add PTI

PTI/.gitignore

@@ -0,0 +1 @@
+

PTI/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2021 Daniel Roich
+
+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.

PTI/README.md

@@ -0,0 +1,229 @@
+# PTI: Pivotal Tuning for Latent-based editing of Real Images
+
+<!-- > Recently, a surge of advanced facial editing techniques have been proposed
+that leverage the generative power of a pre-trained StyleGAN. To successfully
+edit an image this way, one must first project (or invert) the image into
+the pre-trained generator’s domain. As it turns out, however, StyleGAN’s
+latent space induces an inherent tradeoff between distortion and editability,
+i.e. between maintaining the original appearance and convincingly altering
+some of its attributes. Practically, this means it is still challenging to
+apply ID-preserving facial latent-space editing to faces which are out of the
+generator’s domain. In this paper, we present an approach to bridge this
+gap. Our technique slightly alters the generator, so that an out-of-domain
+image is faithfully mapped into an in-domain latent code. The key idea is
+pivotal tuning — a brief training process that preserves the editing quality
+of an in-domain latent region, while changing its portrayed identity and
+appearance. In Pivotal Tuning Inversion (PTI), an initial inverted latent code
+serves as a pivot, around which the generator is fined-tuned. At the same
+time, a regularization term keeps nearby identities intact, to locally contain
+the effect. This surgical training process ends up altering appearance features
+that represent mostly identity, without affecting editing capabilities.
+To supplement this, we further show that pivotal tuning can also adjust the
+generator to accommodate a multitude of faces, while introducing negligible
+distortion on the rest of the domain. We validate our technique through
+inversion and editing metrics, and show preferable scores to state-of-the-art
+methods. We further qualitatively demonstrate our technique by applying
+advanced edits (such as pose, age, or expression) to numerous images of
+well-known and recognizable identities. Finally, we demonstrate resilience
+to harder cases, including heavy make-up, elaborate hairstyles and/or headwear,
+which otherwise could not have been successfully inverted and edited
+by state-of-the-art methods. -->
+
+<a href="https://arxiv.org/abs/2106.05744"><img src="https://img.shields.io/badge/arXiv-2008.00951-b31b1b.svg"></a>
+<a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-yellow.svg"></a>  
+Inference Notebook: <a href="https://colab.research.google.com/github/danielroich/PTI/blob/main/notebooks/inference_playground.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" height=20></a>  
+
+<p align="center">
+<img src="docs/teaser.jpg"/>  
+<br>
+Pivotal Tuning Inversion (PTI) enables employing off-the-shelf latent based
+semantic editing techniques on real images using StyleGAN. 
+PTI excels in identity preserving edits, portrayed through recognizable figures —
+Serena Williams and Robert Downey Jr. (top), and in handling faces which
+are clearly out-of-domain, e.g., due to heavy makeup (bottom).
+</br>
+</p>
+
+## Description   
+Official Implementation of our PTI paper + code for evaluation metrics. PTI introduces an optimization mechanizem for solving the StyleGAN inversion task.
+Providing near-perfect reconstruction results while maintaining the high editing abilitis of the native StyleGAN latent space W. For more details, see <a href="https://arxiv.org/abs/2106.05744"><img src="https://img.shields.io/badge/arXiv-2008.00951-b31b1b.svg"></a>
+
+## Recent Updates
+**2021.07.01**: Fixed files download phase in the inference notebook. Which might caused the notebook not to run smoothly.
+
+**2021.06.29**: Added support for CPU. In order to run PTI on CPU please change `device` parameter under `configs/global_config.py` to "cpu" instead of "cuda".
+
+**2021.06.25** : Adding mohawk edit using StyleCLIP+PTI in inference notebook.
+	      Updating documentation in inference notebook due to Google Drive rate limit reached.
+	      Currently, Google Drive does not allow to download the pretrined models using Colab automatically. Manual intervention might be needed.
+
+## Getting Started
+### Prerequisites
+- Linux or macOS
+- NVIDIA GPU + CUDA CuDNN (Not mandatory bur recommended)
+- Python 3
+
+### Installation
+- Dependencies:  
+	1. lpips
+	2. wandb
+	3. pytorch
+	4. torchvision
+	5. matplotlib
+	6. dlib
+- All dependencies can be installed using *pip install* and the package name
+
+## Pretrained Models
+Please download the pretrained models from the following links.
+
+### Auxiliary Models
+We provide various auxiliary models needed for PTI inversion task.  
+This includes the StyleGAN generator and pre-trained models used for loss computation.
+| Path | Description
+| :--- | :----------
+|[FFHQ StyleGAN](https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/ffhq.pkl) | StyleGAN2-ada model trained on FFHQ with 1024x1024 output resolution.
+|[Dlib alignment](https://drive.google.com/file/d/1HKmjg6iXsWr4aFPuU0gBXPGR83wqMzq7/view?usp=sharing) | Dlib alignment used for images preproccessing.
+|[FFHQ e4e encoder](https://drive.google.com/file/d/1ALC5CLA89Ouw40TwvxcwebhzWXM5YSCm/view?usp=sharing) | Pretrained e4e encoder. Used for StyleCLIP editing.
+
+Note: The StyleGAN model is used directly from the official [stylegan2-ada-pytorch implementation](https://github.com/NVlabs/stylegan2-ada-pytorch).
+For StyleCLIP pretrained mappers, please see [StyleCLIP's official routes](https://github.com/orpatashnik/StyleCLIP/blob/main/utils.py)
+
+
+By default, we assume that all auxiliary models are downloaded and saved to the directory `pretrained_models`. 
+However, you may use your own paths by changing the necessary values in `configs/path_configs.py`. 
+
+
+## Inversion
+### Preparing your Data
+In order to invert a real image and edit it you should first align and crop it to the correct size. To do so you should perform *One* of the following steps: 
+1. Run `notebooks/align_data.ipynb` and change the "images_path" variable to the raw images path
+2. Run `utils/align_data.py` and change the "images_path" variable to the raw images path
+
+
+### Weights And Biases
+The project supports [Weights And Biases](https://wandb.ai/home) framework for experiment tracking. For the inversion task it enables visualization of the losses progression and the generator intermediate results during the initial inversion and the *Pivotal Tuning*(PT) procedure.
+
+The log frequency can be adjusted using the parameters defined at `configs/global_config.py` under the "Logs" subsection.
+
+There is no no need to have an account. However, in order to use the features provided by Weights and Biases you first have to register on their site.
+
+
+### Running PTI
+The main training script is `scripts/run_pti.py`. The script receives aligned and cropped images from paths configured in the "Input info" subscetion in
+ `configs/paths_config.py`. 
+Results are saved to directories found at "Dirs for output files" under `configs/paths_config.py`. This includes inversion latent codes and tuned generators. 
+The hyperparametrs for the inversion task can be found at  `configs/hyperparameters.py`. They are intilized to the default values used in the paper. 
+
+## Editing
+By default, we assume that all auxiliary edit directions are downloaded and saved to the directory `editings`. 
+However, you may use your own paths by changing the necessary values in `configs/path_configs.py` under "Edit directions" subsection.
+
+Example of editing code can be found at `scripts/latent_editor_wrapper.py`
+
+## Inference Notebooks
+To help visualize the results of PTI we provide a Jupyter notebook found in `notebooks/inference_playground.ipynb`.   
+The notebook will download the pretrained models and run inference on a sample image found online or 
+on images of your choosing. It is recommended to run this in [Google Colab](https://colab.research.google.com/github/danielroich/PTI/blob/main/notebooks/inference_playground.ipynb).
+
+The notebook demonstrates how to:
+- Invert an image using PTI
+- Visualise the inversion and use the PTI output
+- Edit the image after PTI using InterfaceGAN and StyleCLIP
+- Compare to other inversion methods
+
+## Evaluation
+Currently the repository supports qualitative evaluation for reconstruction of: PTI, SG2 (*W Space*), e4e, SG2Plus (*W+ Space*). 
+As well as editing using InterfaceGAN and GANSpace for the same inversion methods.
+To run the evaluation please see `evaluation/qualitative_edit_comparison.py`. Examples of the evaluation scripts are:
+
+<p align="center">
+<img src="docs/model_rec.jpg"/>  
+<br>
+Reconsturction comparison between different methods. The images order is: Original image, W+ inversion, e4e inversion, W inversion, PTI inversion
+</br>  
+</p>
+
+<p align="center">
+<img src="docs/stern_rotation.jpg"/>  
+<br>
+InterfaceGAN pose edit comparison between different methods. The images order is: Original, W+, e4e, W, PTI
+</br>  
+</p>
+
+<p align="center">
+<img src="docs/tyron_original.jpg" width="220" height="220"/>  
+<img src="docs/tyron_edit.jpg" width="220" height="220"/>
+<br>
+Image per edit or several edits without comparison
+</br>  
+</p>
+
+###  Coming Soon - Quantitative evaluation and StyleCLIP qualitative evaluation
+
+## Repository structure
+| Path | Description <img width=200>
+| :--- | :---
+| &boxvr;&nbsp; configs | Folder containing configs defining Hyperparameters, paths and logging
+| &boxvr;&nbsp; criteria | Folder containing various loss and regularization criterias for the optimization
+| &boxvr;&nbsp; dnnlib | Folder containing internal utils for StyleGAN2-ada
+| &boxvr;&nbsp; docs | Folder containing the latent space edit directions
+| &boxvr;&nbsp; editings | Folder containing images displayed in the README
+| &boxvr;&nbsp; environment | Folder containing Anaconda environment used in our experiments
+| &boxvr;&nbsp; licenses | Folder containing licenses of the open source projects used in this repository
+| &boxvr;&nbsp; models | Folder containing models used in different editing techniques and first phase inversion
+| &boxvr;&nbsp; notebooks | Folder with jupyter notebooks to demonstrate the usage of PTI end-to-end
+| &boxvr;&nbsp; scripts | Folder with running scripts for inversion, editing and metric computations
+| &boxvr;&nbsp; torch_utils | Folder containing internal utils for StyleGAN2-ada
+| &boxvr;&nbsp; training | Folder containing the core training logic of PTI
+| &boxvr;&nbsp; utils | Folder with various utility functions
+
+
+## Credits
+**StyleGAN2-ada model and implementation:**  
+https://github.com/NVlabs/stylegan2-ada-pytorch
+Copyright © 2021, NVIDIA Corporation.  
+Nvidia Source Code License https://nvlabs.github.io/stylegan2-ada-pytorch/license.html
+
+**LPIPS model and implementation:**  
+https://github.com/richzhang/PerceptualSimilarity  
+Copyright (c) 2020, Sou Uchida  
+License (BSD 2-Clause) https://github.com/richzhang/PerceptualSimilarity/blob/master/LICENSE
+
+**e4e model and implementation:**   
+https://github.com/omertov/encoder4editing
+Copyright (c) 2021 omertov  
+License (MIT) https://github.com/omertov/encoder4editing/blob/main/LICENSE
+
+**StyleCLIP model and implementation:**   
+https://github.com/orpatashnik/StyleCLIP
+Copyright (c) 2021 orpatashnik  
+License (MIT) https://github.com/orpatashnik/StyleCLIP/blob/main/LICENSE
+
+**InterfaceGAN implementation:**   
+https://github.com/genforce/interfacegan
+Copyright (c) 2020 genforce  
+License (MIT) https://github.com/genforce/interfacegan/blob/master/LICENSE
+
+**GANSpace implementation:**   
+https://github.com/harskish/ganspace
+Copyright (c) 2020 harkish  
+License (Apache License 2.0) https://github.com/harskish/ganspace/blob/master/LICENSE
+
+
+## Acknowledgments
+This repository structure is based on [encoder4editing](https://github.com/omertov/encoder4editing) and [ReStyle](https://github.com/yuval-alaluf/restyle-encoder) repositories
+
+## Contact
+For any inquiry please contact us at our email addresses: [email protected] or [email protected]
+
+
+## Citation
+If you use this code for your research, please cite:
+```
+@article{roich2021pivotal,
+  title={Pivotal Tuning for Latent-based Editing of Real Images},
+  author={Roich, Daniel and Mokady, Ron and Bermano, Amit H and Cohen-Or, Daniel},
+  journal={arXiv preprint arXiv:2106.05744},
+  year={2021}
+}
+```

