init

.gitignore

@@ -0,0 +1,143 @@
+.idea/
+examples/
+training/
+lightning_logs/
+image_log/
+
+*.pth
+*.pt
+*.ckpt
+*.safetensors
+
+gradio_pose2image_private.py
+gradio_canny2image_private.py
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/

LICENSE.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 DAMO Vision Intelligence Lab
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,13 @@
+---
+title: AnyDoor Online
+emoji: 👁
+colorFrom: green
+colorTo: blue
+sdk: gradio
+sdk_version: 4.12.0
+app_file: app.py
+pinned: false
+license: apache-2.0
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,297 @@
+import cv2
+import einops
+import numpy as np
+import torch
+import random
+import gradio as gr
+import os
+import albumentations as A
+from PIL import Image
+import torchvision.transforms as T
+from datasets.data_utils import * 
+from cldm.model import create_model, load_state_dict
+from cldm.ddim_hacked import DDIMSampler
+from omegaconf import OmegaConf
+from cldm.hack import disable_verbosity, enable_sliced_attention
+from huggingface_hub import snapshot_download
+
+snapshot_download(repo_id="xichenhku/AnyDoor_models", local_dir="./AnyDoor_models")
+
+
+cv2.setNumThreads(0)
+cv2.ocl.setUseOpenCL(False)
+
+save_memory = False
+disable_verbosity()
+if save_memory:
+    enable_sliced_attention()
+
+
+config = OmegaConf.load('./configs/demo.yaml')
+model_ckpt =  config.pretrained_model
+model_config = config.config_file
+
+
+
+
+model = create_model(model_config ).cpu()
+model.load_state_dict(load_state_dict(model_ckpt, location='cuda'))
+model = model.cuda()
+ddim_sampler = DDIMSampler(model)
+
+
+def crop_back( pred, tar_image,  extra_sizes, tar_box_yyxx_crop):
+    H1, W1, H2, W2 = extra_sizes
+    y1,y2,x1,x2 = tar_box_yyxx_crop    
+    pred = cv2.resize(pred, (W2, H2))
+    m = 3 # maigin_pixel
+
+    if W1 == H1:
+        tar_image[y1+m :y2-m, x1+m:x2-m, :] =  pred[m:-m, m:-m]
+        return tar_image
+
+    if W1 < W2:
+        pad1 = int((W2 - W1) / 2)
+        pad2 = W2 - W1 - pad1
+        pred = pred[:,pad1: -pad2, :]
+    else:
+        pad1 = int((H2 - H1) / 2)
+        pad2 = H2 - H1 - pad1
+        pred = pred[pad1: -pad2, :, :]
+    tar_image[y1+m :y2-m, x1+m:x2-m, :] =  pred[m:-m, m:-m]
+    return tar_image
+
+
+def inference_single_image(ref_image, 
+                           ref_mask, 
+                           tar_image, 
+                           tar_mask, 
+                           num_samples, 
+                           strength, 
+                           ddim_steps, 
+                           scale, 
+                           seed, 
+                           ):
+    raw_background = tar_image.copy()
+    item = process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+
+    ref = item['ref']
+    hint = item['hint']
+    num_samples = 1
+
+    control = torch.from_numpy(hint.copy()).float().cuda() 
+    control = torch.stack([control for _ in range(num_samples)], dim=0)
+    control = einops.rearrange(control, 'b h w c -> b c h w').clone()
+
+
+    clip_input = torch.from_numpy(ref.copy()).float().cuda() 
+    clip_input = torch.stack([clip_input for _ in range(num_samples)], dim=0)
+    clip_input = einops.rearrange(clip_input, 'b h w c -> b c h w').clone()
+
+    H,W = 512,512
+
+    cond = {"c_concat": [control], "c_crossattn": [model.get_learned_conditioning( clip_input )]}
+    un_cond = {"c_concat": [control], 
+               "c_crossattn": [model.get_learned_conditioning([torch.zeros((1,3,224,224))] * num_samples)]}
+    shape = (4, H // 8, W // 8)
+
+    if save_memory:
+        model.low_vram_shift(is_diffusing=True)
+
+    model.control_scales = ([strength] * 13)
+    samples, _ = ddim_sampler.sample(ddim_steps, num_samples,
+                                     shape, cond, verbose=False, eta=0,
+                                     unconditional_guidance_scale=scale,
+                                     unconditional_conditioning=un_cond)
+
+    if save_memory:
+        model.low_vram_shift(is_diffusing=False)
+
+    x_samples = model.decode_first_stage(samples)
+    x_samples = (einops.rearrange(x_samples, 'b c h w -> b h w c') * 127.5 + 127.5).cpu().numpy()
+
+    result = x_samples[0][:,:,::-1]
+    result = np.clip(result,0,255)
+
+    pred = x_samples[0]
+    pred = np.clip(pred,0,255)[1:,:,:]
+    sizes = item['extra_sizes']
+    tar_box_yyxx_crop = item['tar_box_yyxx_crop'] 
+    tar_image = crop_back(pred, tar_image, sizes, tar_box_yyxx_crop) 
+
+    # keep background unchanged
+    y1,y2,x1,x2 = item['tar_box_yyxx']
+    raw_background[y1:y2, x1:x2, :] = tar_image[y1:y2, x1:x2, :]
+    return raw_background
+
+
+def process_pairs(ref_image, ref_mask, tar_image, tar_mask, max_ratio = 0.8):
+    # ========= Reference ===========
+    # ref expand 
+    ref_box_yyxx = get_bbox_from_mask(ref_mask)
+
+    # ref filter mask 
+    ref_mask_3 = np.stack([ref_mask,ref_mask,ref_mask],-1)
+    masked_ref_image = ref_image * ref_mask_3 + np.ones_like(ref_image) * 255 * (1-ref_mask_3)
+
+    y1,y2,x1,x2 = ref_box_yyxx
+    masked_ref_image = masked_ref_image[y1:y2,x1:x2,:]
+    ref_mask = ref_mask[y1:y2,x1:x2]
+
+    ratio = np.random.randint(11, 15) / 10 #11,13
+    masked_ref_image, ref_mask = expand_image_mask(masked_ref_image, ref_mask, ratio=ratio)
+    ref_mask_3 = np.stack([ref_mask,ref_mask,ref_mask],-1)
+
+    # to square and resize
+    masked_ref_image = pad_to_square(masked_ref_image, pad_value = 255, random = False)
+    masked_ref_image = cv2.resize(masked_ref_image.astype(np.uint8), (224,224) ).astype(np.uint8)
+
+    ref_mask_3 = pad_to_square(ref_mask_3 * 255, pad_value = 0, random = False)
+    ref_mask_3 = cv2.resize(ref_mask_3.astype(np.uint8), (224,224) ).astype(np.uint8)
+    ref_mask = ref_mask_3[:,:,0]
+
+    # collage aug 
+    masked_ref_image_compose, ref_mask_compose =  masked_ref_image, ref_mask
+    ref_mask_3 = np.stack([ref_mask_compose,ref_mask_compose,ref_mask_compose],-1)
+    ref_image_collage = sobel(masked_ref_image_compose, ref_mask_compose/255)
+
+    # ========= Target ===========
+    tar_box_yyxx = get_bbox_from_mask(tar_mask)
+    tar_box_yyxx = expand_bbox(tar_mask, tar_box_yyxx, ratio=[1.1,1.2]) #1.1  1.3
+    tar_box_yyxx_full = tar_box_yyxx
+    
+    # crop
+    tar_box_yyxx_crop =  expand_bbox(tar_image, tar_box_yyxx, ratio=[1.3, 3.0])   
+    tar_box_yyxx_crop = box2squre(tar_image, tar_box_yyxx_crop) # crop box
+    y1,y2,x1,x2 = tar_box_yyxx_crop
+
+    cropped_target_image = tar_image[y1:y2,x1:x2,:]
+    cropped_tar_mask = tar_mask[y1:y2,x1:x2]
+
+    tar_box_yyxx = box_in_box(tar_box_yyxx, tar_box_yyxx_crop)
+    y1,y2,x1,x2 = tar_box_yyxx
+
+    # collage
+    ref_image_collage = cv2.resize(ref_image_collage.astype(np.uint8), (x2-x1, y2-y1))
+    ref_mask_compose = cv2.resize(ref_mask_compose.astype(np.uint8), (x2-x1, y2-y1))
+    ref_mask_compose = (ref_mask_compose > 128).astype(np.uint8)
+
+    collage = cropped_target_image.copy() 
+    collage[y1:y2,x1:x2,:] = ref_image_collage
+
+    collage_mask = cropped_target_image.copy() * 0.0
+    collage_mask[y1:y2,x1:x2,:] = 1.0
+    collage_mask = np.stack([cropped_tar_mask,cropped_tar_mask,cropped_tar_mask],-1)
+
+    # the size before pad
+    H1, W1 = collage.shape[0], collage.shape[1]
+
+    cropped_target_image = pad_to_square(cropped_target_image, pad_value = 0, random = False).astype(np.uint8)
+    collage = pad_to_square(collage, pad_value = 0, random = False).astype(np.uint8)
+    collage_mask = pad_to_square(collage_mask, pad_value = 0, random = False).astype(np.uint8)
+
+    # the size after pad
+    H2, W2 = collage.shape[0], collage.shape[1]
+
+    cropped_target_image = cv2.resize(cropped_target_image.astype(np.uint8), (512,512)).astype(np.float32)
+    collage = cv2.resize(collage.astype(np.uint8), (512,512)).astype(np.float32)
+    collage_mask  = (cv2.resize(collage_mask.astype(np.uint8), (512,512)).astype(np.float32) > 0.5).astype(np.float32)
+
+    masked_ref_image = masked_ref_image  / 255 
+    cropped_target_image = cropped_target_image / 127.5 - 1.0
+    collage = collage / 127.5 - 1.0 
+    collage = np.concatenate([collage, collage_mask[:,:,:1]  ] , -1)
+    
+    item = dict(ref=masked_ref_image.copy(), jpg=cropped_target_image.copy(), hint=collage.copy(), 
+                extra_sizes=np.array([H1, W1, H2, W2]), 
+                tar_box_yyxx_crop=np.array( tar_box_yyxx_crop ),
+                tar_box_yyxx=np.array(tar_box_yyxx_full),
+                 ) 
+    return item
+
+
+ref_dir='./examples/Gradio/FG'
+image_dir='./examples/Gradio/BG'
+ref_list=[os.path.join(ref_dir,file) for file in os.listdir(ref_dir) if '.jpg' in file or '.png' in file or '.jpeg' in file ]
+ref_list.sort()
+image_list=[os.path.join(image_dir,file) for file in os.listdir(image_dir) if '.jpg' in file or '.png' in file or '.jpeg' in file]
+image_list.sort()
+
+def mask_image(image, mask):
+    blanc = np.ones_like(image) * 255
+    mask = np.stack([mask,mask,mask],-1) / 255
+    masked_image = mask * ( 0.5 * blanc + 0.5 * image) + (1-mask) * image
+    return masked_image.astype(np.uint8)
+
+def run_local(base,
+              ref,
+              *args):
+    image = base["image"].convert("RGB")
+    mask = base["mask"].convert("L")
+    ref_image = ref["image"].convert("RGB")
+    ref_mask = ref["mask"].convert("L")
+    image = np.asarray(image)
+    mask = np.asarray(mask)
+    mask = np.where(mask > 128, 255, 0).astype(np.uint8)
+    ref_image = np.asarray(ref_image)
+    ref_mask = np.asarray(ref_mask)
+    ref_mask = np.where(ref_mask > 128, 1, 0).astype(np.uint8)
+
+    processed_item = process_pairs(ref_image.copy(), ref_mask.copy(), image.copy(), mask.copy(), max_ratio = 0.8)
+    masked_ref = (processed_item['ref']*255)
+
+    mased_image = mask_image(image, mask)
+    #synthesis = image
+    synthesis = inference_single_image(ref_image.copy(), ref_mask.copy(), image.copy(), mask.copy(), *args)
+    synthesis = torch.from_numpy(synthesis).permute(2, 0, 1)
+    synthesis = synthesis.permute(1, 2, 0).numpy()
+
+    masked_ref = cv2.resize(masked_ref.astype(np.uint8), (512,512))
+    return [synthesis]
+
+
+
+with gr.Blocks() as demo:
+    with gr.Column():
+        gr.Markdown("#  Play with AnyDoor to Teleport your Target Objects! ")
+        with gr.Row():
+            baseline_gallery = gr.Gallery(label='Output', show_label=True, elem_id="gallery", columns=1, height=768)
+            with gr.Accordion("Advanced Option", open=True):
+                num_samples = gr.Slider(label="Images", minimum=1, maximum=12, value=1, step=1)
+                strength = gr.Slider(label="Control Strength", minimum=0.0, maximum=2.0, value=1.0, step=0.01)
+                ddim_steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=30, step=1)
+                scale = gr.Slider(label="Guidance Scale", minimum=0.1, maximum=30.0, value=5.0, step=0.1)
+                seed = gr.Slider(label="Seed", minimum=-1, maximum=999999999, step=1, value=-1)
+                gr.Markdown(" Higher guidance-scale makes higher fidelity, while lower guidance-scale leads to more harmonized blending.")
+    
+
+
+
+
+        gr.Markdown("# Upload / Select Images for the Background (left) and Reference Object (right)")
+        gr.Markdown("### Your could draw coarse masks on the background to indicate the desired location and shape.")
+        gr.Markdown("### <u>Do not forget</u> to annotate the target object on the reference image.")
+        with gr.Row():
+            base = gr.Image(label="Background", source="upload", tool="sketch", type="pil", height=512, brush_color='#FFFFFF', mask_opacity=0.5)
+            ref = gr.Image(label="Reference", source="upload", tool="sketch", type="pil", height=512, brush_color='#FFFFFF', mask_opacity=0.5)
+        run_local_button = gr.Button(label="Generate", value="Run")
+
+        with gr.Row():
+            with gr.Column():
+                gr.Examples(image_list, inputs=[base],label="Examples - Background Image",examples_per_page=16)
+            with gr.Column():
+                gr.Examples(ref_list, inputs=[ref],label="Examples - Reference Object",examples_per_page=16)
+        
+    run_local_button.click(fn=run_local, 
+                           inputs=[base, 
+                                   ref, 
+                                   num_samples, 
+                                   strength, 
+                                   ddim_steps, 
+                                   scale, 
+                                   seed, 
+                                   ], 
+                           outputs=[baseline_gallery]
+                        )
+
+demo.launch(server_name="0.0.0.0")

