding(n_classes, embed_dim) + self.n_classes = n_classes + self.ucg_rate = ucg_rate + + def forward(self, batch, key=None, disable_dropout=False): + if key is None: + key = self.key + # this is for use in crossattn + c = batch[key][:, None] + if self.ucg_rate > 0. and not disable_dropout: + mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate) + c = mask * c + (1-mask) * torch.ones_like(c)*(self.n_classes-1) + c = c.long() + c = self.embedding(c) + return c + + def get_unconditional_conditioning(self, bs, device="cuda"): + uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000) + uc = torch.ones((bs,), device=device) * uc_class + uc = {self.key: uc} + return uc + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +class FrozenT5Embedder(AbstractEncoder): + """Uses the T5 transformer encoder for text""" + def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl + super().__init__() + self.tokenizer = T5Tokenizer.from_pretrained(version) + self.transformer = T5EncoderModel.from_pretrained(version) + self.device = device + self.max_length = max_length # TODO: typical value? + if freeze: + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + #self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens) + + z = outputs.last_hidden_state + return z + + def encode(self, text): + return self(text) + + +class FrozenCLIPEmbedder(AbstractEncoder): + """Uses the CLIP transformer encoder for text (from huggingface)""" + LAYERS = [ + "last", + "pooled", + "hidden" + ] + def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, + freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32 + super().__init__() + assert layer in self.LAYERS + self.tokenizer = CLIPTokenizer.from_pretrained(version) + self.transformer = CLIPTextModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + self.layer_idx = layer_idx + if layer == "hidden": + assert layer_idx is not None + assert 0 <= abs(layer_idx) <= 12 + + def freeze(self): + self.transformer = self.transformer.eval() + #self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer=="hidden") + if self.layer == "last": + z = outputs.last_hidden_state + elif self.layer == "pooled": + z = outputs.pooler_output[:, None, :] + else: + z = outputs.hidden_states[self.layer_idx] + return z + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP transformer encoder for text + """ + LAYERS = [ + #"pooled", + "last", + "penultimate" + ] + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="last"): + super().__init__() + assert layer in self.LAYERS + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version) + del model.visual + self.model = model + + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "last": + self.layer_idx = 0 + elif self.layer == "penultimate": + self.layer_idx = 1 + else: + raise NotImplementedError() + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + tokens = open_clip.tokenize(text) + z = self.encode_with_transformer(tokens.to(self.device)) + return z + + def encode_with_transformer(self, text): + x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.model.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.model.ln_final(x) + return x + + def text_transformer_forward(self, x: torch.Tensor, attn_mask = None): + for i, r in enumerate(self.model.transformer.resblocks): + if i == len(self.model.transformer.resblocks) - self.layer_idx: + break + if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting(): + x = checkpoint(r, x, attn_mask) + else: + x = r(x, attn_mask=attn_mask) + return x + + def encode(self, text): + return self(text) + + +class FrozenCLIPT5Encoder(AbstractEncoder): + def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda", + clip_max_length=77, t5_max_length=77): + super().__init__() + self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length) + self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length) + print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder)*1.e-6:.2f} M parameters, " + f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder)*1.e-6:.2f} M params.") + + def encode(self, text): + return self(text) + + def forward(self, text): + clip_z = self.clip_encoder.encode(text) + t5_z = self.t5_encoder.encode(text) + return [clip_z, t5_z] + + diff --git a/ldm/modules/image_degradation/__init__.py b/ldm/modules/image_degradation/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7836cada81f90ded99c58d5942eea4c3477f58fc --- /dev/null +++ b/ldm/modules/image_degradation/__init__.py @@ -0,0 +1,2 @@ +from ldm.modules.image_degradation.bsrgan import degradation_bsrgan_variant as degradation_fn_bsr +from ldm.modules.image_degradation.bsrgan_light import degradation_bsrgan_variant as degradation_fn_bsr_light diff --git a/ldm/modules/image_degradation/bsrgan.py b/ldm/modules/image_degradation/bsrgan.py new file mode 100644 index 0000000000000000000000000000000000000000..32ef56169978e550090261cddbcf5eb611a6173b --- /dev/null +++ b/ldm/modules/image_degradation/bsrgan.py @@ -0,0 +1,730 @@ +# -*- coding: utf-8 -*- +""" +# -------------------------------------------- +# Super-Resolution +# -------------------------------------------- +# +# Kai Zhang (cskaizhang@gmail.com) +# https://github.com/cszn +# From 2019/03--2021/08 +# -------------------------------------------- +""" + +import numpy as np +import cv2 +import torch + +from functools import partial +import random +from scipy import ndimage +import scipy +import scipy.stats as ss +from scipy.interpolate import interp2d +from scipy.linalg import orth +import albumentations + +import ldm.modules.image_degradation.utils_image as util + + +def modcrop_np(img, sf): + ''' + Args: + img: numpy image, WxH or WxHxC + sf: scale factor + Return: + cropped image + ''' + w, h = img.shape[:2] + im = np.copy(img) + return im[:w - w % sf, :h - h % sf, ...] + + +""" +# -------------------------------------------- +# anisotropic Gaussian kernels +# -------------------------------------------- +""" + + +def analytic_kernel(k): + """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)""" + k_size = k.shape[0] + # Calculate the big kernels size + big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2)) + # Loop over the small kernel to fill the big one + for r in range(k_size): + for c in range(k_size): + big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k + # Crop the edges of the big kernel to ignore very small values and increase run time of SR + crop = k_size // 2 + cropped_big_k = big_k[crop:-crop, crop:-crop] + # Normalize to 1 + return cropped_big_k / cropped_big_k.sum() + + +def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6): + """ generate an anisotropic Gaussian kernel + Args: + ksize : e.g., 15, kernel size + theta : [0, pi], rotation angle range + l1 : [0.1,50], scaling of eigenvalues + l2 : [0.1,l1], scaling of eigenvalues + If l1 = l2, will get an isotropic Gaussian kernel. + Returns: + k : kernel + """ + + v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.])) + V = np.array([[v[0], v[1]], [v[1], -v[0]]]) + D = np.array([[l1, 0], [0, l2]]) + Sigma = np.dot(np.dot(V, D), np.linalg.inv(V)) + k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize) + + return k + + +def gm_blur_kernel(mean, cov, size=15): + center = size / 2.0 + 0.5 + k = np.zeros([size, size]) + for y in range(size): + for x in range(size): + cy = y - center + 1 + cx = x - center + 1 + k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov) + + k = k / np.sum(k) + return k + + +def shift_pixel(x, sf, upper_left=True): + """shift pixel for super-resolution with different scale factors + Args: + x: WxHxC or WxH + sf: scale factor + upper_left: shift direction + """ + h, w = x.shape[:2] + shift = (sf - 1) * 0.5 + xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0) + if upper_left: + x1 = xv + shift + y1 = yv + shift + else: + x1 = xv - shift + y1 = yv - shift + + x1 = np.clip(x1, 0, w - 1) + y1 = np.clip(y1, 0, h - 1) + + if x.ndim == 2: + x = interp2d(xv, yv, x)(x1, y1) + if x.ndim == 3: + for i in range(x.shape[-1]): + x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1) + + return x + + +def blur(x, k): + ''' + x: image, NxcxHxW + k: kernel, Nx1xhxw + ''' + n, c = x.shape[:2] + p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2 + x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate') + k = k.repeat(1, c, 1, 1) + k = k.view(-1, 1, k.shape[2], k.shape[3]) + x = x.view(1, -1, x.shape[2], x.shape[3]) + x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c) + x = x.view(n, c, x.shape[2], x.shape[3]) + + return x + + +def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0): + """" + # modified version of https://github.com/assafshocher/BlindSR_dataset_generator + # Kai Zhang + # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var + # max_var = 2.5 * sf + """ + # Set random eigen-vals (lambdas) and angle (theta) for COV matrix + lambda_1 = min_var + np.random.rand() * (max_var - min_var) + lambda_2 = min_var + np.random.rand() * (max_var - min_var) + theta = np.random.rand() * np.pi # random theta + noise = -noise_level + np.random.rand(*k_size) * noise_level * 2 + + # Set COV matrix using Lambdas and Theta + LAMBDA = np.diag([lambda_1, lambda_2]) + Q = np.array([[np.cos(theta), -np.sin(theta)], + [np.sin(theta), np.cos(theta)]]) + SIGMA = Q @ LAMBDA @ Q.T + INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :] + + # Set expectation position (shifting kernel for aligned image) + MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2) + MU = MU[None, None, :, None] + + # Create meshgrid for Gaussian + [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1])) + Z = np.stack([X, Y], 2)[:, :, :, None] + + # Calcualte Gaussian for every pixel of the kernel + ZZ = Z - MU + ZZ_t = ZZ.transpose(0, 1, 3, 2) + raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise) + + # shift the kernel so it will be centered + # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor) + + # Normalize the kernel and return + # kernel = raw_kernel_centered / np.sum(raw_kernel_centered) + kernel = raw_kernel / np.sum(raw_kernel) + return kernel + + +def fspecial_gaussian(hsize, sigma): + hsize = [hsize, hsize] + siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0] + std = sigma + [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1)) + arg = -(x * x + y * y) / (2 * std * std) + h = np.exp(arg) + h[h < scipy.finfo(float).eps * h.max()] = 0 + sumh = h.sum() + if sumh != 0: + h = h / sumh + return h + + +def fspecial_laplacian(alpha): + alpha = max([0, min([alpha, 1])]) + h1 = alpha / (alpha + 1) + h2 = (1 - alpha) / (alpha + 1) + h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]] + h = np.array(h) + return h + + +def fspecial(filter_type, *args, **kwargs): + ''' + python code from: + https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py + ''' + if filter_type == 'gaussian': + return fspecial_gaussian(*args, **kwargs) + if filter_type == 'laplacian': + return fspecial_laplacian(*args, **kwargs) + + +""" +# -------------------------------------------- +# degradation models +# -------------------------------------------- +""" + + +def bicubic_degradation(x, sf=3): + ''' + Args: + x: HxWxC image, [0, 1] + sf: down-scale factor + Return: + bicubicly downsampled LR image + ''' + x = util.imresize_np(x, scale=1 / sf) + return x + + +def srmd_degradation(x, k, sf=3): + ''' blur + bicubic downsampling + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2018learning, + title={Learning a single convolutional super-resolution network for multiple degradations}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={3262--3271}, + year={2018} + } + ''' + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror' + x = bicubic_degradation(x, sf=sf) + return x + + +def dpsr_degradation(x, k, sf=3): + ''' bicubic downsampling + blur + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2019deep, + title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={1671--1681}, + year={2019} + } + ''' + x = bicubic_degradation(x, sf=sf) + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + return x + + +def classical_degradation(x, k, sf=3): + ''' blur + downsampling + Args: + x: HxWxC image, [0, 1]/[0, 255] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + ''' + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2)) + st = 0 + return x[st::sf, st::sf, ...] + + +def add_sharpening(img, weight=0.5, radius=50, threshold=10): + """USM sharpening. borrowed from real-ESRGAN + Input image: I; Blurry image: B. + 1. K = I + weight * (I - B) + 2. Mask = 1 if abs(I - B) > threshold, else: 0 + 3. Blur mask: + 4. Out = Mask * K + (1 - Mask) * I + Args: + img (Numpy array): Input image, HWC, BGR; float32, [0, 1]. + weight (float): Sharp weight. Default: 1. + radius (float): Kernel size of Gaussian blur. Default: 50. + threshold (int): + """ + if radius % 2 == 0: + radius += 1 + blur = cv2.GaussianBlur(img, (radius, radius), 0) + residual = img - blur + mask = np.abs(residual) * 255 > threshold + mask = mask.astype('float32') + soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0) + + K = img + weight * residual + K = np.clip(K, 0, 1) + return soft_mask * K + (1 - soft_mask) * img + + +def add_blur(img, sf=4): + wd2 = 4.0 + sf + wd = 2.0 + 0.2 * sf + if random.random() < 0.5: + l1 = wd2 * random.random() + l2 = wd2 * random.random() + k = anisotropic_Gaussian(ksize=2 * random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2) + else: + k = fspecial('gaussian', 2 * random.randint(2, 11) + 3, wd * random.random()) + img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror') + + return img + + +def add_resize(img, sf=4): + rnum = np.random.rand() + if rnum > 0.8: # up + sf1 = random.uniform(1, 2) + elif rnum < 0.7: # down + sf1 = random.uniform(0.5 / sf, 1) + else: + sf1 = 1.0 + img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + return img + + +# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): +# noise_level = random.randint(noise_level1, noise_level2) +# rnum = np.random.rand() +# if rnum > 0.6: # add color Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) +# elif rnum < 0.4: # add grayscale Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) +# else: # add noise +# L = noise_level2 / 255. +# D = np.diag(np.random.rand(3)) +# U = orth(np.random.rand(3, 3)) +# conv = np.dot(np.dot(np.transpose(U), D), U) +# img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) +# img = np.clip(img, 0.0, 1.0) +# return img + +def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + rnum = np.random.rand() + if rnum > 0.6: # add color Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: # add grayscale Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: # add noise + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_speckle_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + img = np.clip(img, 0.0, 1.0) + rnum = random.random() + if rnum > 0.6: + img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: + img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_Poisson_noise(img): + img = np.clip((img * 255.0).round(), 0, 255) / 255. + vals = 10 ** (2 * random.random() + 2.0) # [2, 4] + if random.random() < 0.5: + img = np.random.poisson(img * vals).astype(np.float32) / vals + else: + img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114]) + img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255. + noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray + img += noise_gray[:, :, np.newaxis] + img = np.clip(img, 0.0, 1.0) + return img + + +def add_JPEG_noise(img): + quality_factor = random.randint(30, 95) + img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR) + result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]) + img = cv2.imdecode(encimg, 1) + img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB) + return img + + +def random_crop(lq, hq, sf=4, lq_patchsize=64): + h, w = lq.shape[:2] + rnd_h = random.randint(0, h - lq_patchsize) + rnd_w = random.randint(0, w - lq_patchsize) + lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :] + + rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf) + hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :] + return lq, hq + + +def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + hq = img.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + img = util.imresize_np(img, 1 / 2, True) + img = np.clip(img, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + img = add_blur(img, sf=sf) + + elif i == 1: + img = add_blur(img, sf=sf) + + elif i == 2: + a, b = img.shape[1], img.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror') + img = img[0::sf, 0::sf, ...] # nearest downsampling + img = np.clip(img, 0.0, 1.0) + + elif i == 3: + # downsample3 + img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + img = add_JPEG_noise(img) + + elif i == 6: + # add processed camera sensor noise + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf_ori, lq_patchsize) + + return img, hq + + +# todo no isp_model? +def degradation_bsrgan_variant(image, sf=4, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + image = util.uint2single(image) + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = image.shape[:2] + image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = image.shape[:2] + + hq = image.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + image = util.imresize_np(image, 1 / 2, True) + image = np.clip(image, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + image = add_blur(image, sf=sf) + + elif i == 1: + image = add_blur(image, sf=sf) + + elif i == 2: + a, b = image.shape[1], image.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror') + image = image[0::sf, 0::sf, ...] # nearest downsampling + image = np.clip(image, 0.0, 1.0) + + elif i == 3: + # downsample3 + image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + image = np.clip(image, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + image = add_Gaussian_noise(image, noise_level1=2, noise_level2=25) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + image = add_JPEG_noise(image) + + # elif i == 6: + # # add processed camera sensor noise + # if random.random() < isp_prob and isp_model is not None: + # with torch.no_grad(): + # img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + image = add_JPEG_noise(image) + image = util.single2uint(image) + example = {"image":image} + return example + + +# TODO incase there is a pickle error one needs to replace a += x with a = a + x in add_speckle_noise etc... +def degradation_bsrgan_plus(img, sf=4, shuffle_prob=0.5, use_sharp=True, lq_patchsize=64, isp_model=None): + """ + This is an extended degradation model by combining + the degradation models of BSRGAN and Real-ESRGAN + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + use_shuffle: the degradation shuffle + use_sharp: sharpening the img + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + if use_sharp: + img = add_sharpening(img) + hq = img.copy() + + if random.random() < shuffle_prob: + shuffle_order = random.sample(range(13), 13) + else: + shuffle_order = list(range(13)) + # local shuffle for noise, JPEG is always the last one + shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6))) + shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13))) + + poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1 + + for i in shuffle_order: + if i == 0: + img = add_blur(img, sf=sf) + elif i == 1: + img = add_resize(img, sf=sf) + elif i == 2: + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + elif i == 3: + if random.random() < poisson_prob: + img = add_Poisson_noise(img) + elif i == 4: + if random.random() < speckle_prob: + img = add_speckle_noise(img) + elif i == 5: + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + elif i == 6: + img = add_JPEG_noise(img) + elif i == 7: + img = add_blur(img, sf=sf) + elif i == 8: + img = add_resize(img, sf=sf) + elif i == 9: + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + elif i == 10: + if random.random() < poisson_prob: + img = add_Poisson_noise(img) + elif i == 11: + if random.random() < speckle_prob: + img = add_speckle_noise(img) + elif i == 12: + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + else: + print('check the shuffle!') + + # resize to desired size + img = cv2.resize(img, (int(1 / sf * hq.shape[1]), int(1 / sf * hq.shape[0])), + interpolation=random.choice([1, 2, 3])) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf, lq_patchsize) + + return img, hq + + +if __name__ == '__main__': + print("hey") + img = util.imread_uint('utils/test.png', 3) + print(img) + img = util.uint2single(img) + print(img) + img = img[:448, :448] + h = img.shape[0] // 4 + print("resizing to", h) + sf = 4 + deg_fn = partial(degradation_bsrgan_variant, sf=sf) + for i in range(20): + print(i) + img_lq = deg_fn(img) + print(img_lq) + img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img)["image"] + print(img_lq.shape) + print("bicubic", img_lq_bicubic.shape) + print(img_hq.shape) + lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1) + util.imsave(img_concat, str(i) + '.png') + + diff --git a/ldm/modules/image_degradation/bsrgan_light.py b/ldm/modules/image_degradation/bsrgan_light.py new file mode 100644 index 0000000000000000000000000000000000000000..808c7f882cb75e2ba2340d5b55881d11927351f0 --- /dev/null +++ b/ldm/modules/image_degradation/bsrgan_light.py @@ -0,0 +1,651 @@ +# -*- coding: utf-8 -*- +import numpy as np +import cv2 +import torch + +from functools import partial +import random +from scipy import ndimage +import scipy +import scipy.stats as ss +from scipy.interpolate import interp2d +from scipy.linalg import orth +import albumentations + +import ldm.modules.image_degradation.utils_image as util + +""" +# -------------------------------------------- +# Super-Resolution +# -------------------------------------------- +# +# Kai Zhang (cskaizhang@gmail.com) +# https://github.com/cszn +# From 2019/03--2021/08 +# -------------------------------------------- +""" + +def modcrop_np(img, sf): + ''' + Args: + img: numpy image, WxH or WxHxC + sf: scale factor + Return: + cropped image + ''' + w, h = img.shape[:2] + im = np.copy(img) + return im[:w - w % sf, :h - h % sf, ...] + + +""" +# -------------------------------------------- +# anisotropic Gaussian kernels +# -------------------------------------------- +""" + + +def analytic_kernel(k): + """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)""" + k_size = k.shape[0] + # Calculate the big kernels size + big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2)) + # Loop over the small kernel to fill the big one + for r in range(k_size): + for c in range(k_size): + big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k + # Crop the edges of the big kernel to ignore very small values and increase run time of SR + crop = k_size // 2 + cropped_big_k = big_k[crop:-crop, crop:-crop] + # Normalize to 1 + return cropped_big_k / cropped_big_k.sum() + + +def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6): + """ generate an anisotropic Gaussian kernel + Args: + ksize : e.g., 15, kernel size + theta : [0, pi], rotation angle range + l1 : [0.1,50], scaling of eigenvalues + l2 : [0.1,l1], scaling of eigenvalues + If l1 = l2, will get an isotropic Gaussian kernel. + Returns: + k : kernel + """ + + v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.])) + V = np.array([[v[0], v[1]], [v[1], -v[0]]]) + D = np.array([[l1, 0], [0, l2]]) + Sigma = np.dot(np.dot(V, D), np.linalg.inv(V)) + k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize) + + return k + + +def gm_blur_kernel(mean, cov, size=15): + center = size / 2.0 + 0.5 + k = np.zeros([size, size]) + for y in range(size): + for x in range(size): + cy = y - center + 1 + cx = x - center + 1 + k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov) + + k = k / np.sum(k) + return k + + +def shift_pixel(x, sf, upper_left=True): + """shift pixel for super-resolution with different scale factors + Args: + x: WxHxC or WxH + sf: scale factor + upper_left: shift direction + """ + h, w = x.shape[:2] + shift = (sf - 1) * 0.5 + xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0) + if upper_left: + x1 = xv + shift + y1 = yv + shift + else: + x1 = xv - shift + y1 = yv - shift + + x1 = np.clip(x1, 0, w - 1) + y1 = np.clip(y1, 0, h - 1) + + if x.ndim == 2: + x = interp2d(xv, yv, x)(x1, y1) + if x.ndim == 3: + for i in range(x.shape[-1]): + x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1) + + return x + + +def blur(x, k): + ''' + x: image, NxcxHxW + k: kernel, Nx1xhxw + ''' + n, c = x.shape[:2] + p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2 + x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate') + k = k.repeat(1, c, 1, 1) + k = k.view(-1, 1, k.shape[2], k.shape[3]) + x = x.view(1, -1, x.shape[2], x.shape[3]) + x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c) + x = x.view(n, c, x.shape[2], x.shape[3]) + + return x + + +def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0): + """" + # modified version of https://github.com/assafshocher/BlindSR_dataset_generator + # Kai Zhang + # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var + # max_var = 2.5 * sf + """ + # Set random eigen-vals (lambdas) and angle (theta) for COV matrix + lambda_1 = min_var + np.random.rand() * (max_var - min_var) + lambda_2 = min_var + np.random.rand() * (max_var - min_var) + theta = np.random.rand() * np.pi # random theta + noise = -noise_level + np.random.rand(*k_size) * noise_level * 2 + + # Set COV matrix using Lambdas and Theta + LAMBDA = np.diag([lambda_1, lambda_2]) + Q = np.array([[np.cos(theta), -np.sin(theta)], + [np.sin(theta), np.cos(theta)]]) + SIGMA = Q @ LAMBDA @ Q.T + INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :] + + # Set expectation position (shifting kernel for aligned image) + MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2) + MU = MU[None, None, :, None] + + # Create meshgrid for Gaussian + [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1])) + Z = np.stack([X, Y], 2)[:, :, :, None] + + # Calcualte Gaussian for every pixel of the kernel + ZZ = Z - MU + ZZ_t = ZZ.transpose(0, 1, 3, 2) + raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise) + + # shift the kernel so it will be centered + # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor) + + # Normalize the kernel and return + # kernel = raw_kernel_centered / np.sum(raw_kernel_centered) + kernel = raw_kernel / np.sum(raw_kernel) + return kernel + + +def fspecial_gaussian(hsize, sigma): + hsize = [hsize, hsize] + siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0] + std = sigma + [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1)) + arg = -(x * x + y * y) / (2 * std * std) + h = np.exp(arg) + h[h < scipy.finfo(float).eps * h.max()] = 0 + sumh = h.sum() + if sumh != 0: + h = h / sumh + return h + + +def fspecial_laplacian(alpha): + alpha = max([0, min([alpha, 1])]) + h1 = alpha / (alpha + 1) + h2 = (1 - alpha) / (alpha + 1) + h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]] + h = np.array(h) + return h + + +def fspecial(filter_type, *args, **kwargs): + ''' + python code from: + https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py + ''' + if filter_type == 'gaussian': + return fspecial_gaussian(*args, **kwargs) + if filter_type == 'laplacian': + return fspecial_laplacian(*args, **kwargs) + + +""" +# -------------------------------------------- +# degradation models +# -------------------------------------------- +""" + + +def bicubic_degradation(x, sf=3): + ''' + Args: + x: HxWxC image, [0, 1] + sf: down-scale factor + Return: + bicubicly downsampled LR image + ''' + x = util.imresize_np(x, scale=1 / sf) + return x + + +def srmd_degradation(x, k, sf=3): + ''' blur + bicubic downsampling + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2018learning, + title={Learning a single convolutional super-resolution network for multiple degradations}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={3262--3271}, + year={2018} + } + ''' + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror' + x = bicubic_degradation(x, sf=sf) + return x + + +def dpsr_degradation(x, k, sf=3): + ''' bicubic downsampling + blur + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2019deep, + title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={1671--1681}, + year={2019} + } + ''' + x = bicubic_degradation(x, sf=sf) + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + return x + + +def classical_degradation(x, k, sf=3): + ''' blur + downsampling + Args: + x: HxWxC image, [0, 1]/[0, 255] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + ''' + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2)) + st = 0 + return x[st::sf, st::sf, ...] + + +def add_sharpening(img, weight=0.5, radius=50, threshold=10): + """USM sharpening. borrowed from real-ESRGAN + Input image: I; Blurry image: B. + 1. K = I + weight * (I - B) + 2. Mask = 1 if abs(I - B) > threshold, else: 0 + 3. Blur mask: + 4. Out = Mask * K + (1 - Mask) * I + Args: + img (Numpy array): Input image, HWC, BGR; float32, [0, 1]. + weight (float): Sharp weight. Default: 1. + radius (float): Kernel size of Gaussian blur. Default: 50. + threshold (int): + """ + if radius % 2 == 0: + radius += 1 + blur = cv2.GaussianBlur(img, (radius, radius), 0) + residual = img - blur + mask = np.abs(residual) * 255 > threshold + mask = mask.astype('float32') + soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0) + + K = img + weight * residual + K = np.clip(K, 0, 1) + return soft_mask * K + (1 - soft_mask) * img + + +def add_blur(img, sf=4): + wd2 = 4.0 + sf + wd = 2.0 + 0.2 * sf + + wd2 = wd2/4 + wd = wd/4 + + if random.random() < 0.5: + l1 = wd2 * random.random() + l2 = wd2 * random.random() + k = anisotropic_Gaussian(ksize=random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2) + else: + k = fspecial('gaussian', random.randint(2, 4) + 3, wd * random.random()) + img = ndimage.convolve(img, np.expand_dims(k, axis=2), mode='mirror') + + return img + + +def add_resize(img, sf=4): + rnum = np.random.rand() + if rnum > 0.8: # up + sf1 = random.uniform(1, 2) + elif rnum < 0.7: # down + sf1 = random.uniform(0.5 / sf, 1) + else: + sf1 = 1.0 + img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + return img + + +# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): +# noise_level = random.randint(noise_level1, noise_level2) +# rnum = np.random.rand() +# if rnum > 0.6: # add color Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) +# elif rnum < 0.4: # add grayscale Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) +# else: # add noise +# L = noise_level2 / 255. +# D = np.diag(np.random.rand(3)) +# U = orth(np.random.rand(3, 3)) +# conv = np.dot(np.dot(np.transpose(U), D), U) +# img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) +# img = np.clip(img, 0.0, 1.0) +# return img + +def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + rnum = np.random.rand() + if rnum > 0.6: # add color Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: # add grayscale Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: # add noise + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_speckle_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + img = np.clip(img, 0.0, 1.0) + rnum = random.random() + if rnum > 0.6: + img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: + img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_Poisson_noise(img): + img = np.clip((img * 255.0).round(), 0, 255) / 255. + vals = 10 ** (2 * random.random() + 2.0) # [2, 4] + if random.random() < 0.5: + img = np.random.poisson(img * vals).astype(np.float32) / vals + else: + img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114]) + img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255. + noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray + img += noise_gray[:, :, np.newaxis] + img = np.clip(img, 0.0, 1.0) + return img + + +def add_JPEG_noise(img): + quality_factor = random.randint(80, 95) + img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR) + result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]) + img = cv2.imdecode(encimg, 1) + img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB) + return img + + +def random_crop(lq, hq, sf=4, lq_patchsize=64): + h, w = lq.shape[:2] + rnd_h = random.randint(0, h - lq_patchsize) + rnd_w = random.randint(0, w - lq_patchsize) + lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :] + + rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf) + hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :] + return lq, hq + + +def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + hq = img.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + img = util.imresize_np(img, 1 / 2, True) + img = np.clip(img, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + img = add_blur(img, sf=sf) + + elif i == 1: + img = add_blur(img, sf=sf) + + elif i == 2: + a, b = img.shape[1], img.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + img = ndimage.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror') + img = img[0::sf, 0::sf, ...] # nearest downsampling + img = np.clip(img, 0.0, 1.0) + + elif i == 3: + # downsample3 + img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=8) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + img = add_JPEG_noise(img) + + elif i == 6: + # add processed camera sensor noise + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf_ori, lq_patchsize) + + return img, hq + + +# todo no isp_model? +def degradation_bsrgan_variant(image, sf=4, isp_model=None, up=False): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + image = util.uint2single(image) + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = image.shape[:2] + image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = image.shape[:2] + + hq = image.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + image = util.imresize_np(image, 1 / 2, True) + image = np.clip(image, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + image = add_blur(image, sf=sf) + + # elif i == 1: + # image = add_blur(image, sf=sf) + + if i == 0: + pass + + elif i == 2: + a, b = image.shape[1], image.shape[0] + # downsample2 + if random.random() < 0.8: + sf1 = random.uniform(1, 2 * sf) + image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + image = ndimage.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror') + image = image[0::sf, 0::sf, ...] # nearest downsampling + + image = np.clip(image, 0.0, 1.0) + + elif i == 3: + # downsample3 + image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + image = np.clip(image, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + image = add_Gaussian_noise(image, noise_level1=1, noise_level2=2) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + image = add_JPEG_noise(image) + # + # elif i == 6: + # # add processed camera sensor noise + # if random.random() < isp_prob and isp_model is not None: + # with torch.no_grad(): + # img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + image = add_JPEG_noise(image) + image = util.single2uint(image) + if up: + image = cv2.resize(image, (w1, h1), interpolation=cv2.INTER_CUBIC) # todo: random, as above? want to condition on it then + example = {"image": image} + return example + + + + +if __name__ == '__main__': + print("hey") + img = util.imread_uint('utils/test.png', 3) + img = img[:448, :448] + h = img.shape[0] // 4 + print("resizing to", h) + sf = 4 + deg_fn = partial(degradation_bsrgan_variant, sf=sf) + for i in range(20): + print(i) + img_hq = img + img_lq = deg_fn(img)["image"] + img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq) + print(img_lq) + img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img_hq)["image"] + print(img_lq.shape) + print("bicubic", img_lq_bicubic.shape) + print(img_hq.shape) + lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), + (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1) + util.imsave(img_concat, str(i) + '.png') diff --git a/ldm/modules/image_degradation/utils/test.png b/ldm/modules/image_degradation/utils/test.png new file mode 100644 index 0000000000000000000000000000000000000000..4249b43de0f22707758d13c240268a401642f6e6 Binary files /dev/null and b/ldm/modules/image_degradation/utils/test.png differ diff --git a/ldm/modules/image_degradation/utils_image.py b/ldm/modules/image_degradation/utils_image.py new file mode 100644 index 0000000000000000000000000000000000000000..0175f155ad900ae33c3c46ed87f49b352e3faf98 --- /dev/null +++ b/ldm/modules/image_degradation/utils_image.py @@ -0,0 +1,916 @@ +import os +import math +import random +import numpy as np +import torch +import cv2 +from torchvision.utils import make_grid +from datetime import datetime +#import matplotlib.pyplot as plt # TODO: check with Dominik, also bsrgan.py vs bsrgan_light.py + + +os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE" + + +''' +# -------------------------------------------- +# Kai Zhang (github: https://github.com/cszn) +# 03/Mar/2019 +# -------------------------------------------- +# https://github.com/twhui/SRGAN-pyTorch +# https://github.com/xinntao/BasicSR +# -------------------------------------------- +''' + + +IMG_EXTENSIONS = ['.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tif'] + + +def is_image_file(filename): + return any(filename.endswith(extension) for extension in IMG_EXTENSIONS) + + +def get_timestamp(): + return datetime.now().strftime('%y%m%d-%H%M%S') + + +def imshow(x, title=None, cbar=False, figsize=None): + plt.figure(figsize=figsize) + plt.imshow(np.squeeze(x), interpolation='nearest', cmap='gray') + if title: + plt.title(title) + if cbar: + plt.colorbar() + plt.show() + + +def surf(Z, cmap='rainbow', figsize=None): + plt.figure(figsize=figsize) + ax3 = plt.axes(projection='3d') + + w, h = Z.shape[:2] + xx = np.arange(0,w,1) + yy = np.arange(0,h,1) + X, Y = np.meshgrid(xx, yy) + ax3.plot_surface(X,Y,Z,cmap=cmap) + #ax3.contour(X,Y,Z, zdim='z',offset=-2，cmap=cmap) + plt.show() + + +''' +# -------------------------------------------- +# get image pathes +# -------------------------------------------- +''' + + +def get_image_paths(dataroot): + paths = None # return None if dataroot is None + if dataroot is not None: + paths = sorted(_get_paths_from_images(dataroot)) + return paths + + +def _get_paths_from_images(path): + assert os.path.isdir(path), '{:s} is not a valid directory'.format(path) + images = [] + for dirpath, _, fnames in sorted(os.walk(path)): + for fname in sorted(fnames): + if is_image_file(fname): + img_path = os.path.join(dirpath, fname) + images.append(img_path) + assert images, '{:s} has no valid image file'.format(path) + return images + + +''' +# -------------------------------------------- +# split large images into small images +# -------------------------------------------- +''' + + +def patches_from_image(img, p_size=512, p_overlap=64, p_max=800): + w, h = img.shape[:2] + patches = [] + if w > p_max and h > p_max: + w1 = list(np.arange(0, w-p_size, p_size-p_overlap, dtype=np.int)) + h1 = list(np.arange(0, h-p_size, p_size-p_overlap, dtype=np.int)) + w1.append(w-p_size) + h1.append(h-p_size) +# print(w1) +# print(h1) + for i in w1: + for j in h1: + patches.append(img[i:i+p_size, j:j+p_size,:]) + else: + patches.append(img) + + return patches + + +def imssave(imgs, img_path): + """ + imgs: list, N images of size WxHxC + """ + img_name, ext = os.path.splitext(os.path.basename(img_path)) + + for i, img in enumerate(imgs): + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + new_path = os.path.join(os.path.dirname(img_path), img_name+str('_s{:04d}'.format(i))+'.png') + cv2.imwrite(new_path, img) + + +def split_imageset(original_dataroot, taget_dataroot, n_channels=3, p_size=800, p_overlap=96, p_max=1000): + """ + split the large images from original_dataroot into small overlapped images with size (p_size)x(p_size), + and save them into taget_dataroot; only the images with larger size than (p_max)x(p_max) + will be splitted. + Args: + original_dataroot: + taget_dataroot: + p_size: size of small images + p_overlap: patch size in training is a good choice + p_max: images with smaller size than (p_max)x(p_max) keep unchanged. + """ + paths = get_image_paths(original_dataroot) + for img_path in paths: + # img_name, ext = os.path.splitext(os.path.basename(img_path)) + img = imread_uint(img_path, n_channels=n_channels) + patches = patches_from_image(img, p_size, p_overlap, p_max) + imssave(patches, os.path.join(taget_dataroot,os.path.basename(img_path))) + #if original_dataroot == taget_dataroot: + #del img_path + +''' +# -------------------------------------------- +# makedir +# -------------------------------------------- +''' + + +def mkdir(path): + if not os.path.exists(path): + os.makedirs(path) + + +def mkdirs(paths): + if isinstance(paths, str): + mkdir(paths) + else: + for path in paths: + mkdir(path) + + +def mkdir_and_rename(path): + if os.path.exists(path): + new_name = path + '_archived_' + get_timestamp() + print('Path already exists. Rename it to [{:s}]'.format(new_name)) + os.rename(path, new_name) + os.makedirs(path) + + +''' +# -------------------------------------------- +# read image from path +# opencv is fast, but read BGR numpy image +# -------------------------------------------- +''' + + +# -------------------------------------------- +# get uint8 image of size HxWxn_channles (RGB) +# -------------------------------------------- +def imread_uint(path, n_channels=3): + # input: path + # output: HxWx3(RGB or GGG), or HxWx1 (G) + if n_channels == 1: + img = cv2.imread(path, 0) # cv2.IMREAD_GRAYSCALE + img = np.expand_dims(img, axis=2) # HxWx1 + elif n_channels == 3: + img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # BGR or G + if img.ndim == 2: + img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB) # GGG + else: + img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # RGB + return img + + +# -------------------------------------------- +# matlab's imwrite +# -------------------------------------------- +def imsave(img, img_path): + img = np.squeeze(img) + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + cv2.imwrite(img_path, img) + +def imwrite(img, img_path): + img = np.squeeze(img) + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + cv2.imwrite(img_path, img) + + + +# -------------------------------------------- +# get single image of size HxWxn_channles (BGR) +# -------------------------------------------- +def read_img(path): + # read image by cv2 + # return: Numpy float32, HWC, BGR, [0,1] + img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # cv2.IMREAD_GRAYSCALE + img = img.astype(np.float32) / 255. + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + # some images have 4 channels + if img.shape[2] > 3: + img = img[:, :, :3] + return img + + +''' +# -------------------------------------------- +# image format conversion +# -------------------------------------------- +# numpy(single) <---> numpy(unit) +# numpy(single) <---> tensor +# numpy(unit) <---> tensor +# -------------------------------------------- +''' + + +# -------------------------------------------- +# numpy(single) [0, 1] <---> numpy(unit) +# -------------------------------------------- + + +def uint2single(img): + + return np.float32(img/255.) + + +def single2uint(img): + + return np.uint8((img.clip(0, 1)*255.).round()) + + +def uint162single(img): + + return np.float32(img/65535.) + + +def single2uint16(img): + + return np.uint16((img.clip(0, 1)*65535.).round()) + + +# -------------------------------------------- +# numpy(unit) (HxWxC or HxW) <---> tensor +# -------------------------------------------- + + +# convert uint to 4-dimensional torch tensor +def uint2tensor4(img): + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.).unsqueeze(0) + + +# convert uint to 3-dimensional torch tensor +def uint2tensor3(img): + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.) + + +# convert 2/3/4-dimensional torch tensor to uint +def tensor2uint(img): + img = img.data.squeeze().float().clamp_(0, 1).cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + return np.uint8((img*255.0).round()) + + +# -------------------------------------------- +# numpy(single) (HxWxC) <---> tensor +# -------------------------------------------- + + +# convert single (HxWxC) to 3-dimensional torch tensor +def single2tensor3(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float() + + +# convert single (HxWxC) to 4-dimensional torch tensor +def single2tensor4(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().unsqueeze(0) + + +# convert torch tensor to single +def tensor2single(img): + img = img.data.squeeze().float().cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + + return img + +# convert torch tensor to single +def tensor2single3(img): + img = img.data.squeeze().float().cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + elif img.ndim == 2: + img = np.expand_dims(img, axis=2) + return img + + +def single2tensor5(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float().unsqueeze(0) + + +def single32tensor5(img): + return torch.from_numpy(np.ascontiguousarray(img)).float().unsqueeze(0).unsqueeze(0) + + +def single42tensor4(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float() + + +# from skimage.io import imread, imsave +def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)): + ''' + Converts a torch Tensor into an image Numpy array of BGR channel order + Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order + Output: 3D(H,W,C) or 2D(H,W), [0,255], np.uint8 (default) + ''' + tensor = tensor.squeeze().float().cpu().clamp_(*min_max) # squeeze first, then clamp + tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0]) # to range [0,1] + n_dim = tensor.dim() + if n_dim == 4: + n_img = len(tensor) + img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy() + img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR + elif n_dim == 3: + img_np = tensor.numpy() + img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR + elif n_dim == 2: + img_np = tensor.numpy() + else: + raise TypeError( + 'Only support 4D, 3D and 2D tensor. But received with dimension: {:d}'.format(n_dim)) + if out_type == np.uint8: + img_np = (img_np * 255.0).round() + # Important. Unlike matlab, numpy.unit8() WILL NOT round by default. + return img_np.astype(out_type) + + +''' +# -------------------------------------------- +# Augmentation, flipe and/or rotate +# -------------------------------------------- +# The following two are enough. +# (1) augmet_img: numpy image of WxHxC or WxH +# (2) augment_img_tensor4: tensor image 1xCxWxH +# -------------------------------------------- +''' + + +def augment_img(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + if mode == 0: + return img + elif mode == 1: + return np.flipud(np.rot90(img)) + elif mode == 2: + return np.flipud(img) + elif mode == 3: + return np.rot90(img, k=3) + elif mode == 4: + return np.flipud(np.rot90(img, k=2)) + elif mode == 5: + return np.rot90(img) + elif mode == 6: + return np.rot90(img, k=2) + elif mode == 7: + return np.flipud(np.rot90(img, k=3)) + + +def augment_img_tensor4(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + if mode == 0: + return img + elif mode == 1: + return img.rot90(1, [2, 3]).flip([2]) + elif mode == 2: + return img.flip([2]) + elif mode == 3: + return img.rot90(3, [2, 3]) + elif mode == 4: + return img.rot90(2, [2, 3]).flip([2]) + elif mode == 5: + return img.rot90(1, [2, 3]) + elif mode == 6: + return img.rot90(2, [2, 3]) + elif mode == 7: + return img.rot90(3, [2, 3]).flip([2]) + + +def augment_img_tensor(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + img_size = img.size() + img_np = img.data.cpu().numpy() + if len(img_size) == 3: + img_np = np.transpose(img_np, (1, 2, 0)) + elif len(img_size) == 4: + img_np = np.transpose(img_np, (2, 3, 1, 0)) + img_np = augment_img(img_np, mode=mode) + img_tensor = torch.from_numpy(np.ascontiguousarray(img_np)) + if len(img_size) == 3: + img_tensor = img_tensor.permute(2, 0, 1) + elif len(img_size) == 4: + img_tensor = img_tensor.permute(3, 2, 0, 1) + + return img_tensor.type_as(img) + + +def augment_img_np3(img, mode=0): + if mode == 0: + return img + elif mode == 1: + return img.transpose(1, 0, 2) + elif mode == 2: + return img[::-1, :, :] + elif mode == 3: + img = img[::-1, :, :] + img = img.transpose(1, 0, 2) + return img + elif mode == 4: + return img[:, ::-1, :] + elif mode == 5: + img = img[:, ::-1, :] + img = img.transpose(1, 0, 2) + return img + elif mode == 6: + img = img[:, ::-1, :] + img = img[::-1, :, :] + return img + elif mode == 7: + img = img[:, ::-1, :] + img = img[::-1, :, :] + img = img.transpose(1, 0, 2) + return img + + +def augment_imgs(img_list, hflip=True, rot=True): + # horizontal flip OR rotate + hflip = hflip and random.random() < 0.5 + vflip = rot and random.random() < 0.5 + rot90 = rot and random.random() < 0.5 + + def _augment(img): + if hflip: + img = img[:, ::-1, :] + if vflip: + img = img[::-1, :, :] + if rot90: + img = img.transpose(1, 0, 2) + return img + + return [_augment(img) for img in