PTI/color_transfer_loss.py

@@ -0,0 +1,60 @@
+from typing import List, Optional
+
+import torch
+from torch import nn
+from torch.nn.functional import (
+    smooth_l1_loss,
+)
+
+
+def flatten_CHW(im: torch.Tensor) -> torch.Tensor:
+    """
+    (B, C, H, W) -> (B, -1)
+    """
+    B = im.shape[0]
+    return im.reshape(B, -1)
+
+
+def stddev(x: torch.Tensor) -> torch.Tensor:
+    """
+    x: (B, -1), assume with mean normalized
+    Retuens:
+        stddev: (B)
+    """
+    return torch.sqrt(torch.mean(x * x, dim=-1))
+
+
+def gram_matrix(input_):
+    B, C = input_.shape[:2]
+    features = input_.view(B, C, -1)
+    N = features.shape[-1]
+    G = torch.bmm(features, features.transpose(1, 2))  # C x C
+    return G.div(C * N)
+
+
+class ColorTransferLoss(nn.Module):
+    """Penalize the gram matrix difference between StyleGAN2's ToRGB outputs"""
+    def __init__(
+        self,
+        init_rgbs,
+        scale_rgb: bool = False
+    ):
+        super().__init__()
+
+        with torch.no_grad():
+            init_feats = [x.detach() for x in init_rgbs]
+            self.stds = [stddev(flatten_CHW(rgb)) if scale_rgb else 1 for rgb in init_feats]  # (B, 1, 1, 1) or scalar
+            self.grams = [gram_matrix(rgb / std) for rgb, std in zip(init_feats, self.stds)]
+
+    def forward(self, rgbs: List[torch.Tensor], level: int = None):
+        if level is None:
+            level = len(self.grams)
+
+        feats = rgbs
+        loss = 0
+        for i, (rgb, std) in enumerate(zip(feats[:level], self.stds[:level])):
+            G = gram_matrix(rgb / std)
+            loss = loss + smooth_l1_loss(G, self.grams[i])
+
+        return loss
+

PTI/configs/evaluation_config.py

@@ -0,0 +1 @@
+evaluated_methods = ['e4e', 'SG2', 'SG2Plus']

PTI/configs/global_config.py

@@ -0,0 +1,12 @@
+## Device
+cuda_visible_devices = "1"
+device = "cuda:0"
+
+## Logs
+training_step = 1
+image_rec_result_log_snapshot = 100
+pivotal_training_steps = 0
+model_snapshot_interval = 400
+
+## Run name to be updated during PTI
+run_name = ""

PTI/configs/hyperparameters.py

@@ -0,0 +1,31 @@
+## Architechture
+lpips_type = "alex"
+first_inv_type = "w"
+optim_type = "adam"
+
+## Locality regularization
+latent_ball_num_of_samples = 1
+locality_regularization_interval = 1
+use_locality_regularization = False
+regulizer_l2_lambda = 0.1
+regulizer_lpips_lambda = 0.1
+regulizer_alpha = 30
+
+## Loss
+use_mask = True
+pt_l2_lambda = 0.7
+pt_lpips_lambda = 1
+color_transfer_lambda = 0  # 1e6
+id_lambda = 1
+
+## Steps
+LPIPS_value_threshold = 0.01 # 0.06
+max_pti_steps = 350
+first_inv_steps = 450
+max_images_to_invert = 10
+
+## Optimization
+pti_learning_rate = 3e-4
+first_inv_lr = 5e-3
+train_batch_size = 1
+use_last_w_pivots = True