cldm/cldm.py

@@ -0,0 +1,442 @@
+import einops
+import torch
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+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 UNetModel, TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock
+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=None, context=None, control=None, only_mid_control=False, **kwargs):
+        hs = []
+        with torch.no_grad():
+            t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+            emb = self.time_embed(t_emb)
+            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)
+
+        if control is not None:
+            h += control.pop()
+
+        for i, module in enumerate(self.output_blocks):
+            if only_mid_control or 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)
+
+        h = h.type(x.dtype)
+        return self.out(h)
+
+
+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) # 1,1280
+        
+        # 1,320,64,64
+        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:
+                # skip the first layer
+                h = guided_hint
+                guided_hint = None
+            else:
+                h_new = module(h, emb, context) 
+                h =  h_new 
+            outs.append(zero_conv(h, emb, context))
+
+        h_new = self.middle_block(h, emb, context)  
+        outs.append(self.middle_block_out(h_new, 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()
+        self.time_steps = batch['time_steps']
+        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):
+        uncond =  self.get_learned_conditioning([ torch.zeros((1,3,224,224)) ] * N)
+        return uncond
+
+    @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) 
+
+        # ==== visualize the shape mask or the high-frequency map ====
+        guide_mask = (c_cat[:,-1,:,:].unsqueeze(1) + 1) * 0.5
+        guide_mask = torch.cat([guide_mask,guide_mask,guide_mask],1)
+        HF_map  = c_cat[:,:3,:,:] #* 2.0 - 1.0
+
+        log["control"] = HF_map
+
+        cond_image = batch[self.cond_stage_key].cpu().numpy().copy() 
+        log["conditioning"] = torch.permute( torch.tensor(cond_image), (0,3,1,2)) * 2.0 - 1.0  
+        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 #* 2.0 - 1.0
+        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())
+        params += list(self.cond_stage_model.projector.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()

cldm/ddim_hacked.py

@@ -0,0 +1,318 @@
+"""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]
+        #x_T 1,4,64,64
+        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

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)

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")

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

configs/anydoor.yaml

@@ -0,0 +1,85 @@
+model:
+  target: cldm.cldm.ControlLDM
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "ref"
+    control_key: "hint"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+    only_mid_control: False
+
+    control_stage_config:
+      target: cldm.cldm.ControlNet
+      params:
+        use_checkpoint: True
+        image_size: 32 # unused
+        in_channels: 4 
+        hint_channels: 4 #3
+        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
+
+    unet_config:
+      target: cldm.cldm.ControlledUnetModel
+      params:
+        use_checkpoint: True
+        image_size: 32 # unused
+        in_channels: 4
+        out_channels: 4
+        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
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          #attn_type: "vanilla-xformers"
+          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: ldm.modules.encoders.modules.FrozenDinoV2Encoder
+      weight: path/dinov2_vitg14_pretrain.pth
+    
+

configs/datasets.yaml

@@ -0,0 +1,68 @@
+Train:
+  YoutubeVOS:
+    image_dir: path/YTBVOS/train/JPEGImages/
+    anno: path/YTBVOS/train/Annotations
+    meta: path/YTBVOS/train/meta.json
+
+  YoutubeVIS:
+    image_dir: path/youtubevis/train/JPEGImages/
+    anno: path/youtubevis/train/Annotations/
+    meta: path/youtubevis/train/meta.json
+
+  VIPSeg:
+    image_dir: path/VIPSeg/VIPSeg_720P/images/
+    anno: path/VIPSeg/VIPSeg_720P/panomasksRGB/
+
+  UVO:
+    train:
+      image_dir: path/UVO/uvo_frames_sparse
+      video_json: path/UVO/UVO_sparse_train_video_with_interpolation.json
+      image_json: path/UVO/UVO_sparse_train_video_with_interpolation_reorg.json
+    val:
+      image_dir: path/UVO/uvo_frames_sparse
+      video_json: path/UVO/VideoSparseSet/UVO_sparse_val_video_with_interpolation.json
+      image_json: path/UVO/VideoSparseSet/UVO_sparse_val_video_interpolation_reorg.json
+
+  Mose:
+    image_dir: path/MOSE/train/JPEGImages/
+    anno: path/MOSE/train/Annotations/
+
+  MVImageNet:
+    txt: ./datasets/Preprocess/mvimagenet.txt
+    image_dir: /mnt/workspace/xizhi/data/MVImgNet/
+
+  VitonHD:
+    image_dir: path/TryOn/VitonHD/train/cloth/
+
+  Dresscode:
+    image_dir: /mnt/workspace/xizhi/data/dresscode/DressCode/upper_body/label_maps/
+
+  FashionTryon:
+    image_dir: path/TryOn/FashionTryOn/train
+
+  Lvis:
+    image_dir: path/COCO/train2017
+    json_path: path/lvis_v1/lvis_v1_train.json
+
+  SAM:
+    sub1: path/SAM/0000
+    sub2: path/SAM/0001
+    sub3: path/SAM/0002
+    sub4: path/SAM/0004
+
+  Saliency:
+    MSRA_root: path/Saliency/MSRA10K_Imgs_GT/
+    TR_root: path/Saliency/DUTS-TR/DUTS-TR-Image/
+    TE_root: path/Saliency/DUTS-TE/DUTS-TE-Image/
+    HFlickr_root: path/HFlickr/masks/
+
+Test:
+  DreamBooth:
+    fg_dir: path/DreamBooth/AnyDoor_DreamBooth
+    bg_dir: path/DreamBooth/v1_800
+
+  VitonHDTest:
+    image_dir: path/TryOn/VitonHD/test/cloth
+
+
+

configs/demo.yaml

@@ -0,0 +1,3 @@
+pretrained_model: ./AnyDoor_models/general_v0.1/general_v0.1.ckpt
+config_file: configs/anydoor.yaml
+save_memory: False

configs/inference.yaml

@@ -0,0 +1,3 @@
+pretrained_model: path/epoch=1-step=8687.ckpt
+config_file: configs/anydoor.yaml
+save_memory: False

datasets/Preprocess/uvo_process.py

@@ -0,0 +1,29 @@
+import cv2 
+import json 
+import os
+from pycocotools import mask as mask_utils
+import numpy as np
+from tqdm import tqdm
+
+json_path = 'path/UVO/UVO_sparse_train_video_with_interpolation.json'
+output_path = "path/UVO/UVO_sparse_train_video_with_interpolation_reorg.json"
+
+with open(json_path, 'r') as fcc_file:
+    data = json.load(fcc_file)
+
+info = data['info']
+videos = data['videos']
+print(len(videos))
+
+
+uvo_dict = {}
+for video in tqdm(videos):
+    vid = video['id']
+    file_names = video['file_names']
+    uvo_dict[vid] = file_names
+
+
+with open(output_path,"w") as f:
+    json.dump(uvo_dict,f)
+    print('finish')
+