img_list] + + +''' +# -------------------------------------------- +# modcrop and shave +# -------------------------------------------- +''' + + +def modcrop(img_in, scale): + # img_in: Numpy, HWC or HW + img = np.copy(img_in) + if img.ndim == 2: + H, W = img.shape + H_r, W_r = H % scale, W % scale + img = img[:H - H_r, :W - W_r] + elif img.ndim == 3: + H, W, C = img.shape + H_r, W_r = H % scale, W % scale + img = img[:H - H_r, :W - W_r, :] + else: + raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim)) + return img + + +def shave(img_in, border=0): + # img_in: Numpy, HWC or HW + img = np.copy(img_in) + h, w = img.shape[:2] + img = img[border:h-border, border:w-border] + return img + + +''' +# -------------------------------------------- +# image processing process on numpy image +# channel_convert(in_c, tar_type, img_list): +# rgb2ycbcr(img, only_y=True): +# bgr2ycbcr(img, only_y=True): +# ycbcr2rgb(img): +# -------------------------------------------- +''' + + +def rgb2ycbcr(img, only_y=True): + '''same as matlab rgb2ycbcr + only_y: only return Y channel + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + if only_y: + rlt = np.dot(img, [65.481, 128.553, 24.966]) / 255.0 + 16.0 + else: + rlt = np.matmul(img, [[65.481, -37.797, 112.0], [128.553, -74.203, -93.786], + [24.966, 112.0, -18.214]]) / 255.0 + [16, 128, 128] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def ycbcr2rgb(img): + '''same as matlab ycbcr2rgb + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + rlt = np.matmul(img, [[0.00456621, 0.00456621, 0.00456621], [0, -0.00153632, 0.00791071], + [0.00625893, -0.00318811, 0]]) * 255.0 + [-222.921, 135.576, -276.836] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def bgr2ycbcr(img, only_y=True): + '''bgr version of rgb2ycbcr + only_y: only return Y channel + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + if only_y: + rlt = np.dot(img, [24.966, 128.553, 65.481]) / 255.0 + 16.0 + else: + rlt = np.matmul(img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786], + [65.481, -37.797, 112.0]]) / 255.0 + [16, 128, 128] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def channel_convert(in_c, tar_type, img_list): + # conversion among BGR, gray and y + if in_c == 3 and tar_type == 'gray': # BGR to gray + gray_list = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img in img_list] + return [np.expand_dims(img, axis=2) for img in gray_list] + elif in_c == 3 and tar_type == 'y': # BGR to y + y_list = [bgr2ycbcr(img, only_y=True) for img in img_list] + return [np.expand_dims(img, axis=2) for img in y_list] + elif in_c == 1 and tar_type == 'RGB': # gray/y to BGR + return [cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) for img in img_list] + else: + return img_list + + +''' +# -------------------------------------------- +# metric, PSNR and SSIM +# -------------------------------------------- +''' + + +# -------------------------------------------- +# PSNR +# -------------------------------------------- +def calculate_psnr(img1, img2, border=0): + # img1 and img2 have range [0, 255] + #img1 = img1.squeeze() + #img2 = img2.squeeze() + if not img1.shape == img2.shape: + raise ValueError('Input images must have the same dimensions.') + h, w = img1.shape[:2] + img1 = img1[border:h-border, border:w-border] + img2 = img2[border:h-border, border:w-border] + + img1 = img1.astype(np.float64) + img2 = img2.astype(np.float64) + mse = np.mean((img1 - img2)**2) + if mse == 0: + return float('inf') + return 20 * math.log10(255.0 / math.sqrt(mse)) + + +# -------------------------------------------- +# SSIM +# -------------------------------------------- +def calculate_ssim(img1, img2, border=0): + '''calculate SSIM + the same outputs as MATLAB's + img1, img2: [0, 255] + ''' + #img1 = img1.squeeze() + #img2 = img2.squeeze() + if not img1.shape == img2.shape: + raise ValueError('Input images must have the same dimensions.') + h, w = img1.shape[:2] + img1 = img1[border:h-border, border:w-border] + img2 = img2[border:h-border, border:w-border] + + if img1.ndim == 2: + return ssim(img1, img2) + elif img1.ndim == 3: + if img1.shape[2] == 3: + ssims = [] + for i in range(3): + ssims.append(ssim(img1[:,:,i], img2[:,:,i])) + return np.array(ssims).mean() + elif img1.shape[2] == 1: + return ssim(np.squeeze(img1), np.squeeze(img2)) + else: + raise ValueError('Wrong input image dimensions.') + + +def ssim(img1, img2): + C1 = (0.01 * 255)**2 + C2 = (0.03 * 255)**2 + + img1 = img1.astype(np.float64) + img2 = img2.astype(np.float64) + kernel = cv2.getGaussianKernel(11, 1.5) + window = np.outer(kernel, kernel.transpose()) + + mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid + mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5] + mu1_sq = mu1**2 + mu2_sq = mu2**2 + mu1_mu2 = mu1 * mu2 + sigma1_sq = cv2.filter2D(img1**2, -1, window)[5:-5, 5:-5] - mu1_sq + sigma2_sq = cv2.filter2D(img2**2, -1, window)[5:-5, 5:-5] - mu2_sq + sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2 + + ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * + (sigma1_sq + sigma2_sq + C2)) + return ssim_map.mean() + + +''' +# -------------------------------------------- +# matlab's bicubic imresize (numpy and torch) [0, 1] +# -------------------------------------------- +''' + + +# matlab 'imresize' function, now only support 'bicubic' +def cubic(x): + absx = torch.abs(x) + absx2 = absx**2 + absx3 = absx**3 + return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \ + (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx)) + + +def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing): + if (scale < 1) and (antialiasing): + # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width + kernel_width = kernel_width / scale + + # Output-space coordinates + x = torch.linspace(1, out_length, out_length) + + # Input-space coordinates. Calculate the inverse mapping such that 0.5 + # in output space maps to 0.5 in input space, and 0.5+scale in output + # space maps to 1.5 in input space. + u = x / scale + 0.5 * (1 - 1 / scale) + + # What is the left-most pixel that can be involved in the computation? + left = torch.floor(u - kernel_width / 2) + + # What is the maximum number of pixels that can be involved in the + # computation? Note: it's OK to use an extra pixel here; if the + # corresponding weights are all zero, it will be eliminated at the end + # of this function. + P = math.ceil(kernel_width) + 2 + + # The indices of the input pixels involved in computing the k-th output + # pixel are in row k of the indices matrix. + indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view( + 1, P).expand(out_length, P) + + # The weights used to compute the k-th output pixel are in row k of the + # weights matrix. + distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices + # apply cubic kernel + if (scale < 1) and (antialiasing): + weights = scale * cubic(distance_to_center * scale) + else: + weights = cubic(distance_to_center) + # Normalize the weights matrix so that each row sums to 1. + weights_sum = torch.sum(weights, 1).view(out_length, 1) + weights = weights / weights_sum.expand(out_length, P) + + # If a column in weights is all zero, get rid of it. only consider the first and last column. + weights_zero_tmp = torch.sum((weights == 0), 0) + if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6): + indices = indices.narrow(1, 1, P - 2) + weights = weights.narrow(1, 1, P - 2) + if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6): + indices = indices.narrow(1, 0, P - 2) + weights = weights.narrow(1, 0, P - 2) + weights = weights.contiguous() + indices = indices.contiguous() + sym_len_s = -indices.min() + 1 + sym_len_e = indices.max() - in_length + indices = indices + sym_len_s - 1 + return weights, indices, int(sym_len_s), int(sym_len_e) + + +# -------------------------------------------- +# imresize for tensor image [0, 1] +# -------------------------------------------- +def imresize(img, scale, antialiasing=True): + # Now the scale should be the same for H and W + # input: img: pytorch tensor, CHW or HW [0,1] + # output: CHW or HW [0,1] w/o round + need_squeeze = True if img.dim() == 2 else False + if need_squeeze: + img.unsqueeze_(0) + in_C, in_H, in_W = img.size() + out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) + kernel_width = 4 + kernel = 'cubic' + + # Return the desired dimension order for performing the resize. The + # strategy is to perform the resize first along the dimension with the + # smallest scale factor. + # Now we do not support this. + + # get weights and indices + weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( + in_H, out_H, scale, kernel, kernel_width, antialiasing) + weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( + in_W, out_W, scale, kernel, kernel_width, antialiasing) + # process H dimension + # symmetric copying + img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W) + img_aug.narrow(1, sym_len_Hs, in_H).copy_(img) + + sym_patch = img[:, :sym_len_Hs, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv) + + sym_patch = img[:, -sym_len_He:, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) + + out_1 = torch.FloatTensor(in_C, out_H, in_W) + kernel_width = weights_H.size(1) + for i in range(out_H): + idx = int(indices_H[i][0]) + for j in range(out_C): + out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i]) + + # process W dimension + # symmetric copying + out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We) + out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1) + + sym_patch = out_1[:, :, :sym_len_Ws] + inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(2, inv_idx) + out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv) + + sym_patch = out_1[:, :, -sym_len_We:] + inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(2, inv_idx) + out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) + + out_2 = torch.FloatTensor(in_C, out_H, out_W) + kernel_width = weights_W.size(1) + for i in range(out_W): + idx = int(indices_W[i][0]) + for j in range(out_C): + out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i]) + if need_squeeze: + out_2.squeeze_() + return out_2 + + +# -------------------------------------------- +# imresize for numpy image [0, 1] +# -------------------------------------------- +def imresize_np(img, scale, antialiasing=True): + # Now the scale should be the same for H and W + # input: img: Numpy, HWC or HW [0,1] + # output: HWC or HW [0,1] w/o round + img = torch.from_numpy(img) + need_squeeze = True if img.dim() == 2 else False + if need_squeeze: + img.unsqueeze_(2) + + in_H, in_W, in_C = img.size() + out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) + kernel_width = 4 + kernel = 'cubic' + + # Return the desired dimension order for performing the resize. The + # strategy is to perform the resize first along the dimension with the + # smallest scale factor. + # Now we do not support this. + + # get weights and indices + weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( + in_H, out_H, scale, kernel, kernel_width, antialiasing) + weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( + in_W, out_W, scale, kernel, kernel_width, antialiasing) + # process H dimension + # symmetric copying + img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C) + img_aug.narrow(0, sym_len_Hs, in_H).copy_(img) + + sym_patch = img[:sym_len_Hs, :, :] + inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(0, inv_idx) + img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv) + + sym_patch = img[-sym_len_He:, :, :] + inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(0, inv_idx) + img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) + + out_1 = torch.FloatTensor(out_H, in_W, in_C) + kernel_width = weights_H.size(1) + for i in range(out_H): + idx = int(indices_H[i][0]) + for j in range(out_C): + out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i]) + + # process W dimension + # symmetric copying + out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C) + out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1) + + sym_patch = out_1[:, :sym_len_Ws, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv) + + sym_patch = out_1[:, -sym_len_We:, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) + + out_2 = torch.FloatTensor(out_H, out_W, in_C) + kernel_width = weights_W.size(1) + for i in range(out_W): + idx = int(indices_W[i][0]) + for j in range(out_C): + out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i]) + if need_squeeze: + out_2.squeeze_() + + return out_2.numpy() + + +if __name__ == '__main__': + print('---') +# img = imread_uint('test.bmp', 3) +# img = uint2single(img) +# img_bicubic = imresize_np(img, 1/4) \ No newline at end of file diff --git a/ldm/modules/midas/__init__.py b/ldm/modules/midas/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/midas/api.py b/ldm/modules/midas/api.py new file mode 100644 index 0000000000000000000000000000000000000000..b58ebbffd942a2fc22264f0ab47e400c26b9f41c --- /dev/null +++ b/ldm/modules/midas/api.py @@ -0,0 +1,170 @@ +# based on https://github.com/isl-org/MiDaS + +import cv2 +import torch +import torch.nn as nn +from torchvision.transforms import Compose + +from ldm.modules.midas.midas.dpt_depth import DPTDepthModel +from ldm.modules.midas.midas.midas_net import MidasNet +from ldm.modules.midas.midas.midas_net_custom import MidasNet_small +from ldm.modules.midas.midas.transforms import Resize, NormalizeImage, PrepareForNet + + +ISL_PATHS = { + "dpt_large": "midas_models/dpt_large-midas-2f21e586.pt", + "dpt_hybrid": "midas_models/dpt_hybrid-midas-501f0c75.pt", + "midas_v21": "", + "midas_v21_small": "", +} + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +def load_midas_transform(model_type): + # https://github.com/isl-org/MiDaS/blob/master/run.py + # load transform only + if model_type == "dpt_large": # DPT-Large + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "dpt_hybrid": # DPT-Hybrid + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "midas_v21": + net_w, net_h = 384, 384 + resize_mode = "upper_bound" + normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) + + elif model_type == "midas_v21_small": + net_w, net_h = 256, 256 + resize_mode = "upper_bound" + normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) + + else: + assert False, f"model_type '{model_type}' not implemented, use: --model_type large" + + transform = Compose( + [ + Resize( + net_w, + net_h, + resize_target=None, + keep_aspect_ratio=True, + ensure_multiple_of=32, + resize_method=resize_mode, + image_interpolation_method=cv2.INTER_CUBIC, + ), + normalization, + PrepareForNet(), + ] + ) + + return transform + + +def load_model(model_type): + # https://github.com/isl-org/MiDaS/blob/master/run.py + # load network + model_path = ISL_PATHS[model_type] + if model_type == "dpt_large": # DPT-Large + model = DPTDepthModel( + path=model_path, + backbone="vitl16_384", + non_negative=True, + ) + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "dpt_hybrid": # DPT-Hybrid + model = DPTDepthModel( + path=model_path, + backbone="vitb_rn50_384", + non_negative=True, + ) + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "midas_v21": + model = MidasNet(model_path, non_negative=True) + net_w, net_h = 384, 384 + resize_mode = "upper_bound" + normalization = NormalizeImage( + mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] + ) + + elif model_type == "midas_v21_small": + model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True, + non_negative=True, blocks={'expand': True}) + net_w, net_h = 256, 256 + resize_mode = "upper_bound" + normalization = NormalizeImage( + mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] + ) + + else: + print(f"model_type '{model_type}' not implemented, use: --model_type large") + assert False + + transform = Compose( + [ + Resize( + net_w, + net_h, + resize_target=None, + keep_aspect_ratio=True, + ensure_multiple_of=32, + resize_method=resize_mode, + image_interpolation_method=cv2.INTER_CUBIC, + ), + normalization, + PrepareForNet(), + ] + ) + + return model.eval(), transform + + +class MiDaSInference(nn.Module): + MODEL_TYPES_TORCH_HUB = [ + "DPT_Large", + "DPT_Hybrid", + "MiDaS_small" + ] + MODEL_TYPES_ISL = [ + "dpt_large", + "dpt_hybrid", + "midas_v21", + "midas_v21_small", + ] + + def __init__(self, model_type): + super().__init__() + assert (model_type in self.MODEL_TYPES_ISL) + model, _ = load_model(model_type) + self.model = model + self.model.train = disabled_train + + def forward(self, x): + # x in 0..1 as produced by calling self.transform on a 0..1 float64 numpy array + # NOTE: we expect that the correct transform has been called during dataloading. + with torch.no_grad(): + prediction = self.model(x) + prediction = torch.nn.functional.interpolate( + prediction.unsqueeze(1), + size=x.shape[2:], + mode="bicubic", + align_corners=False, + ) + assert prediction.shape == (x.shape[0], 1, x.shape[2], x.shape[3]) + return prediction + diff --git a/ldm/modules/midas/midas/__init__.py b/ldm/modules/midas/midas/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/midas/midas/base_model.py b/ldm/modules/midas/midas/base_model.py new file mode 100644 index 0000000000000000000000000000000000000000..5cf430239b47ec5ec07531263f26f5c24a2311cd --- /dev/null +++ b/ldm/modules/midas/midas/base_model.py @@ -0,0 +1,16 @@ +import torch + + +class BaseModel(torch.nn.Module): + def load(self, path): + """Load model from file. + + Args: + path (str): file path + """ + parameters = torch.load(path, map_location=torch.device('cpu')) + + if "optimizer" in parameters: + parameters = parameters["model"] + + self.load_state_dict(parameters) diff --git a/ldm/modules/midas/midas/blocks.py b/ldm/modules/midas/midas/blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..2145d18fa98060a618536d9a64fe6589e9be4f78 --- /dev/null +++ b/ldm/modules/midas/midas/blocks.py @@ -0,0 +1,342 @@ +import torch +import torch.nn as nn + +from .vit import ( + _make_pretrained_vitb_rn50_384, + _make_pretrained_vitl16_384, + _make_pretrained_vitb16_384, + forward_vit, +) + +def _make_encoder(backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore",): + if backbone == "vitl16_384": + pretrained = _make_pretrained_vitl16_384( + use_pretrained, hooks=hooks, use_readout=use_readout + ) + scratch = _make_scratch( + [256, 512, 1024, 1024], features, groups=groups, expand=expand + ) # ViT-L/16 - 85.0% Top1 (backbone) + elif backbone == "vitb_rn50_384": + pretrained = _make_pretrained_vitb_rn50_384( + use_pretrained, + hooks=hooks, + use_vit_only=use_vit_only, + use_readout=use_readout, + ) + scratch = _make_scratch( + [256, 512, 768, 768], features, groups=groups, expand=expand + ) # ViT-H/16 - 85.0% Top1 (backbone) + elif backbone == "vitb16_384": + pretrained = _make_pretrained_vitb16_384( + use_pretrained, hooks=hooks, use_readout=use_readout + ) + scratch = _make_scratch( + [96, 192, 384, 768], features, groups=groups, expand=expand + ) # ViT-B/16 - 84.6% Top1 (backbone) + elif backbone == "resnext101_wsl": + pretrained = _make_pretrained_resnext101_wsl(use_pretrained) + scratch = _make_scratch([256, 512, 1024, 2048], features, groups=groups, expand=expand) # efficientnet_lite3 + elif backbone == "efficientnet_lite3": + pretrained = _make_pretrained_efficientnet_lite3(use_pretrained, exportable=exportable) + scratch = _make_scratch([32, 48, 136, 384], features, groups=groups, expand=expand) # efficientnet_lite3 + else: + print(f"Backbone '{backbone}' not implemented") + assert False + + return pretrained, scratch + + +def _make_scratch(in_shape, out_shape, groups=1, expand=False): + scratch = nn.Module() + + out_shape1 = out_shape + out_shape2 = out_shape + out_shape3 = out_shape + out_shape4 = out_shape + if expand==True: + out_shape1 = out_shape + out_shape2 = out_shape*2 + out_shape3 = out_shape*4 + out_shape4 = out_shape*8 + + scratch.layer1_rn = nn.Conv2d( + in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer2_rn = nn.Conv2d( + in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer3_rn = nn.Conv2d( + in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer4_rn = nn.Conv2d( + in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + + return scratch + + +def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False): + efficientnet = torch.hub.load( + "rwightman/gen-efficientnet-pytorch", + "tf_efficientnet_lite3", + pretrained=use_pretrained, + exportable=exportable + ) + return _make_efficientnet_backbone(efficientnet) + + +def _make_efficientnet_backbone(effnet): + pretrained = nn.Module() + + pretrained.layer1 = nn.Sequential( + effnet.conv_stem, effnet.bn1, effnet.act1, *effnet.blocks[0:2] + ) + pretrained.layer2 = nn.Sequential(*effnet.blocks[2:3]) + pretrained.layer3 = nn.Sequential(*effnet.blocks[3:5]) + pretrained.layer4 = nn.Sequential(*effnet.blocks[5:9]) + + return pretrained + + +def _make_resnet_backbone(resnet): + pretrained = nn.Module() + pretrained.layer1 = nn.Sequential( + resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1 + ) + + pretrained.layer2 = resnet.layer2 + pretrained.layer3 = resnet.layer3 + pretrained.layer4 = resnet.layer4 + + return pretrained + + +def _make_pretrained_resnext101_wsl(use_pretrained): + resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl") + return _make_resnet_backbone(resnet) + + + +class Interpolate(nn.Module): + """Interpolation module. + """ + + def __init__(self, scale_factor, mode, align_corners=False): + """Init. + + Args: + scale_factor (float): scaling + mode (str): interpolation mode + """ + super(Interpolate, self).__init__() + + self.interp = nn.functional.interpolate + self.scale_factor = scale_factor + self.mode = mode + self.align_corners = align_corners + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: interpolated data + """ + + x = self.interp( + x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners + ) + + return x + + +class ResidualConvUnit(nn.Module): + """Residual convolution module. + """ + + def __init__(self, features): + """Init. + + Args: + features (int): number of features + """ + super().__init__() + + self.conv1 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True + ) + + self.conv2 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True + ) + + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: output + """ + out = self.relu(x) + out = self.conv1(out) + out = self.relu(out) + out = self.conv2(out) + + return out + x + + +class FeatureFusionBlock(nn.Module): + """Feature fusion block. + """ + + def __init__(self, features): + """Init. + + Args: + features (int): number of features + """ + super(FeatureFusionBlock, self).__init__() + + self.resConfUnit1 = ResidualConvUnit(features) + self.resConfUnit2 = ResidualConvUnit(features) + + def forward(self, *xs): + """Forward pass. + + Returns: + tensor: output + """ + output = xs[0] + + if len(xs) == 2: + output += self.resConfUnit1(xs[1]) + + output = self.resConfUnit2(output) + + output = nn.functional.interpolate( + output, scale_factor=2, mode="bilinear", align_corners=True + ) + + return output + + + + +class ResidualConvUnit_custom(nn.Module): + """Residual convolution module. + """ + + def __init__(self, features, activation, bn): + """Init. + + Args: + features (int): number of features + """ + super().__init__() + + self.bn = bn + + self.groups=1 + + self.conv1 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups + ) + + self.conv2 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups + ) + + if self.bn==True: + self.bn1 = nn.BatchNorm2d(features) + self.bn2 = nn.BatchNorm2d(features) + + self.activation = activation + + self.skip_add = nn.quantized.FloatFunctional() + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: output + """ + + out = self.activation(x) + out = self.conv1(out) + if self.bn==True: + out = self.bn1(out) + + out = self.activation(out) + out = self.conv2(out) + if self.bn==True: + out = self.bn2(out) + + if self.groups > 1: + out = self.conv_merge(out) + + return self.skip_add.add(out, x) + + # return out + x + + +class FeatureFusionBlock_custom(nn.Module): + """Feature fusion block. + """ + + def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True): + """Init. + + Args: + features (int): number of features + """ + super(FeatureFusionBlock_custom, self).__init__() + + self.deconv = deconv + self.align_corners = align_corners + + self.groups=1 + + self.expand = expand + out_features = features + if self.expand==True: + out_features = features//2 + + self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1) + + self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn) + self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn) + + self.skip_add = nn.quantized.FloatFunctional() + + def forward(self, *xs): + """Forward pass. + + Returns: + tensor: output + """ + output = xs[0] + + if len(xs) == 2: + res = self.resConfUnit1(xs[1]) + output = self.skip_add.add(output, res) + # output += res + + output = self.resConfUnit2(output) + + output = nn.functional.interpolate( + output, scale_factor=2, mode="bilinear", align_corners=self.align_corners + ) + + output = self.out_conv(output) + + return output + diff --git a/ldm/modules/midas/midas/dpt_depth.py b/ldm/modules/midas/midas/dpt_depth.py new file mode 100644 index 0000000000000000000000000000000000000000..4e9aab5d2767dffea39da5b3f30e2798688216f1 --- /dev/null +++ b/ldm/modules/midas/midas/dpt_depth.py @@ -0,0 +1,109 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F + +from .base_model import BaseModel +from .blocks import ( + FeatureFusionBlock, + FeatureFusionBlock_custom, + Interpolate, + _make_encoder, + forward_vit, +) + + +def _make_fusion_block(features, use_bn): + return FeatureFusionBlock_custom( + features, + nn.ReLU(False), + deconv=False, + bn=use_bn, + expand=False, + align_corners=True, + ) + + +class DPT(BaseModel): + def __init__( + self, + head, + features=256, + backbone="vitb_rn50_384", + readout="project", + channels_last=False, + use_bn=False, + ): + + super(DPT, self).__init__() + + self.channels_last = channels_last + + hooks = { + "vitb_rn50_384": [0, 1, 8, 11], + "vitb16_384": [2, 5, 8, 11], + "vitl16_384": [5, 11, 17, 23], + } + + # Instantiate backbone and reassemble blocks + self.pretrained, self.scratch = _make_encoder( + backbone, + features, + False, # Set to true of you want to train from scratch, uses ImageNet weights + groups=1, + expand=False, + exportable=False, + hooks=hooks[backbone], + use_readout=readout, + ) + + self.scratch.refinenet1 = _make_fusion_block(features, use_bn) + self.scratch.refinenet2 = _make_fusion_block(features, use_bn) + self.scratch.refinenet3 = _make_fusion_block(features, use_bn) + self.scratch.refinenet4 = _make_fusion_block(features, use_bn) + + self.scratch.output_conv = head + + + def forward(self, x): + if self.channels_last == True: + x.contiguous(memory_format=torch.channels_last) + + layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return out + + +class DPTDepthModel(DPT): + def __init__(self, path=None, non_negative=True, **kwargs): + features = kwargs["features"] if "features" in kwargs else 256 + + head = nn.Sequential( + nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1), + Interpolate(scale_factor=2, mode="bilinear", align_corners=True), + nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1), + nn.ReLU(True), + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + nn.Identity(), + ) + + super().__init__(head, **kwargs) + + if path is not None: + self.load(path) + + def forward(self, x): + return super().forward(x).squeeze(dim=1) + diff --git a/ldm/modules/midas/midas/midas_net.py b/ldm/modules/midas/midas/midas_net.py new file mode 100644 index 0000000000000000000000000000000000000000..8a954977800b0a0f48807e80fa63041910e33c1f --- /dev/null +++ b/ldm/modules/midas/midas/midas_net.py @@ -0,0 +1,76 @@ +"""MidashNet: Network for monocular depth estimation trained by mixing several datasets. +This file contains code that is adapted from +https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py +""" +import torch +import torch.nn as nn + +from .base_model import BaseModel +from .blocks import FeatureFusionBlock, Interpolate, _make_encoder + + +class MidasNet(BaseModel): + """Network for monocular depth estimation. + """ + + def __init__(self, path=None, features=256, non_negative=True): + """Init. + + Args: + path (str, optional): Path to saved model. Defaults to None. + features (int, optional): Number of features. Defaults to 256. + backbone (str, optional): Backbone network for encoder. Defaults to resnet50 + """ + print("Loading weights: ", path) + + super(MidasNet, self).__init__() + + use_pretrained = False if path is None else True + + self.pretrained, self.scratch = _make_encoder(backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained) + + self.scratch.refinenet4 = FeatureFusionBlock(features) + self.scratch.refinenet3 = FeatureFusionBlock(features) + self.scratch.refinenet2 = FeatureFusionBlock(features) + self.scratch.refinenet1 = FeatureFusionBlock(features) + + self.scratch.output_conv = nn.Sequential( + nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1), + Interpolate(scale_factor=2, mode="bilinear"), + nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1), + nn.ReLU(True), + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + ) + + if path: + self.load(path) + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input data (image) + + Returns: + tensor: depth + """ + + layer_1 = self.pretrained.layer1(x) + layer_2 = self.pretrained.layer2(layer_1) + layer_3 = self.pretrained.layer3(layer_2) + layer_4 = self.pretrained.layer4(layer_3) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return torch.squeeze(out, dim=1) diff --git a/ldm/modules/midas/midas/midas_net_custom.py b/ldm/modules/midas/midas/midas_net_custom.py new file mode 100644 index 0000000000000000000000000000000000000000..50e4acb5e53d5fabefe3dde16ab49c33c2b7797c --- /dev/null +++ b/ldm/modules/midas/midas/midas_net_custom.py @@ -0,0 +1,128 @@ +"""MidashNet: Network for monocular depth estimation trained by mixing several datasets. +This file contains code that is adapted from +https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py +""" +import torch +import torch.nn as nn + +from .base_model import BaseModel +from .blocks import FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder + + +class MidasNet_small(BaseModel): + """Network for monocular depth estimation. + """ + + def __init__(self, path=None, features=64, backbone="efficientnet_lite3", non_negative=True, exportable=True, channels_last=False, align_corners=True, + blocks={'expand': True}): + """Init. + + Args: + path (str, optional): Path to saved model. Defaults to None. + features (int, optional): Number of features. Defaults to 256. + backbone (str, optional): Backbone network for encoder. Defaults to resnet50 + """ + print("Loading weights: ", path) + + super(MidasNet_small, self).__init__() + + use_pretrained = False if path else True + + self.channels_last = channels_last + self.blocks = blocks + self.backbone = backbone + + self.groups = 1 + + features1=features + features2=features + features3=features + features4=features + self.expand = False + if "expand" in self.blocks and self.blocks['expand'] == True: + self.expand = True + features1=features + features2=features*2 + features3=features*4 + features4=features*8 + + self.pretrained, self.scratch = _make_encoder(self.backbone, features, use_pretrained, groups=self.groups, expand=self.expand, exportable=exportable) + + self.scratch.activation = nn.ReLU(False) + + self.scratch.refinenet4 = FeatureFusionBlock_custom(features4, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet3 = FeatureFusionBlock_custom(features3, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet2 = FeatureFusionBlock_custom(features2, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet1 = FeatureFusionBlock_custom(features1, self.scratch.activation, deconv=False, bn=False, align_corners=align_corners) + + + self.scratch.output_conv = nn.Sequential( + nn.Conv2d(features, features//2, kernel_size=3, stride=1, padding=1, groups=self.groups), + Interpolate(scale_factor=2, mode="bilinear"), + nn.Conv2d(features//2, 32, kernel_size=3, stride=1, padding=1), + self.scratch.activation, + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + nn.Identity(), + ) + + if path: + self.load(path) + + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input data (image) + + Returns: + tensor: depth + """ + if self.channels_last==True: + print("self.channels_last = ", self.channels_last) + x.contiguous(memory_format=torch.channels_last) + + + layer_1 = self.pretrained.layer1(x) + layer_2 = self.pretrained.layer2(layer_1) + layer_3 = self.pretrained.layer3(layer_2) + layer_4 = self.pretrained.layer4(layer_3) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return torch.squeeze(out, dim=1) + + + +def fuse_model(m): + prev_previous_type = nn.Identity() + prev_previous_name = '' + previous_type = nn.Identity() + previous_name = '' + for name, module in m.named_modules(): + if prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d and type(module) == nn.ReLU: + # print("FUSED ", prev_previous_name, previous_name, name) + torch.quantization.fuse_modules(m, [prev_previous_name, previous_name, name], inplace=True) + elif prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d: + # print("FUSED ", prev_previous_name, previous_name) + torch.quantization.fuse_modules(m, [prev_previous_name, previous_name], inplace=True) + # elif previous_type == nn.Conv2d and type(module) == nn.ReLU: + # print("FUSED ", previous_name, name) + # torch.quantization.fuse_modules(m, [previous_name, name], inplace=True) + + prev_previous_type = previous_type + prev_previous_name = previous_name + previous_type = type(module) + previous_name = name \ No newline at end of file diff --git a/ldm/modules/midas/midas/transforms.py b/ldm/modules/midas/midas/transforms.py new file mode 100644 index 0000000000000000000000000000000000000000..350cbc11662633ad7f8968eb10be2e7de6e384e9 --- /dev/null +++ b/ldm/modules/midas/midas/transforms.py @@ -0,0 +1,234 @@ +import numpy as np +import cv2 +import math + + +def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA): + """Rezise the sample to ensure the given size. Keeps aspect ratio. + + Args: + sample (dict): sample + size (tuple): image size + + Returns: + tuple: new size + """ + shape = list(sample["disparity"].shape) + + if shape[0] >= size[0] and shape[1] >= size[1]: + return sample + + scale = [0, 0] + scale[0] = size[0] / shape[0] + scale[1] = size[1] / shape[1] + + scale = max(scale) + + shape[0] = math.ceil(scale * shape[0]) + shape[1] = math.ceil(scale * shape[1]) + + # resize + sample["image"] = cv2.resize( + sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method + ) + + sample["disparity"] = cv2.resize( + sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST + ) + sample["mask"] = cv2.resize( + sample["mask"].astype(np.float32), + tuple(shape[::-1]), + interpolation=cv2.INTER_NEAREST, + ) + sample["mask"] = sample["mask"].astype(bool) + + return tuple(shape) + + +class Resize(object): + """Resize sample to given size (width, height). + """ + + def __init__( + self, + width, + height, + resize_target=True, + keep_aspect_ratio=False, + ensure_multiple_of=1, + resize_method="lower_bound", + image_interpolation_method=cv2.INTER_AREA, + ): + """Init. + + Args: + width (int): desired output width + height (int): desired output height + resize_target (bool, optional): + True: Resize the full sample (image, mask, target). + False: Resize image only. + Defaults to True. + keep_aspect_ratio (bool, optional): + True: Keep the aspect ratio of the input sample. + Output sample might not have the given width and height, and + resize behaviour depends on the parameter 'resize_method'. + Defaults to False. + ensure_multiple_of (int, optional): + Output width and height is constrained to be multiple of this parameter. + Defaults to 1. + resize_method (str, optional): + "lower_bound": Output will be at least as large as the given size. + "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.) + "minimal": Scale as least as possible. (Output size might be smaller than given size.) + Defaults to "lower_bound". + """ + self.__width = width + self.__height = height + + self.__resize_target = resize_target + self.__keep_aspect_ratio = keep_aspect_ratio + self.__multiple_of = ensure_multiple_of + self.__resize_method = resize_method + self.__image_interpolation_method = image_interpolation_method + + def constrain_to_multiple_of(self, x, min_val=0, max_val=None): + y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int) + + if max_val is not None and y > max_val: + y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int) + + if y < min_val: + y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int) + + return y + + def get_size(self, width, height): + # determine new height and width + scale_height = self.__height / height + scale_width = self.__width / width + + if self.__keep_aspect_ratio: + if self.__resize_method == "lower_bound": + # scale such that output size is lower bound + if scale_width > scale_height: + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + elif self.__resize_method == "upper_bound": + # scale such that output size is upper bound + if scale_width < scale_height: + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + elif self.__resize_method == "minimal": + # scale as least as possbile + if abs(1 - scale_width) < abs(1 - scale_height): + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + else: + raise ValueError( + f"resize_method {self.__resize_method} not implemented" + ) + + if self.__resize_method == "lower_bound": + new_height = self.constrain_to_multiple_of( + scale_height * height, min_val=self.__height + ) + new_width = self.constrain_to_multiple_of( + scale_width * width, min_val=self.__width + ) + elif self.__resize_method == "upper_bound": + new_height = self.constrain_to_multiple_of( + scale_height * height, max_val=self.__height + ) + new_width = self.constrain_to_multiple_of( + scale_width * width, max_val=self.__width + ) + elif self.__resize_method == "minimal": + new_height = self.constrain_to_multiple_of(scale_height * height) + new_width = self.constrain_to_multiple_of(scale_width * width) + else: + raise ValueError(f"resize_method {self.__resize_method} not implemented") + + return (new_width, new_height) + + def __call__(self, sample): + width, height = self.get_size( + sample["image"].shape[1], sample["image"].shape[0] + ) + + # resize sample + sample["image"] = cv2.resize( + sample["image"], + (width, height), + interpolation=self.__image_interpolation_method, + ) + + if self.__resize_target: + if "disparity" in sample: + sample["disparity"] = cv2.resize( + sample["disparity"], + (width, height), + interpolation=cv2.INTER_NEAREST, + ) + + if "depth" in sample: + sample["depth"] = cv2.resize( + sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST + ) + + sample["mask"] = cv2.resize( + sample["mask"].astype(np.float32), + (width, height), + interpolation=cv2.INTER_NEAREST, + ) + sample["mask"] = sample["mask"].astype(bool) + + return sample + + +class NormalizeImage(object): + """Normlize image by given mean and std. + """ + + def __init__(self, mean, std): + self.__mean = mean + self.__std = std + + def __call__(self, sample): + sample["image"] = (sample["image"] - self.__mean) / self.__std + + return sample + + +class PrepareForNet(object): + """Prepare sample for usage as network input. + """ + + def __init__(self): + pass + + def __call__(self, sample): + image = np.transpose(sample["image"], (2, 0, 1)) + sample["image"] = np.ascontiguousarray(image).astype(np.float32) + + if "mask" in sample: + sample["mask"] = sample["mask"].astype(np.float32) + sample["mask"] = np.ascontiguousarray(sample["mask"]) + + if "disparity" in sample: + disparity = sample["disparity"].astype(np.float32) + sample["disparity"] = np.ascontiguousarray(disparity) + + if "depth" in sample: + depth = sample["depth"].astype(np.float32) + sample["depth"] = np.ascontiguousarray(depth) + + return sample diff --git a/ldm/modules/midas/midas/vit.py b/ldm/modules/midas/midas/vit.py new file mode 100644 index 0000000000000000000000000000000000000000..ea46b1be88b261b0dec04f3da0256f5f66f88a74 --- /dev/null +++ b/ldm/modules/midas/midas/vit.py @@ -0,0 +1,491 @@ +import torch +import torch.nn as nn +import timm +import types +import math +import torch.nn.functional as F + + +class Slice(nn.Module): + def __init__(self, start_index=1): + super(Slice, self).__init__() + self.start_index = start_index + + def forward(self, x): + return x[:, self.start_index :] + + +class AddReadout(nn.Module): + def __init__(self, start_index=1): + super(AddReadout, self).__init__() + self.start_index = start_index + + def forward(self, x): + if self.start_index == 2: + readout = (x[:, 0] + x[:, 1]) / 2 + else: + readout = x[:, 0] + return x[:, self.start_index :] + readout.unsqueeze(1) + + +class ProjectReadout(nn.Module): + def __init__(self, in_features, start_index=1): + super(ProjectReadout, self).__init__() + self.start_index = start_index + + self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU()) + + def forward(self, x): + readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :]) + features = torch.cat((x[:, self.start_index :], readout), -1) + + return self.project(features) + + +class Transpose(nn.Module): + def __init__(self, dim0, dim1): + super(Transpose, self).__init__() + self.dim0 = dim0 + self.dim1 = dim1 + + def forward(self, x): + x = x.transpose(self.dim0, self.dim1) + return x + + +def forward_vit(pretrained, x): + b, c, h, w = x.shape + + glob = pretrained.model.forward_flex(x) + + layer_1 = pretrained.activations["1"] + layer_2 = pretrained.activations["2"] + layer_3 = pretrained.activations["3"] + layer_4 = pretrained.activations["4"] + + layer_1 = pretrained.act_postprocess1[0:2](layer_1) + layer_2 = pretrained.act_postprocess2[0:2](layer_2) + layer_3 = pretrained.act_postprocess3[0:2](layer_3) + layer_4 = pretrained.act_postprocess4[0:2](layer_4) + + unflatten = nn.Sequential( + nn.Unflatten( + 2, + torch.Size( + [ + h // pretrained.model.patch_size[1], + w // pretrained.model.patch_size[0], + ] + ), + ) + ) + + if layer_1.ndim == 3: + layer_1 = unflatten(layer_1) + if layer_2.ndim == 3: + layer_2 = unflatten(layer_2) + if layer_3.ndim == 3: + layer_3 = unflatten(layer_3) + if layer_4.ndim == 3: + layer_4 = unflatten(layer_4) + + layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1) + layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2) + layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3) + layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4) + + return layer_1, layer_2, layer_3, layer_4 + + +def _resize_pos_embed(self, posemb, gs_h, gs_w): + posemb_tok, posemb_grid = ( + posemb[:, : self.start_index], + posemb[0, self.start_index :], + ) + + gs_old = int(math.sqrt(len(posemb_grid))) + + posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2) + posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear") + posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1) + + posemb = torch.cat([posemb_tok, posemb_grid], dim=1) + + return posemb + + +def forward_flex(self, x): + b, c, h, w = x.shape + + pos_embed = self._resize_pos_embed( + self.pos_embed, h // self.patch_size[1], w // self.patch_size[0] + ) + + B = x.shape[0] + + if hasattr(self.patch_embed, "backbone"): + x = self.patch_embed.backbone(x) + if isinstance(x, (list, tuple)): + x = x[-1] # last feature if backbone outputs list/tuple of features + + x = self.patch_embed.proj(x).flatten(2).transpose(1, 2) + + if getattr(self, "dist_token", None) is not None: + cls_tokens = self.cls_token.expand( + B, -1, -1 + ) # stole cls_tokens impl from Phil Wang, thanks + dist_token = self.dist_token.expand(B, -1, -1) + x = torch.cat((cls_tokens, dist_token, x), dim=1) + else: + cls_tokens = self.cls_token.expand( + B, -1, -1 + ) # stole cls_tokens impl from Phil Wang, thanks + x = torch.cat((cls_tokens, x), dim=1) + + x = x + pos_embed + x = self.pos_drop(x) + + for blk in self.blocks: + x = blk(x) + + x = self.norm(x) + + return x + + +activations = {} + + +def get_activation(name): + def hook(model, input, output): + activations[name] = output + + return hook + + +def get_readout_oper(vit_features, features, use_readout, start_index=1): + if use_readout == "ignore": + readout_oper = [Slice(start_index)] * len(features) + elif use_readout == "add": + readout_oper = [AddReadout(start_index)] * len(features) + elif use_readout == "project": + readout_oper = [ + ProjectReadout(vit_features, start_index) for out_feat in features + ] + else: + assert ( + False + ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'" + + return readout_oper + + +def _make_vit_b16_backbone( + model, + features=[96, 192, 384, 768], + size=[384, 384], + hooks=[2, 5, 8, 11], + vit_features=768, + use_readout="ignore", + start_index=1, +): + pretrained = nn.Module() + + pretrained.model = model + pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1")) + pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2")) + pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3")) + pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4")) + + pretrained.activations = activations + + readout_oper = get_readout_oper(vit_features, features, use_readout, start_index) + + # 32, 48, 136, 384 + pretrained.act_postprocess1 = nn.Sequential( + readout_oper[0], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[0], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[0], + out_channels=features[0], + kernel_size=4, + stride=4, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess2 = nn.Sequential( + readout_oper[1], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[1], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[1], + out_channels=features[1], + kernel_size=2, + stride=2, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess3 = nn.Sequential( + readout_oper[2], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[2], + kernel_size=1, + stride=1, + padding=0, + ), + ) + + pretrained.act_postprocess4 = nn.Sequential( + readout_oper[3], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[3], + kernel_size=1, + stride=1, + padding=0, + ), + nn.Conv2d( + in_channels=features[3], + out_channels=features[3], + kernel_size=3, + stride=2, + padding=1, + ), + ) + + pretrained.model.start_index = start_index + pretrained.model.patch_size = [16, 16] + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model) + pretrained.model._resize_pos_embed = types.MethodType( + _resize_pos_embed, pretrained.model + ) + + return pretrained + + +def _make_pretrained_vitl16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_large_patch16_384", pretrained=pretrained) + + hooks = [5, 11, 17, 23] if hooks == None else hooks + return _make_vit_b16_backbone( + model, + features=[256, 512, 1024, 1024], + hooks=hooks, + vit_features=1024, + use_readout=use_readout, + ) + + +def _make_pretrained_vitb16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_base_patch16_384", pretrained=pretrained) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout + ) + + +def _make_pretrained_deitb16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout + ) + + +def _make_pretrained_deitb16_distil_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model( + "vit_deit_base_distilled_patch16_384", pretrained=pretrained + ) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, + features=[96, 192, 384, 768], + hooks=hooks, + use_readout=use_readout, + start_index=2, + ) + + +def _make_vit_b_rn50_backbone( + model, + features=[256, 512, 768, 768], + size=[384, 384], + hooks=[0, 1, 8, 11], + vit_features=768, + use_vit_only=False, + use_readout="ignore", + start_index=1, +): + pretrained = nn.Module() + + pretrained.model = model + + if use_vit_only == True: + pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1")) + pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2")) + else: + pretrained.model.patch_embed.backbone.stages[0].register_forward_hook( + get_activation("1") + ) + pretrained.model.patch_embed.backbone.stages[1].register_forward_hook( + get_activation("2") + ) + + pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3")) + pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4")) + + pretrained.activations = activations + + readout_oper = get_readout_oper(vit_features, features, use_readout, start_index) + + if use_vit_only == True: + pretrained.act_postprocess1 = nn.Sequential( + readout_oper[0], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[0], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[0], + out_channels=features[0], + kernel_size=4, + stride=4, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess2 = nn.Sequential( + readout_oper[1], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[1], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[1], + out_channels=features[1], + kernel_size=2, + stride=2, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + else: + pretrained.act_postprocess1 = nn.Sequential( + nn.Identity(), nn.Identity(), nn.Identity() + ) + pretrained.act_postprocess2 = nn.Sequential( + nn.Identity(), nn.Identity(), nn.Identity() + ) + + pretrained.act_postprocess3 = nn.Sequential( + readout_oper[2], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[2], + kernel_size=1, + stride=1, + padding=0, + ), + ) + + pretrained.act_postprocess4 = nn.Sequential( + readout_oper[3], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[3], + kernel_size=1, + stride=1, + padding=0, + ), + nn.Conv2d( + in_channels=features[3], + out_channels=features[3], + kernel_size=3, + stride=2, + padding=1, + ), + ) + + pretrained.model.start_index = start_index + pretrained.model.patch_size = [16, 16] + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model) + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model._resize_pos_embed = types.MethodType( + _resize_pos_embed, pretrained.model + ) + + return pretrained + + +def _make_pretrained_vitb_rn50_384( + pretrained, use_readout="ignore", hooks=None, use_vit_only=False +): + model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained) + + hooks = [0, 1, 8, 11] if hooks == None else hooks + return _make_vit_b_rn50_backbone( + model, + features=[256, 512, 768, 768], + size=[384, 384], + hooks=hooks, + use_vit_only=use_vit_only, + use_readout=use_readout, + ) diff --git a/ldm/modules/midas/utils.py b/ldm/modules/midas/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..9a9d3b5b66370fa98da9e067ba53ead848ea9a59 --- /dev/null +++ b/ldm/modules/midas/utils.py @@ -0,0 +1,189 @@ +"""Utils for monoDepth.""" +import sys +import re +import numpy as np +import cv2 +import torch + + +def read_pfm(path): + """Read pfm file. + + Args: + path (str): path to file + + Returns: + tuple: (data, scale) + """ + with open(path, "rb") as file: + + color = None + width = None + height = None + scale = None + endian = None + + header = file.readline().rstrip() + if header.decode("ascii") == "PF": + color = True + elif header.decode("ascii") == "Pf": + color = False + else: + raise Exception("Not a PFM file: " + path) + + dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii")) + if dim_match: + width, height = list(map(int, dim_match.groups())) + else: + raise Exception("Malformed PFM header.") + + scale = float(file.readline().decode("ascii").rstrip()) + if scale < 0: + # little-endian + endian = "<" + scale = -scale + else: + # big-endian + endian = ">" + + data = np.fromfile(file, endian + "f") + shape = (height, width, 3) if color else (height, width) + + data = np.reshape(data, shape) + data = np.flipud(data) + + return data, scale + + +def write_pfm(path, image, scale=1): + """Write pfm file. + + Args: + path (str): pathto file + image (array): data + scale (int, optional): Scale. Defaults to 1. + """ + + with open(path, "wb") as file: + color = None + + if image.dtype.name != "float32": + raise Exception("Image dtype must be float32.") + + image = np.flipud(image) + + if len(image.shape) == 3 and image.shape[2] == 3: # color image + color = True + elif ( + len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1 + ): # greyscale + color = False + else: + raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.") + + file.write("PF\n" if color else "Pf\n".encode()) + file.write("%d %d\n".encode() % (image.shape[1], image.shape[0])) + + endian = image.dtype.byteorder + + if endian == "<" or endian == "=" and sys.byteorder == "little": + scale = -scale + + file.write("%f\n".encode() % scale) + + image.tofile(file) + + +def read_image(path): + """Read image and output RGB image (0-1). + + Args: + path (str): path to file + + Returns: + array: RGB image (0-1) + """ + img = cv2.imread(path) + + if img.ndim == 2: + img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) + + img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0 + + return img + + +def resize_image(img): + """Resize image and make it fit for network. + + Args: + img (array): image + + Returns: + tensor: data ready for network + """ + height_orig = img.shape[0] + width_orig = img.shape[1] + + if width_orig > height_orig: + scale = width_orig / 384 + else: + scale = height_orig / 384 + + height = (np.ceil(height_orig / scale / 32) * 32).astype(int) + width = (np.ceil(width_orig / scale / 32) * 32).astype(int) + + img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA) + + img_resized = ( + torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float() + ) + img_resized = img_resized.unsqueeze(0) + + return img_resized + + +def resize_depth(depth, width, height): + """Resize depth map and bring to CPU (numpy). + + Args: + depth (tensor): depth + width (int): image width + height (int): image height + + Returns: + array: processed depth + """ + depth = torch.squeeze(depth[0, :, :, :]).to("cpu") + + depth_resized = cv2.resize( + depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC + ) + + return depth_resized + +def write_depth(path, depth, bits=1): + """Write depth map to pfm and png file. + + Args: + path (str): filepath without extension + depth (array): depth + """ + write_pfm(path + ".pfm", depth.astype(np.float32)) + + depth_min = depth.min() + depth_max = depth.max() + + max_val = (2**(8*bits))-1 + + if depth_max - depth_min > np.finfo("float").eps: + out = max_val * (depth - depth_min) / (depth_max - depth_min) + else: + out = np.zeros(depth.shape, dtype=depth.type) + + if bits == 1: + cv2.imwrite(path + ".png", out.astype("uint8")) + elif bits == 2: + cv2.imwrite(path + ".png", out.astype("uint16")) + + return diff --git a/ldm/util.py b/ldm/util.py new file mode 100644 index 0000000000000000000000000000000000000000..45cb050ece6f401a22dde098ce3f1ff663c5eb6a --- /dev/null +++ b/ldm/util.py @@ -0,0 +1,197 @@ +import importlib + +import torch +from torch import optim +import numpy as np + +from inspect import isfunction +from PIL import Image, ImageDraw, ImageFont + + +def log_txt_as_img(wh, xc, size=10): + # wh a tuple of (width, height) + # xc a list of captions to plot + b = len(xc) + txts = list() + for bi in range(b): + txt = Image.new("RGB", wh, color="white") + draw = ImageDraw.Draw(txt) + font = ImageFont.truetype('font/DejaVuSans.ttf', size=size) + nc = int(40 * (wh[0] / 256)) + lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc)) + + try: + draw.text((0, 0), lines, fill="black", font=font) + except UnicodeEncodeError: + print("Cant encode string for logging. Skipping.") + + txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0 + txts.append(txt) + txts = np.stack(txts) + txts = torch.tensor(txts) + return txts + + +def ismap(x): + if not isinstance(x, torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] > 3) + + +def isimage(x): + if not isinstance(x,torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1) + + +def exists(x): + return x is not None + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def mean_flat(tensor): + """ + https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86 + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def count_params(model, verbose=False): + total_params = sum(p.numel() for p in model.parameters()) + if verbose: + print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.") + return total_params + + +def instantiate_from_config(config): + if not "target" in config: + if config == '__is_first_stage__': + return None + elif config == "__is_unconditional__": + return None + raise KeyError("Expected key `target` to instantiate.") + return get_obj_from_str(config["target"])(**config.get("params", dict())) + + +def get_obj_from_str(string, reload=False): + module, cls = string.rsplit(".", 1) + if reload: + module_imp = importlib.import_module(module) + importlib.reload(module_imp) + return getattr(importlib.import_module(module, package=None), cls) + + +class AdamWwithEMAandWings(optim.Optimizer): + # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298 + def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8, # TODO: check hyperparameters before using + weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999, # ema decay to match previous code + ema_power=1., param_names=()): + """AdamW that saves EMA versions of the parameters.""" + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0.0 <= weight_decay: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + if not 0.0 <= ema_decay <= 1.0: + raise ValueError("Invalid ema_decay value: {}".format(ema_decay)) + defaults = dict(lr=lr, betas=betas, eps=eps, + weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay, + ema_power=ema_power, param_names=param_names) + super().__init__(params, defaults) + + def __setstate__(self, state): + super().__setstate__(state) + for group in self.param_groups: + group.setdefault('amsgrad', False) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + params_with_grad = [] + grads = [] + exp_avgs = [] + exp_avg_sqs = [] + ema_params_with_grad = [] + state_sums = [] + max_exp_avg_sqs = [] + state_steps = [] + amsgrad = group['amsgrad'] + beta1, beta2 = group['betas'] + ema_decay = group['ema_decay'] + ema_power = group['ema_power'] + + for p in group['params']: + if p.grad is None: + continue + params_with_grad.append(p) + if p.grad.is_sparse: + raise RuntimeError('AdamW does not support sparse gradients') + grads.append(p.grad) + + state = self.state[p] + + # State initialization + if len(state) == 0: + state['step'] = 0 + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format) + # Exponential moving average of squared gradient values + state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) + if amsgrad: + # Maintains max of all exp. moving avg. of sq. grad. values + state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) + # Exponential moving average of parameter values + state['param_exp_avg'] = p.detach().float().clone() + + exp_avgs.append(state['exp_avg']) + exp_avg_sqs.append(state['exp_avg_sq']) + ema_params_with_grad.append(state['param_exp_avg']) + + if amsgrad: + max_exp_avg_sqs.append(state['max_exp_avg_sq']) + + # update the steps for each param group update + state['step'] += 1 + # record the step after step update + state_steps.append(state['step']) + + optim._functional.adamw(params_with_grad, + grads, + exp_avgs, + exp_avg_sqs, + max_exp_avg_sqs, + state_steps, + amsgrad=amsgrad, + beta1=beta1, + beta2=beta2, + lr=group['lr'], + weight_decay=group['weight_decay'], + eps=group['eps'], + maximize=False) + + cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power) + for param, ema_param in zip(params_with_grad, ema_params_with_grad): + ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay) + + return loss \ No newline at end of file diff --git a/lvdm/basics.py b/lvdm/basics.py new file mode 100644 index 0000000000000000000000000000000000000000..65c771d13a7f4a932ac370f08797a8b6ba9e85ff --- /dev/null +++ b/lvdm/basics.py @@ -0,0 +1,100 @@ +# adopted from +# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# and +# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +# and +# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py +# +# thanks! + +import torch.nn as nn +from utils.utils import instantiate_from_config + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + +def scale_module(module, scale): + """ + Scale the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().mul_(scale) + return module + + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def linear(*args, **kwargs): + """ + Create a linear module. + """ + return nn.Linear(*args, **kwargs) + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def nonlinearity(type='silu'): + if type == 'silu': + return nn.SiLU() + elif type == 'leaky_relu': + return nn.LeakyReLU() + + +class GroupNormSpecific(nn.GroupNorm): + def forward(self, x): + return super().forward(x.float()).type(x.dtype) + + +def normalization(channels, num_groups=32): + """ + Make a standard normalization layer. + :param channels: number of input channels. + :return: an nn.Module for normalization. + """ + return GroupNormSpecific(num_groups, channels) + + +class HybridConditioner(nn.Module): + + def __init__(self, c_concat_config, c_crossattn_config): + super().__init__() + self.concat_conditioner = instantiate_from_config(c_concat_config) + self.crossattn_conditioner = instantiate_from_config(c_crossattn_config) + + def forward(self, c_concat, c_crossattn): + c_concat = self.concat_conditioner(c_concat) + c_crossattn = self.crossattn_conditioner(c_crossattn) + return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]} \ No newline at end of file diff --git a/lvdm/common.py b/lvdm/common.py new file mode 100644 index 0000000000000000000000000000000000000000..55a150b618e275f01d3a59ad9c7579176c4ea1b8 --- /dev/null +++ b/lvdm/common.py @@ -0,0 +1,94 @@ +import math +from inspect import isfunction +import torch +from torch import nn +import torch.distributed as dist + + +def gather_data(data, return_np=True): + ''' gather data from multiple processes to one list ''' + data_list = [torch.zeros_like(data) for _ in range(dist.get_world_size())] + dist.all_gather(data_list, data) # gather not supported with NCCL + if return_np: + data_list = [data.cpu().numpy() for data in data_list] + return data_list + +def autocast(f): + def do_autocast(*args, **kwargs): + with torch.cuda.amp.autocast(enabled=True, + dtype=torch.get_autocast_gpu_dtype(), + cache_enabled=torch.is_autocast_cache_enabled()): + return f(*args, **kwargs) + return do_autocast + + +def extract_into_tensor(a, t, x_shape): + b, *_ = t.shape + out = a.gather(-1, t) + return out.reshape(b, *((1,) * (len(x_shape) - 1))) + + +def noise_like(shape, device, repeat=False): + repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1))) + noise = lambda: torch.randn(shape, device=device) + return repeat_noise() if repeat else noise() + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + +def exists(val): + return val is not None + +def identity(*args, **kwargs): + return nn.Identity() + +def uniq(arr): + return{el: True for el in arr}.keys() + +def mean_flat(tensor): + """ + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + +def ismap(x): + if not isinstance(x, torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] > 3) + +def isimage(x): + if not isinstance(x,torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1) + +def max_neg_value(t): + return -torch.finfo(t.dtype).max + +def shape_to_str(x): + shape_str = "x".join([str(x) for x in x.shape]) + return shape_str + +def init_(tensor): + dim = tensor.shape[-1] + std = 1 / math.sqrt(dim) + tensor.uniform_(-std, std) + return tensor + +ckpt = torch.utils.checkpoint.checkpoint +def checkpoint(func, inputs, params, flag): + """ + Evaluate a function without caching intermediate activations, allowing for + reduced memory at the expense of extra compute in the backward pass. + :param func: the function to evaluate. + :param inputs: the argument sequence to pass to `func`. + :param params: a sequence of parameters `func` depends on but does not + explicitly take as arguments. + :param flag: if False, disable gradient checkpointing. + """ + if flag: + return ckpt(func, *inputs, use_reentrant=False) + else: + return func(*inputs) \ No newline at end of file diff --git a/lvdm/data/base.py b/lvdm/data/base.py new file mode 100644 index 0000000000000000000000000000000000000000..01aac7a81cf970c702d452aeaad7b6ff979d97d1 --- /dev/null +++ b/lvdm/data/base.py @@ -0,0 +1,23 @@ +from abc import abstractmethod +from torch.utils.data import IterableDataset + + +class Txt2ImgIterableBaseDataset(IterableDataset): + ''' + Define an interface to make the IterableDatasets for text2img data chainable + ''' + def __init__(self, num_records=0, valid_ids=None, size=256): + super().__init__() + self.num_records = num_records + self.valid_ids = valid_ids + self.sample_ids = valid_ids + self.size = size + + print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.') + + def __len__(self): + return self.num_records + + @abstractmethod + def __iter__(self): + pass \ No newline at end of file diff --git a/lvdm/data/webvid.py b/lvdm/data/webvid.py new file mode 100644 index 0000000000000000000000000000000000000000..d8f10692d45eadf2df4440cdb5d0e4418af386b8 --- /dev/null +++ b/lvdm/data/webvid.py @@ -0,0 +1,202 @@ +import os +import random +from tqdm import tqdm +import pandas as pd +from decord import VideoReader, cpu + +import torch +from torch.utils.data import Dataset +from torch.utils.data import DataLoader +from torchvision import transforms + + +class WebVid(Dataset): + """ + WebVid Dataset. + Assumes webvid data is structured as follows. + Webvid/ + videos/ + 000001_000050/ ($page_dir) + 1.mp4 (videoid.mp4) + ... + 5000.mp4 + ... + """ + def __init__(self, + meta_path, + data_dir, + subsample=None, + video_length=16, + resolution=[256, 512], + frame_stride=1, + frame_stride_min=1, + spatial_transform=None, + crop_resolution=None, + fps_max=None, + load_raw_resolution=False, + fixed_fps=None, + random_fs=False, + ): + self.meta_path = meta_path + self.data_dir = data_dir + self.subsample = subsample + self.video_length = video_length + self.resolution = [resolution, resolution] if isinstance(resolution, int) else resolution + self.fps_max = fps_max + self.frame_stride = frame_stride + self.frame_stride_min = frame_stride_min + self.fixed_fps = fixed_fps + self.load_raw_resolution = load_raw_resolution + self.random_fs = random_fs + self._load_metadata() + if spatial_transform is not None: + if spatial_transform == "random_crop": + self.spatial_transform = transforms.RandomCrop(crop_resolution) + elif spatial_transform == "center_crop": + self.spatial_transform = transforms.Compose([ + transforms.CenterCrop(resolution), + ]) + elif spatial_transform == "resize_center_crop": + # assert(self.resolution[0] == self.resolution[1]) + self.spatial_transform = transforms.Compose([ + transforms.Resize(min(self.resolution)), + transforms.CenterCrop(self.resolution), + ]) + elif spatial_transform == "resize": + self.spatial_transform = transforms.Resize(self.resolution) + else: + raise NotImplementedError + else: + self.spatial_transform = None + + def _load_metadata(self): + metadata = pd.read_csv(self.meta_path) + print(f'>>> {len(metadata)} data samples loaded.') + if self.subsample is not None: + metadata = metadata.sample(self.subsample, random_state=0) + + metadata['caption'] = metadata['name'] + del metadata['name'] + self.metadata = metadata + self.metadata.dropna(inplace=True) + + def _get_video_path(self, sample): + rel_video_fp = os.path.join(sample['page_dir'], str(sample['videoid']) + '.mp4') + full_video_fp = os.path.join(self.data_dir, 'videos', rel_video_fp) + return full_video_fp + + def __getitem__(self, index): + if self.random_fs: + frame_stride = random.randint(self.frame_stride_min, self.frame_stride) + else: + frame_stride = self.frame_stride + + ## get frames until success + while True: + index = index % len(self.metadata) + sample = self.metadata.iloc[index] + video_path = self._get_video_path(sample) + ## video_path should be in the format of "....../WebVid/videos/$page_dir/$videoid.mp4" + caption = sample['caption'] + + try: + if self.load_raw_resolution: + video_reader = VideoReader(video_path, ctx=cpu(0)) + else: + video_reader = VideoReader(video_path, ctx=cpu(0), width=530, height=300) + if len(video_reader) < self.video_length: + print(f"video length ({len(video_reader)}) is smaller than target length({self.video_length})") + index += 1 + continue + else: + pass + except: + index += 1 + print(f"Load video failed! path = {video_path}") + continue + + fps_ori = video_reader.get_avg_fps() + if self.fixed_fps is not None: + frame_stride = int(frame_stride * (1.0 * fps_ori / self.fixed_fps)) + + ## to avoid extreme cases when fixed_fps is used + frame_stride = max(frame_stride, 1) + + ## get valid range (adapting case by case) + required_frame_num = frame_stride * (self.video_length-1) + 1 + frame_num = len(video_reader) + if frame_num < required_frame_num: + ## drop extra samples if fixed fps is required + if self.fixed_fps is not None and frame_num < required_frame_num * 0.5: + index += 1 + continue + else: + frame_stride = frame_num // self.video_length + required_frame_num = frame_stride * (self.video_length-1) + 1 + + ## select a random clip + random_range = frame_num - required_frame_num + start_idx = random.randint(0, random_range) if random_range > 0 else 0 + + ## calculate frame indices + frame_indices = [start_idx + frame_stride*i for i in range(self.video_length)] + try: + frames = video_reader.get_batch(frame_indices) + break + except: + print(f"Get frames failed! path = {video_path}; [max_ind vs frame_total:{max(frame_indices)} / {frame_num}]") + index += 1 + continue + + ## process data + assert(frames.shape[0] == self.video_length),f'{len(frames)}, self.video_length={self.video_length}' + frames = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float() # [t,h,w,c] -> [c,t,h,w] + + if self.spatial_transform is not None: + frames = self.spatial_transform(frames) + + if self.resolution is not None: + assert (frames.shape[2], frames.shape[3]) == (self.resolution[0], self.resolution[1]), f'frames={frames.shape}, self.resolution={self.resolution}' + + ## turn frames tensors to [-1,1] + frames = (frames / 255 - 0.5) * 2 + fps_clip = fps_ori // frame_stride + if self.fps_max is not None and fps_clip > self.fps_max: + fps_clip = self.fps_max + + data = {'video': frames, 'caption': caption, 'path': video_path, 'fps': fps_clip, 'frame_stride': frame_stride} + return data + + def __len__(self): + return len(self.metadata) + + +if __name__== "__main__": + meta_path = "" ## path to the meta file + data_dir = "" ## path to the data directory + save_dir = "" ## path to the save directory + dataset = WebVid(meta_path, + data_dir, + subsample=None, + video_length=16, + resolution=[256,448], + frame_stride=4, + spatial_transform="resize_center_crop", + crop_resolution=None, + fps_max=None, + load_raw_resolution=True + ) + dataloader = DataLoader(dataset, + batch_size=1, + num_workers=0, + shuffle=False) + + + import sys + sys.path.insert(1, os.path.join(sys.path[0], '..', '..')) + from utils.save_video import tensor_to_mp4 + for i, batch in tqdm(enumerate(dataloader), desc="Data Batch"): + video = batch['video'] + name = batch['path'][0].split('videos/')[-1].replace('/','_') + tensor_to_mp4(video, save_dir+'/'+name, fps=8) + diff --git a/lvdm/distributions.py b/lvdm/distributions.py new file mode 100644 index 0000000000000000000000000000000000000000..9a2a82ecace3ce27fb7816ddaf088e179c2d5ffd --- /dev/null +++ b/lvdm/distributions.py @@ -0,0 +1,95 @@ +import torch +import numpy as np + + +class AbstractDistribution: + def sample(self): + raise NotImplementedError() + + def mode(self): + raise NotImplementedError() + + +class DiracDistribution(AbstractDistribution): + def __init__(self, value): + self.value = value + + def sample(self): + return self.value + + def mode(self): + return self.value + + +class DiagonalGaussianDistribution(object): + def __init__(self, parameters, deterministic=False): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device) + + def sample(self, noise=None): + if noise is None: + noise = torch.randn(self.mean.shape) + + x = self.mean + self.std * noise.to(device=self.parameters.device) + return x + + def kl(self, other=None): + if self.deterministic: + return torch.Tensor([0.]) + else: + if other is None: + return 0.5 * torch.sum(torch.pow(self.mean, 2) + + self.var - 1.0 - self.logvar, + dim=[1, 2, 3]) + else: + return 0.5 * torch.sum( + torch.pow(self.mean - other.mean, 2) / other.var + + self.var / other.var - 1.0 - self.logvar + other.logvar, + dim=[1, 2, 3]) + + def nll(self, sample, dims=[1,2,3]): + if self.deterministic: + return torch.Tensor([0.]) + logtwopi = np.log(2.0 * np.pi) + return 0.5 * torch.sum( + logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, + dim=dims) + + def mode(self): + return self.mean + + +def normal_kl(mean1, logvar1, mean2, logvar2): + """ + source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12 + Compute the KL divergence between two gaussians. + Shapes are automatically broadcasted, so batches can be compared to + scalars, among other use cases. + """ + tensor = None + for obj in (mean1, logvar1, mean2, logvar2): + if isinstance(obj, torch.Tensor): + tensor = obj + break + assert tensor is not None, "at least one argument must be a Tensor" + + # Force variances to be Tensors. Broadcasting helps convert scalars to + # Tensors, but it does not work for torch.exp(). + logvar1, logvar2 = [ + x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor) + for x in (logvar1, logvar2) + ] + + return 0.5 * ( + -1.0 + + logvar2 + - logvar1 + + torch.exp(logvar1 - logvar2) + + ((mean1 - mean2) ** 2) * torch.exp(-logvar2) + ) \ No newline at end of file diff --git a/lvdm/ema.py b/lvdm/ema.py new file mode 100644 index 0000000000000000000000000000000000000000..cd2f8e3115f816b4cac674397238cd8c22de9bc2 --- /dev/null +++ b/lvdm/ema.py @@ -0,0 +1,76 @@ +import torch +from torch import nn + + +class LitEma(nn.Module): + def __init__(self, model, decay=0.9999, use_num_upates=True): + super().__init__() + if decay < 0.0 or decay > 1.0: + raise ValueError('Decay must be between 0 and 1') + + self.m_name2s_name = {} + self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32)) + self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates + else torch.tensor(-1,dtype=torch.int)) + + for name, p in model.named_parameters(): + if p.requires_grad: + #remove as '.'-character is not allowed in buffers + s_name = name.replace('.','') + self.m_name2s_name.update({name:s_name}) + self.register_buffer(s_name,p.clone().detach().data) + + self.collected_params = [] + + def forward(self,model): + decay = self.decay + + if self.num_updates >= 0: + self.num_updates += 1 + decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates)) + + one_minus_decay = 1.0 - decay + + with torch.no_grad(): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + + for key in m_param: + if m_param[key].requires_grad: + sname = self.m_name2s_name[key] + shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) + shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) + else: + assert not key in self.m_name2s_name + + def copy_to(self, model): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + for key in m_param: + if m_param[key].requires_grad: + m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) + else: + assert not key in self.m_name2s_name + + def store(self, parameters): + """ + Save the current parameters for restoring later. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + temporarily stored. + """ + self.collected_params = [param.clone() for param in parameters] + + def restore(self, parameters): + """ + Restore the parameters stored with the `store` method. + Useful to validate the model with EMA parameters without affecting the + original optimization process. Store the parameters before the + `copy_to` method. After validation (or model saving), use this to + restore the former parameters. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. + """ + for c_param, param in zip(self.collected_params, parameters): + param.data.copy_(c_param.data) \ No newline at end of file diff --git a/lvdm/models/autoencoder.py b/lvdm/models/autoencoder.py new file mode 100644 index 0000000000000000000000000000000000000000..799d05c3967075be312df0dcd82da68999c9d201 --- /dev/null +++ b/lvdm/models/autoencoder.py @@ -0,0 +1,275 @@ +import os +from contextlib import contextmanager +import torch +import numpy as np +from einops import rearrange +import torch.nn.functional as F +import pytorch_lightning as pl +from lvdm.modules.networks.ae_modules import Encoder, Decoder +from lvdm.distributions import DiagonalGaussianDistribution +from utils.utils import instantiate_from_config + +TIMESTEPS=16 +class AutoencoderKL(pl.LightningModule): + def __init__(self, + ddconfig, + lossconfig, + embed_dim, + ckpt_path=None, + ignore_keys=[], + image_key="image", + colorize_nlabels=None, + monitor=None, + test=False, + logdir=None, + input_dim=4, + test_args=None, + additional_decode_keys=None, + use_checkpoint=False, + diff_boost_factor=3.0, + ): + super().__init__() + self.image_key = image_key + self.encoder = Encoder(**ddconfig) + self.decoder = Decoder(**ddconfig) + self.loss = instantiate_from_config(lossconfig) + assert ddconfig["double_z"] + self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1) + self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) + self.embed_dim = embed_dim + self.input_dim = input_dim + self.test = test + self.test_args = test_args + self.logdir = logdir + if colorize_nlabels is not None: + assert type(colorize_nlabels)==int + self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1)) + if monitor is not None: + self.monitor = monitor + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) + if self.test: + self.init_test() + + def init_test(self,): + self.test = True + save_dir = os.path.join(self.logdir, "test") + if 'ckpt' in self.test_args: + ckpt_name = os.path.basename(self.test_args.ckpt).split('.ckpt')[0] + f'_epoch{self._cur_epoch}' + self.root = os.path.join(save_dir, ckpt_name) + else: + self.root = save_dir + if 'test_subdir' in self.test_args: + self.root = os.path.join(save_dir, self.test_args.test_subdir) + + self.root_zs = os.path.join(self.root, "zs") + self.root_dec = os.path.join(self.root, "reconstructions") + self.root_inputs = os.path.join(self.root, "inputs") + os.makedirs(self.root, exist_ok=True) + + if self.test_args.save_z: + os.makedirs(self.root_zs, exist_ok=True) + if self.test_args.save_reconstruction: + os.makedirs(self.root_dec, exist_ok=True) + if self.test_args.save_input: + os.makedirs(self.root_inputs, exist_ok=True) + assert(self.test_args is not None) + self.test_maximum = getattr(self.test_args, 'test_maximum', None) + self.count = 0 + self.eval_metrics = {} + self.decodes = [] + self.save_decode_samples = 2048 + + def init_from_ckpt(self, path, ignore_keys=list()): + sd = torch.load(path, map_location="cpu") + try: + self._cur_epoch = sd['epoch'] + sd = sd["state_dict"] + except: + self._cur_epoch = 'null' + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + self.load_state_dict(sd, strict=False) + # self.load_state_dict(sd, strict=True) + print(f"Restored from {path}") + + def encode(self, x, return_hidden_states=False, **kwargs): + if return_hidden_states: + h, hidden = self.encoder(x, return_hidden_states) + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + return posterior, hidden + else: + h = self.encoder(x) + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + return posterior + + def decode(self, z, **kwargs): + if len(kwargs) == 0: ## use the original decoder in AutoencoderKL + z = self.post_quant_conv(z) + dec = self.decoder(z, **kwargs) ##change for SVD decoder by adding **kwargs + return dec + + def forward(self, input, sample_posterior=True, **additional_decode_kwargs): + input_tuple = (input, ) + forward_temp = partial(self._forward, sample_posterior=sample_posterior, **additional_decode_kwargs) + return checkpoint(forward_temp, input_tuple, self.parameters(), self.use_checkpoint) + + + def _forward(self, input, sample_posterior=True, **additional_decode_kwargs): + posterior = self.encode(input) + if sample_posterior: + z = posterior.sample() + else: + z = posterior.mode() + dec = self.decode(z, **additional_decode_kwargs) + ## print(input.shape, dec.shape) torch.Size([16, 3, 256, 256]) torch.Size([16, 3, 256, 256]) + return dec, posterior + + def get_input(self, batch, k): + x = batch[k] + if x.dim() == 5 and self.input_dim == 4: + b,c,t,h,w = x.shape + self.b = b + self.t = t + x = rearrange(x, 'b c t h w -> (b t) c h w') + + return x + + def training_step(self, batch, batch_idx, optimizer_idx): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + + if optimizer_idx == 0: + # train encoder+decoder+logvar + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return aeloss + + if optimizer_idx == 1: + # train the discriminator + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + + self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return discloss + + def validation_step(self, batch, batch_idx): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step, + last_layer=self.get_last_layer(), split="val") + + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step, + last_layer=self.get_last_layer(), split="val") + + self.log("val/rec_loss", log_dict_ae["val/rec_loss"]) + self.log_dict(log_dict_ae) + self.log_dict(log_dict_disc) + return self.log_dict + + def configure_optimizers(self): + lr = self.learning_rate + opt_ae = torch.optim.Adam(list(self.encoder.parameters())+ + list(self.decoder.parameters())+ + list(self.quant_conv.parameters())+ + list(self.post_quant_conv.parameters()), + lr=lr, betas=(0.5, 0.9)) + opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(), + lr=lr, betas=(0.5, 0.9)) + return [opt_ae, opt_disc], [] + + def get_last_layer(self): + return self.decoder.conv_out.weight + + @torch.no_grad() + def log_images(self, batch, only_inputs=False, **kwargs): + log = dict() + x = self.get_input(batch, self.image_key) + x = x.to(self.device) + if not only_inputs: + xrec, posterior = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec.shape[1] > 3 + x = self.to_rgb(x) + xrec = self.to_rgb(xrec) + log["samples"] = self.decode(torch.randn_like(posterior.sample())) + log["reconstructions"] = xrec + log["inputs"] = x + return log + + def to_rgb(self, x): + assert self.image_key == "segmentation" + if not hasattr(self, "colorize"): + self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x)) + x = F.conv2d(x, weight=self.colorize) + x = 2.*(x-x.min())/(x.max()-x.min()) - 1. + return x + +class IdentityFirstStage(torch.nn.Module): + def __init__(self, *args, vq_interface=False, **kwargs): + self.vq_interface = vq_interface # TODO: Should be true by default but check to not break older stuff + super().__init__() + + def encode(self, x, *args, **kwargs): + return x + + def decode(self, x, *args, **kwargs): + return x + + def quantize(self, x, *args, **kwargs): + if self.vq_interface: + return x, None, [None, None, None] + return x + + def forward(self, x, *args, **kwargs): + return x + +from lvdm.models.autoencoder_dualref import VideoDecoder +class AutoencoderKL_Dualref(AutoencoderKL): + def __init__(self, + ddconfig, + lossconfig, + embed_dim, + ckpt_path=None, + ignore_keys=[], + image_key="image", + colorize_nlabels=None, + monitor=None, + test=False, + logdir=None, + input_dim=4, + test_args=None, + additional_decode_keys=None, + use_checkpoint=False, + diff_boost_factor=3.0, + ): + super().__init__(ddconfig, lossconfig, embed_dim, ckpt_path, ignore_keys, image_key, colorize_nlabels, monitor, test, logdir, input_dim, test_args, additional_decode_keys, use_checkpoint, diff_boost_factor) + self.decoder = VideoDecoder(**ddconfig) + + def _forward(self, input, sample_posterior=True, **additional_decode_kwargs): + posterior, hidden_states = self.encode(input, return_hidden_states=True) + + hidden_states_first_last = [] + ### use only the first and last hidden states + for hid in hidden_states: + hid = rearrange(hid, '(b t) c h w -> b c t h w', t=TIMESTEPS) + hid_new = torch.cat([hid[:, :, 0:1], hid[:, :, -1:]], dim=2) + hidden_states_first_last.append(hid_new) + + if sample_posterior: + z = posterior.sample() + else: + z = posterior.mode() + dec = self.decode(z, ref_context=hidden_states_first_last, **additional_decode_kwargs) + ## print(input.shape, dec.shape) torch.Size([16, 3, 256, 256]) torch.Size([16, 3, 256, 256]) + return dec, posterior \ No newline at end of file diff --git a/lvdm/models/autoencoder_dualref.py b/lvdm/models/autoencoder_dualref.py new file mode 100644 index 0000000000000000000000000000000000000000..8529799f9aa63ef355fa53f9b9d787a812936555 --- /dev/null +++ b/lvdm/models/autoencoder_dualref.py @@ -0,0 +1,1177 @@ +#### https://github.com/Stability-AI/generative-models +from einops import rearrange, repeat +import logging +from typing import Any, Callable, Optional, Iterable, Union + +import numpy as np +import torch +import torch.nn as nn +from packaging import version +logpy = logging.getLogger(__name__) + +try: + import xformers + import xformers.ops + + XFORMERS_IS_AVAILABLE = True +except: + XFORMERS_IS_AVAILABLE = False + logpy.warning("no module 'xformers'. Processing without...") + +from lvdm.modules.attention_svd import LinearAttention, MemoryEfficientCrossAttention + + +def nonlinearity(x): + # swish + return x * torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32): + return torch.nn.GroupNorm( + num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True + ) + + +class ResnetBlock(nn.Module): + def __init__( + self, + *, + in_channels, + out_channels=None, + conv_shortcut=False, + dropout, + temb_channels=512, + ): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels) + self.conv1 = torch.nn.Conv2d( + in_channels, out_channels, kernel_size=3, stride=1, padding=1 + ) + if temb_channels > 0: + self.temb_proj = torch.nn.Linear(temb_channels, out_channels) + self.norm2 = Normalize(out_channels) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d( + out_channels, out_channels, kernel_size=3, stride=1, padding=1 + ) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = torch.nn.Conv2d( + in_channels, out_channels, kernel_size=3, stride=1, padding=1 + ) + else: + self.nin_shortcut = torch.nn.Conv2d( + in_channels, out_channels, kernel_size=1, stride=1, padding=0 + ) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x + h + + +class LinAttnBlock(LinearAttention): + """to match AttnBlock usage""" + + def __init__(self, in_channels): + super().__init__(dim=in_channels, heads=1, dim_head=in_channels) + + +class AttnBlock(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.k = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.v = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.proj_out = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + + def attention(self, h_: torch.Tensor) -> torch.Tensor: + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + b, c, h, w = q.shape + q, k, v = map( + lambda x: rearrange(x, "b c h w -> b 1 (h w) c").contiguous(), (q, k, v) + ) + h_ = torch.nn.functional.scaled_dot_product_attention( + q, k, v + ) # scale is dim ** -0.5 per default + # compute attention + + return rearrange(h_, "b 1 (h w) c -> b c h w", h=h, w=w, c=c, b=b) + + def forward(self, x, **kwargs): + h_ = x + h_ = self.attention(h_) + h_ = self.proj_out(h_) + return x + h_ + + +class MemoryEfficientAttnBlock(nn.Module): + """ + Uses xformers efficient implementation, + see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + Note: this is a single-head self-attention operation + """ + + # + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.k = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.v = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.proj_out = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.attention_op: Optional[Any] = None + + def attention(self, h_: torch.Tensor) -> torch.Tensor: + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + B, C, H, W = q.shape + q, k, v = map(lambda x: rearrange(x, "b c h w -> b (h w) c"), (q, k, v)) + + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(B, t.shape[1], 1, C) + .permute(0, 2, 1, 3) + .reshape(B * 1, t.shape[1], C) + .contiguous(), + (q, k, v), + ) + out = xformers.ops.memory_efficient_attention( + q, k, v, attn_bias=None, op=self.attention_op + ) + + out = ( + out.unsqueeze(0) + .reshape(B, 1, out.shape[1], C) + .permute(0, 2, 1, 3) + .reshape(B, out.shape[1], C) + ) + return rearrange(out, "b (h w) c -> b c h w", b=B, h=H, w=W, c=C) + + def forward(self, x, **kwargs): + h_ = x + h_ = self.attention(h_) + h_ = self.proj_out(h_) + return x + h_ + + +class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention): + def forward(self, x, context=None, mask=None, **unused_kwargs): + b, c, h, w = x.shape + x = rearrange(x, "b c h w -> b (h w) c") + out = super().forward(x, context=context, mask=mask) + out = rearrange(out, "b (h w) c -> b c h w", h=h, w=w, c=c) + return x + out + + +def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None): + assert attn_type in [ + "vanilla", + "vanilla-xformers", + "memory-efficient-cross-attn", + "linear", + "none", + "memory-efficient-cross-attn-fusion", + ], f"attn_type {attn_type} unknown" + if ( + version.parse(torch.__version__) < version.parse("2.0.0") + and attn_type != "none" + ): + assert XFORMERS_IS_AVAILABLE, ( + f"We do not support vanilla attention in {torch.__version__} anymore, " + f"as it is too expensive. Please install xformers via e.g. 'pip install xformers==0.0.16'" + ) + # attn_type = "vanilla-xformers" + logpy.info(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + assert attn_kwargs is None + return AttnBlock(in_channels) + elif attn_type == "vanilla-xformers": + logpy.info( + f"building MemoryEfficientAttnBlock with {in_channels} in_channels..." + ) + return MemoryEfficientAttnBlock(in_channels) + elif attn_type == "memory-efficient-cross-attn": + attn_kwargs["query_dim"] = in_channels + return MemoryEfficientCrossAttentionWrapper(**attn_kwargs) + elif attn_type == "memory-efficient-cross-attn-fusion": + attn_kwargs["query_dim"] = in_channels + return MemoryEfficientCrossAttentionWrapperFusion(**attn_kwargs) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + return LinAttnBlock(in_channels) + +class MemoryEfficientCrossAttentionWrapperFusion(MemoryEfficientCrossAttention): + # print('x.shape: ',x.shape, 'context.shape: ',context.shape) ##torch.Size([8, 128, 256, 256]) torch.Size([1, 128, 2, 256, 256]) + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0, **kwargs): + super().__init__(query_dim, context_dim, heads, dim_head, dropout, **kwargs) + self.norm = Normalize(query_dim) + nn.init.zeros_(self.to_out[0].weight) + nn.init.zeros_(self.to_out[0].bias) + + def forward(self, x, context=None, mask=None): + if self.training: + return checkpoint(self._forward, x, context, mask, use_reentrant=False) + else: + return self._forward(x, context, mask) + + def _forward( + self, + x, + context=None, + mask=None, + ): + bt, c, h, w = x.shape + h_ = self.norm(x) + h_ = rearrange(h_, "b c h w -> b (h w) c") + q = self.to_q(h_) + + + b, c, l, h, w = context.shape + context = rearrange(context, "b c l h w -> (b l) (h w) c") + k = self.to_k(context) + v = self.to_v(context) + k = rearrange(k, "(b l) d c -> b l d c", l=l) + k = torch.cat([k[:, [0] * (bt//b)], k[:, [1]*(bt//b)]], dim=2) + k = rearrange(k, "b l d c -> (b l) d c") + + v = rearrange(v, "(b l) d c -> b l d c", l=l) + v = torch.cat([v[:, [0] * (bt//b)], v[:, [1]*(bt//b)]], dim=2) + v = rearrange(v, "b l d c -> (b l) d c") + + + b, _, _ = q.shape ##actually bt + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + + # actually compute the attention, what we cannot get enough of + if version.parse(xformers.__version__) >= version.parse("0.0.21"): + # NOTE: workaround for + # https://github.com/facebookresearch/xformers/issues/845 + max_bs = 32768 + N = q.shape[0] + n_batches = math.ceil(N / max_bs) + out = list() + for i_batch in range(n_batches): + batch = slice(i_batch * max_bs, (i_batch + 1) * max_bs) + out.append( + xformers.ops.memory_efficient_attention( + q[batch], + k[batch], + v[batch], + attn_bias=None, + op=self.attention_op, + ) + ) + out = torch.cat(out, 0) + else: + out = xformers.ops.memory_efficient_attention( + q, k, v, attn_bias=None, op=self.attention_op + ) + + # TODO: Use this directly in the attention operation, as a bias + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + out = self.to_out(out) + out = rearrange(out, "bt (h w) c -> bt c h w", h=h, w=w, c=c) + return x + out + +class Combiner(nn.Module): + def __init__(self, ch) -> None: + super().__init__() + self.conv = nn.Conv2d(ch,ch,1,padding=0) + + nn.init.zeros_(self.conv.weight) + nn.init.zeros_(self.conv.bias) + + def forward(self, x, context): + if self.training: + return checkpoint(self._forward, x, context, use_reentrant=False) + else: + return self._forward(x, context) + + def _forward(self, x, context): + ## x: b c h w, context: b c 2 h w + b, c, l, h, w = context.shape + bt, c, h, w = x.shape + context = rearrange(context, "b c l h w -> (b l) c h w") + context = self.conv(context) + context = rearrange(context, "(b l) c h w -> b c l h w", l=l) + x = rearrange(x, "(b t) c h w -> b c t h w", t=bt//b) + x[:,:,0] = x[:,:,0] + context[:,:,0] + x[:,:,-1] = x[:,:,-1] + context[:,:,1] + x = rearrange(x, "b c t h w -> (b t) c h w") + return x + + +class Decoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + give_pre_end=False, + tanh_out=False, + use_linear_attn=False, + attn_type="vanilla-xformers", + attn_level=[2,3], + **ignorekwargs, + ): + super().__init__() + if use_linear_attn: + attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + self.attn_level = attn_level + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,) + tuple(ch_mult) + block_in = ch * ch_mult[self.num_resolutions - 1] + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.z_shape = (1, z_channels, curr_res, curr_res) + logpy.info( + "Working with z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape) + ) + ) + + make_attn_cls = self._make_attn() + make_resblock_cls = self._make_resblock() + make_conv_cls = self._make_conv() + # z to block_in + self.conv_in = torch.nn.Conv2d( + z_channels, block_in, kernel_size=3, stride=1, padding=1 + ) + + # middle + self.mid = nn.Module() + self.mid.block_1 = make_resblock_cls( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + ) + self.mid.attn_1 = make_attn_cls(block_in, attn_type=attn_type) + self.mid.block_2 = make_resblock_cls( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + ) + + # upsampling + self.up = nn.ModuleList() + self.attn_refinement = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + block.append( + make_resblock_cls( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn_cls(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + if i_level in self.attn_level: + self.attn_refinement.insert(0, make_attn_cls(block_in, attn_type='memory-efficient-cross-attn-fusion', attn_kwargs={})) + else: + self.attn_refinement.insert(0, Combiner(block_in)) + # end + self.norm_out = Normalize(block_in) + self.attn_refinement.append(Combiner(block_in)) + self.conv_out = make_conv_cls( + block_in, out_ch, kernel_size=3, stride=1, padding=1 + ) + + def _make_attn(self) -> Callable: + return make_attn + + def _make_resblock(self) -> Callable: + return ResnetBlock + + def _make_conv(self) -> Callable: + return torch.nn.Conv2d + + def get_last_layer(self, **kwargs): + return self.conv_out.weight + + def forward(self, z, ref_context=None, **kwargs): + ## ref_context: b c 2 h w, 2 means starting and ending frame + # assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + + # middle + h = self.mid.block_1(h, temb, **kwargs) + h = self.mid.attn_1(h, **kwargs) + h = self.mid.block_2(h, temb, **kwargs) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.up[i_level].block[i_block](h, temb, **kwargs) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h, **kwargs) + if ref_context: + h = self.attn_refinement[i_level](x=h, context=ref_context[i_level]) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + if ref_context: + # print(h.shape, ref_context[i_level].shape) #torch.Size([8, 128, 256, 256]) torch.Size([1, 128, 2, 256, 256]) + h = self.attn_refinement[-1](x=h, context=ref_context[-1]) + h = self.conv_out(h, **kwargs) + if self.tanh_out: + h = torch.tanh(h) + return h + +##### + + +from abc import abstractmethod +from lvdm.models.utils_diffusion import timestep_embedding + +from torch.utils.checkpoint import checkpoint +from lvdm.basics import ( + zero_module, + conv_nd, + linear, + normalization, +) +from lvdm.modules.networks.openaimodel3d import Upsample, Downsample +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + + @abstractmethod + def forward(self, x: torch.Tensor, emb: torch.Tensor): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels: int, + emb_channels: int, + dropout: float, + out_channels: Optional[int] = None, + use_conv: bool = False, + use_scale_shift_norm: bool = False, + dims: int = 2, + use_checkpoint: bool = False, + up: bool = False, + down: bool = False, + kernel_size: int = 3, + exchange_temb_dims: bool = False, + skip_t_emb: bool = False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + self.exchange_temb_dims = exchange_temb_dims + + if isinstance(kernel_size, Iterable): + padding = [k // 2 for k in kernel_size] + else: + padding = kernel_size // 2 + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, kernel_size, padding=padding), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.skip_t_emb = skip_t_emb + self.emb_out_channels = ( + 2 * self.out_channels if use_scale_shift_norm else self.out_channels + ) + if self.skip_t_emb: + # print(f"Skipping timestep embedding in {self.