PTI/configs/paths_config.py

@@ -0,0 +1,41 @@
+year = "2010"
+
+## Pretrained models paths
+e4e = "./pretrained_models/e4e_ffhq_encode.pt"
+
+
+stylegan2_ada_ffhq = f"../pretrained_models/{year}.pkl"
+
+style_clip_pretrained_mappers = ""
+ir_se50 = "/share/phoenix/nfs04/S7/wikitime_models/model_ir_se50.pth"
+dlib = "./pretrained_models/align.dat"
+deeplab = "/share/phoenix/nfs04/S7/wikitime_models/deeplab_model/deeplab_model.pth"
+
+## Dirs for output files
+checkpoints_dir = "./checkpoints"
+embedding_base_dir = "./embeddings"
+styleclip_output_dir = "./StyleCLIP_results"
+experiments_output_dir = "./output"
+
+## Input info
+### Input dir, where the images reside
+input_data_path = (
+    f"/share/phoenix/nfs04/S7/emc348/WikiFaces/datasets/new_crops/test/{year}"
+)
+input_data_id = f"{year}"
+
+
+
+
+## Keywords
+pti_results_keyword = "PTI"
+e4e_results_keyword = "e4e"
+sg2_results_keyword = "SG2"
+sg2_plus_results_keyword = "SG2_plus"
+multi_id_model_type = "multi_id_fine"
+
+## Edit directions
+interfacegan_age = "editings/interfacegan_directions/age.pt"
+interfacegan_smile = "editings/interfacegan_directions/smile.pt"
+interfacegan_rotation = "editings/interfacegan_directions/rotation.pt"
+ffhq_pca = "editings/ganspace_pca/ffhq_pca.pt"

PTI/criteria/backbones/__init__.py

@@ -0,0 +1,25 @@
+from .iresnet import iresnet18, iresnet34, iresnet50, iresnet100, iresnet200
+from .mobilefacenet import get_mbf
+
+
+def get_model(name, **kwargs):
+    # resnet
+    if name == "r18":
+        return iresnet18(False, **kwargs)
+    elif name == "r34":
+        return iresnet34(False, **kwargs)
+    elif name == "r50":
+        return iresnet50(False, **kwargs)
+    elif name == "r100":
+        return iresnet100(False, **kwargs)
+    elif name == "r200":
+        return iresnet200(False, **kwargs)
+    elif name == "r2060":
+        from .iresnet2060 import iresnet2060
+        return iresnet2060(False, **kwargs)
+    elif name == "mbf":
+        fp16 = kwargs.get("fp16", False)
+        num_features = kwargs.get("num_features", 512)
+        return get_mbf(fp16=fp16, num_features=num_features)
+    else:
+        raise ValueError()

PTI/criteria/backbones/iresnet.py

@@ -0,0 +1,186 @@
+import torch
+from torch import nn
+
+__all__ = ['iresnet18', 'iresnet34', 'iresnet50', 'iresnet100', 'iresnet200']
+
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05,)
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+    def __init__(self,
+                 block, layers, dropout=0, num_features=512, zero_init_residual=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
+        super(IResNet, self).__init__()
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.PReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2])
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05,)
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
+            )
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = self.layer1(x)
+            x = self.layer2(x)
+            x = self.layer3(x)
+            x = self.layer4(x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            x = self.dropout(x)
+        x = self.fc(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet18(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet18', IBasicBlock, [2, 2, 2, 2], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet34(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet34', IBasicBlock, [3, 4, 6, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet50(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet50', IBasicBlock, [3, 4, 14, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet100(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet100', IBasicBlock, [3, 13, 30, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet200(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet200', IBasicBlock, [6, 26, 60, 6], pretrained,
+                    progress, **kwargs)

PTI/criteria/backbones/iresnet2060.py

@@ -0,0 +1,176 @@
+import torch
+from torch import nn
+
+assert torch.__version__ >= "1.8.1"
+from torch.utils.checkpoint import checkpoint_sequential
+
+__all__ = ['iresnet2060']
+
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05, )
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05, )
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05, )
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+
+    def __init__(self,
+                 block, layers, dropout=0, num_features=512, zero_init_residual=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
+        super(IResNet, self).__init__()
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.PReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2])
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05, )
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
+            )
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation))
+
+        return nn.Sequential(*layers)
+
+    def checkpoint(self, func, num_seg, x):
+        if self.training:
+            return checkpoint_sequential(func, num_seg, x)
+        else:
+            return func(x)
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = self.layer1(x)
+            x = self.checkpoint(self.layer2, 20, x)
+            x = self.checkpoint(self.layer3, 100, x)
+            x = self.layer4(x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            x = self.dropout(x)
+        x = self.fc(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet2060(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet2060', IBasicBlock, [3, 128, 1024 - 128, 3], pretrained, progress, **kwargs)

PTI/criteria/backbones/mobilefacenet.py

@@ -0,0 +1,130 @@
+'''
+Adapted from https://github.com/cavalleria/cavaface.pytorch/blob/master/backbone/mobilefacenet.py
+Original author cavalleria
+'''
+
+import torch.nn as nn
+from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, Sequential, Module
+import torch
+
+
+class Flatten(Module):
+    def forward(self, x):
+        return x.view(x.size(0), -1)
+
+
+class ConvBlock(Module):
+    def __init__(self, in_c, out_c, kernel=(1, 1), stride=(1, 1), padding=(0, 0), groups=1):
+        super(ConvBlock, self).__init__()
+        self.layers = nn.Sequential(
+            Conv2d(in_c, out_c, kernel, groups=groups, stride=stride, padding=padding, bias=False),
+            BatchNorm2d(num_features=out_c),
+            PReLU(num_parameters=out_c)
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class LinearBlock(Module):
+    def __init__(self, in_c, out_c, kernel=(1, 1), stride=(1, 1), padding=(0, 0), groups=1):
+        super(LinearBlock, self).__init__()
+        self.layers = nn.Sequential(
+            Conv2d(in_c, out_c, kernel, stride, padding, groups=groups, bias=False),
+            BatchNorm2d(num_features=out_c)
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class DepthWise(Module):
+    def __init__(self, in_c, out_c, residual=False, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=1):
+        super(DepthWise, self).__init__()
+        self.residual = residual
+        self.layers = nn.Sequential(
+            ConvBlock(in_c, out_c=groups, kernel=(1, 1), padding=(0, 0), stride=(1, 1)),
+            ConvBlock(groups, groups, groups=groups, kernel=kernel, padding=padding, stride=stride),
+            LinearBlock(groups, out_c, kernel=(1, 1), padding=(0, 0), stride=(1, 1))
+        )
+
+    def forward(self, x):
+        short_cut = None
+        if self.residual:
+            short_cut = x
+        x = self.layers(x)
+        if self.residual:
+            output = short_cut + x
+        else:
+            output = x
+        return output
+
+
+class Residual(Module):
+    def __init__(self, c, num_block, groups, kernel=(3, 3), stride=(1, 1), padding=(1, 1)):
+        super(Residual, self).__init__()
+        modules = []
+        for _ in range(num_block):
+            modules.append(DepthWise(c, c, True, kernel, stride, padding, groups))
+        self.layers = Sequential(*modules)
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class GDC(Module):
+    def __init__(self, embedding_size):
+        super(GDC, self).__init__()
+        self.layers = nn.Sequential(
+            LinearBlock(512, 512, groups=512, kernel=(7, 7), stride=(1, 1), padding=(0, 0)),
+            Flatten(),
+            Linear(512, embedding_size, bias=False),
+            BatchNorm1d(embedding_size))
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class MobileFaceNet(Module):
+    def __init__(self, fp16=False, num_features=512):
+        super(MobileFaceNet, self).__init__()
+        scale = 2
+        self.fp16 = fp16
+        self.layers = nn.Sequential(
+            ConvBlock(3, 64 * scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1)),
+            ConvBlock(64 * scale, 64 * scale, kernel=(3, 3), stride=(1, 1), padding=(1, 1), groups=64),
+            DepthWise(64 * scale, 64 * scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=128),
+            Residual(64 * scale, num_block=4, groups=128, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+            DepthWise(64 * scale, 128 * scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=256),
+            Residual(128 * scale, num_block=6, groups=256, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+            DepthWise(128 * scale, 128 * scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=512),
+            Residual(128 * scale, num_block=2, groups=256, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+        )
+        self.conv_sep = ConvBlock(128 * scale, 512, kernel=(1, 1), stride=(1, 1), padding=(0, 0))
+        self.features = GDC(num_features)
+        self._initialize_weights()
+
+    def _initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+                if m.bias is not None:
+                    m.bias.data.zero_()
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+            elif isinstance(m, nn.Linear):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+                if m.bias is not None:
+                    m.bias.data.zero_()
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.layers(x)
+        x = self.conv_sep(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def get_mbf(fp16, num_features):
+    return MobileFaceNet(fp16, num_features)