datasets/base.py

@@ -0,0 +1,220 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+cv2.setNumThreads(0)
+cv2.ocl.setUseOpenCL(False)
+import albumentations as A
+
+
+class BaseDataset(Dataset):
+    def __init__(self):
+        image_mask_dict = {}
+        self.data = []
+
+    def __len__(self):
+        # We adjust the ratio of different dataset by setting the length.
+        pass
+
+    
+    def aug_data_back(self, image):
+        transform = A.Compose([
+            A.ColorJitter(p=0.5, brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5),
+            A.ChannelShuffle()
+            ])
+        transformed = transform(image=image.astype(np.uint8))
+        transformed_image = transformed["image"]
+        return transformed_image
+    
+    def aug_data_mask(self, image, mask):
+        transform = A.Compose([
+            A.HorizontalFlip(p=0.5),
+            A.RandomBrightnessContrast(p=0.5),
+            #A.Rotate(limit=20, border_mode=cv2.BORDER_CONSTANT,  value=(0,0,0)),
+            ])
+
+        transformed = transform(image=image.astype(np.uint8), mask = mask)
+        transformed_image = transformed["image"]
+        transformed_mask = transformed["mask"]
+        return transformed_image, transformed_mask
+
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H or w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H or w < W:
+                pass_flag = False
+        return pass_flag
+
+
+    def __getitem__(self, idx):
+        while(True):
+            try:
+                idx = np.random.randint(0, len(self.data)-1)
+                item = self.get_sample(idx)
+                return item
+            except:
+                idx = np.random.randint(0, len(self.data)-1)
+                
+    def get_sample(self, idx):
+        # Implemented for each specific dataset
+        pass
+
+    def sample_timestep(self, max_step =1000):
+        if np.random.rand() < 0.3:
+            step = np.random.randint(0,max_step)
+            return np.array([step])
+
+        if self.dynamic == 1:
+            # coarse videos
+            step_start = max_step // 2
+            step_end = max_step
+        elif self.dynamic == 0:
+            # static images
+            step_start = 0 
+            step_end = max_step // 2
+        else:
+            # fine multi-view images/videos/3Ds
+            step_start = 0
+            step_end = max_step
+        step = np.random.randint(step_start, step_end)
+        return np.array([step])
+
+    def check_mask_area(self, mask):
+        H,W = mask.shape[0], mask.shape[1]
+        ratio = mask.sum() / (H * W)
+        if ratio > 0.8 * 0.8  or ratio < 0.1 * 0.1:
+            return False
+        else:
+            return True 
+    
+
+    def process_pairs(self, ref_image, ref_mask, tar_image, tar_mask, max_ratio = 0.8):
+        assert mask_score(ref_mask) > 0.90
+        assert self.check_mask_area(ref_mask) == True
+        assert self.check_mask_area(tar_mask)  == True
+
+        # ========= Reference ===========
+        '''
+        # similate the case that the mask for reference object is coarse. Seems useless :(
+
+        if np.random.uniform(0, 1) < 0.7: 
+            ref_mask_clean = ref_mask.copy()
+            ref_mask_clean = np.stack([ref_mask_clean,ref_mask_clean,ref_mask_clean],-1)
+            ref_mask = perturb_mask(ref_mask, 0.6, 0.9)
+            
+            # select a fake bg to avoid the background leakage
+            fake_target = tar_image.copy()
+            h,w = ref_image.shape[0], ref_image.shape[1]
+            fake_targe = cv2.resize(fake_target, (w,h))
+            fake_back = np.fliplr(np.flipud(fake_target))
+            fake_back = self.aug_data_back(fake_back)
+            ref_image = ref_mask_clean * ref_image + (1-ref_mask_clean) * fake_back
+        '''
+
+        # Get the outline Box of the reference image
+        ref_box_yyxx = get_bbox_from_mask(ref_mask)
+        assert self.check_region_size(ref_mask, ref_box_yyxx, ratio = 0.10, mode = 'min') == True
+        
+        # Filtering background for the reference image
+        ref_mask_3 = np.stack([ref_mask,ref_mask,ref_mask],-1)
+        masked_ref_image = ref_image * ref_mask_3 + np.ones_like(ref_image) * 255 * (1-ref_mask_3)
+
+        y1,y2,x1,x2 = ref_box_yyxx
+        masked_ref_image = masked_ref_image[y1:y2,x1:x2,:]
+        ref_mask = ref_mask[y1:y2,x1:x2]
+
+        ratio = np.random.randint(11, 15) / 10 
+        masked_ref_image, ref_mask = expand_image_mask(masked_ref_image, ref_mask, ratio=ratio)
+        ref_mask_3 = np.stack([ref_mask,ref_mask,ref_mask],-1)
+
+        # Padding reference image to square and resize to 224
+        masked_ref_image = pad_to_square(masked_ref_image, pad_value = 255, random = False)
+        masked_ref_image = cv2.resize(masked_ref_image.astype(np.uint8), (224,224) ).astype(np.uint8)
+
+        ref_mask_3 = pad_to_square(ref_mask_3 * 255, pad_value = 0, random = False)
+        ref_mask_3 = cv2.resize(ref_mask_3.astype(np.uint8), (224,224) ).astype(np.uint8)
+        ref_mask = ref_mask_3[:,:,0]
+
+        # Augmenting reference image
+        #masked_ref_image_aug = self.aug_data(masked_ref_image) 
+        
+        # Getting for high-freqency map
+        masked_ref_image_compose, ref_mask_compose =  self.aug_data_mask(masked_ref_image, ref_mask) 
+        masked_ref_image_aug = masked_ref_image_compose.copy()
+
+        ref_mask_3 = np.stack([ref_mask_compose,ref_mask_compose,ref_mask_compose],-1)
+        ref_image_collage = sobel(masked_ref_image_compose, ref_mask_compose/255)
+        
+
+        # ========= Training Target ===========
+        tar_box_yyxx = get_bbox_from_mask(tar_mask)
+        tar_box_yyxx = expand_bbox(tar_mask, tar_box_yyxx, ratio=[1.1,1.2]) #1.1  1.3
+        assert self.check_region_size(tar_mask, tar_box_yyxx, ratio = max_ratio, mode = 'max') == True
+        
+        # Cropping around the target object 
+        tar_box_yyxx_crop =  expand_bbox(tar_image, tar_box_yyxx, ratio=[1.3, 3.0])   
+        tar_box_yyxx_crop = box2squre(tar_image, tar_box_yyxx_crop) # crop box
+        y1,y2,x1,x2 = tar_box_yyxx_crop
+        cropped_target_image = tar_image[y1:y2,x1:x2,:]
+        cropped_tar_mask = tar_mask[y1:y2,x1:x2]
+        tar_box_yyxx = box_in_box(tar_box_yyxx, tar_box_yyxx_crop)
+        y1,y2,x1,x2 = tar_box_yyxx
+
+        # Prepairing collage image
+        ref_image_collage = cv2.resize(ref_image_collage.astype(np.uint8), (x2-x1, y2-y1))
+        ref_mask_compose = cv2.resize(ref_mask_compose.astype(np.uint8), (x2-x1, y2-y1))
+        ref_mask_compose = (ref_mask_compose > 128).astype(np.uint8)
+
+        collage = cropped_target_image.copy() 
+        collage[y1:y2,x1:x2,:] = ref_image_collage
+
+        collage_mask = cropped_target_image.copy() * 0.0
+        collage_mask[y1:y2,x1:x2,:] = 1.0
+
+        if np.random.uniform(0, 1) < 0.7: 
+            cropped_tar_mask = perturb_mask(cropped_tar_mask)
+            collage_mask = np.stack([cropped_tar_mask,cropped_tar_mask,cropped_tar_mask],-1)
+
+        H1, W1 = collage.shape[0], collage.shape[1]
+
+        cropped_target_image = pad_to_square(cropped_target_image, pad_value = 0, random = False).astype(np.uint8)
+        collage = pad_to_square(collage, pad_value = 0, random = False).astype(np.uint8)
+        collage_mask = pad_to_square(collage_mask, pad_value = 2, random = False).astype(np.uint8)
+        H2, W2 = collage.shape[0], collage.shape[1]
+
+        cropped_target_image = cv2.resize(cropped_target_image.astype(np.uint8), (512,512)).astype(np.float32)
+        collage = cv2.resize(collage.astype(np.uint8), (512,512)).astype(np.float32)
+        collage_mask  = cv2.resize(collage_mask.astype(np.uint8), (512,512),  interpolation = cv2.INTER_NEAREST).astype(np.float32)
+        collage_mask[collage_mask == 2] = -1
+        
+        # Prepairing dataloader items
+        masked_ref_image_aug = masked_ref_image_aug  / 255 
+        cropped_target_image = cropped_target_image / 127.5 - 1.0
+        collage = collage / 127.5 - 1.0 
+        collage = np.concatenate([collage, collage_mask[:,:,:1]  ] , -1)
+        
+        item = dict(
+                ref=masked_ref_image_aug.copy(), 
+                jpg=cropped_target_image.copy(), 
+                hint=collage.copy(), 
+                extra_sizes=np.array([H1, W1, H2, W2]), 
+                tar_box_yyxx_crop=np.array(tar_box_yyxx_crop) 
+                ) 
+        return item
+
+
+
+
+