__class__.__name__}") + assert not self.use_scale_shift_norm + self.emb_layers = None + self.exchange_temb_dims = False + else: + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + self.emb_out_channels, + ), + ) + + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd( + dims, + self.out_channels, + self.out_channels, + kernel_size, + padding=padding, + ) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, kernel_size, padding=padding + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor: + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + if self.use_checkpoint: + return checkpoint(self._forward, x, emb, use_reentrant=False) + else: + return self._forward(x, emb) + + def _forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor: + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + + if self.skip_t_emb: + emb_out = torch.zeros_like(h) + else: + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = torch.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + if self.exchange_temb_dims: + emb_out = rearrange(emb_out, "b t c ... -> b c t ...") + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h +##### + +##### +from lvdm.modules.attention_svd import * +class VideoTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, + "softmax-xformers": MemoryEfficientCrossAttention, + } + + def __init__( + self, + dim, + n_heads, + d_head, + dropout=0.0, + context_dim=None, + gated_ff=True, + checkpoint=True, + timesteps=None, + ff_in=False, + inner_dim=None, + attn_mode="softmax", + disable_self_attn=False, + disable_temporal_crossattention=False, + switch_temporal_ca_to_sa=False, + ): + super().__init__() + + attn_cls = self.ATTENTION_MODES[attn_mode] + + self.ff_in = ff_in or inner_dim is not None + if inner_dim is None: + inner_dim = dim + + assert int(n_heads * d_head) == inner_dim + + self.is_res = inner_dim == dim + + if self.ff_in: + self.norm_in = nn.LayerNorm(dim) + self.ff_in = FeedForward( + dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff + ) + + self.timesteps = timesteps + self.disable_self_attn = disable_self_attn + if self.disable_self_attn: + self.attn1 = attn_cls( + query_dim=inner_dim, + heads=n_heads, + dim_head=d_head, + context_dim=context_dim, + dropout=dropout, + ) # is a cross-attention + else: + self.attn1 = attn_cls( + query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout + ) # is a self-attention + + self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff) + + if disable_temporal_crossattention: + if switch_temporal_ca_to_sa: + raise ValueError + else: + self.attn2 = None + else: + self.norm2 = nn.LayerNorm(inner_dim) + if switch_temporal_ca_to_sa: + self.attn2 = attn_cls( + query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout + ) # is a self-attention + else: + self.attn2 = attn_cls( + query_dim=inner_dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + ) # is self-attn if context is none + + self.norm1 = nn.LayerNorm(inner_dim) + self.norm3 = nn.LayerNorm(inner_dim) + self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa + + self.checkpoint = checkpoint + if self.checkpoint: + print(f"====>{self.__class__.__name__} is using checkpointing") + else: + print(f"====>{self.__class__.__name__} is NOT using checkpointing") + + def forward( + self, x: torch.Tensor, context: torch.Tensor = None, timesteps: int = None + ) -> torch.Tensor: + if self.checkpoint: + return checkpoint(self._forward, x, context, timesteps, use_reentrant=False) + else: + return self._forward(x, context, timesteps=timesteps) + + def _forward(self, x, context=None, timesteps=None): + assert self.timesteps or timesteps + assert not (self.timesteps and timesteps) or self.timesteps == timesteps + timesteps = self.timesteps or timesteps + B, S, C = x.shape + x = rearrange(x, "(b t) s c -> (b s) t c", t=timesteps) + + if self.ff_in: + x_skip = x + x = self.ff_in(self.norm_in(x)) + if self.is_res: + x += x_skip + + if self.disable_self_attn: + x = self.attn1(self.norm1(x), context=context) + x + else: + x = self.attn1(self.norm1(x)) + x + + if self.attn2 is not None: + if self.switch_temporal_ca_to_sa: + x = self.attn2(self.norm2(x)) + x + else: + x = self.attn2(self.norm2(x), context=context) + x + x_skip = x + x = self.ff(self.norm3(x)) + if self.is_res: + x += x_skip + + x = rearrange( + x, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps + ) + return x + + def get_last_layer(self): + return self.ff.net[-1].weight + +##### + +##### +import functools +def partialclass(cls, *args, **kwargs): + class NewCls(cls): + __init__ = functools.partialmethod(cls.__init__, *args, **kwargs) + + return NewCls +###### + +class VideoResBlock(ResnetBlock): + def __init__( + self, + out_channels, + *args, + dropout=0.0, + video_kernel_size=3, + alpha=0.0, + merge_strategy="learned", + **kwargs, + ): + super().__init__(out_channels=out_channels, dropout=dropout, *args, **kwargs) + if video_kernel_size is None: + video_kernel_size = [3, 1, 1] + self.time_stack = ResBlock( + channels=out_channels, + emb_channels=0, + dropout=dropout, + dims=3, + use_scale_shift_norm=False, + use_conv=False, + up=False, + down=False, + kernel_size=video_kernel_size, + use_checkpoint=True, + skip_t_emb=True, + ) + + self.merge_strategy = merge_strategy + if self.merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([alpha])) + elif self.merge_strategy == "learned": + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([alpha])) + ) + else: + raise ValueError(f"unknown merge strategy {self.merge_strategy}") + + def get_alpha(self, bs): + if self.merge_strategy == "fixed": + return self.mix_factor + elif self.merge_strategy == "learned": + return torch.sigmoid(self.mix_factor) + else: + raise NotImplementedError() + + def forward(self, x, temb, skip_video=False, timesteps=None): + if timesteps is None: + timesteps = self.timesteps + + b, c, h, w = x.shape + + x = super().forward(x, temb) + + if not skip_video: + x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) + + x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) + + x = self.time_stack(x, temb) + + alpha = self.get_alpha(bs=b // timesteps) + x = alpha * x + (1.0 - alpha) * x_mix + + x = rearrange(x, "b c t h w -> (b t) c h w") + return x + + +class AE3DConv(torch.nn.Conv2d): + def __init__(self, in_channels, out_channels, video_kernel_size=3, *args, **kwargs): + super().__init__(in_channels, out_channels, *args, **kwargs) + if isinstance(video_kernel_size, Iterable): + padding = [int(k // 2) for k in video_kernel_size] + else: + padding = int(video_kernel_size // 2) + + self.time_mix_conv = torch.nn.Conv3d( + in_channels=out_channels, + out_channels=out_channels, + kernel_size=video_kernel_size, + padding=padding, + ) + + def forward(self, input, timesteps, skip_video=False): + x = super().forward(input) + if skip_video: + return x + x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) + x = self.time_mix_conv(x) + return rearrange(x, "b c t h w -> (b t) c h w") + + +class VideoBlock(AttnBlock): + def __init__( + self, in_channels: int, alpha: float = 0, merge_strategy: str = "learned" + ): + super().__init__(in_channels) + # no context, single headed, as in base class + self.time_mix_block = VideoTransformerBlock( + dim=in_channels, + n_heads=1, + d_head=in_channels, + checkpoint=True, + ff_in=True, + attn_mode="softmax", + ) + + time_embed_dim = self.in_channels * 4 + self.video_time_embed = torch.nn.Sequential( + torch.nn.Linear(self.in_channels, time_embed_dim), + torch.nn.SiLU(), + torch.nn.Linear(time_embed_dim, self.in_channels), + ) + + self.merge_strategy = merge_strategy + if self.merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([alpha])) + elif self.merge_strategy == "learned": + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([alpha])) + ) + else: + raise ValueError(f"unknown merge strategy {self.merge_strategy}") + + def forward(self, x, timesteps, skip_video=False): + if skip_video: + return super().forward(x) + + x_in = x + x = self.attention(x) + h, w = x.shape[2:] + x = rearrange(x, "b c h w -> b (h w) c") + + x_mix = x + num_frames = torch.arange(timesteps, device=x.device) + num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps) + num_frames = rearrange(num_frames, "b t -> (b t)") + t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False) + emb = self.video_time_embed(t_emb) # b, n_channels + emb = emb[:, None, :] + x_mix = x_mix + emb + + alpha = self.get_alpha() + x_mix = self.time_mix_block(x_mix, timesteps=timesteps) + x = alpha * x + (1.0 - alpha) * x_mix # alpha merge + + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w) + x = self.proj_out(x) + + return x_in + x + + def get_alpha( + self, + ): + if self.merge_strategy == "fixed": + return self.mix_factor + elif self.merge_strategy == "learned": + return torch.sigmoid(self.mix_factor) + else: + raise NotImplementedError(f"unknown merge strategy {self.merge_strategy}") + + +class MemoryEfficientVideoBlock(MemoryEfficientAttnBlock): + def __init__( + self, in_channels: int, alpha: float = 0, merge_strategy: str = "learned" + ): + super().__init__(in_channels) + # no context, single headed, as in base class + self.time_mix_block = VideoTransformerBlock( + dim=in_channels, + n_heads=1, + d_head=in_channels, + checkpoint=True, + ff_in=True, + attn_mode="softmax-xformers", + ) + + time_embed_dim = self.in_channels * 4 + self.video_time_embed = torch.nn.Sequential( + torch.nn.Linear(self.in_channels, time_embed_dim), + torch.nn.SiLU(), + torch.nn.Linear(time_embed_dim, self.in_channels), + ) + + self.merge_strategy = merge_strategy + if self.merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([alpha])) + elif self.merge_strategy == "learned": + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([alpha])) + ) + else: + raise ValueError(f"unknown merge strategy {self.merge_strategy}") + + def forward(self, x, timesteps, skip_time_block=False): + if skip_time_block: + return super().forward(x) + + x_in = x + x = self.attention(x) + h, w = x.shape[2:] + x = rearrange(x, "b c h w -> b (h w) c") + + x_mix = x + num_frames = torch.arange(timesteps, device=x.device) + num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps) + num_frames = rearrange(num_frames, "b t -> (b t)") + t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False) + emb = self.video_time_embed(t_emb) # b, n_channels + emb = emb[:, None, :] + x_mix = x_mix + emb + + alpha = self.get_alpha() + x_mix = self.time_mix_block(x_mix, timesteps=timesteps) + x = alpha * x + (1.0 - alpha) * x_mix # alpha merge + + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w) + x = self.proj_out(x) + + return x_in + x + + def get_alpha( + self, + ): + if self.merge_strategy == "fixed": + return self.mix_factor + elif self.merge_strategy == "learned": + return torch.sigmoid(self.mix_factor) + else: + raise NotImplementedError(f"unknown merge strategy {self.merge_strategy}") + + +def make_time_attn( + in_channels, + attn_type="vanilla", + attn_kwargs=None, + alpha: float = 0, + merge_strategy: str = "learned", +): + assert attn_type in [ + "vanilla", + "vanilla-xformers", + ], f"attn_type {attn_type} not supported for spatio-temporal attention" + print( + f"making spatial and temporal attention of type '{attn_type}' with {in_channels} in_channels" + ) + if not XFORMERS_IS_AVAILABLE and attn_type == "vanilla-xformers": + print( + f"Attention mode '{attn_type}' is not available. Falling back to vanilla attention. " + f"This is not a problem in Pytorch >= 2.0. FYI, you are running with PyTorch version {torch.__version__}" + ) + attn_type = "vanilla" + + if attn_type == "vanilla": + assert attn_kwargs is None + return partialclass( + VideoBlock, in_channels, alpha=alpha, merge_strategy=merge_strategy + ) + elif attn_type == "vanilla-xformers": + print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...") + return partialclass( + MemoryEfficientVideoBlock, + in_channels, + alpha=alpha, + merge_strategy=merge_strategy, + ) + else: + return NotImplementedError() + + +class Conv2DWrapper(torch.nn.Conv2d): + def forward(self, input: torch.Tensor, **kwargs) -> torch.Tensor: + return super().forward(input) + + +class VideoDecoder(Decoder): + available_time_modes = ["all", "conv-only", "attn-only"] + + def __init__( + self, + *args, + video_kernel_size: Union[int, list] = [3,1,1], + alpha: float = 0.0, + merge_strategy: str = "learned", + time_mode: str = "conv-only", + **kwargs, + ): + self.video_kernel_size = video_kernel_size + self.alpha = alpha + self.merge_strategy = merge_strategy + self.time_mode = time_mode + assert ( + self.time_mode in self.available_time_modes + ), f"time_mode parameter has to be in {self.available_time_modes}" + super().__init__(*args, **kwargs) + + def get_last_layer(self, skip_time_mix=False, **kwargs): + if self.time_mode == "attn-only": + raise NotImplementedError("TODO") + else: + return ( + self.conv_out.time_mix_conv.weight + if not skip_time_mix + else self.conv_out.weight + ) + + def _make_attn(self) -> Callable: + if self.time_mode not in ["conv-only", "only-last-conv"]: + return partialclass( + make_time_attn, + alpha=self.alpha, + merge_strategy=self.merge_strategy, + ) + else: + return super()._make_attn() + + def _make_conv(self) -> Callable: + if self.time_mode != "attn-only": + return partialclass(AE3DConv, video_kernel_size=self.video_kernel_size) + else: + return Conv2DWrapper + + def _make_resblock(self) -> Callable: + if self.time_mode not in ["attn-only", "only-last-conv"]: + return partialclass( + VideoResBlock, + video_kernel_size=self.video_kernel_size, + alpha=self.alpha, + merge_strategy=self.merge_strategy, + ) + else: + return super()._make_resblock() \ No newline at end of file diff --git a/lvdm/models/ddpm3d.py b/lvdm/models/ddpm3d.py new file mode 100644 index 0000000000000000000000000000000000000000..cec0591b1caf557012ba0c156017b709a4bafda3 --- /dev/null +++ b/lvdm/models/ddpm3d.py @@ -0,0 +1,1325 @@ +""" +wild mixture of +https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py +https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +https://github.com/CompVis/taming-transformers +-- merci +""" + +from functools import partial +from contextlib import contextmanager +import numpy as np +from tqdm import tqdm +from einops import rearrange, repeat +import logging +mainlogger = logging.getLogger('mainlogger') +import random +import torch +import torch.nn as nn +from torch.optim.lr_scheduler import LambdaLR, CosineAnnealingLR +from torchvision.utils import make_grid +import pytorch_lightning as pl +from pytorch_lightning.utilities import rank_zero_only +from utils.utils import instantiate_from_config +from lvdm.ema import LitEma +from lvdm.models.samplers.ddim import DDIMSampler +from lvdm.distributions import DiagonalGaussianDistribution +from lvdm.models.utils_diffusion import make_beta_schedule, rescale_zero_terminal_snr +from lvdm.basics import disabled_train +from lvdm.common import ( + extract_into_tensor, + noise_like, + exists, + default +) +import math +from lvdm.models.autoencoder_dualref import VideoDecoder +__conditioning_keys__ = {'concat': 'c_concat', + 'crossattn': 'c_crossattn', + 'adm': 'y'} + +class DDPM(pl.LightningModule): + # classic DDPM with Gaussian diffusion, in image space + def __init__(self, + unet_config, + timesteps=1000, + beta_schedule="linear", + loss_type="l2", + ckpt_path=None, + ignore_keys=[], + load_only_unet=False, + monitor=None, + use_ema=True, + first_stage_key="image", + image_size=256, + channels=3, + log_every_t=100, + clip_denoised=True, + linear_start=1e-4, + linear_end=2e-2, + cosine_s=8e-3, + given_betas=None, + original_elbo_weight=0., + v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta + l_simple_weight=1., + conditioning_key=None, + parameterization="eps", # all assuming fixed variance schedules + scheduler_config=None, + use_positional_encodings=False, + learn_logvar=False, + logvar_init=0., + rescale_betas_zero_snr=False, + ): + super().__init__() + assert parameterization in ["eps", "x0", "v"], 'currently only supporting "eps" and "x0" and "v"' + self.parameterization = parameterization + mainlogger.info(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode") + self.cond_stage_model = None + self.clip_denoised = clip_denoised + self.log_every_t = log_every_t + self.first_stage_key = first_stage_key + self.channels = channels + self.temporal_length = unet_config.params.temporal_length + self.image_size = image_size # try conv? + if isinstance(self.image_size, int): + self.image_size = [self.image_size, self.image_size] + self.use_positional_encodings = use_positional_encodings + self.model = DiffusionWrapper(unet_config, conditioning_key) + #count_params(self.model, verbose=True) + self.use_ema = use_ema + self.rescale_betas_zero_snr = rescale_betas_zero_snr + if self.use_ema: + self.model_ema = LitEma(self.model) + mainlogger.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + self.use_scheduler = scheduler_config is not None + if self.use_scheduler: + self.scheduler_config = scheduler_config + + self.v_posterior = v_posterior + self.original_elbo_weight = original_elbo_weight + self.l_simple_weight = l_simple_weight + + if monitor is not None: + self.monitor = monitor + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet) + + self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps, + linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s) + + ## for reschedule + self.given_betas = given_betas + self.beta_schedule = beta_schedule + self.timesteps = timesteps + self.cosine_s = cosine_s + + self.loss_type = loss_type + + self.learn_logvar = learn_logvar + self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,)) + if self.learn_logvar: + self.logvar = nn.Parameter(self.logvar, requires_grad=True) + + def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if exists(given_betas): + betas = given_betas + else: + betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, + cosine_s=cosine_s) + if self.rescale_betas_zero_snr: + betas = rescale_zero_terminal_snr(betas) + + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + + if self.parameterization != 'v': + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + else: + self.register_buffer('sqrt_recip_alphas_cumprod', torch.zeros_like(to_torch(alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', torch.zeros_like(to_torch(alphas_cumprod))) + + # calculations for posterior q(x_{t-1} | x_t, x_0) + posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / ( + 1. - alphas_cumprod) + self.v_posterior * betas + # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) + self.register_buffer('posterior_variance', to_torch(posterior_variance)) + # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain + self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20)))) + self.register_buffer('posterior_mean_coef1', to_torch( + betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))) + self.register_buffer('posterior_mean_coef2', to_torch( + (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod))) + + if self.parameterization == "eps": + lvlb_weights = self.betas ** 2 / ( + 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)) + elif self.parameterization == "x0": + lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod)) + elif self.parameterization == "v": + lvlb_weights = torch.ones_like(self.betas ** 2 / ( + 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))) + else: + raise NotImplementedError("mu not supported") + # TODO how to choose this term + lvlb_weights[0] = lvlb_weights[1] + self.register_buffer('lvlb_weights', lvlb_weights, persistent=False) + assert not torch.isnan(self.lvlb_weights).all() + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.model.parameters()) + self.model_ema.copy_to(self.model) + if context is not None: + mainlogger.info(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.model.parameters()) + if context is not None: + mainlogger.info(f"{context}: Restored training weights") + + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + mainlogger.info("Deleting key {} from state_dict.".format(k)) + del sd[k] + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + mainlogger.info(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + mainlogger.info(f"Missing Keys: {missing}") + if len(unexpected) > 0: + mainlogger.info(f"Unexpected Keys: {unexpected}") + + def q_mean_variance(self, x_start, t): + """ + Get the distribution q(x_t | x_0). + :param x_start: the [N x C x ...] tensor of noiseless inputs. + :param t: the number of diffusion steps (minus 1). Here, 0 means one step. + :return: A tuple (mean, variance, log_variance), all of x_start's shape. + """ + mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start) + variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape) + log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape) + return mean, variance, log_variance + + def predict_start_from_noise(self, x_t, t, noise): + return ( + extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - + extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise + ) + + def predict_start_from_z_and_v(self, x_t, t, v): + # self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + # self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t - + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v + ) + + def predict_eps_from_z_and_v(self, x_t, t, v): + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * v + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * x_t + ) + + def q_posterior(self, x_start, x_t, t): + posterior_mean = ( + extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start + + extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t + ) + posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape) + posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape) + return posterior_mean, posterior_variance, posterior_log_variance_clipped + + def p_mean_variance(self, x, t, clip_denoised: bool): + model_out = self.model(x, t) + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + if clip_denoised: + x_recon.clamp_(-1., 1.) + + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, t, clip_denoised=True, repeat_noise=False): + b, *_, device = *x.shape, x.device + model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised) + noise = noise_like(x.shape, device, repeat_noise) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def p_sample_loop(self, shape, return_intermediates=False): + device = self.betas.device + b = shape[0] + img = torch.randn(shape, device=device) + intermediates = [img] + for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps): + img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long), + clip_denoised=self.clip_denoised) + if i % self.log_every_t == 0 or i == self.num_timesteps - 1: + intermediates.append(img) + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, batch_size=16, return_intermediates=False): + image_size = self.image_size + channels = self.channels + return self.p_sample_loop((batch_size, channels, image_size, image_size), + return_intermediates=return_intermediates) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def get_v(self, x, noise, t): + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise - + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x + ) + + def get_loss(self, pred, target, mean=True): + if self.loss_type == 'l1': + loss = (target - pred).abs() + if mean: + loss = loss.mean() + elif self.loss_type == 'l2': + if mean: + loss = torch.nn.functional.mse_loss(target, pred) + else: + loss = torch.nn.functional.mse_loss(target, pred, reduction='none') + else: + raise NotImplementedError("unknown loss type '{loss_type}'") + + return loss + + def p_losses(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + model_out = self.model(x_noisy, t) + + loss_dict = {} + if self.parameterization == "eps": + target = noise + elif self.parameterization == "x0": + target = x_start + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported") + + loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3]) + + log_prefix = 'train' if self.training else 'val' + + loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()}) + loss_simple = loss.mean() * self.l_simple_weight + + loss_vlb = (self.lvlb_weights[t] * loss).mean() + loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb}) + + loss = loss_simple + self.original_elbo_weight * loss_vlb + + loss_dict.update({f'{log_prefix}/loss': loss}) + + return loss, loss_dict + + def forward(self, x, *args, **kwargs): + # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size + # assert h == img_size and w == img_size, f'height and width of image must be {img_size}' + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + return self.p_losses(x, t, *args, **kwargs) + + def get_input(self, batch, k): + x = batch[k] + ''' + if len(x.shape) == 3: + x = x[..., None] + x = rearrange(x, 'b h w c -> b c h w') + ''' + x = x.to(memory_format=torch.contiguous_format).float() + return x + + def shared_step(self, batch): + x = self.get_input(batch, self.first_stage_key) + loss, loss_dict = self(x) + return loss, loss_dict + + def training_step(self, batch, batch_idx): + loss, loss_dict = self.shared_step(batch) + + self.log_dict(loss_dict, prog_bar=True, + logger=True, on_step=True, on_epoch=True) + + self.log("global_step", self.global_step, + prog_bar=True, logger=True, on_step=True, on_epoch=False) + + if self.use_scheduler: + lr = self.optimizers().param_groups[0]['lr'] + self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False) + + return loss + + @torch.no_grad() + def validation_step(self, batch, batch_idx): + _, loss_dict_no_ema = self.shared_step(batch) + with self.ema_scope(): + _, loss_dict_ema = self.shared_step(batch) + loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema} + self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True) + self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True) + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self.model) + + def _get_rows_from_list(self, samples): + n_imgs_per_row = len(samples) + denoise_grid = rearrange(samples, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs): + log = dict() + x = self.get_input(batch, self.first_stage_key) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + x = x.to(self.device)[:N] + log["inputs"] = x + + # get diffusion row + diffusion_row = list() + x_start = x[:n_row] + + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(x_start) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + diffusion_row.append(x_noisy) + + log["diffusion_row"] = self._get_rows_from_list(diffusion_row) + + if sample: + # get denoise row + with self.ema_scope("Plotting"): + samples, denoise_row = self.sample(batch_size=N, return_intermediates=True) + + log["samples"] = samples + log["denoise_row"] = self._get_rows_from_list(denoise_row) + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + if self.learn_logvar: + params = params + [self.logvar] + opt = torch.optim.AdamW(params, lr=lr) + return opt + +class LatentDiffusion(DDPM): + """main class""" + def __init__(self, + first_stage_config, + cond_stage_config, + num_timesteps_cond=None, + cond_stage_key="caption", + cond_stage_trainable=False, + cond_stage_forward=None, + conditioning_key=None, + uncond_prob=0.2, + uncond_type="empty_seq", + scale_factor=1.0, + scale_by_std=False, + encoder_type="2d", + only_model=False, + noise_strength=0, + use_dynamic_rescale=False, + base_scale=0.7, + turning_step=400, + loop_video=False, + fps_condition_type='fs', + perframe_ae=False, + # added + logdir=None, + rand_cond_frame=False, + en_and_decode_n_samples_a_time=None, + *args, **kwargs): + self.num_timesteps_cond = default(num_timesteps_cond, 1) + self.scale_by_std = scale_by_std + assert self.num_timesteps_cond <= kwargs['timesteps'] + # for backwards compatibility after implementation of DiffusionWrapper + ckpt_path = kwargs.pop("ckpt_path", None) + ignore_keys = kwargs.pop("ignore_keys", []) + conditioning_key = default(conditioning_key, 'crossattn') + super().__init__(conditioning_key=conditioning_key, *args, **kwargs) + + self.cond_stage_trainable = cond_stage_trainable + self.cond_stage_key = cond_stage_key + self.noise_strength = noise_strength + self.use_dynamic_rescale = use_dynamic_rescale + self.loop_video = loop_video + self.fps_condition_type = fps_condition_type + self.perframe_ae = perframe_ae + + self.logdir = logdir + self.rand_cond_frame = rand_cond_frame + self.en_and_decode_n_samples_a_time = en_and_decode_n_samples_a_time + + try: + self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1 + except: + self.num_downs = 0 + if not scale_by_std: + self.scale_factor = scale_factor + else: + self.register_buffer('scale_factor', torch.tensor(scale_factor)) + + if use_dynamic_rescale: + scale_arr1 = np.linspace(1.0, base_scale, turning_step) + scale_arr2 = np.full(self.num_timesteps, base_scale) + scale_arr = np.concatenate((scale_arr1, scale_arr2)) + to_torch = partial(torch.tensor, dtype=torch.float32) + self.register_buffer('scale_arr', to_torch(scale_arr)) + + self.instantiate_first_stage(first_stage_config) + self.instantiate_cond_stage(cond_stage_config) + self.first_stage_config = first_stage_config + self.cond_stage_config = cond_stage_config + self.clip_denoised = False + + self.cond_stage_forward = cond_stage_forward + self.encoder_type = encoder_type + assert(encoder_type in ["2d", "3d"]) + self.uncond_prob = uncond_prob + self.classifier_free_guidance = True if uncond_prob > 0 else False + assert(uncond_type in ["zero_embed", "empty_seq"]) + self.uncond_type = uncond_type + + self.restarted_from_ckpt = False + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys, only_model=only_model) + self.restarted_from_ckpt = True + + def make_cond_schedule(self, ): + self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long) + ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long() + self.cond_ids[:self.num_timesteps_cond] = ids + + @rank_zero_only + @torch.no_grad() + def on_train_batch_start(self, batch, batch_idx, dataloader_idx=None): + # only for very first batch, reset the self.scale_factor + if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and \ + not self.restarted_from_ckpt: + assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously' + # set rescale weight to 1./std of encodings + mainlogger.info("### USING STD-RESCALING ###") + x = super().get_input(batch, self.first_stage_key) + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + del self.scale_factor + self.register_buffer('scale_factor', 1. / z.flatten().std()) + mainlogger.info(f"setting self.scale_factor to {self.scale_factor}") + mainlogger.info("### USING STD-RESCALING ###") + mainlogger.info(f"std={z.flatten().std()}") + + def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s) + + self.shorten_cond_schedule = self.num_timesteps_cond > 1 + if self.shorten_cond_schedule: + self.make_cond_schedule() + + def instantiate_first_stage(self, config): + model = instantiate_from_config(config) + self.first_stage_model = model.eval() + self.first_stage_model.train = disabled_train + for param in self.first_stage_model.parameters(): + param.requires_grad = False + + def instantiate_cond_stage(self, config): + if not self.cond_stage_trainable: + model = instantiate_from_config(config) + self.cond_stage_model = model.eval() + self.cond_stage_model.train = disabled_train + for param in self.cond_stage_model.parameters(): + param.requires_grad = False + else: + model = instantiate_from_config(config) + self.cond_stage_model = model + + def get_learned_conditioning(self, c): + if self.cond_stage_forward is None: + if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode): + c = self.cond_stage_model.encode(c) + if isinstance(c, DiagonalGaussianDistribution): + c = c.mode() + else: + c = self.cond_stage_model(c) + else: + assert hasattr(self.cond_stage_model, self.cond_stage_forward) + c = getattr(self.cond_stage_model, self.cond_stage_forward)(c) + return c + + def get_first_stage_encoding(self, encoder_posterior, noise=None): + if isinstance(encoder_posterior, DiagonalGaussianDistribution): + z = encoder_posterior.sample(noise=noise) + elif isinstance(encoder_posterior, torch.Tensor): + z = encoder_posterior + else: + raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented") + return self.scale_factor * z + + @torch.no_grad() + def encode_first_stage(self, x): + if self.encoder_type == "2d" and x.dim() == 5: + b, _, t, _, _ = x.shape + x = rearrange(x, 'b c t h w -> (b t) c h w') + reshape_back = True + else: + reshape_back = False + + ## consume more GPU memory but faster + if not self.perframe_ae: + encoder_posterior = self.first_stage_model.encode(x) + results = self.get_first_stage_encoding(encoder_posterior).detach() + else: ## consume less GPU memory but slower + results = [] + for index in range(x.shape[0]): + frame_batch = self.first_stage_model.encode(x[index:index+1,:,:,:]) + frame_result = self.get_first_stage_encoding(frame_batch).detach() + results.append(frame_result) + results = torch.cat(results, dim=0) + + if reshape_back: + results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t) + + return results + + def decode_core(self, z, **kwargs): + if self.encoder_type == "2d" and z.dim() == 5: + b, _, t, _, _ = z.shape + z = rearrange(z, 'b c t h w -> (b t) c h w') + reshape_back = True + else: + reshape_back = False + + z = 1. / self.scale_factor * z + if not self.perframe_ae: + results = self.first_stage_model.decode(z, **kwargs) + else: + + results = [] + + n_samples = default(self.en_and_decode_n_samples_a_time, self.temporal_length) + n_rounds = math.ceil(z.shape[0] / n_samples) + with torch.autocast("cuda", enabled=True): + for n in range(n_rounds): + if isinstance(self.first_stage_model.decoder, VideoDecoder): + kwargs.update({"timesteps": len(z[n * n_samples : (n + 1) * n_samples])}) + else: + kwargs = {} + + out = self.first_stage_model.decode( + z[n * n_samples : (n + 1) * n_samples], **kwargs + ) + results.append(out) + results = torch.cat(results, dim=0) + + if reshape_back: + results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t) + return results + + @torch.no_grad() + def decode_first_stage(self, z, **kwargs): + return self.decode_core(z, **kwargs) + + # same as above but without decorator + def differentiable_decode_first_stage(self, z, **kwargs): + return self.decode_core(z, **kwargs) + + @torch.no_grad() + def get_batch_input(self, batch, random_uncond, return_first_stage_outputs=False, return_original_cond=False): + ## video shape: b, c, t, h, w + x = super().get_input(batch, self.first_stage_key) + + ## encode video frames x to z via a 2D encoder + z = self.encode_first_stage(x) + + ## get caption condition + cond = batch[self.cond_stage_key] + if random_uncond and self.uncond_type == 'empty_seq': + for i, ci in enumerate(cond): + if random.random() < self.uncond_prob: + cond[i] = "" + if isinstance(cond, dict) or isinstance(cond, list): + cond_emb = self.get_learned_conditioning(cond) + else: + cond_emb = self.get_learned_conditioning(cond.to(self.device)) + if random_uncond and self.uncond_type == 'zero_embed': + for i, ci in enumerate(cond): + if random.random() < self.uncond_prob: + cond_emb[i] = torch.zeros_like(cond_emb[i]) + + out = [z, cond_emb] + ## optional output: self-reconst or caption + if return_first_stage_outputs: + xrec = self.decode_first_stage(z) + out.extend([xrec]) + + if return_original_cond: + out.append(cond) + + return out + + def forward(self, x, c, **kwargs): + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + if self.use_dynamic_rescale: + x = x * extract_into_tensor(self.scale_arr, t, x.shape) + return self.p_losses(x, c, t, **kwargs) + + def shared_step(self, batch, random_uncond, **kwargs): + x, c = self.get_batch_input(batch, random_uncond=random_uncond) + loss, loss_dict = self(x, c, **kwargs) + + return loss, loss_dict + + def apply_model(self, x_noisy, t, cond, **kwargs): + if isinstance(cond, dict): + # hybrid case, cond is exptected to be a dict + pass + else: + if not isinstance(cond, list): + cond = [cond] + key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn' + cond = {key: [cond[0]]} + + control_cond = cond["control_cond"] + + if control_cond is not None: + control_cond = rearrange(control_cond, 'b c t h w-> (b t) c h w') + control_x = rearrange(x_noisy, 'b c t h w-> (b t) c h w') + control_context = repeat(cond["c_crossattn"][0], "b c l-> (repeat b) c l", repeat=16) + control = self.control_model(x=control_x, hint=control_cond, timesteps=t, context=control_context) + control = [c * self.control_scale for c in control] + else: + control = None + + x_recon = self.model(x_noisy, t, c_crossattn=cond["c_crossattn"], c_concat=cond["c_concat"], control=control, **kwargs) + + if isinstance(x_recon, tuple): + return x_recon[0] + else: + return x_recon + + + def p_losses(self, x_start, cond, t, noise=None, **kwargs): + if self.noise_strength > 0: + b, c, f, _, _ = x_start.shape + offset_noise = torch.randn(b, c, f, 1, 1, device=x_start.device) + noise = default(noise, lambda: torch.randn_like(x_start) + self.noise_strength * offset_noise) + else: + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + + model_output = self.apply_model(x_noisy, t, cond, **kwargs) + + loss_dict = {} + prefix = 'train' if self.training else 'val' + + if self.parameterization == "x0": + target = x_start + elif self.parameterization == "eps": + target = noise + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError() + + loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3, 4]) + loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()}) + + if self.logvar.device is not self.device: + self.logvar = self.logvar.to(self.device) + logvar_t = self.logvar[t] + # logvar_t = self.logvar[t.item()].to(self.device) # device conflict when ddp shared + loss = loss_simple / torch.exp(logvar_t) + logvar_t + # loss = loss_simple / torch.exp(self.logvar) + self.logvar + if self.learn_logvar: + loss_dict.update({f'{prefix}/loss_gamma': loss.mean()}) + loss_dict.update({'logvar': self.logvar.data.mean()}) + + loss = self.l_simple_weight * loss.mean() + + loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3, 4)) + loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean() + loss_dict.update({f'{prefix}/loss_vlb': loss_vlb}) + loss += (self.original_elbo_weight * loss_vlb) + loss_dict.update({f'{prefix}/loss': loss}) + + return loss, loss_dict + + def training_step(self, batch, batch_idx): + loss, loss_dict = self.shared_step(batch, random_uncond=self.classifier_free_guidance) + ## sync_dist | rank_zero_only + self.log_dict(loss_dict, prog_bar=True, logger=True, on_step=True, on_epoch=True, sync_dist=False) + #self.log("epoch/global_step", self.global_step.float(), prog_bar=True, logger=True, on_step=True, on_epoch=False) + ''' + if self.use_scheduler: + lr = self.optimizers().param_groups[0]['lr'] + self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False, rank_zero_only=True) + ''' + if (batch_idx+1) % self.log_every_t == 0: + mainlogger.info(f"batch:{batch_idx}|epoch:{self.current_epoch} [globalstep:{self.global_step}]: loss={loss}") + return loss + + def _get_denoise_row_from_list(self, samples, desc=''): + denoise_row = [] + for zd in tqdm(samples, desc=desc): + denoise_row.append(self.decode_first_stage(zd.to(self.device))) + n_log_timesteps = len(denoise_row) + + denoise_row = torch.stack(denoise_row) # n_log_timesteps, b, C, H, W + + if denoise_row.dim() == 5: + denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_log_timesteps) + elif denoise_row.dim() == 6: + # video, grid_size=[n_log_timesteps*bs, t] + video_length = denoise_row.shape[3] + denoise_grid = rearrange(denoise_row, 'n b c t h w -> b n c t h w') + denoise_grid = rearrange(denoise_grid, 'b n c t h w -> (b n) c t h w') + denoise_grid = rearrange(denoise_grid, 'n c t h w -> (n t) c h w') + denoise_grid = make_grid(denoise_grid, nrow=video_length) + else: + raise ValueError + + return denoise_grid + + @torch.no_grad() + def log_images(self, batch, sample=True, ddim_steps=200, ddim_eta=1., plot_denoise_rows=False, \ + unconditional_guidance_scale=1.0, **kwargs): + """ log images for LatentDiffusion """ + ##### control sampled imgae for logging, larger value may cause OOM + sampled_img_num = 2 + for key in batch.keys(): + batch[key] = batch[key][:sampled_img_num] + + ## TBD: currently, classifier_free_guidance sampling is only supported by DDIM + use_ddim = ddim_steps is not None + log = dict() + z, c, xrec, xc = self.get_batch_input(batch, random_uncond=False, + return_first_stage_outputs=True, + return_original_cond=True) + + N = xrec.shape[0] + log["reconst"] = xrec + log["condition"] = xc + + + if sample: + # get uncond embedding for classifier-free guidance sampling + if unconditional_guidance_scale != 1.0: + if isinstance(c, dict): + c_cat, c_emb = c["c_concat"][0], c["c_crossattn"][0] + log["condition_cat"] = c_cat + else: + c_emb = c + + if self.uncond_type == "empty_seq": + prompts = N * [""] + uc = self.get_learned_conditioning(prompts) + elif self.uncond_type == "zero_embed": + uc = torch.zeros_like(c_emb) + ## hybrid case + if isinstance(c, dict): + uc_hybrid = {"c_concat": [c_cat], "c_crossattn": [uc]} + uc = uc_hybrid + else: + uc = None + + with self.ema_scope("Plotting"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps,eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, x0=z, **kwargs) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + return log + + def p_mean_variance(self, x, c, t, clip_denoised: bool, return_x0=False, score_corrector=None, corrector_kwargs=None, **kwargs): + t_in = t + model_out = self.apply_model(x, t_in, c, **kwargs) + + if score_corrector is not None: + assert self.parameterization == "eps" + model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs) + + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + else: + raise NotImplementedError() + + if clip_denoised: + x_recon.clamp_(-1., 1.) + + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + + if return_x0: + return model_mean, posterior_variance, posterior_log_variance, x_recon + else: + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, return_x0=False, \ + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, **kwargs): + b, *_, device = *x.shape, x.device + outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, return_x0=return_x0, \ + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs, **kwargs) + if return_x0: + model_mean, _, model_log_variance, x0 = outputs + else: + model_mean, _, model_log_variance = outputs + + noise = noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + + if return_x0: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0 + else: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def p_sample_loop(self, cond, shape, return_intermediates=False, x_T=None, verbose=True, callback=None, \ + timesteps=None, mask=None, x0=None, img_callback=None, start_T=None, log_every_t=None, **kwargs): + + if not log_every_t: + log_every_t = self.log_every_t + device = self.betas.device + b = shape[0] + # sample an initial noise + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + intermediates = [img] + if timesteps is None: + timesteps = self.num_timesteps + if start_T is not None: + timesteps = min(timesteps, start_T) + + iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(range(0, timesteps)) + + if mask is not None: + assert x0 is not None + assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match + + for i in iterator: + ts = torch.full((b,), i, device=device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img = self.p_sample(img, cond, ts, clip_denoised=self.clip_denoised, **kwargs) + if mask is not None: + img_orig = self.q_sample(x0, ts) + img = img_orig * mask + (1. - mask) * img + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(img) + if callback: callback(i) + if img_callback: img_callback(img, i) + + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None, \ + verbose=True, timesteps=None, mask=None, x0=None, shape=None, **kwargs): + if shape is None: + shape = (batch_size, self.channels, self.temporal_length, *self.image_size) + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + return self.p_sample_loop(cond, + shape, + return_intermediates=return_intermediates, x_T=x_T, + verbose=verbose, timesteps=timesteps, + mask=mask, x0=x0, **kwargs) + + @torch.no_grad() + def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs): + if ddim: + ddim_sampler = DDIMSampler(self) + shape = (self.channels, self.temporal_length, *self.image_size) + samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, shape, cond, verbose=False, **kwargs) + + else: + samples, intermediates = self.sample(cond=cond, batch_size=batch_size, return_intermediates=True, **kwargs) + + return samples, intermediates + + def configure_schedulers(self, optimizer): + assert 'target' in self.scheduler_config + scheduler_name = self.scheduler_config.target.split('.')[-1] + interval = self.scheduler_config.interval + frequency = self.scheduler_config.frequency + if scheduler_name == "LambdaLRScheduler": + scheduler = instantiate_from_config(self.scheduler_config) + scheduler.start_step = self.global_step + lr_scheduler = { + 'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule), + 'interval': interval, + 'frequency': frequency + } + elif scheduler_name == "CosineAnnealingLRScheduler": + scheduler = instantiate_from_config(self.scheduler_config) + decay_steps = scheduler.decay_steps + last_step = -1 if self.global_step == 0 else scheduler.start_step + lr_scheduler = { + 'scheduler': CosineAnnealingLR(optimizer, T_max=decay_steps, last_epoch=last_step), + 'interval': interval, + 'frequency': frequency + } + else: + raise NotImplementedError + return lr_scheduler + +class LatentVisualDiffusion(LatentDiffusion): + def __init__(self, img_cond_stage_config, image_proj_stage_config, freeze_embedder=True, image_proj_model_trainable=True, *args, **kwargs): + super().