PTI/criteria/deeplab.py

@@ -0,0 +1,353 @@
+# Taken from the https://github.com/chenxi116/DeepLabv3.pytorch repository.
+
+import torch
+import torch.nn as nn
+import math
+import torch.utils.model_zoo as model_zoo
+from torch.nn import functional as F
+import os
+
+
+__all__ = ["ResNet", "resnet50", "resnet101", "resnet152"]
+
+
+model_urls = {
+    "resnet50": "https://download.pytorch.org/models/resnet50-19c8e357.pth",
+    "resnet101": "https://download.pytorch.org/models/resnet101-5d3b4d8f.pth",
+    "resnet152": "https://download.pytorch.org/models/resnet152-b121ed2d.pth",
+}
+
+
+class Conv2d(nn.Conv2d):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        padding=0,
+        dilation=1,
+        groups=1,
+        bias=True,
+    ):
+        super(Conv2d, self).__init__(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride,
+            padding,
+            dilation,
+            groups,
+            bias,
+        )
+
+    def forward(self, x):
+        # return super(Conv2d, self).forward(x)
+        weight = self.weight
+        weight_mean = (
+            weight.mean(dim=1, keepdim=True)
+            .mean(dim=2, keepdim=True)
+            .mean(dim=3, keepdim=True)
+        )
+        weight = weight - weight_mean
+        std = weight.view(weight.size(0), -1).std(dim=1).view(-1, 1, 1, 1) + 1e-5
+        weight = weight / std.expand_as(weight)
+        return F.conv2d(
+            x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups
+        )
+
+
+class ASPP(nn.Module):
+    def __init__(
+        self,
+        C,
+        depth,
+        num_classes,
+        conv=nn.Conv2d,
+        norm=nn.BatchNorm2d,
+        momentum=0.0003,
+        mult=1,
+    ):
+        super(ASPP, self).__init__()
+        self._C = C
+        self._depth = depth
+        self._num_classes = num_classes
+
+        self.global_pooling = nn.AdaptiveAvgPool2d(1)
+        self.relu = nn.ReLU(inplace=True)
+        self.aspp1 = conv(C, depth, kernel_size=1, stride=1, bias=False)
+        self.aspp2 = conv(
+            C,
+            depth,
+            kernel_size=3,
+            stride=1,
+            dilation=int(6 * mult),
+            padding=int(6 * mult),
+            bias=False,
+        )
+        self.aspp3 = conv(
+            C,
+            depth,
+            kernel_size=3,
+            stride=1,
+            dilation=int(12 * mult),
+            padding=int(12 * mult),
+            bias=False,
+        )
+        self.aspp4 = conv(
+            C,
+            depth,
+            kernel_size=3,
+            stride=1,
+            dilation=int(18 * mult),
+            padding=int(18 * mult),
+            bias=False,
+        )
+        self.aspp5 = conv(C, depth, kernel_size=1, stride=1, bias=False)
+        self.aspp1_bn = norm(depth, momentum)
+        self.aspp2_bn = norm(depth, momentum)
+        self.aspp3_bn = norm(depth, momentum)
+        self.aspp4_bn = norm(depth, momentum)
+        self.aspp5_bn = norm(depth, momentum)
+        self.conv2 = conv(depth * 5, depth, kernel_size=1, stride=1, bias=False)
+        self.bn2 = norm(depth, momentum)
+        self.conv3 = nn.Conv2d(depth, num_classes, kernel_size=1, stride=1)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x1 = self.aspp1_bn(x1)
+        x1 = self.relu(x1)
+        x2 = self.aspp2(x)
+        x2 = self.aspp2_bn(x2)
+        x2 = self.relu(x2)
+        x3 = self.aspp3(x)
+        x3 = self.aspp3_bn(x3)
+        x3 = self.relu(x3)
+        x4 = self.aspp4(x)
+        x4 = self.aspp4_bn(x4)
+        x4 = self.relu(x4)
+        x5 = self.global_pooling(x)
+        x5 = self.aspp5(x5)
+        x5 = self.aspp5_bn(x5)
+        x5 = self.relu(x5)
+        x5 = nn.Upsample((x.shape[2], x.shape[3]), mode="bilinear", align_corners=True)(
+            x5
+        )
+        x = torch.cat((x1, x2, x3, x4, x5), 1)
+        x = self.conv2(x)
+        x = self.bn2(x)
+        x = self.relu(x)
+        x = self.conv3(x)
+
+        return x
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(
+        self,
+        inplanes,
+        planes,
+        stride=1,
+        downsample=None,
+        dilation=1,
+        conv=None,
+        norm=None,
+    ):
+        super(Bottleneck, self).__init__()
+        self.conv1 = conv(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = norm(planes)
+        self.conv2 = conv(
+            planes,
+            planes,
+            kernel_size=3,
+            stride=stride,
+            dilation=dilation,
+            padding=dilation,
+            bias=False,
+        )
+        self.bn2 = norm(planes)
+        self.conv3 = conv(planes, planes * self.expansion, kernel_size=1, bias=False)
+        self.bn3 = norm(planes * self.expansion)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class ResNet(nn.Module):
+    def __init__(
+        self, block, layers, num_classes, num_groups=None, weight_std=False, beta=False
+    ):
+        self.inplanes = 64
+        self.norm = (
+            lambda planes, momentum=0.05: nn.BatchNorm2d(planes, momentum=momentum)
+            if num_groups is None
+            else nn.GroupNorm(num_groups, planes)
+        )
+        self.conv = Conv2d if weight_std else nn.Conv2d
+
+        super(ResNet, self).__init__()
+        if not beta:
+            self.conv1 = self.conv(
+                3, 64, kernel_size=7, stride=2, padding=3, bias=False
+            )
+        else:
+            self.conv1 = nn.Sequential(
+                self.conv(3, 64, 3, stride=2, padding=1, bias=False),
+                self.conv(64, 64, 3, stride=1, padding=1, bias=False),
+                self.conv(64, 64, 3, stride=1, padding=1, bias=False),
+            )
+        self.bn1 = self.norm(64)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=1, dilation=2)
+        self.aspp = ASPP(
+            512 * block.expansion, 256, num_classes, conv=self.conv, norm=self.norm
+        )
+
+        for m in self.modules():
+            if isinstance(m, self.conv):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                m.weight.data.normal_(0, math.sqrt(2.0 / n))
+            elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.GroupNorm):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilation=1):
+        downsample = None
+        if stride != 1 or dilation != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                self.conv(
+                    self.inplanes,
+                    planes * block.expansion,
+                    kernel_size=1,
+                    stride=stride,
+                    dilation=max(1, dilation / 2),
+                    bias=False,
+                ),
+                self.norm(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes,
+                planes,
+                stride,
+                downsample,
+                dilation=max(1, dilation / 2),
+                conv=self.conv,
+                norm=self.norm,
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    dilation=dilation,
+                    conv=self.conv,
+                    norm=self.norm,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        size = (x.shape[2], x.shape[3])
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.aspp(x)
+        x = nn.Upsample(size, mode="bilinear", align_corners=True)(x)
+        return x
+
+
+def resnet50(pretrained=False, **kwargs):
+    """Constructs a ResNet-50 model.
+
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
+    if pretrained:
+        model.load_state_dict(model_zoo.load_url(model_urls["resnet50"]))
+    return model
+
+
+def resnet101(path=None, pretrained=False, num_groups=None, weight_std=False, **kwargs):
+    """Constructs a ResNet-101 model.
+
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(
+        Bottleneck,
+        [3, 4, 23, 3],
+        num_groups=num_groups,
+        weight_std=weight_std,
+        **kwargs
+    )
+    if pretrained:
+        model_dict = model.state_dict()
+        if num_groups and weight_std:
+            path = os.path.join(os.path.dirname(path), "R-101-GN-WS.pth.tar")
+            pretrained_dict = torch.load(path)
+            overlap_dict = {
+                k[7:]: v for k, v in pretrained_dict.items() if k[7:] in model_dict
+            }
+            assert len(overlap_dict) == 312
+        elif not num_groups and not weight_std:
+            pretrained_dict = model_zoo.load_url(model_urls["resnet101"])
+            overlap_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
+        else:
+            raise ValueError("Currently only support BN or GN+WS")
+        model_dict.update(overlap_dict)
+        model.load_state_dict(model_dict)
+    return model
+
+
+def resnet152(pretrained=False, **kwargs):
+    """Constructs a ResNet-152 model.
+
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
+    if pretrained:
+        model.load_state_dict(model_zoo.load_url(model_urls["resnet152"]))
+    return model