datasets/data_utils.py

@@ -0,0 +1,356 @@
+import numpy as np
+import torch 
+import cv2
+
+
+def mask_score(mask):
+    '''Scoring the mask according to connectivity.'''
+    mask = mask.astype(np.uint8)
+    if mask.sum() < 10:
+        return 0
+    contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
+    cnt_area = [cv2.contourArea(cnt) for cnt in contours]
+    conc_score = np.max(cnt_area) / sum(cnt_area)
+    return conc_score
+
+
+def sobel(img, mask, thresh = 50):
+    '''Calculating the high-frequency map.'''
+    H,W = img.shape[0], img.shape[1]
+    img = cv2.resize(img,(256,256))
+    mask = (cv2.resize(mask,(256,256)) > 0.5).astype(np.uint8)
+    kernel = np.ones((5,5),np.uint8)
+    mask = cv2.erode(mask, kernel, iterations = 2)
+    
+    Ksize = 3
+    sobelx = cv2.Sobel(img, cv2.CV_64F, 1, 0, ksize=Ksize)
+    sobely = cv2.Sobel(img, cv2.CV_64F, 0, 1, ksize=Ksize)
+    sobel_X = cv2.convertScaleAbs(sobelx)
+    sobel_Y = cv2.convertScaleAbs(sobely)
+    scharr = cv2.addWeighted(sobel_X, 0.5, sobel_Y, 0.5, 0)
+    scharr = np.max(scharr,-1) * mask    
+    
+    scharr[scharr < thresh] = 0.0
+    scharr = np.stack([scharr,scharr,scharr],-1)
+    scharr = (scharr.astype(np.float32)/255 * img.astype(np.float32) ).astype(np.uint8)
+    scharr = cv2.resize(scharr,(W,H))
+    return scharr
+
+
+def resize_and_pad(image, box):
+    '''Fitting an image to the box region while keeping the aspect ratio.'''
+    y1,y2,x1,x2 = box
+    H,W = y2-y1, x2-x1
+    h,w =  image.shape[0], image.shape[1]
+    r_box = W / H 
+    r_image = w / h
+    if r_box >= r_image:
+        h_target = H
+        w_target = int(w * H / h) 
+        image = cv2.resize(image, (w_target, h_target))
+
+        w1 = (W - w_target) // 2
+        w2 = W - w_target - w1
+        pad_param = ((0,0),(w1,w2),(0,0))
+        image = np.pad(image, pad_param, 'constant', constant_values=255)
+    else:
+        w_target = W 
+        h_target = int(h * W / w)
+        image = cv2.resize(image, (w_target, h_target))
+
+        h1 = (H-h_target) // 2 
+        h2 = H - h_target - h1
+        pad_param =((h1,h2),(0,0),(0,0))
+        image = np.pad(image, pad_param, 'constant', constant_values=255)
+    return image
+
+
+
+def expand_image_mask(image, mask, ratio=1.4):
+    h,w = image.shape[0], image.shape[1]
+    H,W = int(h * ratio), int(w * ratio) 
+    h1 = int((H - h) // 2)
+    h2 = H - h - h1
+    w1 = int((W -w) // 2)
+    w2 = W -w - w1
+
+    pad_param_image = ((h1,h2),(w1,w2),(0,0))
+    pad_param_mask = ((h1,h2),(w1,w2))
+    image = np.pad(image, pad_param_image, 'constant', constant_values=255)
+    mask = np.pad(mask, pad_param_mask, 'constant', constant_values=0)
+    return image, mask
+
+
+def resize_box(yyxx, H,W,h,w):
+    y1,y2,x1,x2 = yyxx
+    y1,y2 = int(y1/H * h), int(y2/H * h)
+    x1,x2 = int(x1/W * w), int(x2/W * w)
+    y1,y2 = min(y1,h), min(y2,h)
+    x1,x2 = min(x1,w), min(x2,w)
+    return (y1,y2,x1,x2)
+
+
+def get_bbox_from_mask(mask):
+    h,w = mask.shape[0],mask.shape[1]
+
+    if mask.sum() < 10:
+        return 0,h,0,w
+    rows = np.any(mask,axis=1)
+    cols = np.any(mask,axis=0)
+    y1,y2 = np.where(rows)[0][[0,-1]]
+    x1,x2 = np.where(cols)[0][[0,-1]]
+    return (y1,y2,x1,x2)
+
+
+def expand_bbox(mask,yyxx,ratio=[1.2,2.0], min_crop=0):
+    y1,y2,x1,x2 = yyxx
+    ratio = np.random.randint( ratio[0] * 10,  ratio[1] * 10 ) / 10
+    H,W = mask.shape[0], mask.shape[1]
+    xc, yc = 0.5 * (x1 + x2), 0.5 * (y1 + y2)
+    h = ratio * (y2-y1+1)
+    w = ratio * (x2-x1+1)
+    h = max(h,min_crop)
+    w = max(w,min_crop)
+
+    x1 = int(xc - w * 0.5)
+    x2 = int(xc + w * 0.5)
+    y1 = int(yc - h * 0.5)
+    y2 = int(yc + h * 0.5)
+
+    x1 = max(0,x1)
+    x2 = min(W,x2)
+    y1 = max(0,y1)
+    y2 = min(H,y2)
+    return (y1,y2,x1,x2)
+
+
+def box2squre(image, box):
+    H,W = image.shape[0], image.shape[1]
+    y1,y2,x1,x2 = box
+    cx = (x1 + x2) // 2
+    cy = (y1 + y2) // 2
+    h,w = y2-y1, x2-x1
+
+    if h >= w:
+        x1 = cx - h//2
+        x2 = cx + h//2
+    else:
+        y1 = cy - w//2
+        y2 = cy + w//2
+    x1 = max(0,x1)
+    x2 = min(W,x2)
+    y1 = max(0,y1)
+    y2 = min(H,y2)
+    return (y1,y2,x1,x2)
+
+
+def pad_to_square(image, pad_value = 255, random = False):
+    H,W = image.shape[0], image.shape[1]
+    if H == W:
+        return image
+
+    padd = abs(H - W)
+    if random:
+        padd_1 = int(np.random.randint(0,padd))
+    else:
+        padd_1 = int(padd / 2)
+    padd_2 = padd - padd_1
+
+    if H > W:
+        pad_param = ((0,0),(padd_1,padd_2),(0,0))
+    else:
+        pad_param = ((padd_1,padd_2),(0,0),(0,0))
+
+    image = np.pad(image, pad_param, 'constant', constant_values=pad_value)
+    return image
+
+
+
+def box_in_box(small_box, big_box):
+    y1,y2,x1,x2 = small_box
+    y1_b, _, x1_b, _ = big_box
+    y1,y2,x1,x2 = y1 - y1_b ,y2 - y1_b, x1 - x1_b ,x2 - x1_b
+    return (y1,y2,x1,x2 )
+
+
+
+def shuffle_image(image, N):
+    height, width = image.shape[:2]
+    
+    block_height = height // N
+    block_width = width // N
+    blocks = []
+    
+    for i in range(N):
+        for j in range(N):
+            block = image[i*block_height:(i+1)*block_height, j*block_width:(j+1)*block_width]
+            blocks.append(block)
+    
+    np.random.shuffle(blocks)
+    shuffled_image = np.zeros((height, width, 3), dtype=np.uint8)
+
+    for i in range(N):
+        for j in range(N):
+            shuffled_image[i*block_height:(i+1)*block_height, j*block_width:(j+1)*block_width] = blocks[i*N+j]
+    return shuffled_image
+
+
+def get_mosaic_mask(image, fg_mask, N=16, ratio = 0.5):
+    ids = [i for i in range(N * N)]
+    masked_number = int(N * N * ratio)
+    masked_id = np.random.choice(ids, masked_number, replace=False)
+    
+
+    
+    height, width = image.shape[:2]
+    mask = np.ones((height, width))
+    
+    block_height = height // N
+    block_width = width // N
+    
+    b_id = 0
+    for i in range(N):
+        for j in range(N):
+            if b_id in masked_id:
+                mask[i*block_height:(i+1)*block_height, j*block_width:(j+1)*block_width] = mask[i*block_height:(i+1)*block_height, j*block_width:(j+1)*block_width] * 0
+            b_id += 1
+    mask = mask * fg_mask
+    mask3 = np.stack([mask,mask,mask],-1).copy().astype(np.uint8)
+    noise = q_x(image)
+    noise_mask = image * mask3 + noise * (1-mask3)
+    return noise_mask
+
+def extract_canney_noise(image, mask, dilate=True):
+    h,w = image.shape[0],image.shape[1]
+    mask = cv2.resize(mask.astype(np.uint8),(w,h)) > 0.5
+    kernel = np.ones((8, 8), dtype=np.uint8)
+    mask =  cv2.erode(mask.astype(np.uint8), kernel, 10)
+
+    canny = cv2.Canny(image, 50,100) * mask
+    kernel = np.ones((8, 8), dtype=np.uint8)
+    mask = (cv2.dilate(canny, kernel, 5) > 128).astype(np.uint8)
+    mask = np.stack([mask,mask,mask],-1)
+
+    pure_noise = q_x(image, t=1) * 0 + 255
+    canny_noise = mask * image + (1-mask) * pure_noise
+    return canny_noise
+
+
+def get_random_structure(size):
+    choice = np.random.randint(1, 5)
+
+    if choice == 1:
+        return cv2.getStructuringElement(cv2.MORPH_RECT, (size, size))
+    elif choice == 2:
+        return cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (size, size))
+    elif choice == 3:
+        return cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (size, size//2))
+    elif choice == 4:
+        return cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (size//2, size))
+
+def random_dilate(seg, min=3, max=10):
+    size = np.random.randint(min, max)
+    kernel = get_random_structure(size)
+    seg = cv2.dilate(seg,kernel,iterations = 1)
+    return seg
+
+def random_erode(seg, min=3, max=10):
+    size = np.random.randint(min, max)
+    kernel = get_random_structure(size)
+    seg = cv2.erode(seg,kernel,iterations = 1)
+    return seg
+
+def compute_iou(seg, gt):
+    intersection = seg*gt
+    union = seg+gt
+    return (np.count_nonzero(intersection) + 1e-6) / (np.count_nonzero(union) + 1e-6)
+
+
+def select_max_region(mask):
+    nums, labels, stats, centroids = cv2.connectedComponentsWithStats(mask, connectivity=8)
+    background = 0
+    for row in range(stats.shape[0]):
+        if stats[row, :][0] == 0 and stats[row, :][1] == 0:
+            background = row
+    stats_no_bg = np.delete(stats, background, axis=0)
+    max_idx = stats_no_bg[:, 4].argmax()
+    max_region = np.where(labels==max_idx+1, 1, 0)
+
+    return max_region.astype(np.uint8)
+
+
+
+def perturb_mask(gt, min_iou = 0.3,  max_iou = 0.99):
+    iou_target = np.random.uniform(min_iou, max_iou)
+    h, w = gt.shape
+    gt = gt.astype(np.uint8)
+    seg = gt.copy()
+    
+    # Rare case
+    if h <= 2 or w <= 2:
+        print('GT too small, returning original')
+        return seg
+
+    # Do a bunch of random operations
+    for _ in range(250):
+        for _ in range(4):
+            lx, ly = np.random.randint(w), np.random.randint(h)
+            lw, lh = np.random.randint(lx+1,w+1), np.random.randint(ly+1,h+1)
+
+            # Randomly set one pixel to 1/0. With the following dilate/erode, we can create holes/external regions
+            if np.random.rand() < 0.1:
+                cx = int((lx + lw) / 2)
+                cy = int((ly + lh) / 2)
+                seg[cy, cx] = np.random.randint(2) * 255
+
+            # Dilate/erode
+            if np.random.rand() < 0.5:
+                seg[ly:lh, lx:lw] = random_dilate(seg[ly:lh, lx:lw])
+            else:
+                seg[ly:lh, lx:lw] = random_erode(seg[ly:lh, lx:lw])
+            
+            seg = np.logical_or(seg, gt).astype(np.uint8)
+            #seg = select_max_region(seg) 
+
+        if compute_iou(seg, gt) < iou_target:
+            break
+    seg = select_max_region(seg.astype(np.uint8)) 
+    return seg.astype(np.uint8)
+
+
+def q_x(x_0,t=65):
+    '''Adding noise for and given image.'''
+    x_0 = torch.from_numpy(x_0).float() / 127.5 - 1
+    num_steps = 100
+    
+    betas = torch.linspace(-6,6,num_steps)
+    betas = torch.sigmoid(betas)*(0.5e-2 - 1e-5)+1e-5
+
+    alphas = 1-betas
+    alphas_prod = torch.cumprod(alphas,0)
+    
+    alphas_prod_p = torch.cat([torch.tensor([1]).float(),alphas_prod[:-1]],0)
+    alphas_bar_sqrt = torch.sqrt(alphas_prod)
+    one_minus_alphas_bar_log = torch.log(1 - alphas_prod)
+    one_minus_alphas_bar_sqrt = torch.sqrt(1 - alphas_prod)
+    
+    noise = torch.randn_like(x_0)
+    alphas_t = alphas_bar_sqrt[t]
+    alphas_1_m_t = one_minus_alphas_bar_sqrt[t]
+    return (alphas_t * x_0 + alphas_1_m_t * noise).numpy()  * 127.5 + 127.5 
+
+
+def extract_target_boundary(img, target_mask):
+    Ksize = 3
+    sobelx = cv2.Sobel(img, cv2.CV_64F, 1, 0, ksize=Ksize)
+    sobely = cv2.Sobel(img, cv2.CV_64F, 0, 1, ksize=Ksize)
+
+    # sobel-x
+    sobel_X = cv2.convertScaleAbs(sobelx)
+    # sobel-y
+    sobel_Y = cv2.convertScaleAbs(sobely)
+    # sobel-xy
+    scharr = cv2.addWeighted(sobel_X, 0.5, sobel_Y, 0.5, 0)
+    scharr = np.max(scharr,-1).astype(np.float32)/255
+    scharr = scharr *  target_mask.astype(np.float32)
+    return scharr

datasets/dreambooth.py

@@ -0,0 +1,84 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+
+class DreamBoothDataset(BaseDataset):
+    def __init__(self, fg_dir, bg_dir):
+        self.bg_dir = bg_dir
+        bg_data = os.listdir(self.bg_dir)
+        self.bg_data = [i for i in bg_data if 'mask' in i]
+        self.image_dir = fg_dir
+        self.data  = os.listdir(self.image_dir)
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        '''
+         Dynamic:
+            0: Static View, High Quality
+            1: Multi-view, Low Quality
+            2: Multi-view, High Quality
+        '''
+        self.dynamic = 1 
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        idx = np.random.randint(0, len(self.data)-1)
+        item = self.get_sample(idx)
+        return item
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_alpha_mask(self, mask_path):
+        image = cv2.imread( mask_path, cv2.IMREAD_UNCHANGED)
+        mask = (image[:,:,-1] > 128).astype(np.uint8)
+        return mask
+        
+    def get_sample(self, idx):
+        dir_name = self.data[idx]
+        dir_path = os.path.join(self.image_dir, dir_name)
+        images = os.listdir(dir_path)
+        image_name = [i for i in images if '.png' in i][0]
+        image_path = os.path.join(dir_path, image_name)
+
+        image = cv2.imread( image_path, cv2.IMREAD_UNCHANGED)
+        mask = (image[:,:,-1] > 128).astype(np.uint8)
+        image = image[:,:,:-1]
+
+        image = cv2.cvtColor(image.copy(), cv2.COLOR_BGR2RGB)
+        ref_image = image 
+        ref_mask = mask
+        ref_image, ref_mask = expand_image_mask(image, mask, ratio=1.4)
+        bg_idx =  np.random.randint(0, len(self.bg_data)-1)
+        
+        tar_mask_name = self.bg_data[bg_idx]
+        tar_mask_path = os.path.join(self.bg_dir, tar_mask_name)
+        tar_image_path = tar_mask_path.replace('_mask','_GT')
+
+        tar_image = cv2.imread(tar_image_path).astype(np.uint8)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+        tar_mask = (cv2.imread(tar_mask_path) > 128).astype(np.uint8)[:,:,0] 
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+

datasets/dresscode.py

@@ -0,0 +1,61 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+import albumentations as A
+
+class DresscodeDataset(BaseDataset):
+    def __init__(self, image_dir):
+        self.image_root =  image_dir 
+        self.data = os.listdir(self.image_root)
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 2
+
+    def __len__(self):
+        return 20000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_sample(self, idx):
+        tar_mask_path = os.path.join(self.image_root, self.data[idx])
+        tar_image_path = tar_mask_path.replace('label_maps/','images/').replace('_4.png','_0.jpg')
+        ref_image_path = tar_mask_path.replace('label_maps/','images/').replace('_4.png','_1.jpg')
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+
+        tar_image = cv2.imread(tar_image_path)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = (ref_image < 240).astype(np.uint8)[:,:,0]
+        
+
+        tar_mask = Image.open(tar_mask_path ).convert('P')
+        tar_mask= np.array(tar_mask)
+        tar_mask = tar_mask == 4
+
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask, max_ratio = 1.0)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+