__init__(*args, **kwargs) + self.image_proj_model_trainable = image_proj_model_trainable + self._init_embedder(img_cond_stage_config, freeze_embedder) + self._init_img_ctx_projector(image_proj_stage_config, image_proj_model_trainable) + + def _init_img_ctx_projector(self, config, trainable): + self.image_proj_model = instantiate_from_config(config) + if not trainable: + self.image_proj_model.eval() + self.image_proj_model.train = disabled_train + for param in self.image_proj_model.parameters(): + param.requires_grad = False + + def _init_embedder(self, config, freeze=True): + self.embedder = instantiate_from_config(config) + if freeze: + self.embedder.eval() + self.embedder.train = disabled_train + for param in self.embedder.parameters(): + param.requires_grad = False + + def shared_step(self, batch, random_uncond, **kwargs): + x, c, fs = self.get_batch_input(batch, random_uncond=random_uncond, return_fs=True) + kwargs.update({"fs": fs.long()}) + loss, loss_dict = self(x, c, **kwargs) + return loss, loss_dict + + def get_batch_input(self, batch, random_uncond, return_first_stage_outputs=False, return_original_cond=False, return_fs=False, return_cond_frame=False, return_original_input=False, **kwargs): + ## x: b c t h w + x = super().get_input(batch, self.first_stage_key) + ## encode video frames x to z via a 2D encoder + z = self.encode_first_stage(x) + + ## get caption condition + cond_input = batch[self.cond_stage_key] + + if isinstance(cond_input, dict) or isinstance(cond_input, list): + cond_emb = self.get_learned_conditioning(cond_input) + else: + cond_emb = self.get_learned_conditioning(cond_input.to(self.device)) + + cond = {} + ## to support classifier-free guidance, randomly drop out only text conditioning 5%, only image conditioning 5%, and both 5%. + if random_uncond: + random_num = torch.rand(x.size(0), device=x.device) + else: + random_num = torch.ones(x.size(0), device=x.device) ## by doning so, we can get text embedding and complete img emb for inference + prompt_mask = rearrange(random_num < 2 * self.uncond_prob, "n -> n 1 1") + input_mask = 1 - rearrange((random_num >= self.uncond_prob).float() * (random_num < 3 * self.uncond_prob).float(), "n -> n 1 1 1") + + null_prompt = self.get_learned_conditioning([""]) + prompt_imb = torch.where(prompt_mask, null_prompt, cond_emb.detach()) + + ## get conditioning frame + cond_frame_index = 0 + if self.rand_cond_frame: + cond_frame_index = random.randint(0, self.model.diffusion_model.temporal_length-1) + + img = x[:,:,cond_frame_index,...] + img = input_mask * img + ## img: b c h w + img_emb = self.embedder(img) ## b l c + img_emb = self.image_proj_model(img_emb) + + if self.model.conditioning_key == 'hybrid': + ## simply repeat the cond_frame to match the seq_len of z + img_cat_cond = z[:,:,cond_frame_index,:,:] + img_cat_cond = img_cat_cond.unsqueeze(2) + img_cat_cond = repeat(img_cat_cond, 'b c t h w -> b c (repeat t) h w', repeat=z.shape[2]) + + cond["c_concat"] = [img_cat_cond] # b c t h w + cond["c_crossattn"] = [torch.cat([prompt_imb, img_emb], dim=1)] ## concat in the seq_len dim + + out = [z, cond] + if return_first_stage_outputs: + xrec = self.decode_first_stage(z) + out.extend([xrec]) + + if return_original_cond: + out.append(cond_input) + if return_fs: + if self.fps_condition_type == 'fs': + fs = super().get_input(batch, 'frame_stride') + elif self.fps_condition_type == 'fps': + fs = super().get_input(batch, 'fps') + out.append(fs) + if return_cond_frame: + out.append(x[:,:,cond_frame_index,...].unsqueeze(2)) + if return_original_input: + out.append(x) + + return out + + @torch.no_grad() + def log_images(self, batch, sample=True, ddim_steps=50, ddim_eta=1., plot_denoise_rows=False, \ + unconditional_guidance_scale=1.0, mask=None, **kwargs): + """ log images for LatentVisualDiffusion """ + ##### sampled_img_num: control sampled imgae for logging, larger value may cause OOM + sampled_img_num = 1 + for key in batch.keys(): + batch[key] = batch[key][:sampled_img_num] + + ## TBD: currently, classifier_free_guidance sampling is only supported by DDIM + use_ddim = ddim_steps is not None + log = dict() + + z, c, xrec, xc, fs, cond_x = self.get_batch_input(batch, random_uncond=False, + return_first_stage_outputs=True, + return_original_cond=True, + return_fs=True, + return_cond_frame=True) + + N = xrec.shape[0] + log["image_condition"] = cond_x + log["reconst"] = xrec + xc_with_fs = [] + for idx, content in enumerate(xc): + xc_with_fs.append(content + '_fs=' + str(fs[idx].item())) + log["condition"] = xc_with_fs + kwargs.update({"fs": fs.long()}) + + c_cat = None + if sample: + # get uncond embedding for classifier-free guidance sampling + if unconditional_guidance_scale != 1.0: + if isinstance(c, dict): + c_emb = c["c_crossattn"][0] + if 'c_concat' in c.keys(): + c_cat = c["c_concat"][0] + else: + c_emb = c + + if self.uncond_type == "empty_seq": + prompts = N * [""] + uc_prompt = self.get_learned_conditioning(prompts) + elif self.uncond_type == "zero_embed": + uc_prompt = torch.zeros_like(c_emb) + + img = torch.zeros_like(xrec[:,:,0]) ## b c h w + ## img: b c h w + img_emb = self.embedder(img) ## b l c + uc_img = self.image_proj_model(img_emb) + + uc = torch.cat([uc_prompt, uc_img], dim=1) + ## hybrid case + if isinstance(c, dict): + uc_hybrid = {"c_concat": [c_cat], "c_crossattn": [uc]} + uc = uc_hybrid + else: + uc = None + + with self.ema_scope("Plotting"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps,eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, x0=z, **kwargs) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + return log + + def configure_optimizers(self): + """ configure_optimizers for LatentDiffusion """ + lr = self.learning_rate + + params = list(self.model.parameters()) + mainlogger.info(f"@Training [{len(params)}] Full Paramters.") + + if self.cond_stage_trainable: + params_cond_stage = [p for p in self.cond_stage_model.parameters() if p.requires_grad == True] + mainlogger.info(f"@Training [{len(params_cond_stage)}] Paramters for Cond_stage_model.") + params.extend(params_cond_stage) + + if self.image_proj_model_trainable: + mainlogger.info(f"@Training [{len(list(self.image_proj_model.parameters()))}] Paramters for Image_proj_model.") + params.extend(list(self.image_proj_model.parameters())) + + if self.learn_logvar: + mainlogger.info('Diffusion model optimizing logvar') + if isinstance(params[0], dict): + params.append({"params": [self.logvar]}) + else: + params.append(self.logvar) + + ## optimizer + optimizer = torch.optim.AdamW(params, lr=lr) + + ## lr scheduler + if self.use_scheduler: + mainlogger.info("Setting up scheduler...") + lr_scheduler = self.configure_schedulers(optimizer) + return [optimizer], [lr_scheduler] + + return optimizer + + +class DiffusionWrapper(pl.LightningModule): + def __init__(self, diff_model_config, conditioning_key): + super().__init__() + self.diffusion_model = instantiate_from_config(diff_model_config) + self.conditioning_key = conditioning_key + + def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, control = None, + c_adm=None, s=None, mask=None, **kwargs): + # temporal_context = fps is foNone + if self.conditioning_key is None: + out = self.diffusion_model(x, t) + elif self.conditioning_key == 'concat': + xc = torch.cat([x] + c_concat, dim=1) + out = self.diffusion_model(xc, t, **kwargs) + elif self.conditioning_key == 'crossattn': + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(x, t, context=cc, **kwargs) + elif self.conditioning_key == 'hybrid': + ## it is just right [b,c,t,h,w]: concatenate in channel dim + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc, control=control, **kwargs) + + elif self.conditioning_key == 'resblockcond': + cc = c_crossattn[0] + out = self.diffusion_model(x, t, context=cc) + elif self.conditioning_key == 'adm': + cc = c_crossattn[0] + out = self.diffusion_model(x, t, y=cc) + elif self.conditioning_key == 'hybrid-adm': + assert c_adm is not None + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc, y=c_adm, **kwargs) + elif self.conditioning_key == 'hybrid-time': + assert s is not None + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc, s=s) + elif self.conditioning_key == 'concat-time-mask': + # assert s is not None + xc = torch.cat([x] + c_concat, dim=1) + out = self.diffusion_model(xc, t, context=None, s=s, mask=mask) + elif self.conditioning_key == 'concat-adm-mask': + # assert s is not None + if c_concat is not None: + xc = torch.cat([x] + c_concat, dim=1) + else: + xc = x + out = self.diffusion_model(xc, t, context=None, y=s, mask=mask) + elif self.conditioning_key == 'hybrid-adm-mask': + cc = torch.cat(c_crossattn, 1) + if c_concat is not None: + xc = torch.cat([x] + c_concat, dim=1) + else: + xc = x + out = self.diffusion_model(xc, t, context=cc, y=s, mask=mask) + elif self.conditioning_key == 'hybrid-time-adm': # adm means y, e.g., class index + # assert s is not None + assert c_adm is not None + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc, s=s, y=c_adm) + elif self.conditioning_key == 'crossattn-adm': + assert c_adm is not None + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(x, t, context=cc, y=c_adm) + else: + raise NotImplementedError() + + return out \ No newline at end of file diff --git a/lvdm/models/samplers/ddim.py b/lvdm/models/samplers/ddim.py new file mode 100644 index 0000000000000000000000000000000000000000..88006edc2856efc80a5f33839dcaffc27d2930b2 --- /dev/null +++ b/lvdm/models/samplers/ddim.py @@ -0,0 +1,317 @@ +import numpy as np +from tqdm import tqdm +import torch +from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps, rescale_noise_cfg +from lvdm.common import noise_like +from lvdm.common import extract_into_tensor +import copy + + +class DDIMSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + self.counter = 0 + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + if self.model.use_dynamic_rescale: + self.ddim_scale_arr = self.model.scale_arr[self.ddim_timesteps] + self.ddim_scale_arr_prev = torch.cat([self.ddim_scale_arr[0:1], self.ddim_scale_arr[:-1]]) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + schedule_verbose=False, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + precision=None, + fs=None, + timestep_spacing='uniform', #uniform_trailing for starting from last timestep + guidance_rescale=0.0, + **kwargs + ): + + # check condition bs + if conditioning is not None: + if isinstance(conditioning, dict): + try: + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + except: + cbs = conditioning[list(conditioning.keys())[0]][0].shape[0] + + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_discretize=timestep_spacing, ddim_eta=eta, verbose=schedule_verbose) + + # make shape + if len(shape) == 3: + C, H, W = shape + size = (batch_size, C, H, W) + elif len(shape) == 4: + C, T, H, W = shape + size = (batch_size, C, T, H, W) + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + verbose=verbose, + precision=precision, + fs=fs, + guidance_rescale=guidance_rescale, + **kwargs) + return samples, intermediates + + @torch.no_grad() + def ddim_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,precision=None,fs=None,guidance_rescale=0.0, + **kwargs): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + if precision is not None: + if precision == 16: + img = img.to(dtype=torch.float16) + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + if verbose: + iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) + else: + iterator = time_range + + clean_cond = kwargs.pop("clean_cond", False) + + # cond_copy, unconditional_conditioning_copy = copy.deepcopy(cond), copy.deepcopy(unconditional_conditioning) + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + + ## use mask to blend noised original latent (img_orig) & new sampled latent (img) + if mask is not None: + assert x0 is not None + if clean_cond: + img_orig = x0 + else: + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img # keep original & modify use img + + + + + outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + mask=mask,x0=x0,fs=fs,guidance_rescale=guidance_rescale, + **kwargs) + + + img, pred_x0 = outs + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + uc_type=None, conditional_guidance_scale_temporal=None,mask=None,x0=None,guidance_rescale=0.0,**kwargs): + b, *_, device = *x.shape, x.device + if x.dim() == 5: + is_video = True + else: + is_video = False + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + model_output = self.model.apply_model(x, t, c, **kwargs) # unet denoiser + else: + ### do_classifier_free_guidance + if isinstance(c, torch.Tensor) or isinstance(c, dict): + e_t_cond = self.model.apply_model(x, t, c, **kwargs) + e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs) + else: + raise NotImplementedError + + model_output = e_t_uncond + unconditional_guidance_scale * (e_t_cond - e_t_uncond) + + if guidance_rescale > 0.0: + model_output = rescale_noise_cfg(model_output, e_t_cond, guidance_rescale=guidance_rescale) + + if self.model.parameterization == "v": + e_t = self.model.predict_eps_from_z_and_v(x, t, model_output) + else: + e_t = model_output + + if score_corrector is not None: + assert self.model.parameterization == "eps", 'not implemented' + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + # sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + sigmas = self.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + + if is_video: + size = (b, 1, 1, 1, 1) + else: + size = (b, 1, 1, 1) + a_t = torch.full(size, alphas[index], device=device) + a_prev = torch.full(size, alphas_prev[index], device=device) + sigma_t = torch.full(size, sigmas[index], device=device) + sqrt_one_minus_at = torch.full(size, sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + if self.model.parameterization != "v": + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + else: + pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output) + + if self.model.use_dynamic_rescale: + scale_t = torch.full(size, self.ddim_scale_arr[index], device=device) + prev_scale_t = torch.full(size, self.ddim_scale_arr_prev[index], device=device) + rescale = (prev_scale_t / scale_t) + pred_x0 *= rescale + + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + + return x_prev, pred_x0 + + @torch.no_grad() + def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + use_original_steps=False, callback=None): + + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + timesteps = timesteps[:t_start] + + time_range = np.flip(timesteps) + total_steps = timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='Decoding image', total=total_steps) + x_dec = x_latent + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long) + x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning) + if callback: callback(i) + return x_dec + + @torch.no_grad() + def stochastic_encode(self, x0, t, use_original_steps=False, noise=None): + # fast, but does not allow for exact reconstruction + # t serves as an index to gather the correct alphas + if use_original_steps: + sqrt_alphas_cumprod = self.sqrt_alphas_cumprod + sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod + else: + sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas) + sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas + + if noise is None: + noise = torch.randn_like(x0) + return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 + + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise) diff --git a/lvdm/models/samplers/ddim_multiplecond.py b/lvdm/models/samplers/ddim_multiplecond.py new file mode 100644 index 0000000000000000000000000000000000000000..31c3d89aa2df1ad72dbc8533622ba75d8b5feb16 --- /dev/null +++ b/lvdm/models/samplers/ddim_multiplecond.py @@ -0,0 +1,323 @@ +import numpy as np +from tqdm import tqdm +import torch +from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps, rescale_noise_cfg +from lvdm.common import noise_like +from lvdm.common import extract_into_tensor +import copy + + +class DDIMSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + self.counter = 0 + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + if self.model.use_dynamic_rescale: + self.ddim_scale_arr = self.model.scale_arr[self.ddim_timesteps] + self.ddim_scale_arr_prev = torch.cat([self.ddim_scale_arr[0:1], self.ddim_scale_arr[:-1]]) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + schedule_verbose=False, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + precision=None, + fs=None, + timestep_spacing='uniform', #uniform_trailing for starting from last timestep + guidance_rescale=0.0, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + + # check condition bs + if conditioning is not None: + if isinstance(conditioning, dict): + try: + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + except: + cbs = conditioning[list(conditioning.keys())[0]][0].shape[0] + + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + # print('==> timestep_spacing: ', timestep_spacing, guidance_rescale) + self.make_schedule(ddim_num_steps=S, ddim_discretize=timestep_spacing, ddim_eta=eta, verbose=schedule_verbose) + + # make shape + if len(shape) == 3: + C, H, W = shape + size = (batch_size, C, H, W) + elif len(shape) == 4: + C, T, H, W = shape + size = (batch_size, C, T, H, W) + # print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + verbose=verbose, + precision=precision, + fs=fs, + guidance_rescale=guidance_rescale, + **kwargs) + return samples, intermediates + + @torch.no_grad() + def ddim_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,precision=None,fs=None,guidance_rescale=0.0, + **kwargs): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + if precision is not None: + if precision == 16: + img = img.to(dtype=torch.float16) + + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + if verbose: + iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) + else: + iterator = time_range + + clean_cond = kwargs.pop("clean_cond", False) + + # cond_copy, unconditional_conditioning_copy = copy.deepcopy(cond), copy.deepcopy(unconditional_conditioning) + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + + ## use mask to blend noised original latent (img_orig) & new sampled latent (img) + if mask is not None: + assert x0 is not None + if clean_cond: + img_orig = x0 + else: + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img # keep original & modify use img + + + + + outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + mask=mask,x0=x0,fs=fs,guidance_rescale=guidance_rescale, + **kwargs) + + + + img, pred_x0 = outs + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + uc_type=None, cfg_img=None,mask=None,x0=None,guidance_rescale=0.0, **kwargs): + b, *_, device = *x.shape, x.device + if x.dim() == 5: + is_video = True + else: + is_video = False + if cfg_img is None: + cfg_img = unconditional_guidance_scale + + unconditional_conditioning_img_nonetext = kwargs['unconditional_conditioning_img_nonetext'] + + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + model_output = self.model.apply_model(x, t, c, **kwargs) # unet denoiser + else: + ### with unconditional condition + e_t_cond = self.model.apply_model(x, t, c, **kwargs) + e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs) + e_t_uncond_img = self.model.apply_model(x, t, unconditional_conditioning_img_nonetext, **kwargs) + # text cfg + model_output = e_t_uncond + cfg_img * (e_t_uncond_img - e_t_uncond) + unconditional_guidance_scale * (e_t_cond - e_t_uncond_img) + if guidance_rescale > 0.0: + model_output = rescale_noise_cfg(model_output, e_t_cond, guidance_rescale=guidance_rescale) + + if self.model.parameterization == "v": + e_t = self.model.predict_eps_from_z_and_v(x, t, model_output) + else: + e_t = model_output + + if score_corrector is not None: + assert self.model.parameterization == "eps", 'not implemented' + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + + if is_video: + size = (b, 1, 1, 1, 1) + else: + size = (b, 1, 1, 1) + a_t = torch.full(size, alphas[index], device=device) + a_prev = torch.full(size, alphas_prev[index], device=device) + sigma_t = torch.full(size, sigmas[index], device=device) + sqrt_one_minus_at = torch.full(size, sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + if self.model.parameterization != "v": + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + else: + pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output) + + if self.model.use_dynamic_rescale: + scale_t = torch.full(size, self.ddim_scale_arr[index], device=device) + prev_scale_t = torch.full(size, self.ddim_scale_arr_prev[index], device=device) + rescale = (prev_scale_t / scale_t) + pred_x0 *= rescale + + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + + return x_prev, pred_x0 + + @torch.no_grad() + def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + use_original_steps=False, callback=None): + + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + timesteps = timesteps[:t_start] + + time_range = np.flip(timesteps) + total_steps = timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='Decoding image', total=total_steps) + x_dec = x_latent + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long) + x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning) + if callback: callback(i) + return x_dec + + @torch.no_grad() + def stochastic_encode(self, x0, t, use_original_steps=False, noise=None): + # fast, but does not allow for exact reconstruction + # t serves as an index to gather the correct alphas + if use_original_steps: + sqrt_alphas_cumprod = self.sqrt_alphas_cumprod + sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod + else: + sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas) + sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas + + if noise is None: + noise = torch.randn_like(x0) + return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 + + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise) \ No newline at end of file diff --git a/lvdm/models/utils_diffusion.py b/lvdm/models/utils_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..f5265d4af6bcc4acb94b9790b903911ff0710cdf --- /dev/null +++ b/lvdm/models/utils_diffusion.py @@ -0,0 +1,158 @@ +import math +import numpy as np +import torch +import torch.nn.functional as F +from einops import repeat + + +def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False): + """ + Create sinusoidal timestep embeddings. + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an [N x dim] Tensor of positional embeddings. + """ + if not repeat_only: + half = dim // 2 + freqs = torch.exp( + -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half + ).to(device=timesteps.device) + args = timesteps[:, None].float() * freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) + else: + embedding = repeat(timesteps, 'b -> b d', d=dim) + return embedding + + +def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if schedule == "linear": + betas = ( + torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2 + ) + + elif schedule == "cosine": + timesteps = ( + torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s + ) + alphas = timesteps / (1 + cosine_s) * np.pi / 2 + alphas = torch.cos(alphas).pow(2) + alphas = alphas / alphas[0] + betas = 1 - alphas[1:] / alphas[:-1] + betas = np.clip(betas, a_min=0, a_max=0.999) + + elif schedule == "sqrt_linear": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) + elif schedule == "sqrt": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5 + else: + raise ValueError(f"schedule '{schedule}' unknown.") + return betas.numpy() + + +def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True): + if ddim_discr_method == 'uniform': + c = num_ddpm_timesteps // num_ddim_timesteps + ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c))) + steps_out = ddim_timesteps + 1 + elif ddim_discr_method == 'uniform_trailing': + c = num_ddpm_timesteps / num_ddim_timesteps + ddim_timesteps = np.flip(np.round(np.arange(num_ddpm_timesteps, 0, -c))).astype(np.int64) + steps_out = ddim_timesteps - 1 + elif ddim_discr_method == 'quad': + ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int) + steps_out = ddim_timesteps + 1 + else: + raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"') + + # assert ddim_timesteps.shape[0] == num_ddim_timesteps + # add one to get the final alpha values right (the ones from first scale to data during sampling) + # steps_out = ddim_timesteps + 1 + if verbose: + print(f'Selected timesteps for ddim sampler: {steps_out}') + return steps_out + + +def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True): + # select alphas for computing the variance schedule + # print(f'ddim_timesteps={ddim_timesteps}, len_alphacums={len(alphacums)}') + alphas = alphacums[ddim_timesteps] + alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist()) + + # according the formula provided in https://arxiv.org/abs/2010.02502 + sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)) + if verbose: + print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}') + print(f'For the chosen value of eta, which is {eta}, ' + f'this results in the following sigma_t schedule for ddim sampler {sigmas}') + return sigmas, alphas, alphas_prev + + +def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, + which defines the cumulative product of (1-beta) over time from t = [0,1]. + :param num_diffusion_timesteps: the number of betas to produce. + :param alpha_bar: a lambda that takes an argument t from 0 to 1 and + produces the cumulative product of (1-beta) up to that + part of the diffusion process. + :param max_beta: the maximum beta to use; use values lower than 1 to + prevent singularities. + """ + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return np.array(betas) + +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1) + + Args: + betas (`numpy.ndarray`): + the betas that the scheduler is being initialized with. + + Returns: + `numpy.ndarray`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_bar_sqrt = np.sqrt(alphas_cumprod) + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].copy() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].copy() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = np.concatenate([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg \ No newline at end of file diff --git a/lvdm/modules/attention.py b/lvdm/modules/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..176885d9ff6f5675c413523a38b78845ce04bd97 --- /dev/null +++ b/lvdm/modules/attention.py @@ -0,0 +1,514 @@ +import torch +from torch import nn, einsum +import torch.nn.functional as F +from einops import rearrange, repeat +from functools import partial +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False +from lvdm.common import ( + checkpoint, + exists, + default, +) +from lvdm.basics import zero_module + + +class RelativePosition(nn.Module): + """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """ + + def __init__(self, num_units, max_relative_position): + super().__init__() + self.num_units = num_units + self.max_relative_position = max_relative_position + self.embeddings_table = nn.Parameter(torch.Tensor(max_relative_position * 2 + 1, num_units)) + nn.init.xavier_uniform_(self.embeddings_table) + + def forward(self, length_q, length_k): + device = self.embeddings_table.device + range_vec_q = torch.arange(length_q, device=device) + range_vec_k = torch.arange(length_k, device=device) + distance_mat = range_vec_k[None, :] - range_vec_q[:, None] + distance_mat_clipped = torch.clamp(distance_mat, -self.max_relative_position, self.max_relative_position) + final_mat = distance_mat_clipped + self.max_relative_position + final_mat = final_mat.long() + embeddings = self.embeddings_table[final_mat] + return embeddings + + +class CrossAttention(nn.Module): + + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., + relative_position=False, temporal_length=None, video_length=None, image_cross_attention=False, image_cross_attention_scale=1.0, image_cross_attention_scale_learnable=False, text_context_len=77): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.scale = dim_head**-0.5 + self.heads = heads + self.dim_head = dim_head + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)) + + self.relative_position = relative_position + if self.relative_position: + assert(temporal_length is not None) + self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length) + self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length) + else: + ## only used for spatial attention, while NOT for temporal attention + if XFORMERS_IS_AVAILBLE and temporal_length is None: + self.forward = self.efficient_forward + + self.video_length = video_length + self.image_cross_attention = image_cross_attention + self.image_cross_attention_scale = image_cross_attention_scale + self.text_context_len = text_context_len + self.image_cross_attention_scale_learnable = image_cross_attention_scale_learnable + if self.image_cross_attention: + self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False) + if image_cross_attention_scale_learnable: + self.register_parameter('alpha', nn.Parameter(torch.tensor(0.)) ) + + + def forward(self, x, context=None, mask=None): + spatial_self_attn = (context is None) + k_ip, v_ip, out_ip = None, None, None + + h = self.heads + q = self.to_q(x) + context = default(context, x) + + if self.image_cross_attention and not spatial_self_attn: + context, context_image = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:] + k = self.to_k(context) + v = self.to_v(context) + k_ip = self.to_k_ip(context_image) + v_ip = self.to_v_ip(context_image) + else: + if not spatial_self_attn: + context = context[:,:self.text_context_len,:] + k = self.to_k(context) + v = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale + if self.relative_position: + len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1] + k2 = self.relative_position_k(len_q, len_k) + sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check + sim += sim2 + del k + + if exists(mask): + ## feasible for causal attention mask only + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b i j -> (b h) i j', h=h) + sim.masked_fill_(~(mask>0.5), max_neg_value) + + # attention, what we cannot get enough of + sim = sim.softmax(dim=-1) + + out = torch.einsum('b i j, b j d -> b i d', sim, v) + if self.relative_position: + v2 = self.relative_position_v(len_q, len_v) + out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check + out += out2 + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + + + ## for image cross-attention + if k_ip is not None: + k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip)) + sim_ip = torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale + del k_ip + sim_ip = sim_ip.softmax(dim=-1) + out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip) + out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h) + + + if out_ip is not None: + if self.image_cross_attention_scale_learnable: + out = out + self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha)+1) + else: + out = out + self.image_cross_attention_scale * out_ip + + return self.to_out(out) + + def efficient_forward(self, x, context=None, mask=None): + spatial_self_attn = (context is None) + k_ip, v_ip, out_ip = None, None, None + + q = self.to_q(x) + context = default(context, x) + + if self.image_cross_attention and not spatial_self_attn: + context, context_image = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:] + k = self.to_k(context) + v = self.to_v(context) + k_ip = self.to_k_ip(context_image) + v_ip = self.to_v_ip(context_image) + else: + if not spatial_self_attn: + context = context[:,:self.text_context_len,:] + k = self.to_k(context) + v = self.to_v(context) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + # actually compute the attention, what we cannot get enough of + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None) + + ## for image cross-attention + if k_ip is not None: + k_ip, v_ip = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (k_ip, v_ip), + ) + out_ip = xformers.ops.memory_efficient_attention(q, k_ip, v_ip, attn_bias=None, op=None) + out_ip = ( + out_ip.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + if out_ip is not None: + if self.image_cross_attention_scale_learnable: + out = out + self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha)+1) + else: + out = out + self.image_cross_attention_scale * out_ip + + return self.to_out(out) + + +class BasicTransformerBlock(nn.Module): + + def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, + disable_self_attn=False, attention_cls=None, video_length=None, image_cross_attention=False, image_cross_attention_scale=1.0, image_cross_attention_scale_learnable=False, text_context_len=77): + super().__init__() + attn_cls = CrossAttention if attention_cls is None else attention_cls + self.disable_self_attn = disable_self_attn + self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, + context_dim=context_dim if self.disable_self_attn else None) + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout, video_length=video_length, image_cross_attention=image_cross_attention, image_cross_attention_scale=image_cross_attention_scale, image_cross_attention_scale_learnable=image_cross_attention_scale_learnable,text_context_len=text_context_len) + self.image_cross_attention = image_cross_attention + + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.norm3 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + + + def forward(self, x, context=None, mask=None, **kwargs): + ## implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments + input_tuple = (x,) ## should not be (x), otherwise *input_tuple will decouple x into multiple arguments + if context is not None: + input_tuple = (x, context) + if mask is not None: + forward_mask = partial(self._forward, mask=mask) + return checkpoint(forward_mask, (x,), self.parameters(), self.checkpoint) + return checkpoint(self._forward, input_tuple, self.parameters(), self.checkpoint) + + + def _forward(self, x, context=None, mask=None): + x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None, mask=mask) + x + x = self.attn2(self.norm2(x), context=context, mask=mask) + x + x = self.ff(self.norm3(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data in spatial axis. + First, project the input (aka embedding) + and reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + NEW: use_linear for more efficiency instead of the 1x1 convs + """ + + def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None, + use_checkpoint=True, disable_self_attn=False, use_linear=False, video_length=None, + image_cross_attention=False, image_cross_attention_scale_learnable=False): + super().__init__() + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + if not use_linear: + self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + attention_cls = None + self.transformer_blocks = nn.ModuleList([ + BasicTransformerBlock( + inner_dim, + n_heads, + d_head, + dropout=dropout, + context_dim=context_dim, + disable_self_attn=disable_self_attn, + checkpoint=use_checkpoint, + attention_cls=attention_cls, + video_length=video_length, + image_cross_attention=image_cross_attention, + image_cross_attention_scale_learnable=image_cross_attention_scale_learnable, + ) for d in range(depth) + ]) + if not use_linear: + self.proj_out = zero_module(nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)) + else: + self.proj_out = zero_module(nn.Linear(inner_dim, in_channels)) + self.use_linear = use_linear + + + def forward(self, x, context=None, **kwargs): + b, c, h, w = x.shape + x_in = x + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, 'b c h w -> b (h w) c').contiguous() + if self.use_linear: + x = self.proj_in(x) + for i, block in enumerate(self.transformer_blocks): + x = block(x, context=context, **kwargs) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous() + if not self.use_linear: + x = self.proj_out(x) + return x + x_in + + +class TemporalTransformer(nn.Module): + """ + Transformer block for image-like data in temporal axis. + First, reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + """ + def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None, + use_checkpoint=True, use_linear=False, only_self_att=True, causal_attention=False, causal_block_size=1, + relative_position=False, temporal_length=None): + super().__init__() + self.only_self_att = only_self_att + self.relative_position = relative_position + self.causal_attention = causal_attention + self.causal_block_size = causal_block_size + + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + if not use_linear: + self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + if relative_position: + assert(temporal_length is not None) + attention_cls = partial(CrossAttention, relative_position=True, temporal_length=temporal_length) + else: + attention_cls = partial(CrossAttention, temporal_length=temporal_length) + if self.causal_attention: + assert(temporal_length is not None) + self.mask = torch.tril(torch.ones([1, temporal_length, temporal_length])) + + if self.only_self_att: + context_dim = None + self.transformer_blocks = nn.ModuleList([ + BasicTransformerBlock( + inner_dim, + n_heads, + d_head, + dropout=dropout, + context_dim=context_dim, + attention_cls=attention_cls, + checkpoint=use_checkpoint) for d in range(depth) + ]) + if not use_linear: + self.proj_out = zero_module(nn.Conv1d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)) + else: + self.proj_out = zero_module(nn.Linear(inner_dim, in_channels)) + self.use_linear = use_linear + + def forward(self, x, context=None): + b, c, t, h, w = x.shape + x_in = x + x = self.norm(x) + x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous() + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, 'bhw c t -> bhw t c').contiguous() + if self.use_linear: + x = self.proj_in(x) + + temp_mask = None + if self.causal_attention: + # slice the from mask map + temp_mask = self.mask[:,:t,:t].to(x.device) + + if temp_mask is not None: + mask = temp_mask.to(x.device) + mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b*h*w) + else: + mask = None + + if self.only_self_att: + ## note: if no context is given, cross-attention defaults to self-attention + for i, block in enumerate(self.transformer_blocks): + x = block(x, mask=mask) + x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous() + else: + x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous() + context = rearrange(context, '(b t) l con -> b t l con', t=t).contiguous() + for i, block in enumerate(self.transformer_blocks): + # calculate each batch one by one (since number in shape could not greater then 65,535 for some package) + for j in range(b): + context_j = repeat( + context[j], + 't l con -> (t r) l con', r=(h * w) // t, t=t).contiguous() + ## note: causal mask will not applied in cross-attention case + x[j] = block(x[j], context=context_j) + + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous() + if not self.use_linear: + x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous() + x = self.proj_out(x) + x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h, w=w).contiguous() + + return x + x_in + + +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + nn.Linear(dim, inner_dim), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +class LinearAttention(nn.Module): + def __init__(self, dim, heads=4, dim_head=32): + super().__init__() + self.heads = heads + hidden_dim = dim_head * heads + self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False) + self.to_out = nn.Conv2d(hidden_dim, dim, 1) + + def forward(self, x): + b, c, h, w = x.shape + qkv = self.to_qkv(x) + q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3) + k = k.softmax(dim=-1) + context = torch.einsum('bhdn,bhen->bhde', k, v) + out = torch.einsum('bhde,bhdn->bhen', context, q) + out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w) + return self.to_out(out) + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = rearrange(q, 'b c h w -> b (h w) c') + k = rearrange(k, 'b c h w -> b c (h w)') + w_ = torch.einsum('bij,bjk->bik', q, k) + + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, 'b c h w -> b c (h w)') + w_ = rearrange(w_, 'b i j -> b j i') + h_ = torch.einsum('bij,bjk->bik', v, w_) + h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h) + h_ = self.proj_out(h_) + + return x+h_ diff --git a/lvdm/modules/attention_svd.py b/lvdm/modules/attention_svd.py new file mode 100644 index 0000000000000000000000000000000000000000..92ceb3c978025c9bb9a640d63558a20a4989d377 --- /dev/null +++ b/lvdm/modules/attention_svd.py @@ -0,0 +1,759 @@ +import logging +import math +from inspect import isfunction +from typing import Any, Optional + +import torch +import torch.nn.functional as F +from einops import rearrange, repeat +from packaging import version +from torch import nn +from torch.utils.checkpoint import checkpoint + +logpy = logging.getLogger(__name__) + +if version.parse(torch.__version__) >= version.parse("2.0.0"): + SDP_IS_AVAILABLE = True + from torch.backends.cuda import SDPBackend, sdp_kernel + + BACKEND_MAP = { + SDPBackend.MATH: { + "enable_math": True, + "enable_flash": False, + "enable_mem_efficient": False, + }, + SDPBackend.FLASH_ATTENTION: { + "enable_math": False, + "enable_flash": True, + "enable_mem_efficient": False, + }, + SDPBackend.EFFICIENT_ATTENTION: { + "enable_math": False, + "enable_flash": False, + "enable_mem_efficient": True, + }, + None: {"enable_math": True, "enable_flash": True, "enable_mem_efficient": True}, + } +else: + from contextlib import nullcontext + + SDP_IS_AVAILABLE = False + sdp_kernel = nullcontext + BACKEND_MAP = {} + logpy.warn( + f"No SDP backend available, likely because you are running in pytorch " + f"versions < 2.0. In fact, you are using PyTorch {torch.__version__}. " + f"You might want to consider upgrading." + ) + +try: + import xformers + import xformers.ops + + XFORMERS_IS_AVAILABLE = True +except: + XFORMERS_IS_AVAILABLE = False + logpy.warn("no module 'xformers'. Processing without...") + +# from .diffusionmodules.util import mixed_checkpoint as checkpoint + + +def exists(val): + return val is not None + + +def uniq(arr): + return {el: True for el in arr}.keys() + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def max_neg_value(t): + return -torch.finfo(t.dtype).max + + +def init_(tensor): + dim = tensor.shape[-1] + std = 1 / math.sqrt(dim) + tensor.uniform_(-std, std) + return tensor + + +# feedforward +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = ( + nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU()) + if not glu + else GEGLU(dim, inner_dim) + ) + + self.net = nn.Sequential( + project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def Normalize(in_channels): + return torch.nn.GroupNorm( + num_groups=32, num_channels=in_channels, eps=1e-6, affine=True + ) + + +class LinearAttention(nn.Module): + def __init__(self, dim, heads=4, dim_head=32): + super().__init__() + self.heads = heads + hidden_dim = dim_head * heads + self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False) + self.to_out = nn.Conv2d(hidden_dim, dim, 1) + + def forward(self, x): + b, c, h, w = x.shape + qkv = self.to_qkv(x) + q, k, v = rearrange( + qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3 + ) + k = k.softmax(dim=-1) + context = torch.einsum("bhdn,bhen->bhde", k, v) + out = torch.einsum("bhde,bhdn->bhen", context, q) + out = rearrange( + out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w + ) + return self.to_out(out) + + +class SelfAttention(nn.Module): + ATTENTION_MODES = ("xformers", "torch", "math") + + def __init__( + self, + dim: int, + num_heads: int = 8, + qkv_bias: bool = False, + qk_scale: Optional[float] = None, + attn_drop: float = 0.0, + proj_drop: float = 0.0, + attn_mode: str = "xformers", + ): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim**-0.5 + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + assert attn_mode in self.ATTENTION_MODES + self.attn_mode = attn_mode + + def forward(self, x: torch.Tensor) -> torch.Tensor: + B, L, C = x.shape + + qkv = self.qkv(x) + if self.attn_mode == "torch": + qkv = rearrange( + qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads + ).float() + q, k, v = qkv[0], qkv[1], qkv[2] # B H L D + x = torch.nn.functional.scaled_dot_product_attention(q, k, v) + x = rearrange(x, "B H L D -> B L (H D)") + elif self.attn_mode == "xformers": + qkv = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.num_heads) + q, k, v = qkv[0], qkv[1], qkv[2] # B L H D + x = xformers.ops.memory_efficient_attention(q, k, v) + x = rearrange(x, "B L H D -> B L (H D)", H=self.num_heads) + elif self.attn_mode == "math": + qkv = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads) + q, k, v = qkv[0], qkv[1], qkv[2] # B H L D + attn = (q @ k.transpose(-2, -1)) * self.scale + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + x = (attn @ v).transpose(1, 2).reshape(B, L, C) + else: + raise NotImplemented + + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.k = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.v = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.proj_out = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b, c, h, w = q.shape + q = rearrange(q, "b c h w -> b (h w) c") + k = rearrange(k, "b c h w -> b c (h w)") + w_ = torch.einsum("bij,bjk->bik", q, k) + + w_ = w_ * (int(c) ** (-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, "b c h w -> b c (h w)") + w_ = rearrange(w_, "b i j -> b j i") + h_ = torch.einsum("bij,bjk->bik", v, w_) + h_ = rearrange(h_, "b c (h w) -> b c h w", h=h) + h_ = self.proj_out(h_) + + return x + h_ + + +class CrossAttention(nn.Module): + def __init__( + self, + query_dim, + context_dim=None, + heads=8, + dim_head=64, + dropout=0.0, + backend=None, + ): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.scale = dim_head**-0.5 + self.heads = heads + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), nn.Dropout(dropout) + ) + self.backend = backend + + def forward( + self, + x, + context=None, + mask=None, + additional_tokens=None, + n_times_crossframe_attn_in_self=0, + ): + h = self.heads + + if additional_tokens is not None: + # get the number of masked tokens at the beginning of the output sequence + n_tokens_to_mask = additional_tokens.shape[1] + # add additional token + x = torch.cat([additional_tokens, x], dim=1) + + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + if n_times_crossframe_attn_in_self: + # reprogramming cross-frame attention as in https://arxiv.org/abs/2303.13439 + assert x.shape[0] % n_times_crossframe_attn_in_self == 0 + n_cp = x.shape[0] // n_times_crossframe_attn_in_self + k = repeat( + k[::n_times_crossframe_attn_in_self], "b ... -> (b n) ...", n=n_cp + ) + v = repeat( + v[::n_times_crossframe_attn_in_self], "b ... -> (b n) ...", n=n_cp + ) + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v)) + + ## old + """ + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + del q, k + + if exists(mask): + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + sim = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', sim, v) + """ + ## new + with sdp_kernel(**BACKEND_MAP[self.backend]): + # print("dispatching into backend", self.backend, "q/k/v shape: ", q.shape, k.shape, v.shape) + out = F.scaled_dot_product_attention( + q, k, v, attn_mask=mask + ) # scale is dim_head ** -0.5 per default + + del q, k, v + out = rearrange(out, "b h n d -> b n (h d)", h=h) + + if additional_tokens is not None: + # remove additional token + out = out[:, n_tokens_to_mask:] + return self.to_out(out) + + +class MemoryEfficientCrossAttention(nn.Module): + # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + def __init__( + self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0, **kwargs + ): + super().__init__() + logpy.debug( + f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, " + f"context_dim is {context_dim} and using {heads} heads with a " + f"dimension of {dim_head}." + ) + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.heads = heads + self.dim_head = dim_head + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), nn.Dropout(dropout) + ) + self.attention_op: Optional[Any] = None + + def forward( + self, + x, + context=None, + mask=None, + additional_tokens=None, + n_times_crossframe_attn_in_self=0, + ): + if additional_tokens is not None: + # get the number of masked tokens at the beginning of the output sequence + n_tokens_to_mask = additional_tokens.shape[1] + # add additional token + x = torch.cat([additional_tokens, x], dim=1) + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + if n_times_crossframe_attn_in_self: + # reprogramming cross-frame attention as in https://arxiv.org/abs/2303.13439 + assert x.shape[0] % n_times_crossframe_attn_in_self == 0 + # n_cp = x.shape[0]//n_times_crossframe_attn_in_self + k = repeat( + k[::n_times_crossframe_attn_in_self], + "b ... -> (b n) ...", + n=n_times_crossframe_attn_in_self, + ) + v = repeat( + v[::n_times_crossframe_attn_in_self], + "b ... -> (b n) ...", + n=n_times_crossframe_attn_in_self, + ) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + + # actually compute the attention, what we cannot get enough of + if version.parse(xformers.__version__) >= version.parse("0.0.21"): + # NOTE: workaround for + # https://github.com/facebookresearch/xformers/issues/845 + max_bs = 32768 + N = q.shape[0] + n_batches = math.ceil(N / max_bs) + out = list() + for i_batch in range(n_batches): + batch = slice(i_batch * max_bs, (i_batch + 1) * max_bs) + out.append( + xformers.ops.memory_efficient_attention( + q[batch], + k[batch], + v[batch], + attn_bias=None, + op=self.attention_op, + ) + ) + out = torch.cat(out, 0) + else: + out = xformers.ops.memory_efficient_attention( + q, k, v, attn_bias=None, op=self.attention_op + ) + + # TODO: Use this directly in the attention operation, as a bias + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + if additional_tokens is not None: + # remove additional token + out = out[:, n_tokens_to_mask:] + return self.to_out(out) + + +class BasicTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention, # ampere + } + + def __init__( + self, + dim, + n_heads, + d_head, + dropout=0.0, + context_dim=None, + gated_ff=True, + checkpoint=True, + disable_self_attn=False, + attn_mode="softmax", + sdp_backend=None, + ): + super().__init__() + assert attn_mode in self.ATTENTION_MODES + if attn_mode != "softmax" and not XFORMERS_IS_AVAILABLE: + logpy.warn( + f"Attention mode '{attn_mode}' is not available. Falling " + f"back to native attention. This is not a problem in " + f"Pytorch >= 2.0. FYI, you are running with PyTorch " + f"version {torch.__version__}." + ) + attn_mode = "softmax" + elif attn_mode == "softmax" and not SDP_IS_AVAILABLE: + logpy.warn( + "We do not support vanilla attention anymore, as it is too " + "expensive. Sorry." + ) + if not XFORMERS_IS_AVAILABLE: + assert ( + False + ), "Please install xformers via e.g. 'pip install xformers==0.0.16'" + else: + logpy.info("Falling back to xformers efficient attention.") + attn_mode = "softmax-xformers" + attn_cls = self.ATTENTION_MODES[attn_mode] + if version.parse(torch.__version__) >= version.parse("2.0.0"): + assert sdp_backend is None or isinstance(sdp_backend, SDPBackend) + else: + assert sdp_backend is None + self.disable_self_attn = disable_self_attn + self.attn1 = attn_cls( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + context_dim=context_dim if self.disable_self_attn else None, + backend=sdp_backend, + ) # is a self-attention if not self.disable_self_attn + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.attn2 = attn_cls( + query_dim=dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + backend=sdp_backend, + ) # is self-attn if context is none + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.norm3 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + if self.checkpoint: + logpy.debug(f"{self.