PTI/criteria/helpers.py

@@ -0,0 +1,119 @@
+from collections import namedtuple
+import torch
+from torch.nn import Conv2d, BatchNorm2d, PReLU, ReLU, Sigmoid, MaxPool2d, AdaptiveAvgPool2d, Sequential, Module
+
+"""
+ArcFace implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
+"""
+
+
+class Flatten(Module):
+	def forward(self, input):
+		return input.view(input.size(0), -1)
+
+
+def l2_norm(input, axis=1):
+	norm = torch.norm(input, 2, axis, True)
+	output = torch.div(input, norm)
+	return output
+
+
+class Bottleneck(namedtuple('Block', ['in_channel', 'depth', 'stride'])):
+	""" A named tuple describing a ResNet block. """
+
+
+def get_block(in_channel, depth, num_units, stride=2):
+	return [Bottleneck(in_channel, depth, stride)] + [Bottleneck(depth, depth, 1) for i in range(num_units - 1)]
+
+
+def get_blocks(num_layers):
+	if num_layers == 50:
+		blocks = [
+			get_block(in_channel=64, depth=64, num_units=3),
+			get_block(in_channel=64, depth=128, num_units=4),
+			get_block(in_channel=128, depth=256, num_units=14),
+			get_block(in_channel=256, depth=512, num_units=3)
+		]
+	elif num_layers == 100:
+		blocks = [
+			get_block(in_channel=64, depth=64, num_units=3),
+			get_block(in_channel=64, depth=128, num_units=13),
+			get_block(in_channel=128, depth=256, num_units=30),
+			get_block(in_channel=256, depth=512, num_units=3)
+		]
+	elif num_layers == 152:
+		blocks = [
+			get_block(in_channel=64, depth=64, num_units=3),
+			get_block(in_channel=64, depth=128, num_units=8),
+			get_block(in_channel=128, depth=256, num_units=36),
+			get_block(in_channel=256, depth=512, num_units=3)
+		]
+	else:
+		raise ValueError("Invalid number of layers: {}. Must be one of [50, 100, 152]".format(num_layers))
+	return blocks
+
+
+class SEModule(Module):
+	def __init__(self, channels, reduction):
+		super(SEModule, self).__init__()
+		self.avg_pool = AdaptiveAvgPool2d(1)
+		self.fc1 = Conv2d(channels, channels // reduction, kernel_size=1, padding=0, bias=False)
+		self.relu = ReLU(inplace=True)
+		self.fc2 = Conv2d(channels // reduction, channels, kernel_size=1, padding=0, bias=False)
+		self.sigmoid = Sigmoid()
+
+	def forward(self, x):
+		module_input = x
+		x = self.avg_pool(x)
+		x = self.fc1(x)
+		x = self.relu(x)
+		x = self.fc2(x)
+		x = self.sigmoid(x)
+		return module_input * x
+
+
+class bottleneck_IR(Module):
+	def __init__(self, in_channel, depth, stride):
+		super(bottleneck_IR, self).__init__()
+		if in_channel == depth:
+			self.shortcut_layer = MaxPool2d(1, stride)
+		else:
+			self.shortcut_layer = Sequential(
+				Conv2d(in_channel, depth, (1, 1), stride, bias=False),
+				BatchNorm2d(depth)
+			)
+		self.res_layer = Sequential(
+			BatchNorm2d(in_channel),
+			Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), PReLU(depth),
+			Conv2d(depth, depth, (3, 3), stride, 1, bias=False), BatchNorm2d(depth)
+		)
+
+	def forward(self, x):
+		shortcut = self.shortcut_layer(x)
+		res = self.res_layer(x)
+		return res + shortcut
+
+
+class bottleneck_IR_SE(Module):
+	def __init__(self, in_channel, depth, stride):
+		super(bottleneck_IR_SE, self).__init__()
+		if in_channel == depth:
+			self.shortcut_layer = MaxPool2d(1, stride)
+		else:
+			self.shortcut_layer = Sequential(
+				Conv2d(in_channel, depth, (1, 1), stride, bias=False),
+				BatchNorm2d(depth)
+			)
+		self.res_layer = Sequential(
+			BatchNorm2d(in_channel),
+			Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False),
+			PReLU(depth),
+			Conv2d(depth, depth, (3, 3), stride, 1, bias=False),
+			BatchNorm2d(depth),
+			SEModule(depth, 16)
+		)
+
+	def forward(self, x):
+		shortcut = self.shortcut_layer(x)
+		res = self.res_layer(x)
+		return res + shortcut

PTI/criteria/id_loss.py

@@ -0,0 +1,64 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from criteria.model_irse import Backbone
+from criteria.backbones import get_model
+
+
+class IDLoss(nn.Module):
+    """
+    Computes a cosine similarity between people in two images.
+    Taken from TreB1eN's [1] implementation of InsightFace [2, 3], as used in pixel2style2pixel [4].
+
+    [1] https://github.com/TreB1eN/InsightFace_Pytorch
+    [2] https://github.com/deepinsight/insightface
+    [3] Deng, Jiankang and Guo, Jia and Niannan, Xue and Zafeiriou, Stefanos.
+        ArcFace: Additive Angular Margin Loss for Deep Face Recognition. In CVPR, 2019
+    [4] https://github.com/eladrich/pixel2style2pixel
+    """
+
+    def __init__(self, model_path, official=False):
+        """
+        Arguments:
+            model_path (str): Path to IR-SE50 model.
+        """
+        super(IDLoss, self).__init__()
+        print("Loading ResNet ArcFace")
+        self.official = official
+        if official:
+            self.facenet = get_model("r100", fp16=False)
+        else:
+            self.facenet = Backbone(
+                input_size=112, num_layers=50, drop_ratio=0.6, mode="ir_se"
+            )
+
+        self.facenet.load_state_dict(torch.load(model_path))
+        self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112))
+        self.facenet.eval()
+
+    def extract_feats(self, x):
+        x = x[:, :, 35:223, 32:220]  # Crop interesting region
+        x = self.face_pool(x)
+        x_feats = self.facenet(x)
+        return x_feats
+
+    def forward(self, x, y):
+        """
+        Arguments:
+            x (Tensor): The batch of original images
+            y (Tensor): The batch of generated images
+
+        Returns:
+            loss (Tensor): Cosine similarity between the
+                features of the original and generated images.
+
+        """
+
+        x_feats = self.extract_feats(x)
+        y_feats = self.extract_feats(y)
+        if self.official:
+            x_feats = F.normalize(x_feats)
+            y_feats = F.normalize(y_feats)
+
+        loss = (1 - (x_feats * y_feats).sum(dim=1)).mean()
+        return loss

PTI/criteria/l2_loss.py

@@ -0,0 +1,14 @@
+import torch
+import torchvision
+
+l2_criterion = torch.nn.MSELoss(reduction="mean")
+
+
+def l2_loss(real_images, generated_images, gray=False):
+    if gray:
+        real_images = torchvision.transforms.functional.rgb_to_grayscale(real_images)
+        generated_images = torchvision.transforms.functional.rgb_to_grayscale(
+            generated_images
+        )
+    loss = l2_criterion(real_images, generated_images)
+    return loss