datasets/fashiontryon.py

@@ -0,0 +1,75 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+import albumentations as A
+
+class FashionTryonDataset(BaseDataset):
+    def __init__(self, image_dir):
+        self.image_root = image_dir
+        self.data =os.listdir(self.image_root)
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 2
+
+    def __len__(self):
+        return 5000
+
+    def aug_data(self, image):
+        transform = A.Compose([
+            A.RandomBrightnessContrast(p=0.5),
+            ])
+        transformed = transform(image=image.astype(np.uint8))
+        transformed_image = transformed["image"]
+        return transformed_image
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+            
+    def get_sample(self, idx):
+        cloth_dir = os.path.join(self.image_root, self.data[idx])
+        ref_image_path = os.path.join(cloth_dir, 'target.jpg')
+
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image.copy(), cv2.COLOR_BGR2RGB)
+
+        ref_mask_path = os.path.join(cloth_dir,'mask.jpg')
+        ref_mask = cv2.imread(ref_mask_path)[:,:,0] > 128
+
+        target_dirs = [i for i in os.listdir(cloth_dir ) if '.jpg' not in i]
+        target_dir_name = np.random.choice(target_dirs)
+
+        target_image_path = os.path.join(cloth_dir, target_dir_name + '.jpg')
+        target_image= cv2.imread(target_image_path)
+        tar_image = cv2.cvtColor(target_image.copy(), cv2.COLOR_BGR2RGB)
+
+        target_mask_path = os.path.join(cloth_dir, target_dir_name, 'segment.png')
+        tar_mask= cv2.imread(target_mask_path)[:,:,0]
+        target_mask =  tar_mask == 7        
+        kernel = np.ones((3, 3), dtype=np.uint8)
+        tar_mask = cv2.erode(target_mask.astype(np.uint8), kernel, iterations=3)
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask, max_ratio = 1.0)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+
+
+
+ 

datasets/lvis.py

@@ -0,0 +1,77 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+from pycocotools import mask as mask_utils
+from lvis import LVIS
+
+class LvisDataset(BaseDataset):
+    def __init__(self, image_dir, json_path):
+        self.image_dir = image_dir
+        self.json_path = json_path
+        lvis_api = LVIS(json_path)
+        img_ids = sorted(lvis_api.imgs.keys())
+        imgs = lvis_api.load_imgs(img_ids)
+        anns = [lvis_api.img_ann_map[img_id] for img_id in img_ids]
+        self.data = imgs
+        self.annos = anns
+        self.lvis_api = lvis_api
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 0
+
+    def register_subset(self, path):
+        data = os.listdir(path)
+        data = [ os.path.join(path, i) for i in data if '.json' in i]
+        self.data = self.data + data
+
+    def get_sample(self, idx):
+        # ==== get pairs =====
+        image_name = self.data[idx]['coco_url'].split('/')[-1]
+        image_path = os.path.join(self.image_dir, image_name)
+        image = cv2.imread(image_path)
+        ref_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+
+        anno = self.annos[idx]
+        obj_ids = []
+        for i in range(len(anno)):
+            obj = anno[i]
+            area = obj['area']
+            if area > 3600:
+                obj_ids.append(i)
+        assert len(anno) > 0
+        obj_id = np.random.choice(obj_ids)
+        anno = anno[obj_id]
+        ref_mask = self.lvis_api.ann_to_mask(anno)
+
+        tar_image, tar_mask = ref_image.copy(), ref_mask.copy()
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+
+    def __len__(self):
+        return 20000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H or w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H or w < W:
+                pass_flag = False
+        return pass_flag
+
+
+
+        

datasets/mose.py

@@ -0,0 +1,94 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from PIL import Image
+from .base import BaseDataset
+
+class MoseDataset(BaseDataset):
+    def __init__(self, image_dir, anno):
+        self.image_root = image_dir
+        self.anno_root = anno
+
+        video_dirs = []
+        video_dirs = os.listdir(self.image_root)
+        self.data = video_dirs
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 2
+
+    def __len__(self):
+        return 40000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H or w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H or w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_sample(self, idx):
+        video_name = self.data[idx]
+        video_path = os.path.join(self.image_root, video_name)
+        frames = os.listdir(video_path)
+
+        # Sampling frames
+        min_interval = len(frames)  // 10
+        start_frame_index = np.random.randint(low=0, high=len(frames) - min_interval)
+        end_frame_index = start_frame_index + np.random.randint(min_interval,  len(frames) - start_frame_index )
+        end_frame_index = min(end_frame_index, len(frames) - 1)
+
+        # Get image path
+        ref_image_name = frames[start_frame_index]
+        tar_image_name = frames[end_frame_index]
+        ref_image_path = os.path.join(self.image_root, video_name, ref_image_name)
+        tar_image_path = os.path.join(self.image_root, video_name, tar_image_name)
+
+        ref_mask_path = ref_image_path.replace('JPEGImages','Annotations').replace('.jpg', '.png')
+        tar_mask_path = tar_image_path.replace('JPEGImages','Annotations').replace('.jpg', '.png')
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+       
+        tar_image = cv2.imread(tar_image_path)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = Image.open(ref_mask_path ).convert('P')
+        ref_mask= np.array(ref_mask)
+
+        tar_mask = Image.open(tar_mask_path ).convert('P')
+        tar_mask= np.array(tar_mask)
+        
+        ref_ids = np.unique(ref_mask) 
+        tar_ids = np.unique(tar_mask)
+
+        common_ids = list(np.intersect1d(ref_ids, tar_ids))
+        common_ids = [ i  for i in common_ids if i != 0 ]
+        assert len(common_ids) > 0 
+        chosen_id = np.random.choice(common_ids)
+        ref_mask = ref_mask == chosen_id 
+        tar_mask = tar_mask == chosen_id 
+        len_mask = len( self.check_connect( ref_mask.astype(np.uint8) ) )
+        assert len_mask == 1
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+
+    def check_connect(self, mask):
+        contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
+        cnt_area = [cv2.contourArea(cnt) for cnt in contours]
+        return cnt_area
+

datasets/mvimagenet.py

@@ -0,0 +1,81 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+
+class MVImageNetDataset(BaseDataset):
+    def __init__(self, txt, image_dir):
+        with open(txt,"r") as f:
+            data = f.read().split('\n')[:-1]    
+        self.image_dir = image_dir 
+        self.data = data
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 2
+
+    def __len__(self):
+        return 40000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_alpha_mask(self, mask_path):
+        image = cv2.imread( mask_path, cv2.IMREAD_UNCHANGED)
+        mask = (image[:,:,-1] > 128).astype(np.uint8)
+        return mask
+        
+    def get_sample(self, idx):
+        object_dir = self.data[idx].replace('MVDir/', self.image_dir) 
+        frames = os.listdir(object_dir)
+        frames = [ i for i in frames if '.png' in i]
+
+        # Sampling frames
+        min_interval = len(frames)  // 8
+        start_frame_index = np.random.randint(low=0, high=len(frames) - min_interval)
+        end_frame_index = start_frame_index + np.random.randint(min_interval,  len(frames) - start_frame_index )
+        end_frame_index = min(end_frame_index, len(frames) - 1)
+
+        # Get image path
+        ref_mask_name = frames[start_frame_index]
+        tar_mask_name = frames[end_frame_index]
+
+        ref_image_name = ref_mask_name.split('_')[0] + '.jpg'
+        tar_image_name = tar_mask_name.split('_')[0] + '.jpg'
+
+        ref_mask_path = os.path.join(object_dir, ref_mask_name)
+        tar_mask_path = os.path.join(object_dir, tar_mask_name)
+        ref_image_path = os.path.join(object_dir, ref_image_name)
+        tar_image_path = os.path.join(object_dir, tar_image_name) 
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path).astype(np.uint8)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+
+        tar_image = cv2.imread(tar_image_path).astype(np.uint8)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = self.get_alpha_mask(ref_mask_path)
+        tar_mask = self.get_alpha_mask(tar_mask_path)
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+
+        return item_with_collage
+

datasets/saliency_modular.py

@@ -0,0 +1,91 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+
+class SaliencyDataset(BaseDataset):
+    def __init__(self, MSRA_root, TR_root, TE_root, HFlickr_root):
+        image_mask_dict = {}
+
+        # ====== MSRA-10k ======
+        file_lst = os.listdir(MSRA_root)
+        image_lst = [MSRA_root+i for i in file_lst if '.jpg' in i]
+        for i in image_lst:
+            mask_path = i.replace('.jpg','.png')
+            image_mask_dict[i] = mask_path
+
+        # ===== DUT-TR ========
+        file_lst = os.listdir(TR_root)
+        image_lst = [TR_root+i for i in file_lst if '.jpg' in i]
+        for i in image_lst:
+            mask_path = i.replace('.jpg','.png').replace('DUTS-TR-Image','DUTS-TR-Mask')
+            image_mask_dict[i] = mask_path
+        
+        # ===== DUT-TE ========
+        file_lst = os.listdir(TE_root)
+        image_lst = [TE_root+i for i in file_lst if '.jpg' in i]
+        for i in image_lst:
+            mask_path = i.replace('.jpg','.png').replace('DUTS-TE-Image','DUTS-TE-Mask')
+            image_mask_dict[i] = mask_path 
+        
+        # ===== HFlickr =======
+        file_lst = os.listdir(HFlickr_root)
+        mask_list = [HFlickr_root+i for i in file_lst if '.png' in i]
+        for i in file_lst:
+            image_name  = i.split('_')[0] +'.jpg'
+            image_path = HFlickr_root.replace('masks', 'real_images') + image_name
+            mask_path = HFlickr_root + i
+            image_mask_dict[image_path] = mask_path 
+
+        self.image_mask_dict = image_mask_dict
+        self.data = list(self.image_mask_dict.keys() )
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 0
+
+    def __len__(self):
+        return 20000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H or w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H or w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_sample(self, idx):
+
+        # ==== get pairs =====
+        image_path = self.data[idx]
+        mask_path = self.image_mask_dict[image_path]
+
+        instances_mask = cv2.imread(mask_path)
+        if len(instances_mask.shape) == 3:
+            instances_mask = instances_mask[:,:,0]
+        instances_mask = (instances_mask > 128).astype(np.uint8)
+        # ======================
+        ref_image = cv2.imread(image_path)
+        ref_image = cv2.cvtColor(ref_image.copy(), cv2.COLOR_BGR2RGB)
+        tar_image = ref_image
+
+        ref_mask = instances_mask
+        tar_mask = instances_mask 
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+
+
+        

datasets/sam.py

@@ -0,0 +1,78 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+from pycocotools import mask as mask_utils
+
+class SAMDataset(BaseDataset):
+    def __init__(self, sub1, sub2, sub3, sub4):
+        image_mask_dict = {}
+        self.data = []
+        self.register_subset(sub1)
+        self.register_subset(sub2)
+        self.register_subset(sub3)
+        self.register_subset(sub4)
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 0
+
+    def register_subset(self, path):
+        data = os.listdir(path)
+        data = [ os.path.join(path, i) for i in data if '.json' in i]
+        self.data = self.data + data
+
+    def get_sample(self, idx):
+        # ==== get pairs =====
+        json_path = self.data[idx]
+        image_path = json_path.replace('.json', '.jpg')
+
+        with open(json_path, 'r') as json_file:
+            data = json.load(json_file)
+        annotation = data['annotations']
+
+        valid_ids = []
+        for i in range(len(annotation)):
+            area = annotation[i]['area']
+            if area > 100 * 100 * 5:
+                valid_ids.append(i)
+
+        chosen_id = np.random.choice(valid_ids)
+        mask = mask_utils.decode(annotation[chosen_id]["segmentation"] )
+        # ======================
+
+        image = cv2.imread(image_path)
+        ref_image = cv2.cvtColor(image.copy(), cv2.COLOR_BGR2RGB)
+        tar_image = ref_image
+        
+        ref_mask = mask
+        tar_mask = mask
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+
+    def __len__(self):
+        return 20000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H or w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H or w < W:
+                pass_flag = False
+        return pass_flag
+
+
+
+        