__class__.__name__} is using checkpointing") + + def forward( + self, x, context=None, additional_tokens=None, n_times_crossframe_attn_in_self=0 + ): + kwargs = {"x": x} + + if context is not None: + kwargs.update({"context": context}) + + if additional_tokens is not None: + kwargs.update({"additional_tokens": additional_tokens}) + + if n_times_crossframe_attn_in_self: + kwargs.update( + {"n_times_crossframe_attn_in_self": n_times_crossframe_attn_in_self} + ) + + # return mixed_checkpoint(self._forward, kwargs, self.parameters(), self.checkpoint) + if self.checkpoint: + # inputs = {"x": x, "context": context} + return checkpoint(self._forward, x, context) + # return checkpoint(self._forward, inputs, self.parameters(), self.checkpoint) + else: + return self._forward(**kwargs) + + def _forward( + self, x, context=None, additional_tokens=None, n_times_crossframe_attn_in_self=0 + ): + x = ( + self.attn1( + self.norm1(x), + context=context if self.disable_self_attn else None, + additional_tokens=additional_tokens, + n_times_crossframe_attn_in_self=n_times_crossframe_attn_in_self + if not self.disable_self_attn + else 0, + ) + + x + ) + x = ( + self.attn2( + self.norm2(x), context=context, additional_tokens=additional_tokens + ) + + x + ) + x = self.ff(self.norm3(x)) + x + return x + + +class BasicTransformerSingleLayerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention # on the A100s not quite as fast as the above version + # (todo might depend on head_dim, check, falls back to semi-optimized kernels for dim!=[16,32,64,128]) + } + + def __init__( + self, + dim, + n_heads, + d_head, + dropout=0.0, + context_dim=None, + gated_ff=True, + checkpoint=True, + attn_mode="softmax", + ): + super().__init__() + assert attn_mode in self.ATTENTION_MODES + attn_cls = self.ATTENTION_MODES[attn_mode] + self.attn1 = attn_cls( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + context_dim=context_dim, + ) + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + + def forward(self, x, context=None): + # inputs = {"x": x, "context": context} + # return checkpoint(self._forward, inputs, self.parameters(), self.checkpoint) + return checkpoint(self._forward, x, context) + + def _forward(self, x, context=None): + x = self.attn1(self.norm1(x), context=context) + x + x = self.ff(self.norm2(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data. + First, project the input (aka embedding) + and reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + NEW: use_linear for more efficiency instead of the 1x1 convs + """ + + def __init__( + self, + in_channels, + n_heads, + d_head, + depth=1, + dropout=0.0, + context_dim=None, + disable_self_attn=False, + use_linear=False, + attn_type="softmax", + use_checkpoint=True, + # sdp_backend=SDPBackend.FLASH_ATTENTION + sdp_backend=None, + ): + super().__init__() + logpy.debug( + f"constructing {self.__class__.__name__} of depth {depth} w/ " + f"{in_channels} channels and {n_heads} heads." + ) + + if exists(context_dim) and not isinstance(context_dim, list): + context_dim = [context_dim] + if exists(context_dim) and isinstance(context_dim, list): + if depth != len(context_dim): + logpy.warn( + f"{self.__class__.__name__}: Found context dims " + f"{context_dim} of depth {len(context_dim)}, which does not " + f"match the specified 'depth' of {depth}. Setting context_dim " + f"to {depth * [context_dim[0]]} now." + ) + # depth does not match context dims. + assert all( + map(lambda x: x == context_dim[0], context_dim) + ), "need homogenous context_dim to match depth automatically" + context_dim = depth * [context_dim[0]] + elif context_dim is None: + context_dim = [None] * depth + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = Normalize(in_channels) + if not use_linear: + self.proj_in = nn.Conv2d( + in_channels, inner_dim, kernel_size=1, stride=1, padding=0 + ) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + inner_dim, + n_heads, + d_head, + dropout=dropout, + context_dim=context_dim[d], + disable_self_attn=disable_self_attn, + attn_mode=attn_type, + checkpoint=use_checkpoint, + sdp_backend=sdp_backend, + ) + for d in range(depth) + ] + ) + if not use_linear: + self.proj_out = zero_module( + nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + ) + else: + # self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) + self.proj_out = zero_module(nn.Linear(inner_dim, in_channels)) + self.use_linear = use_linear + + def forward(self, x, context=None): + # note: if no context is given, cross-attention defaults to self-attention + if not isinstance(context, list): + context = [context] + b, c, h, w = x.shape + x_in = x + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, "b c h w -> b (h w) c").contiguous() + if self.use_linear: + x = self.proj_in(x) + for i, block in enumerate(self.transformer_blocks): + if i > 0 and len(context) == 1: + i = 0 # use same context for each block + x = block(x, context=context[i]) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w).contiguous() + if not self.use_linear: + x = self.proj_out(x) + return x + x_in + + +class SimpleTransformer(nn.Module): + def __init__( + self, + dim: int, + depth: int, + heads: int, + dim_head: int, + context_dim: Optional[int] = None, + dropout: float = 0.0, + checkpoint: bool = True, + ): + super().__init__() + self.layers = nn.ModuleList([]) + for _ in range(depth): + self.layers.append( + BasicTransformerBlock( + dim, + heads, + dim_head, + dropout=dropout, + context_dim=context_dim, + attn_mode="softmax-xformers", + checkpoint=checkpoint, + ) + ) + + def forward( + self, + x: torch.Tensor, + context: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + for layer in self.layers: + x = layer(x, context) + return x \ No newline at end of file diff --git a/lvdm/modules/encoders/condition.py b/lvdm/modules/encoders/condition.py new file mode 100644 index 0000000000000000000000000000000000000000..443ce9a9e7e015dad5c05db737a189da83dbbb4f --- /dev/null +++ b/lvdm/modules/encoders/condition.py @@ -0,0 +1,389 @@ +import torch +import torch.nn as nn +import kornia +import open_clip +from torch.utils.checkpoint import checkpoint +from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel +from lvdm.common import autocast +from utils.utils import count_params + + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + + +class IdentityEncoder(AbstractEncoder): + def encode(self, x): + return x + + +class ClassEmbedder(nn.Module): + def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1): + super().__init__() + self.key = key + self.embedding = nn.Embedding(n_classes, embed_dim) + self.n_classes = n_classes + self.ucg_rate = ucg_rate + + def forward(self, batch, key=None, disable_dropout=False): + if key is None: + key = self.key + # this is for use in crossattn + c = batch[key][:, None] + if self.ucg_rate > 0. and not disable_dropout: + mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate) + c = mask * c + (1 - mask) * torch.ones_like(c) * (self.n_classes - 1) + c = c.long() + c = self.embedding(c) + return c + + def get_unconditional_conditioning(self, bs, device="cuda"): + uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000) + uc = torch.ones((bs,), device=device) * uc_class + uc = {self.key: uc} + return uc + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +class FrozenT5Embedder(AbstractEncoder): + """Uses the T5 transformer encoder for text""" + + def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, + freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl + super().__init__() + self.tokenizer = T5Tokenizer.from_pretrained(version) + self.transformer = T5EncoderModel.from_pretrained(version) + self.device = device + self.max_length = max_length # TODO: typical value? + if freeze: + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + # self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens) + + z = outputs.last_hidden_state + return z + + def encode(self, text): + return self(text) + + +class FrozenCLIPEmbedder(AbstractEncoder): + """Uses the CLIP transformer encoder for text (from huggingface)""" + LAYERS = [ + "last", + "pooled", + "hidden" + ] + + def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, + freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32 + super().__init__() + assert layer in self.LAYERS + self.tokenizer = CLIPTokenizer.from_pretrained(version) + self.transformer = CLIPTextModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + self.layer_idx = layer_idx + if layer == "hidden": + assert layer_idx is not None + assert 0 <= abs(layer_idx) <= 12 + + def freeze(self): + self.transformer = self.transformer.eval() + # self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer == "hidden") + if self.layer == "last": + z = outputs.last_hidden_state + elif self.layer == "pooled": + z = outputs.pooler_output[:, None, :] + else: + z = outputs.hidden_states[self.layer_idx] + return z + + def encode(self, text): + return self(text) + + +class ClipImageEmbedder(nn.Module): + def __init__( + self, + model, + jit=False, + device='cuda' if torch.cuda.is_available() else 'cpu', + antialias=True, + ucg_rate=0. + ): + super().__init__() + from clip import load as load_clip + self.model, _ = load_clip(name=model, device=device, jit=jit) + + self.antialias = antialias + + self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False) + self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False) + self.ucg_rate = ucg_rate + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize(x, (224, 224), + interpolation='bicubic', align_corners=True, + antialias=self.antialias) + x = (x + 1.) / 2. + # re-normalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def forward(self, x, no_dropout=False): + # x is assumed to be in range [-1,1] + out = self.model.encode_image(self.preprocess(x)) + out = out.to(x.dtype) + if self.ucg_rate > 0. and not no_dropout: + out = torch.bernoulli((1. - self.ucg_rate) * torch.ones(out.shape[0], device=out.device))[:, None] * out + return out + + +class FrozenOpenCLIPEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP transformer encoder for text + """ + LAYERS = [ + # "pooled", + "last", + "penultimate" + ] + + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="last"): + super().__init__() + assert layer in self.LAYERS + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version) + del model.visual + self.model = model + + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "last": + self.layer_idx = 0 + elif self.layer == "penultimate": + self.layer_idx = 1 + else: + raise NotImplementedError() + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + tokens = open_clip.tokenize(text) ## all clip models use 77 as context length + z = self.encode_with_transformer(tokens.to(self.device)) + return z + + def encode_with_transformer(self, text): + x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.model.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.model.ln_final(x) + return x + + def text_transformer_forward(self, x: torch.Tensor, attn_mask=None): + for i, r in enumerate(self.model.transformer.resblocks): + if i == len(self.model.transformer.resblocks) - self.layer_idx: + break + if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting(): + x = checkpoint(r, x, attn_mask) + else: + x = r(x, attn_mask=attn_mask) + return x + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPImageEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP vision transformer encoder for images + """ + + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="pooled", antialias=True, ucg_rate=0.): + super().__init__() + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), + pretrained=version, ) + del model.transformer + self.model = model + # self.mapper = torch.nn.Linear(1280, 1024) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "penultimate": + raise NotImplementedError() + self.layer_idx = 1 + + self.antialias = antialias + + self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False) + self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False) + self.ucg_rate = ucg_rate + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize(x, (224, 224), + interpolation='bicubic', align_corners=True, + antialias=self.antialias) + x = (x + 1.) / 2. + # renormalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def freeze(self): + self.model = self.model.eval() + for param in self.model.parameters(): + param.requires_grad = False + + @autocast + def forward(self, image, no_dropout=False): + z = self.encode_with_vision_transformer(image) + if self.ucg_rate > 0. and not no_dropout: + z = torch.bernoulli((1. - self.ucg_rate) * torch.ones(z.shape[0], device=z.device))[:, None] * z + return z + + def encode_with_vision_transformer(self, img): + img = self.preprocess(img) + x = self.model.visual(img) + return x + + def encode(self, text): + return self(text) + +class FrozenOpenCLIPImageEmbedderV2(AbstractEncoder): + """ + Uses the OpenCLIP vision transformer encoder for images + """ + + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", + freeze=True, layer="pooled", antialias=True): + super().__init__() + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), + pretrained=version, ) + del model.transformer + self.model = model + self.device = device + + if freeze: + self.freeze() + self.layer = layer + if self.layer == "penultimate": + raise NotImplementedError() + self.layer_idx = 1 + + self.antialias = antialias + + self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False) + self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False) + + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize(x, (224, 224), + interpolation='bicubic', align_corners=True, + antialias=self.antialias) + x = (x + 1.) / 2. + # renormalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def freeze(self): + self.model = self.model.eval() + for param in self.model.parameters(): + param.requires_grad = False + + def forward(self, image, no_dropout=False): + ## image: b c h w + z = self.encode_with_vision_transformer(image) + return z + + def encode_with_vision_transformer(self, x): + x = self.preprocess(x) + + # to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1 + if self.model.visual.input_patchnorm: + # einops - rearrange(x, 'b c (h p1) (w p2) -> b (h w) (c p1 p2)') + x = x.reshape(x.shape[0], x.shape[1], self.model.visual.grid_size[0], self.model.visual.patch_size[0], self.model.visual.grid_size[1], self.model.visual.patch_size[1]) + x = x.permute(0, 2, 4, 1, 3, 5) + x = x.reshape(x.shape[0], self.model.visual.grid_size[0] * self.model.visual.grid_size[1], -1) + x = self.model.visual.patchnorm_pre_ln(x) + x = self.model.visual.conv1(x) + else: + x = self.model.visual.conv1(x) # shape = [*, width, grid, grid] + x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2] + x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width] + + # class embeddings and positional embeddings + x = torch.cat( + [self.model.visual.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), + x], dim=1) # shape = [*, grid ** 2 + 1, width] + x = x + self.model.visual.positional_embedding.to(x.dtype) + + # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in + x = self.model.visual.patch_dropout(x) + x = self.model.visual.ln_pre(x) + + x = x.permute(1, 0, 2) # NLD -> LND + x = self.model.visual.transformer(x) + x = x.permute(1, 0, 2) # LND -> NLD + + return x + +class FrozenCLIPT5Encoder(AbstractEncoder): + def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda", + clip_max_length=77, t5_max_length=77): + super().__init__() + self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length) + self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length) + print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, " + f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params.") + + def encode(self, text): + return self(text) + + def forward(self, text): + clip_z = self.clip_encoder.encode(text) + t5_z = self.t5_encoder.encode(text) + return [clip_z, t5_z] diff --git a/lvdm/modules/encoders/resampler.py b/lvdm/modules/encoders/resampler.py new file mode 100644 index 0000000000000000000000000000000000000000..d9abd23f681510d757acdf4dbb31794fe49ab5b1 --- /dev/null +++ b/lvdm/modules/encoders/resampler.py @@ -0,0 +1,145 @@ +# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py +# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py +# and https://github.com/tencent-ailab/IP-Adapter/blob/main/ip_adapter/resampler.py +import math +import torch +import torch.nn as nn + + +class ImageProjModel(nn.Module): + """Projection Model""" + def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4): + super().__init__() + self.cross_attention_dim = cross_attention_dim + self.clip_extra_context_tokens = clip_extra_context_tokens + self.proj = nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim) + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds): + #embeds = image_embeds + embeds = image_embeds.type(list(self.proj.parameters())[0].dtype) + clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim) + clip_extra_context_tokens = self.norm(clip_extra_context_tokens) + return clip_extra_context_tokens + + +# FFN +def FeedForward(dim, mult=4): + inner_dim = int(dim * mult) + return nn.Sequential( + nn.LayerNorm(dim), + nn.Linear(dim, inner_dim, bias=False), + nn.GELU(), + nn.Linear(inner_dim, dim, bias=False), + ) + + +def reshape_tensor(x, heads): + bs, length, width = x.shape + #(bs, length, width) --> (bs, length, n_heads, dim_per_head) + x = x.view(bs, length, heads, -1) + # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head) + x = x.transpose(1, 2) + # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head) + x = x.reshape(bs, heads, length, -1) + return x + + +class PerceiverAttention(nn.Module): + def __init__(self, *, dim, dim_head=64, heads=8): + super().__init__() + self.scale = dim_head**-0.5 + self.dim_head = dim_head + self.heads = heads + inner_dim = dim_head * heads + + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + + self.to_q = nn.Linear(dim, inner_dim, bias=False) + self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False) + self.to_out = nn.Linear(inner_dim, dim, bias=False) + + + def forward(self, x, latents): + """ + Args: + x (torch.Tensor): image features + shape (b, n1, D) + latent (torch.Tensor): latent features + shape (b, n2, D) + """ + x = self.norm1(x) + latents = self.norm2(latents) + + b, l, _ = latents.shape + + q = self.to_q(latents) + kv_input = torch.cat((x, latents), dim=-2) + k, v = self.to_kv(kv_input).chunk(2, dim=-1) + + q = reshape_tensor(q, self.heads) + k = reshape_tensor(k, self.heads) + v = reshape_tensor(v, self.heads) + + # attention + scale = 1 / math.sqrt(math.sqrt(self.dim_head)) + weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards + weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) + out = weight @ v + + out = out.permute(0, 2, 1, 3).reshape(b, l, -1) + + return self.to_out(out) + + +class Resampler(nn.Module): + def __init__( + self, + dim=1024, + depth=8, + dim_head=64, + heads=16, + num_queries=8, + embedding_dim=768, + output_dim=1024, + ff_mult=4, + video_length=None, # using frame-wise version or not + ): + super().__init__() + ## queries for a single frame / image + self.num_queries = num_queries + self.video_length = video_length + + ## queries for each frame + if video_length is not None: + num_queries = num_queries * video_length + + self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5) + self.proj_in = nn.Linear(embedding_dim, dim) + self.proj_out = nn.Linear(dim, output_dim) + self.norm_out = nn.LayerNorm(output_dim) + + self.layers = nn.ModuleList([]) + for _ in range(depth): + self.layers.append( + nn.ModuleList( + [ + PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads), + FeedForward(dim=dim, mult=ff_mult), + ] + ) + ) + + def forward(self, x): + latents = self.latents.repeat(x.size(0), 1, 1) ## B (T L) C + x = self.proj_in(x) + + for attn, ff in self.layers: + latents = attn(x, latents) + latents + latents = ff(latents) + latents + + latents = self.proj_out(latents) + latents = self.norm_out(latents) # B L C or B (T L) C + + return latents \ No newline at end of file diff --git a/lvdm/modules/networks/ae_modules.py b/lvdm/modules/networks/ae_modules.py new file mode 100644 index 0000000000000000000000000000000000000000..f1d52d07354108982f5f6e1e49641bc521c65c49 --- /dev/null +++ b/lvdm/modules/networks/ae_modules.py @@ -0,0 +1,857 @@ +# pytorch_diffusion + derived encoder decoder +import math + +import torch +import numpy as np +import torch.nn as nn +from einops import rearrange + +from utils.utils import instantiate_from_config +from lvdm.modules.attention import LinearAttention + +def nonlinearity(x): + # swish + return x*torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32): + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + + + +class LinAttnBlock(LinearAttention): + """to match AttnBlock usage""" + def __init__(self, in_channels): + super().__init__(dim=in_channels, heads=1, dim_head=in_channels) + + +class AttnBlock(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = q.reshape(b,c,h*w) # bcl + q = q.permute(0,2,1) # bcl -> blc l=hw + k = k.reshape(b,c,h*w) # bcl + + w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b,c,h*w) + w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b,c,h,w) + + h_ = self.proj_out(h_) + + return x+h_ + +def make_attn(in_channels, attn_type="vanilla"): + assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown' + #print(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + return AttnBlock(in_channels) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + return LinAttnBlock(in_channels) + +class Downsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + self.in_channels = in_channels + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=2, + padding=0) + def forward(self, x): + if self.with_conv: + pad = (0,1,0,1) + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + return x + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + self.in_channels = in_channels + if self.with_conv: + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + return x + +def get_timestep_embedding(timesteps, embedding_dim): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: + From Fairseq. + Build sinusoidal embeddings. + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + assert len(timesteps.shape) == 1 + + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) + emb = emb.to(device=timesteps.device) + emb = timesteps.float()[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0,1,0,0)) + return emb + + + +class ResnetBlock(nn.Module): + def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, + dropout, temb_channels=512): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels) + self.conv1 = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if temb_channels > 0: + self.temb_proj = torch.nn.Linear(temb_channels, + out_channels) + self.norm2 = Normalize(out_channels) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d(out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + else: + self.nin_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x+h + +class Model(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = self.ch*4 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + self.use_timestep = use_timestep + if self.use_timestep: + # timestep embedding + self.temb = nn.Module() + self.temb.dense = nn.ModuleList([ + torch.nn.Linear(self.ch, + self.temb_ch), + torch.nn.Linear(self.temb_ch, + self.temb_ch), + ]) + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + skip_in = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + if i_block == self.num_res_blocks: + skip_in = ch*in_ch_mult[i_level] + block.append(ResnetBlock(in_channels=block_in+skip_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x, t=None, context=None): + #assert x.shape[2] == x.shape[3] == self.resolution + if context is not None: + # assume aligned context, cat along channel axis + x = torch.cat((x, context), dim=1) + if self.use_timestep: + # timestep embedding + assert t is not None + temb = get_timestep_embedding(t, self.ch) + temb = self.temb.dense[0](temb) + temb = nonlinearity(temb) + temb = self.temb.dense[1](temb) + else: + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block]( + torch.cat([h, hs.pop()], dim=1), temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + def get_last_layer(self): + return self.conv_out.weight + + +class Encoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla", + **ignore_kwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + 2*z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x, return_hidden_states=False): + # timestep embedding + temb = None + + # print(f'encoder-input={x.shape}') + # downsampling + hs = [self.conv_in(x)] + + ## if we return hidden states for decoder usage, we will store them in a list + if return_hidden_states: + hidden_states = [] + # print(f'encoder-conv in feat={hs[0].shape}') + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + # print(f'encoder-down feat={h.shape}') + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if return_hidden_states: + hidden_states.append(h) + if i_level != self.num_resolutions-1: + # print(f'encoder-downsample (input)={hs[-1].shape}') + hs.append(self.down[i_level].downsample(hs[-1])) + # print(f'encoder-downsample (output)={hs[-1].shape}') + if return_hidden_states: + hidden_states.append(hs[0]) + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + # print(f'encoder-mid1 feat={h.shape}') + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + # print(f'encoder-mid2 feat={h.shape}') + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + # print(f'end feat={h.shape}') + if return_hidden_states: + return h, hidden_states + else: + return h + + +class Decoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, + attn_type="vanilla", **ignorekwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,)+tuple(ch_mult) + block_in = ch*ch_mult[self.num_resolutions-1] + curr_res = resolution // 2**(self.num_resolutions-1) + self.z_shape = (1,z_channels,curr_res,curr_res) + print("AE working on z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape))) + + # z to block_in + self.conv_in = torch.nn.Conv2d(z_channels, + block_in, + kernel_size=3, + stride=1, + padding=1) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, z): + #assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + + # print(f'decoder-input={z.shape}') + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + # print(f'decoder-conv in feat={h.shape}') + + # middle + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + # print(f'decoder-mid feat={h.shape}') + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block](h, temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + # print(f'decoder-up feat={h.shape}') + if i_level != 0: + h = self.up[i_level].upsample(h) + # print(f'decoder-upsample feat={h.shape}') + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + # print(f'decoder-conv_out feat={h.shape}') + if self.tanh_out: + h = torch.tanh(h) + return h + + +class SimpleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, *args, **kwargs): + super().__init__() + self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1), + ResnetBlock(in_channels=in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=2 * in_channels, + out_channels=4 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=4 * in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + nn.Conv2d(2*in_channels, in_channels, 1), + Upsample(in_channels, with_conv=True)]) + # end + self.norm_out = Normalize(in_channels) + self.conv_out = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + for i, layer in enumerate(self.model): + if i in [1,2,3]: + x = layer(x, None) + else: + x = layer(x) + + h = self.norm_out(x) + h = nonlinearity(h) + x = self.conv_out(h) + return x + + +class UpsampleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, + ch_mult=(2,2), dropout=0.0): + super().__init__() + # upsampling + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + block_in = in_channels + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.res_blocks = nn.ModuleList() + self.upsample_blocks = nn.ModuleList() + for i_level in range(self.num_resolutions): + res_block = [] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + res_block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + self.res_blocks.append(nn.ModuleList(res_block)) + if i_level != self.num_resolutions - 1: + self.upsample_blocks.append(Upsample(block_in, True)) + curr_res = curr_res * 2 + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # upsampling + h = x + for k, i_level in enumerate(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.res_blocks[i_level][i_block](h, None) + if i_level != self.num_resolutions - 1: + h = self.upsample_blocks[k](h) + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class LatentRescaler(nn.Module): + def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2): + super().__init__() + # residual block, interpolate, residual block + self.factor = factor + self.conv_in = nn.Conv2d(in_channels, + mid_channels, + kernel_size=3, + stride=1, + padding=1) + self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + self.attn = AttnBlock(mid_channels) + self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + + self.conv_out = nn.Conv2d(mid_channels, + out_channels, + kernel_size=1, + ) + + def forward(self, x): + x = self.conv_in(x) + for block in self.res_block1: + x = block(x, None) + x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor)))) + x = self.attn(x) + for block in self.res_block2: + x = block(x, None) + x = self.conv_out(x) + return x + + +class MergedRescaleEncoder(nn.Module): + def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, + ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + intermediate_chn = ch * ch_mult[-1] + self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult, + z_channels=intermediate_chn, double_z=False, resolution=resolution, + attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv, + out_ch=None) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn, + mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth) + + def forward(self, x): + x = self.encoder(x) + x = self.rescaler(x) + return x + + +class MergedRescaleDecoder(nn.Module): + def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8), + dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + tmp_chn = z_channels*ch_mult[-1] + self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout, + resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks, + ch_mult=ch_mult, resolution=resolution, ch=ch) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn, + out_channels=tmp_chn, depth=rescale_module_depth) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Upsampler(nn.Module): + def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2): + super().__init__() + assert out_size >= in_size + num_blocks = int(np.log2(out_size//in_size))+1 + factor_up = 1.+ (out_size % in_size) + print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}") + self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels, + out_channels=in_channels) + self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2, + attn_resolutions=[], in_channels=None, ch=in_channels, + ch_mult=[ch_mult for _ in range(num_blocks)]) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Resize(nn.Module): + def __init__(self, in_channels=None, learned=False, mode="bilinear"): + super().__init__() + self.with_conv = learned + self.mode = mode + if self.with_conv: + print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode") + raise NotImplementedError() + assert in_channels is not None + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=4, + stride=2, + padding=1) + + def forward(self, x, scale_factor=1.0): + if scale_factor==1.0: + return x + else: + x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor) + return x + +class FirstStagePostProcessor(nn.Module): + + def __init__(self, ch_mult:list, in_channels, + pretrained_model:nn.Module=None, + reshape=False, + n_channels=None, + dropout=0., + pretrained_config=None): + super().__init__() + if pretrained_config is None: + assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None' + self.pretrained_model = pretrained_model + else: + assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None' + self.instantiate_pretrained(pretrained_config) + + self.do_reshape = reshape + + if n_channels is None: + n_channels = self.pretrained_model.encoder.ch + + self.proj_norm = Normalize(in_channels,num_groups=in_channels//2) + self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3, + stride=1,padding=1) + + blocks = [] + downs = [] + ch_in = n_channels + for m in ch_mult: + blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout)) + ch_in = m * n_channels + downs.append(Downsample(ch_in, with_conv=False)) + + self.model = nn.ModuleList(blocks) + self.downsampler = nn.ModuleList(downs) + + + def instantiate_pretrained(self, config): + model = instantiate_from_config(config) + self.pretrained_model = model.eval() + # self.pretrained_model.train = False + for param in self.pretrained_model.parameters(): + param.requires_grad = False + + + @torch.no_grad() + def encode_with_pretrained(self,x): + c = self.pretrained_model.encode(x) + if isinstance(c, DiagonalGaussianDistribution): + c = c.mode() + return c + + def forward(self,x): + z_fs = self.encode_with_pretrained(x) + z = self.proj_norm(z_fs) + z = self.proj(z) + z = nonlinearity(z) + + for submodel, downmodel in zip(self.model,self.downsampler): + z = submodel(z,temb=None) + z = downmodel(z) + + if self.do_reshape: + z = rearrange(z,'b c h w -> b (h w) c') + return z + diff --git a/lvdm/modules/networks/openaimodel3d.py b/lvdm/modules/networks/openaimodel3d.py new file mode 100644 index 0000000000000000000000000000000000000000..49245da8ff896d938cf13c6cf6cb23548383c6dc --- /dev/null +++ b/lvdm/modules/networks/openaimodel3d.py @@ -0,0 +1,603 @@ +from functools import partial +from abc import abstractmethod +import torch +import torch.nn as nn +from einops import rearrange +import torch.nn.functional as F +from lvdm.models.utils_diffusion import timestep_embedding +from lvdm.common import checkpoint +from lvdm.basics import ( + zero_module, + conv_nd, + linear, + avg_pool_nd, + normalization +) +from lvdm.modules.attention import SpatialTransformer, TemporalTransformer + + +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + @abstractmethod + def forward(self, x, emb): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + """ + A sequential module that passes timestep embeddings to the children that + support it as an extra input. + """ + + def forward(self, x, emb, context=None, batch_size=None): + for layer in self: + if isinstance(layer, TimestepBlock): + x = layer(x, emb, batch_size=batch_size) + elif isinstance(layer, SpatialTransformer): + x = layer(x, context) + elif isinstance(layer, TemporalTransformer): + x = rearrange(x, '(b f) c h w -> b c f h w', b=batch_size) + x = layer(x, context) + x = rearrange(x, 'b c f h w -> (b f) c h w') + else: + x = layer(x) + return x + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd( + dims, self.channels, self.out_channels, 3, stride=stride, padding=padding + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + if use_conv: + self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding) + + def forward(self, x): + assert x.shape[1] == self.channels + if self.dims == 3: + x = F.interpolate(x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode='nearest') + else: + x = F.interpolate(x, scale_factor=2, mode='nearest') + if self.use_conv: + x = self.conv(x) + return x + + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + :param use_temporal_conv: if True, use the temporal convolution. + :param use_image_dataset: if True, the temporal parameters will not be optimized. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + use_conv=False, + up=False, + down=False, + use_temporal_conv=False, + tempspatial_aware=False + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + self.use_temporal_conv = use_temporal_conv + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + nn.Linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module(nn.Conv2d(self.out_channels, self.out_channels, 3, padding=1)), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 3, padding=1) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + if self.use_temporal_conv: + self.temopral_conv = TemporalConvBlock( + self.out_channels, + self.out_channels, + dropout=0.1, + spatial_aware=tempspatial_aware + ) + + def forward(self, x, emb, batch_size=None): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + input_tuple = (x, emb) + if batch_size: + forward_batchsize = partial(self._forward, batch_size=batch_size) + return checkpoint(forward_batchsize, input_tuple, self.parameters(), self.use_checkpoint) + return checkpoint(self._forward, input_tuple, self.parameters(), self.use_checkpoint) + + def _forward(self, x, emb, batch_size=None): + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = torch.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + h = self.skip_connection(x) + h + + if self.use_temporal_conv and batch_size: + h = rearrange(h, '(b t) c h w -> b c t h w', b=batch_size) + h = self.temopral_conv(h) + h = rearrange(h, 'b c t h w -> (b t) c h w') + return h + + +class TemporalConvBlock(nn.Module): + """ + Adapted from modelscope: https://github.com/modelscope/modelscope/blob/master/modelscope/models/multi_modal/video_synthesis/unet_sd.py + """ + def __init__(self, in_channels, out_channels=None, dropout=0.0, spatial_aware=False): + super(TemporalConvBlock, self).__init__() + if out_channels is None: + out_channels = in_channels + self.in_channels = in_channels + self.out_channels = out_channels + th_kernel_shape = (3, 1, 1) if not spatial_aware else (3, 3, 1) + th_padding_shape = (1, 0, 0) if not spatial_aware else (1, 1, 0) + tw_kernel_shape = (3, 1, 1) if not spatial_aware else (3, 1, 3) + tw_padding_shape = (1, 0, 0) if not spatial_aware else (1, 0, 1) + + # conv layers + self.conv1 = nn.Sequential( + nn.GroupNorm(32, in_channels), nn.SiLU(), + nn.Conv3d(in_channels, out_channels, th_kernel_shape, padding=th_padding_shape)) + self.conv2 = nn.Sequential( + nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout), + nn.Conv3d(out_channels, in_channels, tw_kernel_shape, padding=tw_padding_shape)) + self.conv3 = nn.Sequential( + nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout), + nn.Conv3d(out_channels, in_channels, th_kernel_shape, padding=th_padding_shape)) + self.conv4 = nn.Sequential( + nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout), + nn.Conv3d(out_channels, in_channels, tw_kernel_shape, padding=tw_padding_shape)) + + # zero out the last layer params,so the conv block is identity + nn.init.zeros_(self.conv4[-1].weight) + nn.init.zeros_(self.conv4[-1].bias) + + def forward(self, x): + identity = x + x = self.conv1(x) + x = self.conv2(x) + x = self.conv3(x) + x = self.conv4(x) + + return identity + x + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + :param in_channels: in_channels in the input Tensor. + :param model_channels: base channel count for the model. + :param out_channels: channels in the output Tensor. + :param num_res_blocks: number of residual blocks per downsample. + :param attention_resolutions: a collection of downsample rates at which + attention will take place. May be a set, list, or tuple. + For example, if this contains 4, then at 4x downsampling, attention + will be used. + :param dropout: the dropout probability. + :param channel_mult: channel multiplier for each level of the UNet. + :param conv_resample: if True, use learned convolutions for upsampling and + downsampling. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param num_classes: if specified (as an int), then this model will be + class-conditional with `num_classes` classes. + :param use_checkpoint: use gradient checkpointing to reduce memory usage. + :param num_heads: the number of attention heads in each attention layer. + :param num_heads_channels: if specified, ignore num_heads and instead use + a fixed channel width per attention head. + :param num_heads_upsample: works with num_heads to set a different number + of heads for upsampling. Deprecated. + :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. + :param resblock_updown: use residual blocks for up/downsampling. + :param use_new_attention_order: use a different attention pattern for potentially + increased efficiency. + """ + + def __init__(self, + in_channels, + model_channels, + out_channels, + num_res_blocks, + attention_resolutions, + dropout=0.0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + context_dim=None, + use_scale_shift_norm=False, + resblock_updown=False, + num_heads=-1, + num_head_channels=-1, + transformer_depth=1, + use_linear=False, + use_checkpoint=False, + temporal_conv=False, + tempspatial_aware=False, + temporal_attention=True, + use_relative_position=True, + use_causal_attention=False, + temporal_length=None, + use_fp16=False, + addition_attention=False, + temporal_selfatt_only=True, + image_cross_attention=False, + image_cross_attention_scale_learnable=False, + default_fs=4, + fs_condition=False, + ): + super(UNetModel, self).__init__() + if num_heads == -1: + assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' + if num_head_channels == -1: + assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' + + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + self.num_res_blocks = num_res_blocks + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.temporal_attention = temporal_attention + time_embed_dim = model_channels * 4 + self.use_checkpoint = use_checkpoint + self.dtype = torch.float16 if use_fp16 else torch.float32 + temporal_self_att_only = True + self.addition_attention = addition_attention + self.temporal_length = temporal_length + self.image_cross_attention = image_cross_attention + self.image_cross_attention_scale_learnable = image_cross_attention_scale_learnable + self.default_fs = default_fs + self.fs_condition = fs_condition + + ## Time embedding blocks + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + if fs_condition: + self.fps_embedding = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + nn.init.zeros_(self.fps_embedding[-1].weight) + nn.init.zeros_(self.fps_embedding[-1].bias) + ## Input Block + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1)) + ] + ) + if self.addition_attention: + self.init_attn=TimestepEmbedSequential( + TemporalTransformer( + model_channels, + n_heads=8, + d_head=num_head_channels, + depth=transformer_depth, + context_dim=context_dim, + use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, + causal_attention=False, relative_position=use_relative_position, + temporal_length=temporal_length)) + + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for _ in range(num_res_blocks): + layers = [ + ResBlock(ch, time_embed_dim, dropout, + out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware, + use_temporal_conv=temporal_conv + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + layers.append( + SpatialTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, disable_self_attn=False, + video_length=temporal_length, image_cross_attention=self.image_cross_attention, + image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable, + ) + ) + if self.temporal_attention: + layers.append( + TemporalTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, + causal_attention=use_causal_attention, relative_position=use_relative_position, + temporal_length=temporal_length + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock(ch, time_embed_dim, dropout, + out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True + ) + if resblock_updown + else Downsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + layers = [ + ResBlock(ch, time_embed_dim, dropout, + dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware, + use_temporal_conv=temporal_conv + ), + SpatialTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, disable_self_attn=False, video_length=temporal_length, + image_cross_attention=self.image_cross_attention,image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable + ) + ] + if self.temporal_attention: + layers.append( + TemporalTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, + causal_attention=use_causal_attention, relative_position=use_relative_position, + temporal_length=temporal_length + ) + ) + layers.append( + ResBlock(ch, time_embed_dim, dropout, + dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware, + use_temporal_conv=temporal_conv + ) + ) + + ## Middle Block + self.middle_block = TimestepEmbedSequential(*layers) + + ## Output Block + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(num_res_blocks + 1): + ich = input_block_chans.pop() + layers = [ + ResBlock(ch + ich, time_embed_dim, dropout, + out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware, + use_temporal_conv=temporal_conv + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + layers.append( + SpatialTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, disable_self_attn=False, video_length=temporal_length, + image_cross_attention=self.image_cross_attention,image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable + ) + ) + if self.temporal_attention: + layers.append( + TemporalTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, + causal_attention=use_causal_attention, relative_position=use_relative_position, + temporal_length=temporal_length + ) + ) + if level and i == num_res_blocks: + out_ch = ch + layers.append( + ResBlock(ch, time_embed_dim, dropout, + out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + + def forward(self, x, timesteps, context=None, features_adapter=None, fs=None, **kwargs): + b,_,t,_,_ = x.shape + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).type(x.dtype) + emb = self.time_embed(t_emb) + + ## repeat t times for context [(b t) 77 768] & time embedding + ## check if we use per-frame image conditioning + _, l_context, _ = context.shape + if l_context == 77 + t*16: ## !!! HARD CODE here + context_text, context_img = context[:,:77,:], context[:,77:,:] + context_text = context_text.repeat_interleave(repeats=t, dim=0) + context_img = rearrange(context_img, 'b (t l) c -> (b t) l c', t=t) + context = torch.cat([context_text, context_img], dim=1) + else: + context = context.repeat_interleave(repeats=t, dim=0) + emb = emb.repeat_interleave(repeats=t, dim=0) + + ## always in shape (b t) c h w, except for temporal layer + x = rearrange(x, 'b c t h w -> (b t) c h w') + + ## combine emb + if self.fs_condition: + if fs is None: + fs = torch.tensor( + [self.default_fs] * b, dtype=torch.long, device=x.device) + fs_emb = timestep_embedding(fs, self.model_channels, repeat_only=False).type(x.dtype) + + fs_embed = self.fps_embedding(fs_emb) + fs_embed = fs_embed.repeat_interleave(repeats=t, dim=0) + emb = emb + fs_embed + + h = x.type(self.dtype) + adapter_idx = 0 + hs = [] + for id, module in enumerate(self.input_blocks): + h = module(h, emb, context=context, batch_size=b) + if id ==0 and self.addition_attention: + h = self.init_attn(h, emb, context=context, batch_size=b) + ## plug-in adapter features + if ((id+1)%3 == 0) and features_adapter is not None: + h = h + features_adapter[adapter_idx] + adapter_idx += 1 + hs.append(h) + if features_adapter is not None: + assert len(features_adapter)==adapter_idx, 'Wrong features_adapter' + + h = self.middle_block(h, emb, context=context, batch_size=b) + for module in self.output_blocks: + h = torch.cat([h, hs.pop()], dim=1) + h = module(h, emb, context=context, batch_size=b) + h = h.type(x.dtype) + y = self.out(h) + + # reshape back to (b c t h w) + y = rearrange(y, '(b t) c h w -> b c t h w', b=b) + return y \ No newline at end of file diff --git a/lvdm/modules/x_transformer.py b/lvdm/modules/x_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..5321012f860a8fb06850c1ddf495db934addecae --- /dev/null +++ b/lvdm/modules/x_transformer.py @@ -0,0 +1,639 @@ +"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers""" +from functools import partial +from inspect import isfunction +from collections import namedtuple +from einops import rearrange, repeat +import torch +from torch import nn, einsum +import torch.nn.functional as F + +# constants +DEFAULT_DIM_HEAD = 64 + +Intermediates = namedtuple('Intermediates', [ + 'pre_softmax_attn', + 'post_softmax_attn' +]) + +LayerIntermediates = namedtuple('Intermediates', [ + 'hiddens', + 'attn_intermediates' +]) + + +class AbsolutePositionalEmbedding(nn.Module): + def __init__(self, dim, max_seq_len): + super().__init__() + self.emb = nn.Embedding(max_seq_len, dim) + self.init_() + + def init_(self): + nn.init.normal_(self.emb.weight, std=0.02) + + def forward(self, x): + n = torch.arange(x.shape[1], device=x.device) + return self.emb(n)[None, :, :] + + +class FixedPositionalEmbedding(nn.Module): + def __init__(self, dim): + super().