PTI/criteria/localitly_regulizer.py

@@ -0,0 +1,59 @@
+import torch
+import numpy as np
+import wandb
+from criteria import l2_loss
+from configs import hyperparameters
+from configs import global_config
+
+
+class Space_Regulizer:
+    def __init__(self, original_G, lpips_net):
+        self.original_G = original_G
+        self.morphing_regulizer_alpha = hyperparameters.regulizer_alpha
+        self.lpips_loss = lpips_net
+
+    def get_morphed_w_code(self, new_w_code, fixed_w):
+        interpolation_direction = new_w_code - fixed_w
+        interpolation_direction_norm = torch.norm(interpolation_direction, p=2)
+        direction_to_move = hyperparameters.regulizer_alpha * interpolation_direction / interpolation_direction_norm
+        result_w = fixed_w + direction_to_move
+        self.morphing_regulizer_alpha * fixed_w + (1 - self.morphing_regulizer_alpha) * new_w_code
+
+        return result_w
+
+    def get_image_from_ws(self, w_codes, G):
+        return torch.cat([G.synthesis(w_code, noise_mode='none', force_fp32=True) for w_code in w_codes])
+
+    def ball_holder_loss_lazy(self, new_G, num_of_sampled_latents, w_batch, use_wandb=False):
+        loss = 0.0
+
+        z_samples = np.random.randn(num_of_sampled_latents, self.original_G.z_dim)
+        w_samples = self.original_G.mapping(torch.from_numpy(z_samples).to(global_config.device), None,
+                                            truncation_psi=0.5)
+        territory_indicator_ws = [self.get_morphed_w_code(w_code.unsqueeze(0), w_batch) for w_code in w_samples]
+
+        for w_code in territory_indicator_ws:
+            new_img = new_G.synthesis(w_code, noise_mode='none', force_fp32=True)
+            with torch.no_grad():
+                old_img = self.original_G.synthesis(w_code, noise_mode='none', force_fp32=True)
+
+            if hyperparameters.regulizer_l2_lambda > 0:
+                l2_loss_val = l2_loss.l2_loss(old_img, new_img)
+                if use_wandb:
+                    wandb.log({f'space_regulizer_l2_loss_val': l2_loss_val.detach().cpu()},
+                              step=global_config.training_step)
+                loss += l2_loss_val * hyperparameters.regulizer_l2_lambda
+
+            if hyperparameters.regulizer_lpips_lambda > 0:
+                loss_lpips = self.lpips_loss(old_img, new_img)
+                loss_lpips = torch.mean(torch.squeeze(loss_lpips))
+                if use_wandb:
+                    wandb.log({f'space_regulizer_lpips_loss_val': loss_lpips.detach().cpu()},
+                              step=global_config.training_step)
+                loss += loss_lpips * hyperparameters.regulizer_lpips_lambda
+
+        return loss / len(territory_indicator_ws)
+
+    def space_regulizer_loss(self, new_G, w_batch, use_wandb):
+        ret_val = self.ball_holder_loss_lazy(new_G, hyperparameters.latent_ball_num_of_samples, w_batch, use_wandb)
+        return ret_val

PTI/criteria/mask.py

@@ -0,0 +1,123 @@
+import torch
+import torchvision.transforms as transforms
+import criteria.deeplab as deeplab
+import PIL.Image as Image
+import torch.nn as nn
+import torch.nn.functional as F
+from configs import paths_config, global_config
+import numpy as np
+
+
+class Mask(nn.Module):
+    def __init__(self):
+        """
+
+        |  Class     | Number | Class | Number |
+        |------------|--------|-------|--------|
+        | background |  0     | mouth |  10    |
+        | skin       |  1     | u_lip |  11    |
+        | nose       |  2     | l_lip |  12    |
+        | eye_g      |  3     | hair  |  13    |
+        | l_eye      |  4     | hat   |  14    |
+        | r_eye      |  5     | ear_r |  15    |
+        | l_brow     |  6     | neck_l|  16    |
+        | r_brow     |  7     | neck  |  17    |
+        | l_ear      |  8     | cloth |  18    |
+        | r_ear      |  9     |
+
+        """
+        super().__init__()
+
+        self.seg_model = (
+            getattr(deeplab, "resnet101")(
+                path=paths_config.deeplab,
+                pretrained=True,
+                num_classes=19,
+                num_groups=32,
+                weight_std=True,
+                beta=False,
+            )
+            .eval()
+            .requires_grad_(False)
+        )
+
+        ckpt = torch.load(paths_config.deeplab, map_location=global_config.device)
+        state_dict = {
+            k[7:]: v for k, v in ckpt["state_dict"].items() if "tracked" not in k
+        }
+        self.seg_model.load_state_dict(state_dict)
+        self.seg_model = self.seg_model.to(global_config.device)
+
+        self.labels = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 15, 16, 17]
+        self.kernel = torch.ones((1, 1, 25, 25), device=global_config.device)
+
+    def get_labels(self, img):
+        """Returns a mask from an input image"""
+        data_transforms = transforms.Compose(
+            [
+                transforms.Resize((513, 513)),
+                transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+            ]
+        )
+        img = data_transforms(img)
+        with torch.no_grad():
+            out = self.seg_model(img)
+        _, label = torch.max(out, 1)
+        label = label.unsqueeze(0).type(torch.float32)
+
+        label = (
+            F.interpolate(label, size=(256, 256), mode="nearest")
+            .squeeze()
+            .type(torch.LongTensor)
+        )
+        return label
+
+    def get_mask(self, label):
+        mask = torch.zeros_like(label, device=global_config.device, dtype=torch.float)
+        for idx in self.labels:
+            mask[label == idx] = 1
+
+        # smooth the mask with a mean convolution
+        """mask = (
+            1
+            - torch.clamp(
+                torch.nn.functional.conv2d(
+                    1 - mask[None, None, :, :], self.kernel, padding="same"
+                ),
+                0,
+                1,
+            ).squeeze()
+        )"""
+        """ mask = torch.clamp(
+            torch.nn.functional.conv2d(
+                mask[None, None, :, :], self.kernel, padding="same"
+            ),
+            0,
+            1,
+        ).squeeze()"""
+        mask[label == 13] = 0.1
+        return mask
+
+    def forward(self, real_imgs, generated_imgs):
+        #return real_imgs, generated_imgs
+        label = self.get_labels(real_imgs)
+        mask = self.get_mask(label)
+        real_imgs = real_imgs * mask
+        generated_imgs = generated_imgs * mask
+
+        """out = (real_imgs * mask).squeeze().detach()
+
+        out = (out.permute(1, 2, 0) * 127.5 + 127.5).clamp(0, 255).to(torch.uint8)
+        Image.fromarray(out.cpu().numpy()).save("real_mask.png")
+
+        out = (generated_imgs).squeeze().detach()
+
+        out = (out.permute(1, 2, 0) * 127.5 + 127.5).clamp(0, 255).to(torch.uint8)
+        Image.fromarray(out.cpu().numpy()).save("generated_mask.png")
+
+        mask = (mask).squeeze().detach()
+        mask = mask.repeat(3, 1, 1)
+        mask = (mask.permute(1, 2, 0) * 127.5 + 127.5).clamp(0, 255).to(torch.uint8)
+        Image.fromarray(mask.cpu().numpy()).save("mask.png")"""
+
+        return real_imgs, generated_imgs

PTI/criteria/model_irse.py

@@ -0,0 +1,115 @@
+from torch.nn import (
+    Linear,
+    Conv2d,
+    BatchNorm1d,
+    BatchNorm2d,
+    PReLU,
+    Dropout,
+    Sequential,
+    Module,
+)
+from criteria.helpers import (
+    get_blocks,
+    Flatten,
+    bottleneck_IR,
+    bottleneck_IR_SE,
+    l2_norm,
+)
+
+"""
+Modified Backbone implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
+"""
+
+
+class Backbone(Module):
+    def __init__(self, input_size, num_layers, mode="ir", drop_ratio=0.4, affine=True):
+        super(Backbone, self).__init__()
+        assert input_size in [112, 224], "input_size should be 112 or 224"
+        assert num_layers in [50, 100, 152], "num_layers should be 50, 100 or 152"
+        assert mode in ["ir", "ir_se"], "mode should be ir or ir_se"
+        blocks = get_blocks(num_layers)
+        if mode == "ir":
+            unit_module = bottleneck_IR
+        elif mode == "ir_se":
+            unit_module = bottleneck_IR_SE
+        self.input_layer = Sequential(
+            Conv2d(3, 64, (3, 3), 1, 1, bias=False), BatchNorm2d(64), PReLU(64)
+        )
+        if input_size == 112:
+            self.output_layer = Sequential(
+                BatchNorm2d(512),
+                Dropout(drop_ratio),
+                Flatten(),
+                Linear(512 * 7 * 7, 512),
+                BatchNorm1d(512, affine=affine),
+            )
+        else:
+            self.output_layer = Sequential(
+                BatchNorm2d(512),
+                Dropout(drop_ratio),
+                Flatten(),
+                Linear(512 * 14 * 14, 512),
+                BatchNorm1d(512, affine=affine),
+            )
+
+        modules = []
+        for block in blocks:
+            for bottleneck in block:
+                modules.append(
+                    unit_module(
+                        bottleneck.in_channel, bottleneck.depth, bottleneck.stride
+                    )
+                )
+        self.body = Sequential(*modules)
+
+    def forward(self, x):
+        x = self.input_layer(x)
+        x = self.body(x)
+        x = self.output_layer(x)
+        return l2_norm(x)
+
+
+def IR_50(input_size):
+    """Constructs a ir-50 model."""
+    model = Backbone(input_size, num_layers=50, mode="ir", drop_ratio=0.4, affine=False)
+    return model
+
+
+def IR_101(input_size):
+    """Constructs a ir-101 model."""
+    model = Backbone(
+        input_size, num_layers=100, mode="ir", drop_ratio=0.4, affine=False
+    )
+    return model
+
+
+def IR_152(input_size):
+    """Constructs a ir-152 model."""
+    model = Backbone(
+        input_size, num_layers=152, mode="ir", drop_ratio=0.4, affine=False
+    )
+    return model
+
+
+def IR_SE_50(input_size):
+    """Constructs a ir_se-50 model."""
+    model = Backbone(
+        input_size, num_layers=50, mode="ir_se", drop_ratio=0.4, affine=False
+    )
+    return model
+
+
+def IR_SE_101(input_size):
+    """Constructs a ir_se-101 model."""
+    model = Backbone(
+        input_size, num_layers=100, mode="ir_se", drop_ratio=0.4, affine=False
+    )
+    return model
+
+
+def IR_SE_152(input_size):
+    """Constructs a ir_se-152 model."""
+    model = Backbone(
+        input_size, num_layers=152, mode="ir_se", drop_ratio=0.4, affine=False
+    )
+    return model