datasets/uvo.py

@@ -0,0 +1,79 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+from pycocotools import mask as mask_utils
+
+class UVODataset(BaseDataset):
+    def __init__(self, image_dir, video_json, image_json):
+        json_path = video_json 
+        with open(json_path, 'r') as fcc_file:
+            data = json.load(fcc_file)
+
+        image_json_path = image_json
+        with open(image_json_path , 'r') as image_file:
+            video_dict = json.load(image_file)
+
+        self.image_root =  image_dir
+        self.data = data['annotations']
+        self.video_dict = video_dict
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 1
+
+    def __len__(self):
+        return 25000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_sample(self, idx):
+        ins_anno = self.data[idx]
+        video_id = str(ins_anno['video_id'])
+        video_names = self.video_dict[video_id]
+        masks = ins_anno['segmentations']
+        frames = video_names
+
+        # Sampling frames
+        min_interval = len(frames)  // 10
+        start_frame_index = np.random.randint(low=0, high=len(frames) - min_interval)
+        end_frame_index = start_frame_index + np.random.randint(min_interval,  len(frames) - start_frame_index )
+        end_frame_index = min(end_frame_index, len(frames) - 1)
+
+        # Get image path
+        ref_image_name = frames[start_frame_index]
+        tar_image_name = frames[end_frame_index]
+        ref_image_path = os.path.join(self.image_root, ref_image_name) 
+        tar_image_path = os.path.join(self.image_root, tar_image_name) 
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+
+        tar_image = cv2.imread(tar_image_path)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = mask_utils.decode(masks[start_frame_index])
+        tar_mask = mask_utils.decode(masks[end_frame_index])
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+

datasets/uvo_val.py

@@ -0,0 +1,87 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+from pycocotools import mask as mask_utils
+
+class UVOValDataset(BaseDataset):
+    def __init__(self, image_dir, video_json, image_json):
+        json_path = video_json
+        with open(json_path, 'r') as fcc_file:
+            data = json.load(fcc_file)
+        image_json_path = image_json
+        with open(image_json_path , 'r') as image_file:
+            video_dict = json.load(image_file)
+        self.image_root = image_dir
+        self.data = data['annotations']
+        self.video_dict = video_dict
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 1
+
+    def __len__(self):
+        return 8000
+
+    def __getitem__(self, idx):
+        while(1):
+            idx = np.random.randint(0, len(self.data)-1)
+            try:
+                item = self.get_sample(idx)
+                return item
+            except:
+                idx = np.random.randint(0, len(self.data)-1)
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_sample(self, idx):
+        ins_anno = self.data[idx]
+        video_id = str(ins_anno['video_id'])
+
+        video_names = self.video_dict[video_id]
+        masks = ins_anno['segmentations']
+        frames = video_names
+
+        # Sampling frames
+        min_interval = len(frames)  // 5
+        start_frame_index = np.random.randint(low=0, high=len(frames) - min_interval)
+        end_frame_index = start_frame_index + np.random.randint(min_interval,  len(frames) - start_frame_index )
+        end_frame_index = min(end_frame_index, len(frames) - 1)
+
+        # Get image path
+        ref_image_name = frames[start_frame_index]
+        tar_image_name = frames[end_frame_index]
+        ref_image_path = os.path.join(self.image_root, ref_image_name) 
+        tar_image_path = os.path.join(self.image_root, tar_image_name) 
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+
+        tar_image = cv2.imread(tar_image_path)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = mask_utils.decode(masks[start_frame_index])
+        tar_mask = mask_utils.decode(masks[end_frame_index])
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+

datasets/vipseg.py

@@ -0,0 +1,96 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from panopticapi.utils import rgb2id
+from PIL import Image
+from .base import BaseDataset
+
+class VIPSegDataset(BaseDataset):
+    def __init__(self, image_dir, anno):
+        self.image_root = image_dir
+        self.anno_root = anno
+        video_dirs = []
+        video_dirs = os.listdir(self.image_root)
+        self.data = video_dirs
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 1
+
+    def __len__(self):
+        return 30000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H or w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H or w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_sample(self, idx):
+        video_name = self.data[idx]
+        video_path = os.path.join(self.image_root, video_name)
+        frames = os.listdir(video_path)
+
+        # Sampling frames
+        min_interval = len(frames)  // 100
+        start_frame_index = np.random.randint(low=0, high=len(frames) - min_interval)
+        end_frame_index = start_frame_index + np.random.randint(min_interval,  len(frames) - start_frame_index )
+        end_frame_index = min(end_frame_index, len(frames) - 1)
+
+        # Get image path
+        ref_image_name = frames[start_frame_index]
+        tar_image_name = frames[end_frame_index]
+        ref_image_path = os.path.join(self.image_root, video_name, ref_image_name)
+        tar_image_path = os.path.join(self.image_root, video_name, tar_image_name)
+
+        ref_mask_path = ref_image_path.replace('images','panomasksRGB').replace('.jpg', '.png')
+        tar_mask_path = tar_image_path.replace('images','panomasksRGB').replace('.jpg', '.png')
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+
+        tar_image = cv2.imread(tar_image_path)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = np.array(Image.open(ref_mask_path).convert('RGB'))
+        ref_mask = rgb2id(ref_mask)
+
+        tar_mask = np.array(Image.open(tar_mask_path).convert('RGB'))
+        tar_mask = rgb2id(tar_mask)
+        
+        ref_ids = np.unique(ref_mask) 
+        tar_ids = np.unique(tar_mask)
+
+        common_ids = list(np.intersect1d(ref_ids, tar_ids))
+        common_ids = [ i  for i in common_ids if i != 0 ]
+
+        chosen_id = np.random.choice(common_ids)
+        ref_mask = ref_mask == chosen_id 
+        tar_mask = tar_mask == chosen_id 
+
+        len_mask = len( self.check_connect( ref_mask.astype(np.uint8) ) )
+        assert len_mask == 1
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+
+    def check_connect(self, mask):
+        contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
+        cnt_area = [cv2.contourArea(cnt) for cnt in contours]
+        return cnt_area
+

datasets/vitonhd.py

@@ -0,0 +1,61 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+import albumentations as A
+
+class VitonHDDataset(BaseDataset):
+    def __init__(self, image_dir):
+        self.image_root = image_dir
+        self.data = os.listdir(self.image_root)
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 2
+
+    def __len__(self):
+        return 20000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+            
+    def get_sample(self, idx):
+
+        ref_image_path = os.path.join(self.image_root, self.data[idx])
+        tar_image_path = ref_image_path.replace('/cloth/', '/image/')
+        ref_mask_path = ref_image_path.replace('/cloth/','/cloth-mask/')
+        tar_mask_path = ref_image_path.replace('/cloth/', '/image-parse-v3/').replace('.jpg','.png')
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+
+        tar_image = cv2.imread(tar_image_path)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = (cv2.imread(ref_mask_path) > 128).astype(np.uint8)[:,:,0]
+
+        tar_mask = Image.open(tar_mask_path ).convert('P')
+        tar_mask= np.array(tar_mask)
+        tar_mask = tar_mask == 5
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask, max_ratio = 1.0)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+

datasets/ytb_vis.py

@@ -0,0 +1,85 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+
+class YoutubeVISDataset(BaseDataset):
+    def __init__(self, image_dir, anno, meta):
+        self.image_root = image_dir
+        self.anno_root = anno 
+        self.meta_file = meta
+
+        video_dirs = []
+        with open(self.meta_file) as f:
+            records = json.load(f)
+            records = records["videos"]
+            for video_id in records:
+                video_dirs.append(video_id)
+
+        self.records = records
+        self.data = video_dirs
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 1
+
+    def __len__(self):
+        return 40000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_sample(self, idx):
+        video_id = list(self.records.keys())[idx]
+        objects_id = np.random.choice( list(self.records[video_id]["objects"].keys()) )
+        frames = self.records[video_id]["objects"][objects_id]["frames"]
+
+        # Sampling frames
+        min_interval = len(frames)  // 10
+        start_frame_index = np.random.randint(low=0, high=len(frames) - min_interval)
+        end_frame_index = start_frame_index + np.random.randint(min_interval,  len(frames) - start_frame_index )
+        end_frame_index = min(end_frame_index, len(frames) - 1)
+
+        # Get image path
+        ref_image_name = frames[start_frame_index]
+        tar_image_name = frames[end_frame_index]
+        ref_image_path = os.path.join(self.image_root, video_id, ref_image_name) + '.jpg'
+        tar_image_path = os.path.join(self.image_root, video_id, tar_image_name) + '.jpg'
+        ref_mask_path = ref_image_path.replace('JPEGImages','Annotations').replace('.jpg', '.png')
+        tar_mask_path = tar_image_path.replace('JPEGImages','Annotations').replace('.jpg', '.png')
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+
+        tar_image = cv2.imread(tar_image_path)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = Image.open(ref_mask_path ).convert('P')
+        ref_mask= np.array(ref_mask)
+        ref_mask = ref_mask == int(objects_id)
+
+        tar_mask = Image.open(tar_mask_path ).convert('P')
+        tar_mask= np.array(tar_mask)
+        tar_mask = tar_mask == int(objects_id)
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+

datasets/ytb_vos.py

@@ -0,0 +1,87 @@
+import json
+import cv2
+import numpy as np
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import cv2
+from .data_utils import * 
+from .base import BaseDataset
+
+class YoutubeVOSDataset(BaseDataset):
+    def __init__(self, image_dir, anno, meta):
+        self.image_root = image_dir
+        self.anno_root = anno
+        self.meta_file = meta
+
+        video_dirs = []
+        with open(self.meta_file) as f:
+            records = json.load(f)
+            records = records["videos"]
+            for video_id in records:
+                video_dirs.append(video_id)
+
+        self.records = records
+        self.data = video_dirs
+        self.size = (512,512)
+        self.clip_size = (224,224)
+        self.dynamic = 1
+
+    def __len__(self):
+        return 40000
+
+    def check_region_size(self, image, yyxx, ratio, mode = 'max'):
+        pass_flag = True
+        H,W = image.shape[0], image.shape[1]
+        H,W = H * ratio, W * ratio
+        y1,y2,x1,x2 = yyxx
+        h,w = y2-y1,x2-x1
+        if mode == 'max':
+            if h > H and w > W:
+                pass_flag = False
+        elif mode == 'min':
+            if h < H and w < W:
+                pass_flag = False
+        return pass_flag
+
+    def get_sample(self, idx):
+        video_id = list(self.records.keys())[idx]
+        objects_id = np.random.choice( list(self.records[video_id]["objects"].keys()) )
+        frames = self.records[video_id]["objects"][objects_id]["frames"]
+
+        # Sampling frames
+        min_interval = len(frames)  // 10
+        start_frame_index = np.random.randint(low=0, high=len(frames) - min_interval)
+        end_frame_index = start_frame_index + np.random.randint(min_interval,  len(frames) - start_frame_index )
+        end_frame_index = min(end_frame_index, len(frames) - 1)
+
+        # Get image path
+        ref_image_name = frames[start_frame_index]
+        tar_image_name = frames[end_frame_index]
+        ref_image_path = os.path.join(self.image_root, video_id, ref_image_name) + '.jpg'
+        tar_image_path = os.path.join(self.image_root, video_id, tar_image_name) + '.jpg'
+        ref_mask_path = ref_image_path.replace('JPEGImages','Annotations').replace('.jpg', '.png')
+        tar_mask_path = tar_image_path.replace('JPEGImages','Annotations').replace('.jpg', '.png')
+
+        # Read Image and Mask
+        ref_image = cv2.imread(ref_image_path)
+        ref_image = cv2.cvtColor(ref_image, cv2.COLOR_BGR2RGB)
+
+        tar_image = cv2.imread(tar_image_path)
+        tar_image = cv2.cvtColor(tar_image, cv2.COLOR_BGR2RGB)
+
+        ref_mask = Image.open(ref_mask_path ).convert('P')
+        ref_mask= np.array(ref_mask)
+        ref_mask = ref_mask == int(objects_id)
+
+        tar_mask = Image.open(tar_mask_path ).convert('P')
+        tar_mask= np.array(tar_mask)
+        tar_mask = tar_mask == int(objects_id)
+
+
+        item_with_collage = self.process_pairs(ref_image, ref_mask, tar_image, tar_mask)
+        sampled_time_steps = self.sample_timestep()
+        item_with_collage['time_steps'] = sampled_time_steps
+        return item_with_collage
+
+

dinov2/.github/workflows/lint.yaml

@@ -0,0 +1,39 @@
+name: Lint
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - master
+      - 'gh/**'
+
+jobs:
+  run-linters:
+    name: Run linters
+    runs-on: ubuntu-20.04
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: 3.9
+          cache: 'pip'
+          cache-dependency-path: '**/requirements*.txt'
+      - name: Install Python (development) dependencies
+        run: |
+          pip install -r requirements-dev.txt
+      - name: Run flake8
+        run: |
+          flake8
+      - name: Run black
+        if: always()
+        run: |
+          black --check dinov2
+      - name: Run pylint
+        if: always()
+        run: |
+          pylint --exit-zero dinov2

dinov2/.gitignore

@@ -0,0 +1,13 @@
+build/
+dist/
+*.egg-info/
+**/__pycache__/
+
+**/.ipynb_checkpoints
+**/.ipynb_checkpoints/**
+
+**/notebooks
+
+*.swp
+
+.vscode/

dinov2/CODE_OF_CONDUCT.md

@@ -0,0 +1,80 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+This Code of Conduct also applies outside the project spaces when there is a
+reasonable belief that an individual's behavior may have a negative impact on
+the project or its community.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <[email protected]>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

dinov2/CONTRIBUTING.md

@@ -0,0 +1,31 @@
+# Contributing to DINOv2
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints.
+6. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Meta's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+Meta has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## License
+By contributing to DINOv2, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.