__init__() + inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim)) + self.register_buffer('inv_freq', inv_freq) + + def forward(self, x, seq_dim=1, offset=0): + t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset + sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq) + emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1) + return emb[None, :, :] + + +# helpers + +def exists(val): + return val is not None + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def always(val): + def inner(*args, **kwargs): + return val + return inner + + +def not_equals(val): + def inner(x): + return x != val + return inner + + +def equals(val): + def inner(x): + return x == val + return inner + + +def max_neg_value(tensor): + return -torch.finfo(tensor.dtype).max + + +# keyword argument helpers + +def pick_and_pop(keys, d): + values = list(map(lambda key: d.pop(key), keys)) + return dict(zip(keys, values)) + + +def group_dict_by_key(cond, d): + return_val = [dict(), dict()] + for key in d.keys(): + match = bool(cond(key)) + ind = int(not match) + return_val[ind][key] = d[key] + return (*return_val,) + + +def string_begins_with(prefix, str): + return str.startswith(prefix) + + +def group_by_key_prefix(prefix, d): + return group_dict_by_key(partial(string_begins_with, prefix), d) + + +def groupby_prefix_and_trim(prefix, d): + kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d) + kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items()))) + return kwargs_without_prefix, kwargs + + +# classes +class Scale(nn.Module): + def __init__(self, value, fn): + super().__init__() + self.value = value + self.fn = fn + + def forward(self, x, **kwargs): + x, *rest = self.fn(x, **kwargs) + return (x * self.value, *rest) + + +class Rezero(nn.Module): + def __init__(self, fn): + super().__init__() + self.fn = fn + self.g = nn.Parameter(torch.zeros(1)) + + def forward(self, x, **kwargs): + x, *rest = self.fn(x, **kwargs) + return (x * self.g, *rest) + + +class ScaleNorm(nn.Module): + def __init__(self, dim, eps=1e-5): + super().__init__() + self.scale = dim ** -0.5 + self.eps = eps + self.g = nn.Parameter(torch.ones(1)) + + def forward(self, x): + norm = torch.norm(x, dim=-1, keepdim=True) * self.scale + return x / norm.clamp(min=self.eps) * self.g + + +class RMSNorm(nn.Module): + def __init__(self, dim, eps=1e-8): + super().__init__() + self.scale = dim ** -0.5 + self.eps = eps + self.g = nn.Parameter(torch.ones(dim)) + + def forward(self, x): + norm = torch.norm(x, dim=-1, keepdim=True) * self.scale + return x / norm.clamp(min=self.eps) * self.g + + +class Residual(nn.Module): + def forward(self, x, residual): + return x + residual + + +class GRUGating(nn.Module): + def __init__(self, dim): + super().__init__() + self.gru = nn.GRUCell(dim, dim) + + def forward(self, x, residual): + gated_output = self.gru( + rearrange(x, 'b n d -> (b n) d'), + rearrange(residual, 'b n d -> (b n) d') + ) + + return gated_output.reshape_as(x) + + +# feedforward + +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + nn.Linear(dim, inner_dim), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +# attention. +class Attention(nn.Module): + def __init__( + self, + dim, + dim_head=DEFAULT_DIM_HEAD, + heads=8, + causal=False, + mask=None, + talking_heads=False, + sparse_topk=None, + use_entmax15=False, + num_mem_kv=0, + dropout=0., + on_attn=False + ): + super().__init__() + if use_entmax15: + raise NotImplementedError("Check out entmax activation instead of softmax activation!") + self.scale = dim_head ** -0.5 + self.heads = heads + self.causal = causal + self.mask = mask + + inner_dim = dim_head * heads + + self.to_q = nn.Linear(dim, inner_dim, bias=False) + self.to_k = nn.Linear(dim, inner_dim, bias=False) + self.to_v = nn.Linear(dim, inner_dim, bias=False) + self.dropout = nn.Dropout(dropout) + + # talking heads + self.talking_heads = talking_heads + if talking_heads: + self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads)) + self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads)) + + # explicit topk sparse attention + self.sparse_topk = sparse_topk + + # entmax + #self.attn_fn = entmax15 if use_entmax15 else F.softmax + self.attn_fn = F.softmax + + # add memory key / values + self.num_mem_kv = num_mem_kv + if num_mem_kv > 0: + self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head)) + self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head)) + + # attention on attention + self.attn_on_attn = on_attn + self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim) + + def forward( + self, + x, + context=None, + mask=None, + context_mask=None, + rel_pos=None, + sinusoidal_emb=None, + prev_attn=None, + mem=None + ): + b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device + kv_input = default(context, x) + + q_input = x + k_input = kv_input + v_input = kv_input + + if exists(mem): + k_input = torch.cat((mem, k_input), dim=-2) + v_input = torch.cat((mem, v_input), dim=-2) + + if exists(sinusoidal_emb): + # in shortformer, the query would start at a position offset depending on the past cached memory + offset = k_input.shape[-2] - q_input.shape[-2] + q_input = q_input + sinusoidal_emb(q_input, offset=offset) + k_input = k_input + sinusoidal_emb(k_input) + + q = self.to_q(q_input) + k = self.to_k(k_input) + v = self.to_v(v_input) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v)) + + input_mask = None + if any(map(exists, (mask, context_mask))): + q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool()) + k_mask = q_mask if not exists(context) else context_mask + k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool()) + q_mask = rearrange(q_mask, 'b i -> b () i ()') + k_mask = rearrange(k_mask, 'b j -> b () () j') + input_mask = q_mask * k_mask + + if self.num_mem_kv > 0: + mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v)) + k = torch.cat((mem_k, k), dim=-2) + v = torch.cat((mem_v, v), dim=-2) + if exists(input_mask): + input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True) + + dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale + mask_value = max_neg_value(dots) + + if exists(prev_attn): + dots = dots + prev_attn + + pre_softmax_attn = dots + + if talking_heads: + dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous() + + if exists(rel_pos): + dots = rel_pos(dots) + + if exists(input_mask): + dots.masked_fill_(~input_mask, mask_value) + del input_mask + + if self.causal: + i, j = dots.shape[-2:] + r = torch.arange(i, device=device) + mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j') + mask = F.pad(mask, (j - i, 0), value=False) + dots.masked_fill_(mask, mask_value) + del mask + + if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]: + top, _ = dots.topk(self.sparse_topk, dim=-1) + vk = top[..., -1].unsqueeze(-1).expand_as(dots) + mask = dots < vk + dots.masked_fill_(mask, mask_value) + del mask + + attn = self.attn_fn(dots, dim=-1) + post_softmax_attn = attn + + attn = self.dropout(attn) + + if talking_heads: + attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous() + + out = einsum('b h i j, b h j d -> b h i d', attn, v) + out = rearrange(out, 'b h n d -> b n (h d)') + + intermediates = Intermediates( + pre_softmax_attn=pre_softmax_attn, + post_softmax_attn=post_softmax_attn + ) + + return self.to_out(out), intermediates + + +class AttentionLayers(nn.Module): + def __init__( + self, + dim, + depth, + heads=8, + causal=False, + cross_attend=False, + only_cross=False, + use_scalenorm=False, + use_rmsnorm=False, + use_rezero=False, + rel_pos_num_buckets=32, + rel_pos_max_distance=128, + position_infused_attn=False, + custom_layers=None, + sandwich_coef=None, + par_ratio=None, + residual_attn=False, + cross_residual_attn=False, + macaron=False, + pre_norm=True, + gate_residual=False, + **kwargs + ): + super().__init__() + ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs) + attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs) + + dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD) + + self.dim = dim + self.depth = depth + self.layers = nn.ModuleList([]) + + self.has_pos_emb = position_infused_attn + self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None + self.rotary_pos_emb = always(None) + + assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance' + self.rel_pos = None + + self.pre_norm = pre_norm + + self.residual_attn = residual_attn + self.cross_residual_attn = cross_residual_attn + + norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm + norm_class = RMSNorm if use_rmsnorm else norm_class + norm_fn = partial(norm_class, dim) + + norm_fn = nn.Identity if use_rezero else norm_fn + branch_fn = Rezero if use_rezero else None + + if cross_attend and not only_cross: + default_block = ('a', 'c', 'f') + elif cross_attend and only_cross: + default_block = ('c', 'f') + else: + default_block = ('a', 'f') + + if macaron: + default_block = ('f',) + default_block + + if exists(custom_layers): + layer_types = custom_layers + elif exists(par_ratio): + par_depth = depth * len(default_block) + assert 1 < par_ratio <= par_depth, 'par ratio out of range' + default_block = tuple(filter(not_equals('f'), default_block)) + par_attn = par_depth // par_ratio + depth_cut = par_depth * 2 // 3 # 2 / 3 attention layer cutoff suggested by PAR paper + par_width = (depth_cut + depth_cut // par_attn) // par_attn + assert len(default_block) <= par_width, 'default block is too large for par_ratio' + par_block = default_block + ('f',) * (par_width - len(default_block)) + par_head = par_block * par_attn + layer_types = par_head + ('f',) * (par_depth - len(par_head)) + elif exists(sandwich_coef): + assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth' + layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef + else: + layer_types = default_block * depth + + self.layer_types = layer_types + self.num_attn_layers = len(list(filter(equals('a'), layer_types))) + + for layer_type in self.layer_types: + if layer_type == 'a': + layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs) + elif layer_type == 'c': + layer = Attention(dim, heads=heads, **attn_kwargs) + elif layer_type == 'f': + layer = FeedForward(dim, **ff_kwargs) + layer = layer if not macaron else Scale(0.5, layer) + else: + raise Exception(f'invalid layer type {layer_type}') + + if isinstance(layer, Attention) and exists(branch_fn): + layer = branch_fn(layer) + + if gate_residual: + residual_fn = GRUGating(dim) + else: + residual_fn = Residual() + + self.layers.append(nn.ModuleList([ + norm_fn(), + layer, + residual_fn + ])) + + def forward( + self, + x, + context=None, + mask=None, + context_mask=None, + mems=None, + return_hiddens=False + ): + hiddens = [] + intermediates = [] + prev_attn = None + prev_cross_attn = None + + mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers + + for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)): + is_last = ind == (len(self.layers) - 1) + + if layer_type == 'a': + hiddens.append(x) + layer_mem = mems.pop(0) + + residual = x + + if self.pre_norm: + x = norm(x) + + if layer_type == 'a': + out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos, + prev_attn=prev_attn, mem=layer_mem) + elif layer_type == 'c': + out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn) + elif layer_type == 'f': + out = block(x) + + x = residual_fn(out, residual) + + if layer_type in ('a', 'c'): + intermediates.append(inter) + + if layer_type == 'a' and self.residual_attn: + prev_attn = inter.pre_softmax_attn + elif layer_type == 'c' and self.cross_residual_attn: + prev_cross_attn = inter.pre_softmax_attn + + if not self.pre_norm and not is_last: + x = norm(x) + + if return_hiddens: + intermediates = LayerIntermediates( + hiddens=hiddens, + attn_intermediates=intermediates + ) + + return x, intermediates + + return x + + +class Encoder(AttentionLayers): + def __init__(self, **kwargs): + assert 'causal' not in kwargs, 'cannot set causality on encoder' + super().__init__(causal=False, **kwargs) + + + +class TransformerWrapper(nn.Module): + def __init__( + self, + *, + num_tokens, + max_seq_len, + attn_layers, + emb_dim=None, + max_mem_len=0., + emb_dropout=0., + num_memory_tokens=None, + tie_embedding=False, + use_pos_emb=True + ): + super().__init__() + assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder' + + dim = attn_layers.dim + emb_dim = default(emb_dim, dim) + + self.max_seq_len = max_seq_len + self.max_mem_len = max_mem_len + self.num_tokens = num_tokens + + self.token_emb = nn.Embedding(num_tokens, emb_dim) + self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if ( + use_pos_emb and not attn_layers.has_pos_emb) else always(0) + self.emb_dropout = nn.Dropout(emb_dropout) + + self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity() + self.attn_layers = attn_layers + self.norm = nn.LayerNorm(dim) + + self.init_() + + self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t() + + # memory tokens (like [cls]) from Memory Transformers paper + num_memory_tokens = default(num_memory_tokens, 0) + self.num_memory_tokens = num_memory_tokens + if num_memory_tokens > 0: + self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim)) + + # let funnel encoder know number of memory tokens, if specified + if hasattr(attn_layers, 'num_memory_tokens'): + attn_layers.num_memory_tokens = num_memory_tokens + + def init_(self): + nn.init.normal_(self.token_emb.weight, std=0.02) + + def forward( + self, + x, + return_embeddings=False, + mask=None, + return_mems=False, + return_attn=False, + mems=None, + **kwargs + ): + b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens + x = self.token_emb(x) + x += self.pos_emb(x) + x = self.emb_dropout(x) + + x = self.project_emb(x) + + if num_mem > 0: + mem = repeat(self.memory_tokens, 'n d -> b n d', b=b) + x = torch.cat((mem, x), dim=1) + + # auto-handle masking after appending memory tokens + if exists(mask): + mask = F.pad(mask, (num_mem, 0), value=True) + + x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs) + x = self.norm(x) + + mem, x = x[:, :num_mem], x[:, num_mem:] + + out = self.to_logits(x) if not return_embeddings else x + + if return_mems: + hiddens = intermediates.hiddens + new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens + new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems)) + return out, new_mems + + if return_attn: + attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates)) + return out, attn_maps + + return out \ No newline at end of file diff --git a/main/callbacks.py b/main/callbacks.py new file mode 100644 index 0000000000000000000000000000000000000000..b03c583acf4f5b8caba85e8272b31ea1b77d80c2 --- /dev/null +++ b/main/callbacks.py @@ -0,0 +1,133 @@ +import os +import time +import logging +mainlogger = logging.getLogger('mainlogger') + +import torch +import torchvision +import pytorch_lightning as pl +from pytorch_lightning.callbacks import Callback +from pytorch_lightning.utilities import rank_zero_only +from pytorch_lightning.utilities import rank_zero_info +from utils.save_video import log_local, prepare_to_log + + +class ImageLogger(Callback): + def __init__(self, batch_frequency, max_images=8, clamp=True, rescale=True, save_dir=None, \ + to_local=False, log_images_kwargs=None): + super().__init__() + self.rescale = rescale + self.batch_freq = batch_frequency + self.max_images = max_images + self.to_local = to_local + self.clamp = clamp + self.log_images_kwargs = log_images_kwargs if log_images_kwargs else {} + if self.to_local: + ## default save dir + self.save_dir = os.path.join(save_dir, "images") + os.makedirs(os.path.join(self.save_dir, "train"), exist_ok=True) + os.makedirs(os.path.join(self.save_dir, "val"), exist_ok=True) + + def log_to_tensorboard(self, pl_module, batch_logs, filename, split, save_fps=8): + """ log images and videos to tensorboard """ + global_step = pl_module.global_step + for key in batch_logs: + value = batch_logs[key] + tag = "gs%d-%s/%s-%s"%(global_step, split, filename, key) + if isinstance(value, list) and isinstance(value[0], str): + captions = ' |------| '.join(value) + pl_module.logger.experiment.add_text(tag, captions, global_step=global_step) + elif isinstance(value, torch.Tensor) and value.dim() == 5: + video = value + n = video.shape[0] + video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w + frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n), padding=0) for framesheet in video] #[3, n*h, 1*w] + grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w] + grid = (grid + 1.0) / 2.0 + grid = grid.unsqueeze(dim=0) + pl_module.logger.experiment.add_video(tag, grid, fps=save_fps, global_step=global_step) + elif isinstance(value, torch.Tensor) and value.dim() == 4: + img = value + grid = torchvision.utils.make_grid(img, nrow=int(n), padding=0) + grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w + pl_module.logger.experiment.add_image(tag, grid, global_step=global_step) + else: + pass + + @rank_zero_only + def log_batch_imgs(self, pl_module, batch, batch_idx, split="train"): + """ generate images, then save and log to tensorboard """ + skip_freq = self.batch_freq if split == "train" else 5 + if (batch_idx+1) % skip_freq == 0: + is_train = pl_module.training + if is_train: + pl_module.eval() + torch.cuda.empty_cache() + with torch.no_grad(): + log_func = pl_module.log_images + batch_logs = log_func(batch, split=split, **self.log_images_kwargs) + + ## process: move to CPU and clamp + batch_logs = prepare_to_log(batch_logs, self.max_images, self.clamp) + torch.cuda.empty_cache() + + filename = "ep{}_idx{}_rank{}".format( + pl_module.current_epoch, + batch_idx, + pl_module.global_rank) + if self.to_local: + mainlogger.info("Log [%s] batch <%s> to local ..."%(split, filename)) + filename = "gs{}_".format(pl_module.global_step) + filename + log_local(batch_logs, os.path.join(self.save_dir, split), filename, save_fps=10) + else: + mainlogger.info("Log [%s] batch <%s> to tensorboard ..."%(split, filename)) + self.log_to_tensorboard(pl_module, batch_logs, filename, split, save_fps=10) + mainlogger.info('Finish!') + + if is_train: + pl_module.train() + + def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=None): + if self.batch_freq != -1 and pl_module.logdir: + self.log_batch_imgs(pl_module, batch, batch_idx, split="train") + + def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=None): + ## different with validation_step() that saving the whole validation set and only keep the latest, + ## it records the performance of every validation (without overwritten) by only keep a subset + if self.batch_freq != -1 and pl_module.logdir: + self.log_batch_imgs(pl_module, batch, batch_idx, split="val") + if hasattr(pl_module, 'calibrate_grad_norm'): + if (pl_module.calibrate_grad_norm and batch_idx % 25 == 0) and batch_idx > 0: + self.log_gradients(trainer, pl_module, batch_idx=batch_idx) + + +class CUDACallback(Callback): + # see https://github.com/SeanNaren/minGPT/blob/master/mingpt/callback.py + def on_train_epoch_start(self, trainer, pl_module): + # Reset the memory use counter + # lightning update + if int((pl.__version__).split('.')[1])>=7: + gpu_index = trainer.strategy.root_device.index + else: + gpu_index = trainer.root_gpu + torch.cuda.reset_peak_memory_stats(gpu_index) + torch.cuda.synchronize(gpu_index) + self.start_time = time.time() + + def on_train_epoch_end(self, trainer, pl_module): + if int((pl.__version__).split('.')[1])>=7: + gpu_index = trainer.strategy.root_device.index + else: + gpu_index = trainer.root_gpu + torch.cuda.synchronize(gpu_index) + max_memory = torch.cuda.max_memory_allocated(gpu_index) / 2 ** 20 + epoch_time = time.time() - self.start_time + + try: + max_memory = trainer.training_type_plugin.reduce(max_memory) + epoch_time = trainer.training_type_plugin.reduce(epoch_time) + + rank_zero_info(f"Average Epoch time: {epoch_time:.2f} seconds") + rank_zero_info(f"Average Peak memory {max_memory:.2f}MiB") + except AttributeError: + pass diff --git a/main/trainer.py b/main/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..8082563975292d528b5dd1ab879310a33899b7d1 --- /dev/null +++ b/main/trainer.py @@ -0,0 +1,168 @@ +import argparse, os, sys, datetime +from omegaconf import OmegaConf +from transformers import logging as transf_logging +import pytorch_lightning as pl +from pytorch_lightning import seed_everything +from pytorch_lightning.trainer import Trainer +import torch +sys.path.insert(1, os.path.join(sys.path[0], '..')) +from utils.utils import instantiate_from_config +from utils_train import get_trainer_callbacks, get_trainer_logger, get_trainer_strategy +from utils_train import set_logger, init_workspace, load_checkpoints + + +def get_parser(**parser_kwargs): + parser = argparse.ArgumentParser(**parser_kwargs) + parser.add_argument("--seed", "-s", type=int, default=20230211, help="seed for seed_everything") + parser.add_argument("--name", "-n", type=str, default="", help="experiment name, as saving folder") + + parser.add_argument("--base", "-b", nargs="*", metavar="base_config.yaml", help="paths to base configs. Loaded from left-to-right. " + "Parameters can be overwritten or added with command-line options of the form `--key value`.", default=list()) + + parser.add_argument("--train", "-t", action='store_true', default=False, help='train') + parser.add_argument("--val", "-v", action='store_true', default=False, help='val') + parser.add_argument("--test", action='store_true', default=False, help='test') + + parser.add_argument("--logdir", "-l", type=str, default="logs", help="directory for logging dat shit") + parser.add_argument("--auto_resume", action='store_true', default=False, help="resume from full-info checkpoint") + parser.add_argument("--auto_resume_weight_only", action='store_true', default=False, help="resume from weight-only checkpoint") + parser.add_argument("--debug", "-d", action='store_true', default=False, help="enable post-mortem debugging") + + return parser + +def get_nondefault_trainer_args(args): + parser = argparse.ArgumentParser() + parser = Trainer.add_argparse_args(parser) + default_trainer_args = parser.parse_args([]) + return sorted(k for k in vars(default_trainer_args) if getattr(args, k) != getattr(default_trainer_args, k)) + + +if __name__ == "__main__": + now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S") + local_rank = int(os.environ.get('LOCAL_RANK')) + global_rank = int(os.environ.get('RANK')) + num_rank = int(os.environ.get('WORLD_SIZE')) + + parser = get_parser() + ## Extends existing argparse by default Trainer attributes + parser = Trainer.add_argparse_args(parser) + args, unknown = parser.parse_known_args() + ## disable transformer warning + transf_logging.set_verbosity_error() + seed_everything(args.seed) + + ## yaml configs: "model" | "data" | "lightning" + configs = [OmegaConf.load(cfg) for cfg in args.base] + cli = OmegaConf.from_dotlist(unknown) + config = OmegaConf.merge(*configs, cli) + lightning_config = config.pop("lightning", OmegaConf.create()) + trainer_config = lightning_config.get("trainer", OmegaConf.create()) + + ## setup workspace directories + workdir, ckptdir, cfgdir, loginfo = init_workspace(args.name, args.logdir, config, lightning_config, global_rank) + logger = set_logger(logfile=os.path.join(loginfo, 'log_%d:%s.txt'%(global_rank, now))) + logger.info("@lightning version: %s [>=1.8 required]"%(pl.__version__)) + + ## MODEL CONFIG >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> + logger.info("***** Configing Model *****") + config.model.params.logdir = workdir + model = instantiate_from_config(config.model) + + ## load checkpoints + model = load_checkpoints(model, config.model) + + ## register_schedule again to make ZTSNR work + if model.rescale_betas_zero_snr: + model.register_schedule(given_betas=model.given_betas, beta_schedule=model.beta_schedule, timesteps=model.timesteps, + linear_start=model.linear_start, linear_end=model.linear_end, cosine_s=model.cosine_s) + + ## update trainer config + for k in get_nondefault_trainer_args(args): + trainer_config[k] = getattr(args, k) + + num_nodes = trainer_config.num_nodes + ngpu_per_node = trainer_config.devices + logger.info(f"Running on {num_rank}={num_nodes}x{ngpu_per_node} GPUs") + + ## setup learning rate + base_lr = config.model.base_learning_rate + bs = config.data.params.batch_size + if getattr(config.model, 'scale_lr', True): + model.learning_rate = num_rank * bs * base_lr + else: + model.learning_rate = base_lr + + + ## DATA CONFIG >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> + logger.info("***** Configing Data *****") + data = instantiate_from_config(config.data) + data.setup() + for k in data.datasets: + logger.info(f"{k}, {data.datasets[k].__class__.__name__}, {len(data.datasets[k])}") + + + ## TRAINER CONFIG >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> + logger.info("***** Configing Trainer *****") + if "accelerator" not in trainer_config: + trainer_config["accelerator"] = "gpu" + + ## setup trainer args: pl-logger and callbacks + trainer_kwargs = dict() + trainer_kwargs["num_sanity_val_steps"] = 0 + logger_cfg = get_trainer_logger(lightning_config, workdir, args.debug) + trainer_kwargs["logger"] = instantiate_from_config(logger_cfg) + + ## setup callbacks + callbacks_cfg = get_trainer_callbacks(lightning_config, config, workdir, ckptdir, logger) + trainer_kwargs["callbacks"] = [instantiate_from_config(callbacks_cfg[k]) for k in callbacks_cfg] + strategy_cfg = get_trainer_strategy(lightning_config) + trainer_kwargs["strategy"] = strategy_cfg if type(strategy_cfg) == str else instantiate_from_config(strategy_cfg) + trainer_kwargs['precision'] = lightning_config.get('precision', 32) + trainer_kwargs["sync_batchnorm"] = False + + ## trainer config: others + + trainer_args = argparse.Namespace(**trainer_config) + trainer = Trainer.from_argparse_args(trainer_args, **trainer_kwargs) + + ## allow checkpointing via USR1 + def melk(*args, **kwargs): + ## run all checkpoint hooks + if trainer.global_rank == 0: + print("Summoning checkpoint.") + ckpt_path = os.path.join(ckptdir, "last_summoning.ckpt") + trainer.save_checkpoint(ckpt_path) + + def divein(*args, **kwargs): + if trainer.global_rank == 0: + import pudb; + pudb.set_trace() + + import signal + signal.signal(signal.SIGUSR1, melk) + signal.signal(signal.SIGUSR2, divein) + + ## Running LOOP >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> + logger.info("***** Running the Loop *****") + if args.train: + try: + if "strategy" in lightning_config and lightning_config['strategy'].startswith('deepspeed'): + logger.info("") + ## deepspeed + if trainer_kwargs['precision'] == 16: + with torch.cuda.amp.autocast(): + trainer.fit(model, data) + else: + trainer.fit(model, data) + else: + logger.info("") ## this is default + ## ddpsharded + trainer.fit(model, data) + except Exception: + #melk() + raise + + # if args.val: + # trainer.validate(model, data) + # if args.test or not trainer.interrupted: + # trainer.test(model, data) \ No newline at end of file diff --git a/main/utils_data.py b/main/utils_data.py new file mode 100644 index 0000000000000000000000000000000000000000..412216a729612230c3492eafe2b5b51666fdc33c --- /dev/null +++ b/main/utils_data.py @@ -0,0 +1,136 @@ +from functools import partial +import numpy as np + +import torch +import pytorch_lightning as pl +from torch.utils.data import DataLoader, Dataset + +import os, sys +os.chdir(sys.path[0]) +sys.path.append("..") +from lvdm.data.base import Txt2ImgIterableBaseDataset +from utils.utils import instantiate_from_config + + +def worker_init_fn(_): + worker_info = torch.utils.data.get_worker_info() + + dataset = worker_info.dataset + worker_id = worker_info.id + + if isinstance(dataset, Txt2ImgIterableBaseDataset): + split_size = dataset.num_records // worker_info.num_workers + # reset num_records to the true number to retain reliable length information + dataset.sample_ids = dataset.valid_ids[worker_id * split_size:(worker_id + 1) * split_size] + current_id = np.random.choice(len(np.random.get_state()[1]), 1) + return np.random.seed(np.random.get_state()[1][current_id] + worker_id) + else: + return np.random.seed(np.random.get_state()[1][0] + worker_id) + + +class WrappedDataset(Dataset): + """Wraps an arbitrary object with __len__ and __getitem__ into a pytorch dataset""" + + def __init__(self, dataset): + self.data = dataset + + def __len__(self): + return len(self.data) + + def __getitem__(self, idx): + return self.data[idx] + + +class DataModuleFromConfig(pl.LightningDataModule): + def __init__(self, batch_size, train=None, validation=None, test=None, predict=None, + wrap=False, num_workers=None, shuffle_test_loader=False, use_worker_init_fn=False, + shuffle_val_dataloader=False, train_img=None, + test_max_n_samples=None): + super().__init__() + self.batch_size = batch_size + self.dataset_configs = dict() + self.num_workers = num_workers if num_workers is not None else batch_size * 2 + self.use_worker_init_fn = use_worker_init_fn + if train is not None: + self.dataset_configs["train"] = train + self.train_dataloader = self._train_dataloader + if validation is not None: + self.dataset_configs["validation"] = validation + self.val_dataloader = partial(self._val_dataloader, shuffle=shuffle_val_dataloader) + if test is not None: + self.dataset_configs["test"] = test + self.test_dataloader = partial(self._test_dataloader, shuffle=shuffle_test_loader) + if predict is not None: + self.dataset_configs["predict"] = predict + self.predict_dataloader = self._predict_dataloader + + self.img_loader = None + self.wrap = wrap + self.test_max_n_samples = test_max_n_samples + self.collate_fn = None + + def prepare_data(self): + pass + + def setup(self, stage=None): + self.datasets = dict((k, instantiate_from_config(self.dataset_configs[k])) for k in self.dataset_configs) + if self.wrap: + for k in self.datasets: + self.datasets[k] = WrappedDataset(self.datasets[k]) + + def _train_dataloader(self): + is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset) + if is_iterable_dataset or self.use_worker_init_fn: + init_fn = worker_init_fn + else: + init_fn = None + loader = DataLoader(self.datasets["train"], batch_size=self.batch_size, + num_workers=self.num_workers, shuffle=False if is_iterable_dataset else True, + worker_init_fn=init_fn, collate_fn=self.collate_fn, + ) + return loader + + def _val_dataloader(self, shuffle=False): + if isinstance(self.datasets['validation'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn: + init_fn = worker_init_fn + else: + init_fn = None + return DataLoader(self.datasets["validation"], + batch_size=self.batch_size, + num_workers=self.num_workers, + worker_init_fn=init_fn, + shuffle=shuffle, + collate_fn=self.collate_fn, + ) + + def _test_dataloader(self, shuffle=False): + try: + is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset) + except: + is_iterable_dataset = isinstance(self.datasets['test'], Txt2ImgIterableBaseDataset) + + if is_iterable_dataset or self.use_worker_init_fn: + init_fn = worker_init_fn + else: + init_fn = None + + # do not shuffle dataloader for iterable dataset + shuffle = shuffle and (not is_iterable_dataset) + if self.test_max_n_samples is not None: + dataset = torch.utils.data.Subset(self.datasets["test"], list(range(self.test_max_n_samples))) + else: + dataset = self.datasets["test"] + return DataLoader(dataset, batch_size=self.batch_size, + num_workers=self.num_workers, worker_init_fn=init_fn, shuffle=shuffle, + collate_fn=self.collate_fn, + ) + + def _predict_dataloader(self, shuffle=False): + if isinstance(self.datasets['predict'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn: + init_fn = worker_init_fn + else: + init_fn = None + return DataLoader(self.datasets["predict"], batch_size=self.batch_size, + num_workers=self.num_workers, worker_init_fn=init_fn, + collate_fn=self.collate_fn, + ) diff --git a/main/utils_train.py b/main/utils_train.py new file mode 100644 index 0000000000000000000000000000000000000000..505d005de928921e68ea32b1e6fdc380d1660c06 --- /dev/null +++ b/main/utils_train.py @@ -0,0 +1,173 @@ +import os, re +from omegaconf import OmegaConf +import logging +mainlogger = logging.getLogger('mainlogger') + +import torch +from collections import OrderedDict + +def init_workspace(name, logdir, model_config, lightning_config, rank=0): + workdir = os.path.join(logdir, name) + ckptdir = os.path.join(workdir, "checkpoints") + cfgdir = os.path.join(workdir, "configs") + loginfo = os.path.join(workdir, "loginfo") + + # Create logdirs and save configs (all ranks will do to avoid missing directory error if rank:0 is slower) + os.makedirs(workdir, exist_ok=True) + os.makedirs(ckptdir, exist_ok=True) + os.makedirs(cfgdir, exist_ok=True) + os.makedirs(loginfo, exist_ok=True) + + if rank == 0: + if "callbacks" in lightning_config and 'metrics_over_trainsteps_checkpoint' in lightning_config.callbacks: + os.makedirs(os.path.join(ckptdir, 'trainstep_checkpoints'), exist_ok=True) + OmegaConf.save(model_config, os.path.join(cfgdir, "model.yaml")) + OmegaConf.save(OmegaConf.create({"lightning": lightning_config}), os.path.join(cfgdir, "lightning.yaml")) + return workdir, ckptdir, cfgdir, loginfo + +def check_config_attribute(config, name): + if name in config: + value = getattr(config, name) + return value + else: + return None + +def get_trainer_callbacks(lightning_config, config, logdir, ckptdir, logger): + default_callbacks_cfg = { + "model_checkpoint": { + "target": "pytorch_lightning.callbacks.ModelCheckpoint", + "params": { + "dirpath": ckptdir, + "filename": "{epoch}", + "verbose": True, + "save_last": False, + } + }, + "batch_logger": { + "target": "callbacks.ImageLogger", + "params": { + "save_dir": logdir, + "batch_frequency": 1000, + "max_images": 4, + "clamp": True, + } + }, + "learning_rate_logger": { + "target": "pytorch_lightning.callbacks.LearningRateMonitor", + "params": { + "logging_interval": "step", + "log_momentum": False + } + }, + "cuda_callback": { + "target": "callbacks.CUDACallback" + }, + } + + ## optional setting for saving checkpoints + monitor_metric = check_config_attribute(config.model.params, "monitor") + if monitor_metric is not None: + mainlogger.info(f"Monitoring {monitor_metric} as checkpoint metric.") + default_callbacks_cfg["model_checkpoint"]["params"]["monitor"] = monitor_metric + default_callbacks_cfg["model_checkpoint"]["params"]["save_top_k"] = 3 + default_callbacks_cfg["model_checkpoint"]["params"]["mode"] = "min" + + if 'metrics_over_trainsteps_checkpoint' in lightning_config.callbacks: + mainlogger.info('Caution: Saving checkpoints every n train steps without deleting. This might require some free space.') + default_metrics_over_trainsteps_ckpt_dict = { + 'metrics_over_trainsteps_checkpoint': {"target": 'pytorch_lightning.callbacks.ModelCheckpoint', + 'params': { + "dirpath": os.path.join(ckptdir, 'trainstep_checkpoints'), + "filename": "{epoch}-{step}", + "verbose": True, + 'save_top_k': -1, + 'every_n_train_steps': 10000, + 'save_weights_only': True + } + } + } + default_callbacks_cfg.update(default_metrics_over_trainsteps_ckpt_dict) + + if "callbacks" in lightning_config: + callbacks_cfg = lightning_config.callbacks + else: + callbacks_cfg = OmegaConf.create() + callbacks_cfg = OmegaConf.merge(default_callbacks_cfg, callbacks_cfg) + + return callbacks_cfg + +def get_trainer_logger(lightning_config, logdir, on_debug): + default_logger_cfgs = { + "tensorboard": { + "target": "pytorch_lightning.loggers.TensorBoardLogger", + "params": { + "save_dir": logdir, + "name": "tensorboard", + } + }, + "testtube": { + "target": "pytorch_lightning.loggers.CSVLogger", + "params": { + "name": "testtube", + "save_dir": logdir, + } + }, + } + os.makedirs(os.path.join(logdir, "tensorboard"), exist_ok=True) + default_logger_cfg = default_logger_cfgs["tensorboard"] + if "logger" in lightning_config: + logger_cfg = lightning_config.logger + else: + logger_cfg = OmegaConf.create() + logger_cfg = OmegaConf.merge(default_logger_cfg, logger_cfg) + return logger_cfg + +def get_trainer_strategy(lightning_config): + default_strategy_dict = { + "target": "pytorch_lightning.strategies.DDPShardedStrategy" + } + if "strategy" in lightning_config: + strategy_cfg = lightning_config.strategy + return strategy_cfg + else: + strategy_cfg = OmegaConf.create() + + strategy_cfg = OmegaConf.merge(default_strategy_dict, strategy_cfg) + return strategy_cfg + +def load_checkpoints(model, model_cfg): + if check_config_attribute(model_cfg, "pretrained_checkpoint"): + pretrained_ckpt = model_cfg.pretrained_checkpoint + assert os.path.exists(pretrained_ckpt), "Error: Pre-trained checkpoint NOT found at:%s"%pretrained_ckpt + mainlogger.info(">>> Load weights from pretrained checkpoint") + + pl_sd = torch.load(pretrained_ckpt, map_location="cpu") + try: + if 'state_dict' in pl_sd.keys(): + model.load_state_dict(pl_sd["state_dict"], strict=True) + mainlogger.info(">>> Loaded weights from pretrained checkpoint: %s"%pretrained_ckpt) + else: + # deepspeed + new_pl_sd = OrderedDict() + for key in pl_sd['module'].keys(): + new_pl_sd[key[16:]]=pl_sd['module'][key] + model.load_state_dict(new_pl_sd, strict=True) + except: + model.load_state_dict(pl_sd) + else: + mainlogger.info(">>> Start training from scratch") + + return model + +def set_logger(logfile, name='mainlogger'): + logger = logging.getLogger(name) + logger.setLevel(logging.INFO) + fh = logging.FileHandler(logfile, mode='w') + fh.setLevel(logging.INFO) + ch = logging.StreamHandler() + ch.setLevel(logging.DEBUG) + fh.setFormatter(logging.Formatter("%(asctime)s-%(levelname)s: %(message)s")) + ch.setFormatter(logging.Formatter("%(message)s")) + logger.addHandler(fh) + logger.addHandler(ch) + return logger \ No newline at end of file diff --git a/prompts/512_interp/74906_1462_frame1.png b/prompts/512_interp/74906_1462_frame1.png new file mode 100644 index 0000000000000000000000000000000000000000..5a49154a22fbbee0072314b7d8a34e925451013b Binary files /dev/null and b/prompts/512_interp/74906_1462_frame1.png differ diff --git a/prompts/512_interp/74906_1462_frame3.png b/prompts/512_interp/74906_1462_frame3.png new file mode 100644 index 0000000000000000000000000000000000000000..0be8d76a166fb80c99e4a1913ec8eadfde1eb123 Binary files /dev/null and b/prompts/512_interp/74906_1462_frame3.png differ diff --git a/prompts/512_interp/Japan_v2_2_062266_s2_frame1.png b/prompts/512_interp/Japan_v2_2_062266_s2_frame1.png new file mode 100644 index 0000000000000000000000000000000000000000..4cc9b95450dda7e9ccf5992fd9f4c77f88fb476b Binary files /dev/null and b/prompts/512_interp/Japan_v2_2_062266_s2_frame1.png differ diff --git a/prompts/512_interp/Japan_v2_2_062266_s2_frame3.png b/prompts/512_interp/Japan_v2_2_062266_s2_frame3.png new file mode 100644 index 0000000000000000000000000000000000000000..e78eeccb4008280eecbdf9b633197d7fe757c2ad Binary files /dev/null and b/prompts/512_interp/Japan_v2_2_062266_s2_frame3.png differ diff --git a/prompts/512_interp/Japan_v2_3_119235_s2_frame1.png b/prompts/512_interp/Japan_v2_3_119235_s2_frame1.png new file mode 100644 index 0000000000000000000000000000000000000000..12b46965b040bfd5d4481138fc2975e08ad41b70 Binary files /dev/null and b/prompts/512_interp/Japan_v2_3_119235_s2_frame1.png differ diff --git a/prompts/512_interp/Japan_v2_3_119235_s2_frame3.png b/prompts/512_interp/Japan_v2_3_119235_s2_frame3.png new file mode 100644 index 0000000000000000000000000000000000000000..4bc78cd19d2ed487c418748cc148918379cbcbaf Binary files /dev/null and b/prompts/512_interp/Japan_v2_3_119235_s2_frame3.png differ diff --git a/prompts/512_interp/prompts.txt b/prompts/512_interp/prompts.txt new file mode 100644 index 0000000000000000000000000000000000000000..21ff33243e165bef9977ae7e389c6a98fa8e4cee --- /dev/null +++ b/prompts/512_interp/prompts.txt @@ -0,0 +1,3 @@ +walking man +an anime scene +an anime scene \ No newline at end of file diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..18a1c2293cd5de29482efe93122789b15966838a --- /dev/null +++ b/requirements.txt @@ -0,0 +1,24 @@ +decord==0.6.0 +einops==0.3.0 +imageio==2.9.0 +numpy==1.24.2 +omegaconf==2.1.1 +opencv_python +pandas==2.0.0 +Pillow==9.5.0 +pytorch_lightning==1.9.3 +PyYAML==6.0 +setuptools==65.6.3 +torch==2.0.0 +torchvision +tqdm==4.65.0 +transformers==4.25.1 +moviepy +av +xformers +gradio +timm +scikit-learn +open_clip_torch==2.22.0 +kornia +httpx[socks] diff --git a/scripts/evaluation/ddp_wrapper.py b/scripts/evaluation/ddp_wrapper.py new file mode 100644 index 0000000000000000000000000000000000000000..3caec49623b852fadd34feead1574df53f077332 --- /dev/null +++ b/scripts/evaluation/ddp_wrapper.py @@ -0,0 +1,47 @@ +import datetime +import argparse, importlib +from pytorch_lightning import seed_everything + +import torch +import torch.distributed as dist + +def setup_dist(local_rank): + if dist.is_initialized(): + return + torch.cuda.set_device(local_rank) + torch.distributed.init_process_group('nccl', init_method='env://') + + +def get_dist_info(): + if dist.is_available(): + initialized = dist.is_initialized() + else: + initialized = False + if initialized: + rank = dist.get_rank() + world_size = dist.get_world_size() + else: + rank = 0 + world_size = 1 + return rank, world_size + + +if __name__ == '__main__': + now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S") + parser = argparse.ArgumentParser() + parser.add_argument("--module", type=str, help="module name", default="inference") + parser.add_argument("--local_rank", type=int, nargs="?", help="for ddp", default=0) + args, unknown = parser.parse_known_args() + inference_api = importlib.import_module(args.module, package=None) + + inference_parser = inference_api.get_parser() + inference_args, unknown = inference_parser.parse_known_args() + + seed_everything(inference_args.seed) + setup_dist(args.local_rank) + torch.backends.cudnn.benchmark = True + rank, gpu_num = get_dist_info() + + # inference_args.savedir = inference_args.savedir+str('_seed')+str(inference_args.seed) + print("@DynamiCrafter Inference [rank%d]: %s"%(rank, now)) + inference_api.run_inference(inference_args, gpu_num, rank) \ No newline at end of file diff --git a/scripts/evaluation/funcs.py b/scripts/evaluation/funcs.py new file mode 100644 index 0000000000000000000000000000000000000000..6afc437fe3cef2503e725d7c2e7f0bf8e5b43173 --- /dev/null +++ b/scripts/evaluation/funcs.py @@ -0,0 +1,240 @@ +import os, sys, glob +import numpy as np +from collections import OrderedDict +from decord import VideoReader, cpu +import cv2 + +import torch +import torchvision +sys.path.insert(1, os.path.join(sys.path[0], '..', '..')) +from lvdm.models.samplers.ddim import DDIMSampler +from einops import rearrange + + +def batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\ + cfg_scale=1.0, hs=None, temporal_cfg_scale=None, **kwargs): + ddim_sampler = DDIMSampler(model) + uncond_type = model.uncond_type + batch_size = noise_shape[0] + fs = cond["fs"] + del cond["fs"] + if noise_shape[-1] == 32: + timestep_spacing = "uniform" + guidance_rescale = 0.0 + else: + timestep_spacing = "uniform_trailing" + guidance_rescale = 0.7 + ## construct unconditional guidance + if cfg_scale != 1.0: + if uncond_type == "empty_seq": + prompts = batch_size * [""] + #prompts = N * T * [""] ## if is_imgbatch=True + uc_emb = model.get_learned_conditioning(prompts) + elif uncond_type == "zero_embed": + c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond + uc_emb = torch.zeros_like(c_emb) + + ## process image embedding token + if hasattr(model, 'embedder'): + uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device) + ## img: b c h w >> b l c + uc_img = model.embedder(uc_img) + uc_img = model.image_proj_model(uc_img) + uc_emb = torch.cat([uc_emb, uc_img], dim=1) + + if isinstance(cond, dict): + uc = {key:cond[key] for key in cond.keys()} + uc.update({'c_crossattn': [uc_emb]}) + else: + uc = uc_emb + else: + uc = None + + + additional_decode_kwargs = {'ref_context': hs} + x_T = None + batch_variants = [] + + for _ in range(n_samples): + if ddim_sampler is not None: + kwargs.update({"clean_cond": True}) + samples, _ = ddim_sampler.sample(S=ddim_steps, + conditioning=cond, + batch_size=noise_shape[0], + shape=noise_shape[1:], + verbose=False, + unconditional_guidance_scale=cfg_scale, + unconditional_conditioning=uc, + eta=ddim_eta, + temporal_length=noise_shape[2], + conditional_guidance_scale_temporal=temporal_cfg_scale, + x_T=x_T, + fs=fs, + timestep_spacing=timestep_spacing, + guidance_rescale=guidance_rescale, + **kwargs + ) + ## reconstruct from latent to pixel space + batch_images = model.decode_first_stage(samples, **additional_decode_kwargs) + + index = list(range(samples.shape[2])) + del index[1] + del index[-2] + samples = samples[:,:,index,:,:] + ## reconstruct from latent to pixel space + batch_images_middle = model.decode_first_stage(samples, **additional_decode_kwargs) + batch_images[:,:,batch_images.shape[2]//2-1:batch_images.shape[2]//2+1] = batch_images_middle[:,:,batch_images.shape[2]//2-2:batch_images.shape[2]//2] + + + + batch_variants.append(batch_images) + ## batch, , c, t, h, w + batch_variants = torch.stack(batch_variants, dim=1) + return batch_variants + + +def get_filelist(data_dir, ext='*'): + file_list = glob.glob(os.path.join(data_dir, '*.%s'%ext)) + file_list.sort() + return file_list + +def get_dirlist(path): + list = [] + if (os.path.exists(path)): + files = os.listdir(path) + for file in files: + m = os.path.join(path,file) + if (os.path.isdir(m)): + list.append(m) + list.sort() + return list + + +def load_model_checkpoint(model, ckpt): + def load_checkpoint(model, ckpt, full_strict): + state_dict = torch.load(ckpt, map_location="cpu") + if "state_dict" in list(state_dict.keys()): + state_dict = state_dict["state_dict"] + try: + model.load_state_dict(state_dict, strict=full_strict) + except: + ## rename the keys for 256x256 model + new_pl_sd = OrderedDict() + for k,v in state_dict.items(): + new_pl_sd[k] = v + + for k in list(new_pl_sd.keys()): + if "framestride_embed" in k: + new_key = k.replace("framestride_embed", "fps_embedding") + new_pl_sd[new_key] = new_pl_sd[k] + del new_pl_sd[k] + model.load_state_dict(new_pl_sd, strict=full_strict) + else: + ## deepspeed + new_pl_sd = OrderedDict() + for key in state_dict['module'].keys(): + new_pl_sd[key[16:]]=state_dict['module'][key] + model.load_state_dict(new_pl_sd, strict=full_strict) + + return model + load_checkpoint(model, ckpt, full_strict=True) + print('>>> model checkpoint loaded.') + return model + + +def load_prompts(prompt_file): + f = open(prompt_file, 'r') + prompt_list = [] + for idx, line in enumerate(f.readlines()): + l = line.strip() + if len(l) != 0: + prompt_list.append(l) + f.close() + return prompt_list + + +def load_video_batch(filepath_list, frame_stride, video_size=(256,256), video_frames=16): + ''' + Notice about some special cases: + 1. video_frames=-1 means to take all the frames (with fs=1) + 2. when the total video frames is less than required, padding strategy will be used (repeated last frame) + ''' + fps_list = [] + batch_tensor = [] + assert frame_stride > 0, "valid frame stride should be a positive interge!" + for filepath in filepath_list: + padding_num = 0 + vidreader = VideoReader(filepath, ctx=cpu(0), width=video_size[1], height=video_size[0]) + fps = vidreader.get_avg_fps() + total_frames = len(vidreader) + max_valid_frames = (total_frames-1) // frame_stride + 1 + if video_frames < 0: + ## all frames are collected: fs=1 is a must + required_frames = total_frames + frame_stride = 1 + else: + required_frames = video_frames + query_frames = min(required_frames, max_valid_frames) + frame_indices = [frame_stride*i for i in range(query_frames)] + + ## [t,h,w,c] -> [c,t,h,w] + frames = vidreader.get_batch(frame_indices) + frame_tensor = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float() + frame_tensor = (frame_tensor / 255. - 0.5) * 2 + if max_valid_frames < required_frames: + padding_num = required_frames - max_valid_frames + frame_tensor = torch.cat([frame_tensor, *([frame_tensor[:,-1:,:,:]]*padding_num)], dim=1) + print(f'{os.path.split(filepath)[1]} is not long enough: {padding_num} frames padded.') + batch_tensor.append(frame_tensor) + sample_fps = int(fps/frame_stride) + fps_list.append(sample_fps) + + return torch.stack(batch_tensor, dim=0) + +from PIL import Image +def load_image_batch(filepath_list, image_size=(256,256)): + batch_tensor = [] + for filepath in filepath_list: + _, filename = os.path.split(filepath) + _, ext = os.path.splitext(filename) + if ext == '.mp4': + vidreader = VideoReader(filepath, ctx=cpu(0), width=image_size[1], height=image_size[0]) + frame = vidreader.get_batch([0]) + img_tensor = torch.tensor(frame.asnumpy()).squeeze(0).permute(2, 0, 1).float() + elif ext == '.png' or ext == '.jpg': + img = Image.open(filepath).convert("RGB") + rgb_img = np.array(img, np.float32) + #bgr_img = cv2.imread(filepath, cv2.IMREAD_COLOR) + #bgr_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2RGB) + rgb_img = cv2.resize(rgb_img, (image_size[1],image_size[0]), interpolation=cv2.INTER_LINEAR) + img_tensor = torch.from_numpy(rgb_img).permute(2, 0, 1).float() + else: + print(f'ERROR: <{ext}> image loading only support format: [mp4], [png], [jpg]') + raise NotImplementedError + img_tensor = (img_tensor / 255. - 0.5) * 2 + batch_tensor.append(img_tensor) + return torch.stack(batch_tensor, dim=0) + + +def save_videos(batch_tensors, savedir, filenames, fps=10): + # b,samples,c,t,h,w + n_samples = batch_tensors.shape[1] + for idx, vid_tensor in enumerate(batch_tensors): + video = vid_tensor.detach().cpu() + video = torch.clamp(video.float(), -1., 1.) + video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w + frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n_samples)) for framesheet in video] #[3, 1*h, n*w] + grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w] + grid = (grid + 1.0) / 2.0 + grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1) + savepath = os.path.join(savedir, f"{filenames[idx]}.mp4") + torchvision.io.write_video(savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'}) + + +def get_latent_z(model, videos): + b, c, t, h, w = videos.shape + x = rearrange(videos, 'b c t h w -> (b t) c h w') + z = model.encode_first_stage(x) + z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t) + return z + diff --git a/scripts/evaluation/inference.py b/scripts/evaluation/inference.py new file mode 100644 index 0000000000000000000000000000000000000000..334a9843376905ecdc1f96db5bbeb289d64acf3d --- /dev/null +++ b/scripts/evaluation/inference.py @@ -0,0 +1,385 @@ +import argparse, os, sys, glob +import datetime, time +from omegaconf import OmegaConf +from tqdm import tqdm +from einops import rearrange, repeat +from collections import OrderedDict + +import torch +import torchvision +import torchvision.transforms as transforms +from pytorch_lightning import seed_everything +from PIL import Image +sys.path.insert(1, os.path.join(sys.path[0], '..', '..')) +from lvdm.models.samplers.ddim import DDIMSampler +from lvdm.models.samplers.ddim_multiplecond import DDIMSampler as DDIMSampler_multicond +from utils.utils import instantiate_from_config + + +def get_filelist(data_dir, postfixes): + patterns = [os.path.join(data_dir, f"*.