PTI/dnnlib/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from .util import EasyDict, make_cache_dir_path

PTI/dnnlib/util.py

@@ -0,0 +1,477 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Miscellaneous utility classes and functions."""
+
+import ctypes
+import fnmatch
+import importlib
+import inspect
+import numpy as np
+import os
+import shutil
+import sys
+import types
+import io
+import pickle
+import re
+import requests
+import html
+import hashlib
+import glob
+import tempfile
+import urllib
+import urllib.request
+import uuid
+
+from distutils.util import strtobool
+from typing import Any, List, Tuple, Union
+
+
+# Util classes
+# ------------------------------------------------------------------------------------------
+
+
+class EasyDict(dict):
+    """Convenience class that behaves like a dict but allows access with the attribute syntax."""
+
+    def __getattr__(self, name: str) -> Any:
+        try:
+            return self[name]
+        except KeyError:
+            raise AttributeError(name)
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        self[name] = value
+
+    def __delattr__(self, name: str) -> None:
+        del self[name]
+
+
+class Logger(object):
+    """Redirect stderr to stdout, optionally print stdout to a file, and optionally force flushing on both stdout and the file."""
+
+    def __init__(self, file_name: str = None, file_mode: str = "w", should_flush: bool = True):
+        self.file = None
+
+        if file_name is not None:
+            self.file = open(file_name, file_mode)
+
+        self.should_flush = should_flush
+        self.stdout = sys.stdout
+        self.stderr = sys.stderr
+
+        sys.stdout = self
+        sys.stderr = self
+
+    def __enter__(self) -> "Logger":
+        return self
+
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        self.close()
+
+    def write(self, text: Union[str, bytes]) -> None:
+        """Write text to stdout (and a file) and optionally flush."""
+        if isinstance(text, bytes):
+            text = text.decode()
+        if len(text) == 0: # workaround for a bug in VSCode debugger: sys.stdout.write(''); sys.stdout.flush() => crash
+            return
+
+        if self.file is not None:
+            self.file.write(text)
+
+        self.stdout.write(text)
+
+        if self.should_flush:
+            self.flush()
+
+    def flush(self) -> None:
+        """Flush written text to both stdout and a file, if open."""
+        if self.file is not None:
+            self.file.flush()
+
+        self.stdout.flush()
+
+    def close(self) -> None:
+        """Flush, close possible files, and remove stdout/stderr mirroring."""
+        self.flush()
+
+        # if using multiple loggers, prevent closing in wrong order
+        if sys.stdout is self:
+            sys.stdout = self.stdout
+        if sys.stderr is self:
+            sys.stderr = self.stderr
+
+        if self.file is not None:
+            self.file.close()
+            self.file = None
+
+
+# Cache directories
+# ------------------------------------------------------------------------------------------
+
+_dnnlib_cache_dir = None
+
+def set_cache_dir(path: str) -> None:
+    global _dnnlib_cache_dir
+    _dnnlib_cache_dir = path
+
+def make_cache_dir_path(*paths: str) -> str:
+    if _dnnlib_cache_dir is not None:
+        return os.path.join(_dnnlib_cache_dir, *paths)
+    if 'DNNLIB_CACHE_DIR' in os.environ:
+        return os.path.join(os.environ['DNNLIB_CACHE_DIR'], *paths)
+    if 'HOME' in os.environ:
+        return os.path.join(os.environ['HOME'], '.cache', 'dnnlib', *paths)
+    if 'USERPROFILE' in os.environ:
+        return os.path.join(os.environ['USERPROFILE'], '.cache', 'dnnlib', *paths)
+    return os.path.join(tempfile.gettempdir(), '.cache', 'dnnlib', *paths)
+
+# Small util functions
+# ------------------------------------------------------------------------------------------
+
+
+def format_time(seconds: Union[int, float]) -> str:
+    """Convert the seconds to human readable string with days, hours, minutes and seconds."""
+    s = int(np.rint(seconds))
+
+    if s < 60:
+        return "{0}s".format(s)
+    elif s < 60 * 60:
+        return "{0}m {1:02}s".format(s // 60, s % 60)
+    elif s < 24 * 60 * 60:
+        return "{0}h {1:02}m {2:02}s".format(s // (60 * 60), (s // 60) % 60, s % 60)
+    else:
+        return "{0}d {1:02}h {2:02}m".format(s // (24 * 60 * 60), (s // (60 * 60)) % 24, (s // 60) % 60)
+
+
+def ask_yes_no(question: str) -> bool:
+    """Ask the user the question until the user inputs a valid answer."""
+    while True:
+        try:
+            print("{0} [y/n]".format(question))
+            return strtobool(input().lower())
+        except ValueError:
+            pass
+
+
+def tuple_product(t: Tuple) -> Any:
+    """Calculate the product of the tuple elements."""
+    result = 1
+
+    for v in t:
+        result *= v
+
+    return result
+
+
+_str_to_ctype = {
+    "uint8": ctypes.c_ubyte,
+    "uint16": ctypes.c_uint16,
+    "uint32": ctypes.c_uint32,
+    "uint64": ctypes.c_uint64,
+    "int8": ctypes.c_byte,
+    "int16": ctypes.c_int16,
+    "int32": ctypes.c_int32,
+    "int64": ctypes.c_int64,
+    "float32": ctypes.c_float,
+    "float64": ctypes.c_double
+}
+
+
+def get_dtype_and_ctype(type_obj: Any) -> Tuple[np.dtype, Any]:
+    """Given a type name string (or an object having a __name__ attribute), return matching Numpy and ctypes types that have the same size in bytes."""
+    type_str = None
+
+    if isinstance(type_obj, str):
+        type_str = type_obj
+    elif hasattr(type_obj, "__name__"):
+        type_str = type_obj.__name__
+    elif hasattr(type_obj, "name"):
+        type_str = type_obj.name
+    else:
+        raise RuntimeError("Cannot infer type name from input")
+
+    assert type_str in _str_to_ctype.keys()
+
+    my_dtype = np.dtype(type_str)
+    my_ctype = _str_to_ctype[type_str]
+
+    assert my_dtype.itemsize == ctypes.sizeof(my_ctype)
+
+    return my_dtype, my_ctype
+
+
+def is_pickleable(obj: Any) -> bool:
+    try:
+        with io.BytesIO() as stream:
+            pickle.dump(obj, stream)
+        return True
+    except:
+        return False
+
+
+# Functionality to import modules/objects by name, and call functions by name
+# ------------------------------------------------------------------------------------------
+
+def get_module_from_obj_name(obj_name: str) -> Tuple[types.ModuleType, str]:
+    """Searches for the underlying module behind the name to some python object.
+    Returns the module and the object name (original name with module part removed)."""
+
+    # allow convenience shorthands, substitute them by full names
+    obj_name = re.sub("^np.", "numpy.", obj_name)
+    obj_name = re.sub("^tf.", "tensorflow.", obj_name)
+
+    # list alternatives for (module_name, local_obj_name)
+    parts = obj_name.split(".")
+    name_pairs = [(".".join(parts[:i]), ".".join(parts[i:])) for i in range(len(parts), 0, -1)]
+
+    # try each alternative in turn
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+            return module, local_obj_name
+        except:
+            pass
+
+    # maybe some of the modules themselves contain errors?
+    for module_name, _local_obj_name in name_pairs:
+        try:
+            importlib.import_module(module_name) # may raise ImportError
+        except ImportError:
+            if not str(sys.exc_info()[1]).startswith("No module named '" + module_name + "'"):
+                raise
+
+    # maybe the requested attribute is missing?
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+        except ImportError:
+            pass
+
+    # we are out of luck, but we have no idea why
+    raise ImportError(obj_name)
+
+
+def get_obj_from_module(module: types.ModuleType, obj_name: str) -> Any:
+    """Traverses the object name and returns the last (rightmost) python object."""
+    if obj_name == '':
+        return module