dinov2/LICENSE

@@ -0,0 +1,400 @@
+
+Attribution-NonCommercial 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public: 
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  j. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  k. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  l. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

dinov2/MODEL_CARD.md

@@ -0,0 +1,201 @@
+# Model Card for DINOv2-S/B/L/g
+
+These are Vision Transformer models trained following the method described in the paper:
+"DINOv2: Learning Robust Visual Features without Supervision"
+
+We provide 4 models: 1 ViT-g trained from scratch, and 3 ViT-S/B/L models distilled from the ViT-g.
+
+## Model Details
+The model takes an image as input and returns a class token and patch tokens.
+
+The embedding dimension is: 
+- 384 for ViT-S.
+- 768 for ViT-B.
+- 1024 for ViT-L.
+- 1536 for ViT-g.
+
+The models follow a Transformer architecture, with a patch size of 14.
+
+For a 224x224 image, this results in 1 class token + 256 patch tokens.
+
+The models can accept larger images provided the image shapes are multiples of the patch size (14). 
+If this condition is not verified, the model will crop to the closest smaller multiple of the patch size.
+
+### Model Description
+
+- **Developed by:** Meta AI
+- **Model type:** Vision Transformer
+- **License:** CC-BY-NC
+
+- **Repository:** https://github.com/facebookresearch/dinov2
+- **Paper:** https://arxiv.org/abs/2304.07193
+- **Demo:** https://dinov2.metademolab.com/
+
+## Uses
+
+The models are vision backbones providing multi-purpose features for downstream tasks.
+
+### Direct Use
+
+The models can be used without fine-tuning, with downstream classifiers as simple as linear layers, to obtain competitive results:
+- on depth estimation, semantic segmentation, using linear layers.
+- on image classification, using k-NN classifiers on the class token.
+- on image classification, with logistic regression classifiers applied on the class token.
+- on image classification, with a linear layer applied on the class token and the average of the patch tokens.
+- on image retrieval using nearest neighbors.
+
+### Downstream Use
+
+It is technically possible to perform fine-tuning on the models, for small gains (we measured +2% on ImageNet-1k classification). 
+We recommend keeping this as a very last step and only when necessary, as the features already provide good performance out-of-the-box.
+
+## Bias, Risks, and Limitations
+
+Despite improvements thanks to the training method not using annotations, we still observe significant biases in our models toward rich households from Western countries.
+
+### Recommendations
+
+We expect fine-tuning will increase the biases in the features produced by the model as they will be tuned to the fine-tuning labels.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+```python
+import torch
+dinov2_vits14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14')
+dinov2_vitb14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14')
+dinov2_vitl14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')
+dinov2_vitg14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitg14')
+```
+
+## Training Details
+
+### Training Data
+
+- **Training data:** LVD-142M (see paper)
+- **Training regime:** fp16 using PyTorch-FSDP mixed-precision.
+
+### Training Procedure 
+
+- **Training objective:**
+  - DINO self-distillation loss with multi-crop
+  - iBOT masked-image modeling loss
+  - KoLeo regularization on [CLS] tokens
+- **Architectures:**
+  - ViT-S (21M params): Patch size 14, embedding dimension 384, 6 heads, MLP FFN
+  - ViT-B (86M params): Patch size 14, embedding dimension 768, 12 heads, MLP FFN
+  - ViT-L (0.3B params): Patch size 14, embedding dimension 1024, 16 heads, MLP FFN
+  - ViT-g (1.1B params): Patch size 14, embedding dimension 1536, 24 heads, SwiGLU FFN
+- **Distillation:**
+  - Distillation follows the standard DINOv2 pretraining procedure, except the teacher is a pretrained ViT-g, frozen.
+
+## Evaluation
+
+We refer users to the associated paper for the evaluation protocols.
+
+<table>
+  <tr>
+    <th>model</th>
+    <th colspan="3">ImageNet-1k</th>
+    <th>NYU-Depth v2</th>
+    <th>SUN-RGBD</th>
+    <th>ADE20k</th>
+    <th>iNaturalist 2018</th>
+    <th>Oxford-H</th>
+  </tr>
+  <tr>
+    <th rowspan="2">task</th>
+    <th>classif. (acc)</th>
+    <th>classif. (acc)</th>
+    <th>classif. V2 (acc)</th>
+    <th>depth (RMSE)</th>
+    <th>depth (RMSE)</th>
+    <th>segm. (mAP)</th>
+    <th>classif. (acc)</th>
+    <th>retrieval (mAP)</th>
+  </tr>
+  <tr>
+    <!-- <th>^</th> -->
+    <th>k-NN</th>
+    <th>linear</th>
+    <th>linear</th>
+    <th>linear<br />4 layers</th>
+    <th>NYU-D transfer</th>
+    <th>multiscale</th>
+    <th>linear</th>
+    <th>nearest neighbor</th>
+  </tr>
+  <tr>
+    <td>ViT-S/14</td>
+    <td align="right">79.0%</td>
+    <td align="right">81.1%</td>
+    <td align="right">70.8%</td> 
+    <td align="right">0.417</td> 
+    <td align="right">0.431</td> 
+    <td align="right">47.2</td> 
+    <td align="right">69.5%</td> 
+    <td align="right">43.2</td> 
+  </tr>
+  <tr>
+    <td>ViT-B/14</td>
+    <td align="right">82.1%</td>
+    <td align="right">84.5%</td>
+    <td align="right">74.9%</td>
+    <td align="right">0.362</td> 
+    <td align="right">0.400</td> 
+    <td align="right">51.3</td> 
+    <td align="right">76.3%</td> 
+    <td align="right">49.5</td> 
+  </tr>
+  <tr>
+    <td>ViT-L/14</td>
+    <td align="right">83.5%</td>
+    <td align="right">86.3%</td>
+    <td align="right">77.6%</td>
+    <td align="right">0.333</td> 
+    <td align="right">0.396</td> 
+    <td align="right">53.1</td> 
+    <td align="right">79.8%</td> 
+    <td align="right">54.0</td> 
+  </tr>
+  <tr>
+    <td>ViT-g/14</td>
+    <td align="right">83.5%</td>
+    <td align="right">86.5%</td>
+    <td align="right">78.4%</td>
+    <td align="right">0.298</td> 
+    <td align="right">0.362</td> 
+    <td align="right">53.0</td> 
+    <td align="right">81.6%</td> 
+    <td align="right">52.3</td> 
+  </tr>
+</table>
+
+## Environmental Impact
+
+- **Hardware Type:** Nvidia A100
+- **Hours used:** 22,000 for ViT-g, 4,500 for ViT-S distillation, 5,300 for ViT-B distillation, 8,000 for ViT-L distillation
+- **Cloud Provider:** Private infra
+- **Compute Region:** USA
+- **Carbon Emitted:** 7t CO2eq
+
+#### Hardware
+
+Nvidia A100 GPUs
+
+#### Software
+
+PyTorch 2.0,
+xFormers 0.0.18
+
+**BibTeX**
+
+```
+@misc{oquab2023dinov2,
+  title={DINOv2: Learning Robust Visual Features without Supervision},
+  author={Oquab, Maxime and Darcet, Timothée and Moutakanni, Theo and Vo, Huy and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and Howes, Russell and Huang, Po-Yao and Xu, Hu and Sharma, Vasu and Li, Shang-Wen and Galuba, Wojciech and Rabbat, Mike and Assran, Mido and Ballas, Nicolas and Synnaeve, Gabriel and Misra, Ishan and Jegou, Herve and Mairal, Julien and Labatut, Patrick and Joulin, Armand and Bojanowski, Piotr},
+  journal={arXiv:2304.07193},
+  year={2023}
+}
+```