{postfix}") for postfix in postfixes] + file_list = [] + for pattern in patterns: + file_list.extend(glob.glob(pattern)) + file_list.sort() + return file_list + +def load_model_checkpoint(model, ckpt): + state_dict = torch.load(ckpt, map_location="cpu") + if "state_dict" in list(state_dict.keys()): + state_dict = state_dict["state_dict"] + try: + model.load_state_dict(state_dict, strict=True) + except: + ## rename the keys for 256x256 model + new_pl_sd = OrderedDict() + for k,v in state_dict.items(): + new_pl_sd[k] = v + + for k in list(new_pl_sd.keys()): + if "framestride_embed" in k: + new_key = k.replace("framestride_embed", "fps_embedding") + new_pl_sd[new_key] = new_pl_sd[k] + del new_pl_sd[k] + model.load_state_dict(new_pl_sd, strict=True) + else: + # deepspeed + new_pl_sd = OrderedDict() + for key in state_dict['module'].keys(): + new_pl_sd[key[16:]]=state_dict['module'][key] + model.load_state_dict(new_pl_sd) + print('>>> model checkpoint loaded.') + return model + +def load_prompts(prompt_file): + f = open(prompt_file, 'r') + prompt_list = [] + for idx, line in enumerate(f.readlines()): + l = line.strip() + if len(l) != 0: + prompt_list.append(l) + f.close() + return prompt_list + +def load_data_prompts(data_dir, video_size=(256,256), video_frames=16, interp=False): + transform = transforms.Compose([ + transforms.Resize(min(video_size)), + transforms.CenterCrop(video_size), + transforms.ToTensor(), + transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))]) + ## load prompts + prompt_file = get_filelist(data_dir, ['txt']) + assert len(prompt_file) > 0, "Error: found NO prompt file!" + ###### default prompt + default_idx = 0 + default_idx = min(default_idx, len(prompt_file)-1) + if len(prompt_file) > 1: + print(f"Warning: multiple prompt files exist. The one {os.path.split(prompt_file[default_idx])[1]} is used.") + ## only use the first one (sorted by name) if multiple exist + + ## load video + file_list = get_filelist(data_dir, ['jpg', 'png', 'jpeg', 'JPEG', 'PNG']) + # assert len(file_list) == n_samples, "Error: data and prompts are NOT paired!" + data_list = [] + filename_list = [] + prompt_list = load_prompts(prompt_file[default_idx]) + n_samples = len(prompt_list) + for idx in range(n_samples): + if interp: + image1 = Image.open(file_list[2*idx]).convert('RGB') + image_tensor1 = transform(image1).unsqueeze(1) # [c,1,h,w] + image2 = Image.open(file_list[2*idx+1]).convert('RGB') + image_tensor2 = transform(image2).unsqueeze(1) # [c,1,h,w] + frame_tensor1 = repeat(image_tensor1, 'c t h w -> c (repeat t) h w', repeat=video_frames//2) + frame_tensor2 = repeat(image_tensor2, 'c t h w -> c (repeat t) h w', repeat=video_frames//2) + frame_tensor = torch.cat([frame_tensor1, frame_tensor2], dim=1) + _, filename = os.path.split(file_list[idx*2]) + else: + image = Image.open(file_list[idx]).convert('RGB') + image_tensor = transform(image).unsqueeze(1) # [c,1,h,w] + frame_tensor = repeat(image_tensor, 'c t h w -> c (repeat t) h w', repeat=video_frames) + _, filename = os.path.split(file_list[idx]) + + data_list.append(frame_tensor) + filename_list.append(filename) + + return filename_list, data_list, prompt_list + + +def save_results(prompt, samples, filename, fakedir, fps=8, loop=False): + filename = filename.split('.')[0]+'.mp4' + prompt = prompt[0] if isinstance(prompt, list) else prompt + + ## save video + videos = [samples] + savedirs = [fakedir] + for idx, video in enumerate(videos): + if video is None: + continue + # b,c,t,h,w + video = video.detach().cpu() + video = torch.clamp(video.float(), -1., 1.) + n = video.shape[0] + video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w + if loop: + video = video[:-1,...] + + frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n), padding=0) for framesheet in video] #[3, 1*h, n*w] + grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, h, n*w] + grid = (grid + 1.0) / 2.0 + grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1) + path = os.path.join(savedirs[idx], filename) + torchvision.io.write_video(path, grid, fps=fps, video_codec='h264', options={'crf': '10'}) ## crf indicates the quality + + +def save_results_seperate(prompt, samples, filename, fakedir, fps=10, loop=False): + prompt = prompt[0] if isinstance(prompt, list) else prompt + + ## save video + videos = [samples] + savedirs = [fakedir] + for idx, video in enumerate(videos): + if video is None: + continue + # b,c,t,h,w + video = video.detach().cpu() + if loop: # remove the last frame + video = video[:,:,:-1,...] + video = torch.clamp(video.float(), -1., 1.) + n = video.shape[0] + for i in range(n): + grid = video[i,...] + grid = (grid + 1.0) / 2.0 + grid = (grid * 255).to(torch.uint8).permute(1, 2, 3, 0) #thwc + path = os.path.join(savedirs[idx].replace('samples', 'samples_separate'), f'{filename.split(".")[0]}_sample{i}.mp4') + torchvision.io.write_video(path, grid, fps=fps, video_codec='h264', options={'crf': '10'}) + +def get_latent_z(model, videos): + b, c, t, h, w = videos.shape + x = rearrange(videos, 'b c t h w -> (b t) c h w') + z = model.encode_first_stage(x) + z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t) + return z + +def get_latent_z_with_hidden_states(model, videos): + b, c, t, h, w = videos.shape + x = rearrange(videos, 'b c t h w -> (b t) c h w') + encoder_posterior, hidden_states = model.first_stage_model.encode(x, return_hidden_states=True) + + hidden_states_first_last = [] + ### use only the first and last hidden states + for hid in hidden_states: + hid = rearrange(hid, '(b t) c h w -> b c t h w', t=t) + hid_new = torch.cat([hid[:, :, 0:1], hid[:, :, -1:]], dim=2) + hidden_states_first_last.append(hid_new) + + z = model.get_first_stage_encoding(encoder_posterior).detach() + z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t) + return z, hidden_states_first_last + +def image_guided_synthesis(model, prompts, videos, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1., \ + unconditional_guidance_scale=1.0, cfg_img=None, fs=None, text_input=False, multiple_cond_cfg=False, loop=False, interp=False, timestep_spacing='uniform', guidance_rescale=0.0, **kwargs): + ddim_sampler = DDIMSampler(model) if not multiple_cond_cfg else DDIMSampler_multicond(model) + batch_size = noise_shape[0] + fs = torch.tensor([fs] * batch_size, dtype=torch.long, device=model.device) + + if not text_input: + prompts = [""]*batch_size + + img = videos[:,:,0] #bchw + img_emb = model.embedder(img) ## blc + img_emb = model.image_proj_model(img_emb) + + cond_emb = model.get_learned_conditioning(prompts) + cond = {"c_crossattn": [torch.cat([cond_emb,img_emb], dim=1)]} + if model.model.conditioning_key == 'hybrid': + z, hs = get_latent_z_with_hidden_states(model, videos) # b c t h w + if loop or interp: + img_cat_cond = torch.zeros_like(z) + img_cat_cond[:,:,0,:,:] = z[:,:,0,:,:] + img_cat_cond[:,:,-1,:,:] = z[:,:,-1,:,:] + else: + img_cat_cond = z[:,:,:1,:,:] + img_cat_cond = repeat(img_cat_cond, 'b c t h w -> b c (repeat t) h w', repeat=z.shape[2]) + cond["c_concat"] = [img_cat_cond] # b c 1 h w + + if unconditional_guidance_scale != 1.0: + if model.uncond_type == "empty_seq": + prompts = batch_size * [""] + uc_emb = model.get_learned_conditioning(prompts) + elif model.uncond_type == "zero_embed": + uc_emb = torch.zeros_like(cond_emb) + uc_img_emb = model.embedder(torch.zeros_like(img)) ## b l c + uc_img_emb = model.image_proj_model(uc_img_emb) + uc = {"c_crossattn": [torch.cat([uc_emb,uc_img_emb],dim=1)]} + if model.model.conditioning_key == 'hybrid': + uc["c_concat"] = [img_cat_cond] + else: + uc = None + + additional_decode_kwargs = {'ref_context': hs} + + ## we need one more unconditioning image=yes, text="" + if multiple_cond_cfg and cfg_img != 1.0: + uc_2 = {"c_crossattn": [torch.cat([uc_emb,img_emb],dim=1)]} + if model.model.conditioning_key == 'hybrid': + uc_2["c_concat"] = [img_cat_cond] + kwargs.update({"unconditional_conditioning_img_nonetext": uc_2}) + else: + kwargs.update({"unconditional_conditioning_img_nonetext": None}) + + z0 = None + cond_mask = None + + batch_variants = [] + for _ in range(n_samples): + + if z0 is not None: + cond_z0 = z0.clone() + kwargs.update({"clean_cond": True}) + else: + cond_z0 = None + if ddim_sampler is not None: + + samples, _ = ddim_sampler.sample(S=ddim_steps, + conditioning=cond, + batch_size=batch_size, + shape=noise_shape[1:], + verbose=False, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + eta=ddim_eta, + cfg_img=cfg_img, + mask=cond_mask, + x0=cond_z0, + fs=fs, + timestep_spacing=timestep_spacing, + guidance_rescale=guidance_rescale, + **kwargs + ) + + ## reconstruct from latent to pixel space + batch_images = model.decode_first_stage(samples, **additional_decode_kwargs) + + index = list(range(samples.shape[2])) + del index[1] + del index[-2] + samples = samples[:,:,index,:,:] + ## reconstruct from latent to pixel space + batch_images_middle = model.decode_first_stage(samples, **additional_decode_kwargs) + batch_images[:,:,batch_images.shape[2]//2-1:batch_images.shape[2]//2+1] = batch_images_middle[:,:,batch_images.shape[2]//2-2:batch_images.shape[2]//2] + + + + batch_variants.append(batch_images) + ## variants, batch, c, t, h, w + batch_variants = torch.stack(batch_variants) + return batch_variants.permute(1, 0, 2, 3, 4, 5) + + +def run_inference(args, gpu_num, gpu_no): + ## model config + config = OmegaConf.load(args.config) + model_config = config.pop("model", OmegaConf.create()) + + ## set use_checkpoint as False as when using deepspeed, it encounters an error "deepspeed backend not set" + model_config['params']['unet_config']['params']['use_checkpoint'] = False + model = instantiate_from_config(model_config) + model = model.cuda(gpu_no) + model.perframe_ae = args.perframe_ae + assert os.path.exists(args.ckpt_path), "Error: checkpoint Not Found!" + model = load_model_checkpoint(model, args.ckpt_path) + model.eval() + + ## run over data + assert (args.height % 16 == 0) and (args.width % 16 == 0), "Error: image size [h,w] should be multiples of 16!" + assert args.bs == 1, "Current implementation only support [batch size = 1]!" + ## latent noise shape + h, w = args.height // 8, args.width // 8 + channels = model.model.diffusion_model.out_channels + n_frames = args.video_length + print(f'Inference with {n_frames} frames') + noise_shape = [args.bs, channels, n_frames, h, w] + + fakedir = os.path.join(args.savedir, "samples") + fakedir_separate = os.path.join(args.savedir, "samples_separate") + + # os.makedirs(fakedir, exist_ok=True) + os.makedirs(fakedir_separate, exist_ok=True) + + ## prompt file setting + assert os.path.exists(args.prompt_dir), "Error: prompt file Not Found!" + filename_list, data_list, prompt_list = load_data_prompts(args.prompt_dir, video_size=(args.height, args.width), video_frames=n_frames, interp=args.interp) + num_samples = len(prompt_list) + samples_split = num_samples // gpu_num + print('Prompts testing [rank:%d] %d/%d samples loaded.'%(gpu_no, samples_split, num_samples)) + #indices = random.choices(list(range(0, num_samples)), k=samples_per_device) + indices = list(range(samples_split*gpu_no, samples_split*(gpu_no+1))) + prompt_list_rank = [prompt_list[i] for i in indices] + data_list_rank = [data_list[i] for i in indices] + filename_list_rank = [filename_list[i] for i in indices] + + start = time.time() + with torch.no_grad(), torch.cuda.amp.autocast(): + for idx, indice in tqdm(enumerate(range(0, len(prompt_list_rank), args.bs)), desc='Sample Batch'): + prompts = prompt_list_rank[indice:indice+args.bs] + videos = data_list_rank[indice:indice+args.bs] + filenames = filename_list_rank[indice:indice+args.bs] + if isinstance(videos, list): + videos = torch.stack(videos, dim=0).to("cuda") + else: + videos = videos.unsqueeze(0).to("cuda") + + batch_samples = image_guided_synthesis(model, prompts, videos, noise_shape, args.n_samples, args.ddim_steps, args.ddim_eta, \ + args.unconditional_guidance_scale, args.cfg_img, args.frame_stride, args.text_input, args.multiple_cond_cfg, args.loop, args.interp, args.timestep_spacing, args.guidance_rescale) + + ## save each example individually + for nn, samples in enumerate(batch_samples): + ## samples : [n_samples,c,t,h,w] + prompt = prompts[nn] + filename = filenames[nn] + # save_results(prompt, samples, filename, fakedir, fps=8, loop=args.loop) + save_results_seperate(prompt, samples, filename, fakedir, fps=8, loop=args.loop) + + print(f"Saved in {args.savedir}. Time used: {(time.time() - start):.2f} seconds") + + +def get_parser(): + parser = argparse.ArgumentParser() + parser.add_argument("--savedir", type=str, default=None, help="results saving path") + parser.add_argument("--ckpt_path", type=str, default=None, help="checkpoint path") + parser.add_argument("--config", type=str, help="config (yaml) path") + parser.add_argument("--prompt_dir", type=str, default=None, help="a data dir containing videos and prompts") + parser.add_argument("--n_samples", type=int, default=1, help="num of samples per prompt",) + parser.add_argument("--ddim_steps", type=int, default=50, help="steps of ddim if positive, otherwise use DDPM",) + parser.add_argument("--ddim_eta", type=float, default=1.0, help="eta for ddim sampling (0.0 yields deterministic sampling)",) + parser.add_argument("--bs", type=int, default=1, help="batch size for inference, should be one") + parser.add_argument("--height", type=int, default=512, help="image height, in pixel space") + parser.add_argument("--width", type=int, default=512, help="image width, in pixel space") + parser.add_argument("--frame_stride", type=int, default=3, help="frame stride control for 256 model (larger->larger motion), FPS control for 512 or 1024 model (smaller->larger motion)") + parser.add_argument("--unconditional_guidance_scale", type=float, default=1.0, help="prompt classifier-free guidance") + parser.add_argument("--seed", type=int, default=123, help="seed for seed_everything") + parser.add_argument("--video_length", type=int, default=16, help="inference video length") + parser.add_argument("--negative_prompt", action='store_true', default=False, help="negative prompt") + parser.add_argument("--text_input", action='store_true', default=False, help="input text to I2V model or not") + parser.add_argument("--multiple_cond_cfg", action='store_true', default=False, help="use multi-condition cfg or not") + parser.add_argument("--cfg_img", type=float, default=None, help="guidance scale for image conditioning") + parser.add_argument("--timestep_spacing", type=str, default="uniform", help="The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.") + parser.add_argument("--guidance_rescale", type=float, default=0.0, help="guidance rescale in [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891)") + parser.add_argument("--perframe_ae", action='store_true', default=False, help="if we use per-frame AE decoding, set it to True to save GPU memory, especially for the model of 576x1024") + + ## currently not support looping video and generative frame interpolation + parser.add_argument("--loop", action='store_true', default=False, help="generate looping videos or not") + parser.add_argument("--interp", action='store_true', default=False, help="generate generative frame interpolation or not") + return parser + + +if __name__ == '__main__': + now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S") + print("@DynamiCrafter cond-Inference: %s"%now) + parser = get_parser() + args = parser.parse_args() + + seed_everything(args.seed) + rank, gpu_num = 0, 1 + run_inference(args, gpu_num, rank) \ No newline at end of file diff --git a/scripts/gradio/i2v_test.py b/scripts/gradio/i2v_test.py new file mode 100644 index 0000000000000000000000000000000000000000..9cb620acafd60a2d3e32c30025cc102429805a72 --- /dev/null +++ b/scripts/gradio/i2v_test.py @@ -0,0 +1,107 @@ +import os +import time +from omegaconf import OmegaConf +import torch +from scripts.evaluation.funcs import load_model_checkpoint, save_videos, batch_ddim_sampling, get_latent_z +from utils.utils import instantiate_from_config +from huggingface_hub import hf_hub_download +from einops import repeat +import torchvision.transforms as transforms +from pytorch_lightning import seed_everything + + +class Image2Video(): + def __init__(self,result_dir='./tmp/',gpu_num=1,resolution='256_256') -> None: + self.resolution = (int(resolution.split('_')[0]), int(resolution.split('_')[1])) #hw + self.download_model() + + self.result_dir = result_dir + if not os.path.exists(self.result_dir): + os.mkdir(self.result_dir) + ckpt_path='checkpoints/dynamicrafter_'+resolution.split('_')[1]+'_v1/model.ckpt' + config_file='configs/inference_'+resolution.split('_')[1]+'_v1.0.yaml' + config = OmegaConf.load(config_file) + model_config = config.pop("model", OmegaConf.create()) + model_config['params']['unet_config']['params']['use_checkpoint']=False + model_list = [] + for gpu_id in range(gpu_num): + model = instantiate_from_config(model_config) + # model = model.cuda(gpu_id) + assert os.path.exists(ckpt_path), "Error: checkpoint Not Found!" + model = load_model_checkpoint(model, ckpt_path) + model.eval() + model_list.append(model) + self.model_list = model_list + self.save_fps = 8 + + def get_image(self, image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123): + seed_everything(seed) + transform = transforms.Compose([ + transforms.Resize(min(self.resolution)), + transforms.CenterCrop(self.resolution), + ]) + torch.cuda.empty_cache() + print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S',time.localtime(time.time()))) + start = time.time() + gpu_id=0 + if steps > 60: + steps = 60 + model = self.model_list[gpu_id] + model = model.cuda() + batch_size=1 + channels = model.model.diffusion_model.out_channels + frames = model.temporal_length + h, w = self.resolution[0] // 8, self.resolution[1] // 8 + noise_shape = [batch_size, channels, frames, h, w] + + # text cond + with torch.no_grad(), torch.cuda.amp.autocast(): + text_emb = model.get_learned_conditioning([prompt]) + + # img cond + img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device) + img_tensor = (img_tensor / 255. - 0.5) * 2 + + image_tensor_resized = transform(img_tensor) #3,h,w + videos = image_tensor_resized.unsqueeze(0) # bchw + + z = get_latent_z(model, videos.unsqueeze(2)) #bc,1,hw + + img_tensor_repeat = repeat(z, 'b c t h w -> b c (repeat t) h w', repeat=frames) + + cond_images = model.embedder(img_tensor.unsqueeze(0)) ## blc + img_emb = model.image_proj_model(cond_images) + + imtext_cond = torch.cat([text_emb, img_emb], dim=1) + + fs = torch.tensor([fs], dtype=torch.long, device=model.device) + cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat]} + + ## inference + batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale) + ## b,samples,c,t,h,w + prompt_str = prompt.replace("/", "_slash_") if "/" in prompt else prompt + prompt_str = prompt_str.replace(" ", "_") if " " in prompt else prompt_str + prompt_str=prompt_str[:40] + if len(prompt_str) == 0: + prompt_str = 'empty_prompt' + + save_videos(batch_samples, self.result_dir, filenames=[prompt_str], fps=self.save_fps) + print(f"Saved in {prompt_str}. Time used: {(time.time() - start):.2f} seconds") + model = model.cpu() + return os.path.join(self.result_dir, f"{prompt_str}.mp4") + + def download_model(self): + REPO_ID = 'Doubiiu/DynamiCrafter_'+str(self.resolution[1]) if self.resolution[1]!=256 else 'Doubiiu/DynamiCrafter' + filename_list = ['model.ckpt'] + if not os.path.exists('./checkpoints/dynamicrafter_'+str(self.resolution[1])+'_v1/'): + os.makedirs('./checkpoints/dynamicrafter_'+str(self.resolution[1])+'_v1/') + for filename in filename_list: + local_file = os.path.join('./checkpoints/dynamicrafter_'+str(self.resolution[1])+'_v1/', filename) + if not os.path.exists(local_file): + hf_hub_download(repo_id=REPO_ID, filename=filename, local_dir='./checkpoints/dynamicrafter_'+str(self.resolution[1])+'_v1/', local_dir_use_symlinks=False) + +if __name__ == '__main__': + i2v = Image2Video() + video_path = i2v.get_image('prompts/art.png','man fishing in a boat at sunset') + print('done', video_path) \ No newline at end of file diff --git a/scripts/gradio/i2v_test_application.py b/scripts/gradio/i2v_test_application.py new file mode 100644 index 0000000000000000000000000000000000000000..fbe0aae905d698ee4aa5ed5b6632cb438ad68515 --- /dev/null +++ b/scripts/gradio/i2v_test_application.py @@ -0,0 +1,216 @@ +import os +import time +from omegaconf import OmegaConf +import torch +from scripts.evaluation.funcs import load_model_checkpoint, save_videos, batch_ddim_sampling, get_latent_z +from utils.utils import instantiate_from_config +from huggingface_hub import hf_hub_download +from einops import repeat +import torchvision.transforms as transforms +from pytorch_lightning import seed_everything +from einops import rearrange +from cldm.model import load_state_dict +import cv2 + + +def extract_frames(video_path): + # 動画フゑむルをθͺ­γΏθΎΌγ‚€ + cap = cv2.VideoCapture(video_path) + + frame_list = [] + frame_num = 0 + + while True: + # フレームをθͺ­γΏθΎΌγ‚€ + ret, frame = cap.read() + if not ret: + break + + # フレームをγƒͺγ‚Ήγƒˆγ«θΏ½εŠ  + frame_list.append(frame) + frame_num += 1 + + # ε‹•η”»γƒ•γ‚‘γ‚€γƒ«γ‚’ι–‰γ˜γ‚‹ + cap.release() + + return frame_list + +class Image2Video(): + def __init__(self,result_dir='./tmp/',gpu_num=1,resolution='256_256') -> None: + self.resolution = (int(resolution.split('_')[0]), int(resolution.split('_')[1])) #hw + self.download_model() + + self.result_dir = result_dir + if not os.path.exists(self.result_dir): + os.mkdir(self.result_dir) + + #ToonCrafterModel + ckpt_path='checkpoints/tooncrafter_'+resolution.split('_')[1]+'_interp_v1/model.ckpt' + config_file='configs/inference_'+resolution.split('_')[1]+'_v1.0.yaml' + config = OmegaConf.load(config_file) + model_config = config.pop("model", OmegaConf.create()) + model_config['params']['unet_config']['params']['use_checkpoint']=False + + #ControlModel + cn_ckpt_path = "control_models/sketch_encoder.ckpt" + cn_config_file = 'configs/cldm_v21.yaml' + cn_config = OmegaConf.load(cn_config_file) + cn_model_config = cn_config.pop("control_stage_config", OmegaConf.create()) + + + model_list = [] + for gpu_id in range(gpu_num): + model = instantiate_from_config(model_config) + cn_model = instantiate_from_config(cn_model_config) + + # model = model.cuda(gpu_id) + assert os.path.exists(ckpt_path), "Error: checkpoint Not Found!" + model = load_model_checkpoint(model, ckpt_path) + model.eval() + + cn_model.load_state_dict(load_state_dict(cn_ckpt_path, location='cuda')) + cn_model.eval() + + model.control_model = cn_model + + model_list.append(model) + self.model_list = model_list + self.save_fps = 8 + + def get_image(self, image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123, image2=None, frame_guides=None,control_scale=0.6): + control_frames = extract_frames(frame_guides) + seed_everything(seed) + transform = transforms.Compose([ + transforms.Resize(min(self.resolution)), + transforms.CenterCrop(self.resolution), + ]) + torch.cuda.empty_cache() + print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S',time.localtime(time.time()))) + start = time.time() + gpu_id=0 + if steps > 60: + steps = 60 + model = self.model_list[gpu_id] + #model = model.cuda() + model = model.half().cuda() + batch_size=1 + channels = model.model.diffusion_model.out_channels + frames = model.temporal_length + h, w = self.resolution[0] // 8, self.resolution[1] // 8 + noise_shape = [batch_size, channels, frames, h, w] + + # text cond + with torch.no_grad(), torch.cuda.amp.autocast(): + text_emb = model.get_learned_conditioning([prompt]) + + #control cond + if frame_guides is not None: + cn_videos = [] + for frame in control_frames: + frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) + frame = cv2.bitwise_not(frame) + #cn_tensor = torch.from_numpy(frame).unsqueeze(2).permute(2, 0, 1).float().to(model.device) + cn_tensor = torch.from_numpy(frame).unsqueeze(2).permute(2, 0, 1).float().half().to(model.device) + + #cn_tensor = (cn_tensor / 255. - 0.5) * 2 + cn_tensor = ( cn_tensor/255.0 ) + cn_tensor_resized = transform(cn_tensor) #3,h,w + + cn_video = cn_tensor_resized.unsqueeze(0).unsqueeze(2) # bc1hw + cn_videos.append(cn_video) + + cn_videos = torch.cat(cn_videos, dim=2) + model_list = [] + for model in self.model_list: + model.control_scale = control_scale + model_list.append(model) + self.model_list = model_list + + else: + cn_videos = None + + + + # img cond + #img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device) + img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().half().to(model.device) + img_tensor = (img_tensor / 255. - 0.5) * 2 + + image_tensor_resized = transform(img_tensor) #3,h,w + videos = image_tensor_resized.unsqueeze(0).unsqueeze(2) # bc1hw + + # z = get_latent_z(model, videos) #bc,1,hw + videos = repeat(videos, 'b c t h w -> b c (repeat t) h w', repeat=frames//2) + + #img_tensor2 = torch.from_numpy(image2).permute(2, 0, 1).float().to(model.device) + img_tensor2 = torch.from_numpy(image2).permute(2, 0, 1).float().half().to(model.device) + img_tensor2 = (img_tensor2 / 255. - 0.5) * 2 + image_tensor_resized2 = transform(img_tensor2) #3,h,w + videos2 = image_tensor_resized2.unsqueeze(0).unsqueeze(2) # bchw + videos2 = repeat(videos2, 'b c t h w -> b c (repeat t) h w', repeat=frames//2) + + + videos = torch.cat([videos, videos2], dim=2) + z, hs = self.get_latent_z_with_hidden_states(model, videos) + + img_tensor_repeat = torch.zeros_like(z) + + img_tensor_repeat[:,:,:1,:,:] = z[:,:,:1,:,:] + img_tensor_repeat[:,:,-1:,:,:] = z[:,:,-1:,:,:] + + + cond_images = model.embedder(img_tensor.unsqueeze(0)) ## blc + img_emb = model.image_proj_model(cond_images) + + imtext_cond = torch.cat([text_emb, img_emb], dim=1) + + fs = torch.tensor([fs], dtype=torch.long, device=model.device) + cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat], "control_cond": cn_videos} + + ## inference + batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale, hs=hs) + + ## remove the last frame + if image2 is None: + batch_samples = batch_samples[:,:,:,:-1,...] + ## b,samples,c,t,h,w + prompt_str = prompt.replace("/", "_slash_") if "/" in prompt else prompt + prompt_str = prompt_str.replace(" ", "_") if " " in prompt else prompt_str + prompt_str=prompt_str[:40] + if len(prompt_str) == 0: + prompt_str = 'empty_prompt' + + save_videos(batch_samples, self.result_dir, filenames=[prompt_str], fps=self.save_fps) + print(f"Saved in {prompt_str}. Time used: {(time.time() - start):.2f} seconds") + model = model.cpu() + return os.path.join(self.result_dir, f"{prompt_str}.mp4") + + def download_model(self): + REPO_ID = 'Doubiiu/ToonCrafter' + filename_list = ['model.ckpt'] + if not os.path.exists('./checkpoints/tooncrafter_'+str(self.resolution[1])+'_interp_v1/'): + os.makedirs('./checkpoints/tooncrafter_'+str(self.resolution[1])+'_interp_v1/') + for filename in filename_list: + local_file = os.path.join('./checkpoints/tooncrafter_'+str(self.resolution[1])+'_interp_v1/', filename) + if not os.path.exists(local_file): + hf_hub_download(repo_id=REPO_ID, filename=filename, local_dir='./checkpoints/tooncrafter_'+str(self.resolution[1])+'_interp_v1/', local_dir_use_symlinks=False) + + def get_latent_z_with_hidden_states(self, model, videos): + b, c, t, h, w = videos.shape + x = rearrange(videos, 'b c t h w -> (b t) c h w') + encoder_posterior, hidden_states = model.first_stage_model.encode(x, return_hidden_states=True) + + hidden_states_first_last = [] + ### use only the first and last hidden states + for hid in hidden_states: + hid = rearrange(hid, '(b t) c h w -> b c t h w', t=t) + hid_new = torch.cat([hid[:, :, 0:1], hid[:, :, -1:]], dim=2) + hidden_states_first_last.append(hid_new) + + z = model.get_first_stage_encoding(encoder_posterior).detach() + z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t) + return z, hidden_states_first_last +if __name__ == '__main__': + i2v = Image2Video() + video_path = i2v.get_image('prompts/art.png','man fishing in a boat at sunset') + print('done', video_path) \ No newline at end of file diff --git a/scripts/gradio/i2v_test_application_pre.py b/scripts/gradio/i2v_test_application_pre.py new file mode 100644 index 0000000000000000000000000000000000000000..227578fad66e46b06a99b195216ce2bb775ccbba --- /dev/null +++ b/scripts/gradio/i2v_test_application_pre.py @@ -0,0 +1,212 @@ +import os +import time +from omegaconf import OmegaConf +import torch +from scripts.evaluation.funcs import load_model_checkpoint, save_videos, batch_ddim_sampling, get_latent_z +from utils.utils import instantiate_from_config +from huggingface_hub import hf_hub_download +from einops import repeat +import torchvision.transforms as transforms +from pytorch_lightning import seed_everything +from einops import rearrange +from cldm.model import load_state_dict +import cv2 + + +def extract_frames(video_path): + # 動画フゑむルをθͺ­γΏθΎΌγ‚€ + cap = cv2.VideoCapture(video_path) + + frame_list = [] + frame_num = 0 + + while True: + # フレームをθͺ­γΏθΎΌγ‚€ + ret, frame = cap.read() + if not ret: + break + + # フレームをγƒͺγ‚Ήγƒˆγ«θΏ½εŠ  + frame_list.append(frame) + frame_num += 1 + + # ε‹•η”»γƒ•γ‚‘γ‚€γƒ«γ‚’ι–‰γ˜γ‚‹ + cap.release() + + return frame_list + +class Image2Video(): + def __init__(self,result_dir='./tmp/',gpu_num=1,resolution='256_256') -> None: + self.resolution = (int(resolution.split('_')[0]), int(resolution.split('_')[1])) #hw + self.download_model() + + self.result_dir = result_dir + if not os.path.exists(self.result_dir): + os.mkdir(self.result_dir) + + #ToonCrafterModel + ckpt_path='checkpoints/tooncrafter_'+resolution.split('_')[1]+'_interp_v1/model.ckpt' + config_file='configs/inference_'+resolution.split('_')[1]+'_v1.0.yaml' + config = OmegaConf.load(config_file) + model_config = config.pop("model", OmegaConf.create()) + model_config['params']['unet_config']['params']['use_checkpoint']=False + + #ControlModel + cn_ckpt_path = "control_models/sketch_encoder.ckpt" + cn_config_file = 'configs/cldm_v21.yaml' + cn_config = OmegaConf.load(cn_config_file) + cn_model_config = cn_config.pop("control_stage_config", OmegaConf.create()) + + + model_list = [] + for gpu_id in range(gpu_num): + model = instantiate_from_config(model_config) + cn_model = instantiate_from_config(cn_model_config) + + # model = model.cuda(gpu_id) + assert os.path.exists(ckpt_path), "Error: checkpoint Not Found!" + model = load_model_checkpoint(model, ckpt_path) + model.eval() + + cn_model.load_state_dict(load_state_dict(cn_ckpt_path, location='cuda')) + cn_model.eval() + + model.control_model = cn_model + + model_list.append(model) + self.model_list = model_list + self.save_fps = 8 + + def get_image(self, image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123, image2=None, frame_guides=None,control_scale=0.6): + control_frames = extract_frames(frame_guides) + seed_everything(seed) + transform = transforms.Compose([ + transforms.Resize(min(self.resolution)), + transforms.CenterCrop(self.resolution), + ]) + torch.cuda.empty_cache() + print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S',time.localtime(time.time()))) + start = time.time() + gpu_id=0 + if steps > 60: + steps = 60 + model = self.model_list[gpu_id] + model = model.cuda() + batch_size=1 + channels = model.model.diffusion_model.out_channels + frames = model.temporal_length + h, w = self.resolution[0] // 8, self.resolution[1] // 8 + noise_shape = [batch_size, channels, frames, h, w] + + # text cond + with torch.no_grad(), torch.cuda.amp.autocast(): + text_emb = model.get_learned_conditioning([prompt]) + + #control cond + if frame_guides is not None: + cn_videos = [] + for frame in control_frames: + frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) + frame = cv2.bitwise_not(frame) + cn_tensor = torch.from_numpy(frame).unsqueeze(2).permute(2, 0, 1).float().to(model.device) + + #cn_tensor = (cn_tensor / 255. - 0.5) * 2 + cn_tensor = ( cn_tensor/255.0 ) + cn_tensor_resized = transform(cn_tensor) #3,h,w + + cn_video = cn_tensor_resized.unsqueeze(0).unsqueeze(2) # bc1hw + cn_videos.append(cn_video) + + cn_videos = torch.cat(cn_videos, dim=2) + model_list = [] + for model in self.model_list: + model.control_scale = control_scale + model_list.append(model) + self.model_list = model_list + + else: + cn_videos = None + + + + # img cond + img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device) + img_tensor = (img_tensor / 255. - 0.5) * 2 + + image_tensor_resized = transform(img_tensor) #3,h,w + videos = image_tensor_resized.unsqueeze(0).unsqueeze(2) # bc1hw + + # z = get_latent_z(model, videos) #bc,1,hw + videos = repeat(videos, 'b c t h w -> b c (repeat t) h w', repeat=frames//2) + + img_tensor2 = torch.from_numpy(image2).permute(2, 0, 1).float().to(model.device) + img_tensor2 = (img_tensor2 / 255. - 0.5) * 2 + image_tensor_resized2 = transform(img_tensor2) #3,h,w + videos2 = image_tensor_resized2.unsqueeze(0).unsqueeze(2) # bchw + videos2 = repeat(videos2, 'b c t h w -> b c (repeat t) h w', repeat=frames//2) + + + videos = torch.cat([videos, videos2], dim=2) + z, hs = self.get_latent_z_with_hidden_states(model, videos) + + img_tensor_repeat = torch.zeros_like(z) + + img_tensor_repeat[:,:,:1,:,:] = z[:,:,:1,:,:] + img_tensor_repeat[:,:,-1:,:,:] = z[:,:,-1:,:,:] + + + cond_images = model.embedder(img_tensor.unsqueeze(0)) ## blc + img_emb = model.image_proj_model(cond_images) + + imtext_cond = torch.cat([text_emb, img_emb], dim=1) + + fs = torch.tensor([fs], dtype=torch.long, device=model.device) + cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat], "control_cond": cn_videos} + + ## inference + batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale, hs=hs) + + ## remove the last frame + if image2 is None: + batch_samples = batch_samples[:,:,:,:-1,...] + ## b,samples,c,t,h,w + prompt_str = prompt.replace("/", "_slash_") if "/" in prompt else prompt + prompt_str = prompt_str.replace(" ", "_") if " " in prompt else prompt_str + prompt_str=prompt_str[:40] + if len(prompt_str) == 0: + prompt_str = 'empty_prompt' + + save_videos(batch_samples, self.result_dir, filenames=[prompt_str], fps=self.save_fps) + print(f"Saved in {prompt_str}. Time used: {(time.time() - start):.2f} seconds") + model = model.cpu() + return os.path.join(self.result_dir, f"{prompt_str}.mp4") + + def download_model(self): + REPO_ID = 'Doubiiu/ToonCrafter' + filename_list = ['model.ckpt'] + if not os.path.exists('./checkpoints/tooncrafter_'+str(self.resolution[1])+'_interp_v1/'): + os.makedirs('./checkpoints/tooncrafter_'+str(self.resolution[1])+'_interp_v1/') + for filename in filename_list: + local_file = os.path.join('./checkpoints/tooncrafter_'+str(self.resolution[1])+'_interp_v1/', filename) + if not os.path.exists(local_file): + hf_hub_download(repo_id=REPO_ID, filename=filename, local_dir='./checkpoints/tooncrafter_'+str(self.resolution[1])+'_interp_v1/', local_dir_use_symlinks=False) + + def get_latent_z_with_hidden_states(self, model, videos): + b, c, t, h, w = videos.shape + x = rearrange(videos, 'b c t h w -> (b t) c h w') + encoder_posterior, hidden_states = model.first_stage_model.encode(x, return_hidden_states=True) + + hidden_states_first_last = [] + ### use only the first and last hidden states + for hid in hidden_states: + hid = rearrange(hid, '(b t) c h w -> b c t h w', t=t) + hid_new = torch.cat([hid[:, :, 0:1], hid[:, :, -1:]], dim=2) + hidden_states_first_last.append(hid_new) + + z = model.get_first_stage_encoding(encoder_posterior).detach() + z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t) + return z, hidden_states_first_last +if __name__ == '__main__': + i2v = Image2Video() + video_path = i2v.get_image('prompts/art.png','man fishing in a boat at sunset') + print('done', video_path) diff --git a/scripts/run.sh b/scripts/run.sh new file mode 100644 index 0000000000000000000000000000000000000000..490fb70c068ecdf323a99117367401ee4ab4c18e --- /dev/null +++ b/scripts/run.sh @@ -0,0 +1,28 @@ + +ckpt=checkpoints/tooncrafter_512_interp_v1/model.ckpt +config=configs/inference_512_v1.0.yaml + +prompt_dir=prompts/512_interp/ +res_dir="results" + +FS=10 ## This model adopts FPS=5, range recommended: 5-30 (smaller value -> larger motion) + + + +seed=123 +name=tooncrafter_512_interp_seed${seed} +CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/inference.py \ +--seed ${seed} \ +--ckpt_path $ckpt \ +--config $config \ +--savedir $res_dir/$name \ +--n_samples 1 \ +--bs 1 --height 320 --width 512 \ +--unconditional_guidance_scale 7.5 \ +--ddim_steps 50 \ +--ddim_eta 1.0 \ +--prompt_dir $prompt_dir \ +--text_input \ +--video_length 16 \ +--frame_stride ${FS} \ +--timestep_spacing 'uniform_trailing' --guidance_rescale 0.7 --perframe_ae --interp diff --git a/utils/save_video.py b/utils/save_video.py new file mode 100644 index 0000000000000000000000000000000000000000..3915f57a7790f3f3a59c513d59794d52158658d4 --- /dev/null +++ b/utils/save_video.py @@ -0,0 +1,195 @@ +import os +import numpy as np +from tqdm import tqdm +from PIL import Image +from einops import rearrange + +import torch +import torchvision +from torch import Tensor +from torchvision.utils import make_grid +from torchvision.transforms.functional import to_tensor + + +def frames_to_mp4(frame_dir,output_path,fps): + def read_first_n_frames(d: os.PathLike, num_frames: int): + if num_frames: + images = [Image.open(os.path.join(d, f)) for f in sorted(os.listdir(d))[:num_frames]] + else: + images = [Image.open(os.path.join(d, f)) for f in sorted(os.listdir(d))] + images = [to_tensor(x) for x in images] + return torch.stack(images) + videos = read_first_n_frames(frame_dir, num_frames=None) + videos = videos.mul(255).to(torch.uint8).permute(0, 2, 3, 1) + torchvision.io.write_video(output_path, videos, fps=fps, video_codec='h264', options={'crf': '10'}) + + +def tensor_to_mp4(video, savepath, fps, rescale=True, nrow=None): + """ + video: torch.Tensor, b,c,t,h,w, 0-1 + if -1~1, enable rescale=True + """ + n = video.shape[0] + video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w + nrow = int(np.sqrt(n)) if nrow is None else nrow + frame_grids = [torchvision.utils.make_grid(framesheet, nrow=nrow, padding=0) for framesheet in video] # [3, grid_h, grid_w] + grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [T, 3, grid_h, grid_w] + grid = torch.clamp(grid.float(), -1., 1.) + if rescale: + grid = (grid + 1.0) / 2.0 + grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1) # [T, 3, grid_h, grid_w] -> [T, grid_h, grid_w, 3] + torchvision.io.write_video(savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'}) + + +def tensor2videogrids(video, root, filename, fps, rescale=True, clamp=True): + assert(video.dim() == 5) # b,c,t,h,w + assert(isinstance(video, torch.Tensor)) + + video = video.detach().cpu() + if clamp: + video = torch.clamp(video, -1., 1.) + n = video.shape[0] + video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w + frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(np.sqrt(n))) for framesheet in video] # [3, grid_h, grid_w] + grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [T, 3, grid_h, grid_w] + if rescale: + grid = (grid + 1.0) / 2.0 + grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1) # [T, 3, grid_h, grid_w] -> [T, grid_h, grid_w, 3] + path = os.path.join(root, filename) + torchvision.io.write_video(path, grid, fps=fps, video_codec='h264', options={'crf': '10'}) + + +def log_local(batch_logs, save_dir, filename, save_fps=10, rescale=True): + if batch_logs is None: + return None + """ save images and videos from images dict """ + def save_img_grid(grid, path, rescale): + if rescale: + grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w + grid = grid.transpose(0, 1).transpose(1, 2).squeeze(-1) + grid = grid.numpy() + grid = (grid * 255).astype(np.uint8) + os.makedirs(os.path.split(path)[0], exist_ok=True) + Image.fromarray(grid).save(path) + + for key in batch_logs: + value = batch_logs[key] + if isinstance(value, list) and isinstance(value[0], str): + ## a batch of captions + path = os.path.join(save_dir, "%s-%s.txt"%(key, filename)) + with open(path, 'w') as f: + for i, txt in enumerate(value): + f.write(f'idx={i}, txt={txt}\n') + f.close() + elif isinstance(value, torch.Tensor) and value.dim() == 5: + ## save video grids + video = value # b,c,t,h,w + ## only save grayscale or rgb mode + if video.shape[1] != 1 and video.shape[1] != 3: + continue + n = video.shape[0] + video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w + frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(1), padding=0) for framesheet in video] #[3, n*h, 1*w] + grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w] + if rescale: + grid = (grid + 1.0) / 2.0 + grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1) + path = os.path.join(save_dir, "%s-%s.mp4"%(key, filename)) + torchvision.io.write_video(path, grid, fps=save_fps, video_codec='h264', options={'crf': '10'}) + + ## save frame sheet + img = value + video_frames = rearrange(img, 'b c t h w -> (b t) c h w') + t = img.shape[2] + grid = torchvision.utils.make_grid(video_frames, nrow=t, padding=0) + path = os.path.join(save_dir, "%s-%s.jpg"%(key, filename)) + #save_img_grid(grid, path, rescale) + elif isinstance(value, torch.Tensor) and value.dim() == 4: + ## save image grids + img = value + ## only save grayscale or rgb mode + if img.shape[1] != 1 and img.shape[1] != 3: + continue + n = img.shape[0] + grid = torchvision.utils.make_grid(img, nrow=1, padding=0) + path = os.path.join(save_dir, "%s-%s.jpg"%(key, filename)) + save_img_grid(grid, path, rescale) + else: + pass + +def prepare_to_log(batch_logs, max_images=100000, clamp=True): + if batch_logs is None: + return None + # process + for key in batch_logs: + N = batch_logs[key].shape[0] if hasattr(batch_logs[key], 'shape') else len(batch_logs[key]) + N = min(N, max_images) + batch_logs[key] = batch_logs[key][:N] + ## in batch_logs: images & caption + if isinstance(batch_logs[key], torch.Tensor): + batch_logs[key] = batch_logs[key].detach().cpu() + if clamp: + try: + batch_logs[key] = torch.clamp(batch_logs[key].float(), -1., 1.) + except RuntimeError: + print("clamp_scalar_cpu not implemented for Half") + return batch_logs + +# ---------------------------------------------------------------------------------------------- + +def fill_with_black_squares(video, desired_len: int) -> Tensor: + if len(video) >= desired_len: + return video + + return torch.cat([ + video, + torch.zeros_like(video[0]).unsqueeze(0).repeat(desired_len - len(video), 1, 1, 1), + ], dim=0) + +# ---------------------------------------------------------------------------------------------- +def load_num_videos(data_path, num_videos): + # first argument can be either data_path of np array + if isinstance(data_path, str): + videos = np.load(data_path)['arr_0'] # NTHWC + elif isinstance(data_path, np.ndarray): + videos = data_path + else: + raise Exception + + if num_videos is not None: + videos = videos[:num_videos, :, :, :, :] + return videos + +def npz_to_video_grid(data_path, out_path, num_frames, fps, num_videos=None, nrow=None, verbose=True): + # videos = torch.tensor(np.load(data_path)['arr_0']).permute(0,1,4,2,3).div_(255).mul_(2) - 1.0 # NTHWC->NTCHW, np int -> torch tensor 0-1 + if isinstance(data_path, str): + videos = load_num_videos(data_path, num_videos) + elif isinstance(data_path, np.ndarray): + videos = data_path + else: + raise Exception + n,t,h,w,c = videos.shape + videos_th = [] + for i in range(n): + video = videos[i, :,:,:,:] + images = [video[j, :,:,:] for j in range(t)] + images = [to_tensor(img) for img in images] + video = torch.stack(images) + videos_th.append(video) + if verbose: + videos = [fill_with_black_squares(v, num_frames) for v in tqdm(videos_th, desc='Adding empty frames')] # NTCHW + else: + videos = [fill_with_black_squares(v, num_frames) for v in videos_th] # NTCHW + + frame_grids = torch.stack(videos).permute(1, 0, 2, 3, 4) # [T, N, C, H, W] + if nrow is None: + nrow = int(np.ceil(np.sqrt(n))) + if verbose: + frame_grids = [make_grid(fs, nrow=nrow) for fs in tqdm(frame_grids, desc='Making grids')] + else: + frame_grids = [make_grid(fs, nrow=nrow) for fs in frame_grids] + + if os.path.dirname(out_path) != "": + os.makedirs(os.path.dirname(out_path), exist_ok=True) + frame_grids = (torch.stack(frame_grids) * 255).to(torch.uint8).permute(0, 2, 3, 1) # [T, H, W, C] + torchvision.io.write_video(out_path, frame_grids, fps=fps, video_codec='h264', options={'crf': '10'}) diff --git a/utils/utils.py b/utils/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..ef116e56f6f61dce0cf4942986ac659bf93fcd24 --- /dev/null +++ b/utils/utils.py @@ -0,0 +1,77 @@ +import importlib +import numpy as np +import cv2 +import torch +import torch.distributed as dist + + +def count_params(model, verbose=False): + total_params = sum(p.numel() for p in model.parameters()) + if verbose: + print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.") + return total_params + + +def check_istarget(name, para_list): + """ + name: full name of source para + para_list: partial name of target para + """ + istarget=False + for para in para_list: + if para in name: + return True + return istarget + + +def instantiate_from_config(config): + if not "target" in config: + if config == '__is_first_stage__': + return None + elif config == "__is_unconditional__": + return None + raise KeyError("Expected key `target` to instantiate.") + return get_obj_from_str(config["target"])(**config.get("params", dict())) + + +def get_obj_from_str(string, reload=False): + module, cls = string.rsplit(".", 1) + if reload: + module_imp = importlib.import_module(module) + importlib.reload(module_imp) + return getattr(importlib.import_module(module, package=None), cls) + + +def load_npz_from_dir(data_dir): + data = [np.load(os.path.join(data_dir, data_name))['arr_0'] for data_name in os.listdir(data_dir)] + data = np.concatenate(data, axis=0) + return data + + +def load_npz_from_paths(data_paths): + data = [np.load(data_path)['arr_0'] for data_path in data_paths] + data = np.concatenate(data, axis=0) + return data + + +def resize_numpy_image(image, max_resolution=512 * 512, resize_short_edge=None): + h, w = image.shape[:2] + if resize_short_edge is not None: + k = resize_short_edge / min(h, w) + else: + k = max_resolution / (h * w) + k = k**0.5 + h = int(np.round(h * k / 64)) * 64 + w = int(np.round(w * k / 64)) * 64 + image = cv2.resize(image, (w, h), interpolation=cv2.INTER_LANCZOS4) + return image + + +def setup_dist(args): + if dist.is_initialized(): + return + torch.cuda.set_device(args.local_rank) + torch.distributed.init_process_group( + 'nccl', + init_method='env://' + ) \ No newline at end of file