dinov2/README.md

@@ -0,0 +1,248 @@
+# DINOv2: Learning Robust Visual Features without Supervision
+
+**[Meta AI Research, FAIR](https://ai.facebook.com/research/)**
+
+Maxime Oquab,
+Timothée Darcet,
+Théo Moutakanni,
+Huy Vo,
+Marc Szafraniec,
+Vasil Khalidov,
+Patrick Labatut,
+Armand Joulin,
+Piotr Bojanowski
+
+[[`Paper`](https://arxiv.org/abs/2304.07193)] [[`Blog`](https://ai.facebook.com/blog/dino-v2-computer-vision-self-supervised-learning/)] [[`Demo`](https://dinov2.metademolab.com)] [[`BibTeX`](#citing-dinov2)]
+
+PyTorch implementation and pretrained models for DINOv2. For details, see the paper: **DINOv2: Learning Robust Visual Features without Supervision**.
+
+DINOv2 models produce high-performance visual features that can be directly employed with classifiers as simple as linear layers on a variety of computer vision tasks; these visual features are robust and perform well across domains without any requirement for fine-tuning. The models were pretrained on a dataset of 142 M images without using any labels or annotations.
+
+
+https://user-images.githubusercontent.com/60359573/230078733-5faffa19-e6ce-4c55-9200-62dd76f8236a.mp4
+
+<div align="center">
+  Visualization of the three first principal components of the patch features of all frames, mapped to RGB values.
+</div>
+
+## Pretrained models
+
+<table>
+  <tr>
+    <th>model</th>
+    <th># of<br />params</th>
+    <th>ImageNet<br />k-NN</th>
+    <th>ImageNet<br />linear</th>
+    <th>download</th>
+  </tr>
+  <tr>
+    <td>ViT-S/14 distilled</td>
+    <td align="right">21 M</td>
+    <td align="right">79.0%</td>
+    <td align="right">81.1%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_pretrain.pth">backbone only</a></td>
+  </tr>
+  <tr>
+    <td>ViT-B/14 distilled</td>
+    <td align="right">86 M</td>
+    <td align="right">82.1%</td>
+    <td align="right">84.5%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_pretrain.pth">backbone only</a></td>
+  </tr>
+  <tr>
+    <td>ViT-L/14 distilled</td>
+    <td align="right">300 M</td>
+    <td align="right">83.5%</td>
+    <td align="right">86.3%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitl14/dinov2_vitl14_pretrain.pth">backbone only</a></td>
+  </tr>
+  <tr>
+    <td>ViT-g/14</td>
+    <td align="right">1,100 M</td>
+    <td align="right">83.5%</td>
+    <td align="right">86.5%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitg14/dinov2_vitg14_pretrain.pth">backbone only</a></td>
+  </tr>
+</table>
+
+
+### Pretrained models via PyTorch Hub
+
+Please follow the instructions [here](https://pytorch.org/get-started/locally/) to install the PyTorch and torchvision dependencies (these are the only required dependencies). Installing both PyTorch and torchvision with CUDA support is strongly recommended.
+
+The corresponding model card can be found in the [[`MODEL_CARD.md`](MODEL_CARD.md)] file.
+
+```python
+import torch
+
+dinov2_vits14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14')
+dinov2_vitb14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14')
+dinov2_vitl14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')
+dinov2_vitg14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitg14')
+```
+
+## Installation
+
+The training and evaluation code requires PyTorch 2.0 and xFormers 0.0.18 as well as a number of other 3rd party packages. To setup all the required dependencies for training and evaluation, please follow the instructions below:
+
+*conda* **(Recommended)** - Create and activate a `dinov2` conda environment using the provided environment definition:
+
+```shell
+conda env create -f conda.yaml
+conda activate dinov2
+```
+
+*pip* - Use the provided `requirements.txt` to install the dependencies:
+
+```shell
+pip install -r requirements.txt
+```
+
+## Data preparation
+
+Expected contents for the ImageNet-1k data folder:
+- `<root>/test/ILSVRC2012_test_00000001.JPEG`
+- `<root>/test/[..]`
+- `<root>/test/ILSVRC2012_test_00100000.JPEG`
+- `<root>/train/n01440764/n01440764_10026.JPEG`
+- `<root>/train/[...]`
+- `<root>/train/n15075141/n15075141_9993.JPEG`
+- `<root>/val/n01440764/ILSVRC2012_val_00000293.JPEG`
+- `<root>/val/[...]`
+- `<root>/val/n15075141/ILSVRC2012_val_00049174.JPEG`
+- `<root>/labels.txt`
+
+For ImageNet-22k, please adapt the Dataset object accordingly.
+
+## Training
+
+### Fast setup: training DINOv2 ViT-L/16 on ImageNet-1k
+
+Run DINOv2 on 4 A100-80GB nodes (32 GPUs) in a SLURM cluster environment with submitit.
+
+```shell
+python dinov2/run/train/train.py \
+    --nodes 4 \
+    --config-file dinov2/configs/train/vitl16_short.yaml \
+    --output-dir <PATH/TO/OUTPUT/DIR> \
+    train.dataset_path=ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+Training time is approximately 1 day and the resulting checkpoint should reach 81.6% on k-NN eval and 82.9% on linear eval.
+
+The training code saves the weights of the teacher in the `eval` folder every 12500 iterations for evaluation.
+
+### Long setup: training DINOv2 ViT-L/14 on ImageNet-22k
+
+Run on 12 A100-80GB nodes (96 GPUs) in a SLURM cluster environment with submitit.
+
+```
+python dinov2/run/train/train.py \
+    --nodes 12 \
+    --config-file dinov2/configs/train/vitl14.yaml \
+    --output-dir <PATH/TO/OUTPUT/DIR> \
+    train.dataset_path=ImageNet22k:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+Training time is approximately 3.3 days and the resulting checkpoint should reach 82.0% on k-NN eval and 84.5% on linear eval.
+
+The training code saves the weights of the teacher in the `eval` folder every 12500 iterations for evaluation.
+
+
+## Evaluation
+
+The training code regularly saves the teacher weights. In order to evaluate the model, run the following evaluation on a single node:
+
+### k-NN classification on ImageNet-1k
+
+```
+python dinov2/run/eval/knn.py \
+    --config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
+    --pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
+    --output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/knn \
+    --train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
+    --val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+### Logistic regression classification on ImageNet-1k
+
+```
+python dinov2/run/eval/log_regression.py \
+    --config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
+    --pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
+    --output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/logreg \
+    --train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
+    --val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+### Linear classification with data augmentation on ImageNet-1k
+
+```
+python dinov2/run/eval/linear.py \
+    --config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
+    --pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
+    --output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/linear \
+    --train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
+    --val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+We release the weights from evaluating the different models:
+
+<table>
+  <tr>
+    <th>model</th>
+    <th>ImageNet<br />top-1</th>
+    <th>linear evaluation</th>
+  </tr>
+  <tr>
+    <td>ViT-S/14 distilled</td>
+    <td align="right">81.1%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_linear_head.pth">linear head weights</a></td>
+  </tr>
+  <tr>
+    <td>ViT-B/14 distilled</td>
+    <td align="right">84.5%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_linear_head.pth">linear head weights</a></td>
+  </tr>
+  <tr>
+    <td>ViT-L/14 distilled</td>
+    <td align="right">86.3%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitl14/dinov2_vitl14_linear_head.pth">linear head weights</a></td>
+  </tr>
+  <tr>
+    <td>ViT-g/14</td>
+    <td align="right">86.5%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitg14/dinov2_vitg14_linear_head.pth">linear head weights</a></td>
+  </tr>
+</table>
+
+The performance of the provided pretrained model weights can be evaluated as follows on ImageNet-1k:
+
+```
+python dinov2/run/eval/linear.py \
+    --config-file dinov2/configs/eval/vitg14_pretrain.yaml \
+    --pretrained-weights https://dl.fbaipublicfiles.com/dinov2/dinov2_vitg14/dinov2_vitg14_pretrain.pth \
+    --train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
+    --val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+## License
+
+This repository and the models are released under the CC-BY-NC as found in the [LICENSE](LICENSE) file.
+
+## Contributing
+
+See [contributing](CONTRIBUTING.md) and the [code of conduct](CODE_OF_CONDUCT.md).
+
+## Citing DINOv2
+
+If you find this repository useful, please consider giving a star :star: and citation :t-rex::
+
+```
+@misc{oquab2023dinov2,
+  title={DINOv2: Learning Robust Visual Features without Supervision},
+  author={Oquab, Maxime and Darcet, Timothée and Moutakanni, Theo and Vo, Huy and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and Howes, Russell and Huang, Po-Yao and Xu, Hu and Sharma, Vasu and Li, Shang-Wen and Galuba, Wojciech and Rabbat, Mike and Assran, Mido and Ballas, Nicolas and Synnaeve, Gabriel and Misra, Ishan and Jegou, Herve and Mairal, Julien and Labatut, Patrick and Joulin, Armand and Bojanowski, Piotr},
+  journal={arXiv:2304.07193},
+  year={2023}
+}
+```

dinov2/conda.yaml

@@ -0,0 +1,22 @@
+name: dinov2
+channels:
+  - defaults
+  - pytorch
+  - nvidia
+  - xformers
+  - conda-forge
+dependencies:
+  - python=3.9
+  - pytorch::pytorch=2.0.0
+  - pytorch::pytorch-cuda=11.7.0
+  - pytorch::torchvision=0.15.0
+  - omegaconf
+  - torchmetrics=0.10.3
+  - fvcore
+  - iopath
+  - xformers::xformers=0.0.18
+  - pip
+  - pip:
+    - git+https://github.com/facebookincubator/submitit
+    - --extra-index-url https://pypi.nvidia.com
+    - cuml-cu11

dinov2/dinov2/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+__version__ = "0.0.1"

dinov2/dinov2/configs/__init__.py

@@ -0,0 +1,23 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import pathlib
+
+from omegaconf import OmegaConf
+
+
+def load_config(config_name: str):
+    config_filename = config_name + ".yaml"
+    return OmegaConf.load(pathlib.Path(__file__).parent.resolve() / config_filename)
+
+
+dinov2_default_config = load_config("ssl_default_config")
+
+
+def load_and_merge_config(config_name: str):
+    default_config = OmegaConf.create(dinov2_default_config)
+    loaded_config = load_config(config_name)
+    return OmegaConf.merge(default_config, loaded_config)

dinov2/dinov2/configs/eval/vitb14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_base
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

dinov2/dinov2/configs/eval/vitg14_pretrain.yaml

@@ -0,0 +1,7 @@
+student:
+  arch: vit_giant2
+  patch_size: 14
+  ffn_layer: swiglufused
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

dinov2/dinov2/configs/eval/vitl14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_large
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

dinov2/dinov2/configs/eval/vits14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_small
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

dinov2/dinov2/configs/ssl_default_config.yaml

@@ -0,0 +1,115 @@
+MODEL:
+  WEIGHTS: ''
+compute_precision:
+  grad_scaler: true
+  teacher:
+    backbone:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    dino_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    ibot_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+  student:
+    backbone:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    dino_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp32
+        buffer_dtype: fp32
+    ibot_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp32
+        buffer_dtype: fp32
+dino:
+  loss_weight: 1.0
+  head_n_prototypes: 65536
+  head_bottleneck_dim: 256
+  head_nlayers: 3
+  head_hidden_dim: 2048
+  koleo_loss_weight: 0.1
+ibot:
+  loss_weight: 1.0
+  mask_sample_probability: 0.5
+  mask_ratio_min_max:
+  - 0.1
+  - 0.5
+  separate_head: false
+  head_n_prototypes: 65536
+  head_bottleneck_dim: 256
+  head_nlayers: 3
+  head_hidden_dim: 2048
+train:
+  batch_size_per_gpu: 64
+  dataset_path: ImageNet:split=TRAIN
+  output_dir: .
+  saveckp_freq: 20
+  seed: 0
+  num_workers: 10
+  OFFICIAL_EPOCH_LENGTH: 1250
+  cache_dataset: true
+  centering: "centering" # or "sinkhorn_knopp"
+student:
+  arch: vit_large
+  patch_size: 16
+  drop_path_rate: 0.3
+  layerscale: 1.0e-05
+  drop_path_uniform: true
+  pretrained_weights: ''
+  ffn_layer: "mlp"
+  block_chunks: 0
+  qkv_bias: true
+  proj_bias: true
+  ffn_bias: true
+teacher:
+  momentum_teacher: 0.992
+  final_momentum_teacher: 1
+  warmup_teacher_temp: 0.04
+  teacher_temp: 0.07
+  warmup_teacher_temp_epochs: 30
+optim:
+  epochs: 100
+  weight_decay: 0.04
+  weight_decay_end: 0.4
+  base_lr: 0.004  # learning rate for a batch size of 1024
+  lr: 0.  # will be set after applying scaling rule
+  warmup_epochs: 10
+  min_lr: 1.0e-06
+  clip_grad: 3.0
+  freeze_last_layer_epochs: 1
+  scaling_rule: sqrt_wrt_1024
+  patch_embed_lr_mult: 0.2
+  layerwise_decay: 0.9
+  adamw_beta1: 0.9
+  adamw_beta2: 0.999
+crops:
+  global_crops_scale:
+  - 0.32
+  - 1.0
+  local_crops_number: 8
+  local_crops_scale:
+  - 0.05
+  - 0.32
+  global_crops_size: 224
+  local_crops_size: 96
+evaluation:
+  eval_period_iterations: 12500

dinov2/dinov2/configs/train/vitg14.yaml

@@ -0,0 +1,26 @@
+dino:
+  head_n_prototypes: 131072
+  head_bottleneck_dim: 384
+ibot:
+  separate_head: true
+  head_n_prototypes: 131072
+train:
+  batch_size_per_gpu: 12
+  dataset_path: ImageNet22k
+  centering: sinkhorn_knopp
+student:
+  arch: vit_giant2
+  patch_size: 14
+  drop_path_rate: 0.4
+  ffn_layer: swiglufused
+  block_chunks: 4
+teacher:
+  momentum_teacher: 0.994
+optim:
+  epochs: 500
+  weight_decay_end: 0.2
+  base_lr: 2.0e-04  # learning rate for a batch size of 1024
+  warmup_epochs: 80
+  layerwise_decay: 1.0
+crops:
+  local_crops_size: 98

dinov2/dinov2/configs/train/vitl14.yaml

@@ -0,0 +1,26 @@
+dino:
+  head_n_prototypes: 131072
+  head_bottleneck_dim: 384
+ibot:
+  separate_head: true
+  head_n_prototypes: 131072
+train:
+  batch_size_per_gpu: 32
+  dataset_path: ImageNet22k
+  centering: sinkhorn_knopp
+student:
+  arch: vit_large
+  patch_size: 14
+  drop_path_rate: 0.4
+  ffn_layer: swiglufused
+  block_chunks: 4
+teacher:
+  momentum_teacher: 0.994
+optim:
+  epochs: 500
+  weight_decay_end: 0.2
+  base_lr: 2.0e-04  # learning rate for a batch size of 1024
+  warmup_epochs: 80
+  layerwise_decay: 1.0
+crops:
+  local_crops_size: 98

dinov2/dinov2/configs/train/vitl16_short.yaml

@@ -0,0 +1,6 @@
+# this corresponds to the default config
+train:
+  dataset_path: ImageNet:split=TRAIN
+  batch_size_per_gpu: 64
+student:
+  block_chunks: 4

dinov2/dinov2/data/__init__.py

@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .adapters import DatasetWithEnumeratedTargets
+from .loaders import make_data_loader, make_dataset, SamplerType
+from .collate import collate_data_and_cast
+from .masking import MaskingGenerator
+from .augmentations import DataAugmentationDINO