+    obj = module
+    for part in obj_name.split("."):
+        obj = getattr(obj, part)
+    return obj
+
+
+def get_obj_by_name(name: str) -> Any:
+    """Finds the python object with the given name."""
+    module, obj_name = get_module_from_obj_name(name)
+    return get_obj_from_module(module, obj_name)
+
+
+def call_func_by_name(*args, func_name: str = None, **kwargs) -> Any:
+    """Finds the python object with the given name and calls it as a function."""
+    assert func_name is not None
+    func_obj = get_obj_by_name(func_name)
+    assert callable(func_obj)
+    return func_obj(*args, **kwargs)
+
+
+def construct_class_by_name(*args, class_name: str = None, **kwargs) -> Any:
+    """Finds the python class with the given name and constructs it with the given arguments."""
+    return call_func_by_name(*args, func_name=class_name, **kwargs)
+
+
+def get_module_dir_by_obj_name(obj_name: str) -> str:
+    """Get the directory path of the module containing the given object name."""
+    module, _ = get_module_from_obj_name(obj_name)
+    return os.path.dirname(inspect.getfile(module))
+
+
+def is_top_level_function(obj: Any) -> bool:
+    """Determine whether the given object is a top-level function, i.e., defined at module scope using 'def'."""
+    return callable(obj) and obj.__name__ in sys.modules[obj.__module__].__dict__
+
+
+def get_top_level_function_name(obj: Any) -> str:
+    """Return the fully-qualified name of a top-level function."""
+    assert is_top_level_function(obj)
+    module = obj.__module__
+    if module == '__main__':
+        module = os.path.splitext(os.path.basename(sys.modules[module].__file__))[0]
+    return module + "." + obj.__name__
+
+
+# File system helpers
+# ------------------------------------------------------------------------------------------
+
+def list_dir_recursively_with_ignore(dir_path: str, ignores: List[str] = None, add_base_to_relative: bool = False) -> List[Tuple[str, str]]:
+    """List all files recursively in a given directory while ignoring given file and directory names.
+    Returns list of tuples containing both absolute and relative paths."""
+    assert os.path.isdir(dir_path)
+    base_name = os.path.basename(os.path.normpath(dir_path))
+
+    if ignores is None:
+        ignores = []
+
+    result = []
+
+    for root, dirs, files in os.walk(dir_path, topdown=True):
+        for ignore_ in ignores:
+            dirs_to_remove = [d for d in dirs if fnmatch.fnmatch(d, ignore_)]
+
+            # dirs need to be edited in-place
+            for d in dirs_to_remove:
+                dirs.remove(d)
+
+            files = [f for f in files if not fnmatch.fnmatch(f, ignore_)]
+
+        absolute_paths = [os.path.join(root, f) for f in files]
+        relative_paths = [os.path.relpath(p, dir_path) for p in absolute_paths]
+
+        if add_base_to_relative:
+            relative_paths = [os.path.join(base_name, p) for p in relative_paths]
+
+        assert len(absolute_paths) == len(relative_paths)
+        result += zip(absolute_paths, relative_paths)
+
+    return result
+
+
+def copy_files_and_create_dirs(files: List[Tuple[str, str]]) -> None:
+    """Takes in a list of tuples of (src, dst) paths and copies files.
+    Will create all necessary directories."""
+    for file in files:
+        target_dir_name = os.path.dirname(file[1])
+
+        # will create all intermediate-level directories
+        if not os.path.exists(target_dir_name):
+            os.makedirs(target_dir_name)
+
+        shutil.copyfile(file[0], file[1])
+
+
+# URL helpers
+# ------------------------------------------------------------------------------------------
+
+def is_url(obj: Any, allow_file_urls: bool = False) -> bool:
+    """Determine whether the given object is a valid URL string."""
+    if not isinstance(obj, str) or not "://" in obj:
+        return False
+    if allow_file_urls and obj.startswith('file://'):
+        return True
+    try:
+        res = requests.compat.urlparse(obj)
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+        res = requests.compat.urlparse(requests.compat.urljoin(obj, "/"))
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+    except:
+        return False
+    return True
+
+
+def open_url(url: str, cache_dir: str = None, num_attempts: int = 10, verbose: bool = True, return_filename: bool = False, cache: bool = True) -> Any:
+    """Download the given URL and return a binary-mode file object to access the data."""
+    assert num_attempts >= 1
+    assert not (return_filename and (not cache))
+
+    # Doesn't look like an URL scheme so interpret it as a local filename.
+    if not re.match('^[a-z]+://', url):
+        return url if return_filename else open(url, "rb")
+
+    # Handle file URLs.  This code handles unusual file:// patterns that
+    # arise on Windows:
+    #
+    # file:///c:/foo.txt
+    #
+    # which would translate to a local '/c:/foo.txt' filename that's
+    # invalid.  Drop the forward slash for such pathnames.
+    #
+    # If you touch this code path, you should test it on both Linux and
+    # Windows.
+    #
+    # Some internet resources suggest using urllib.request.url2pathname() but
+    # but that converts forward slashes to backslashes and this causes
+    # its own set of problems.
+    if url.startswith('file://'):
+        filename = urllib.parse.urlparse(url).path
+        if re.match(r'^/[a-zA-Z]:', filename):
+            filename = filename[1:]
+        return filename if return_filename else open(filename, "rb")
+
+    assert is_url(url)
+
+    # Lookup from cache.
+    if cache_dir is None:
+        cache_dir = make_cache_dir_path('downloads')
+
+    url_md5 = hashlib.md5(url.encode("utf-8")).hexdigest()
+    if cache:
+        cache_files = glob.glob(os.path.join(cache_dir, url_md5 + "_*"))
+        if len(cache_files) == 1:
+            filename = cache_files[0]
+            return filename if return_filename else open(filename, "rb")
+
+    # Download.
+    url_name = None
+    url_data = None
+    with requests.Session() as session:
+        if verbose:
+            print("Downloading %s ..." % url, end="", flush=True)
+        for attempts_left in reversed(range(num_attempts)):
+            try:
+                with session.get(url) as res:
+                    res.raise_for_status()
+                    if len(res.content) == 0:
+                        raise IOError("No data received")
+
+                    if len(res.content) < 8192:
+                        content_str = res.content.decode("utf-8")
+                        if "download_warning" in res.headers.get("Set-Cookie", ""):
+                            links = [html.unescape(link) for link in content_str.split('"') if "export=download" in link]
+                            if len(links) == 1:
+                                url = requests.compat.urljoin(url, links[0])
+                                raise IOError("Google Drive virus checker nag")
+                        if "Google Drive - Quota exceeded" in content_str:
+                            raise IOError("Google Drive download quota exceeded -- please try again later")
+
+                    match = re.search(r'filename="([^"]*)"', res.headers.get("Content-Disposition", ""))
+                    url_name = match[1] if match else url
+                    url_data = res.content
+                    if verbose:
+                        print(" done")
+                    break
+            except KeyboardInterrupt:
+                raise
+            except:
+                if not attempts_left:
+                    if verbose:
+                        print(" failed")
+                    raise
+                if verbose:
+                    print(".", end="", flush=True)
+
+    # Save to cache.
+    if cache:
+        safe_name = re.sub(r"[^0-9a-zA-Z-._]", "_", url_name)
+        cache_file = os.path.join(cache_dir, url_md5 + "_" + safe_name)
+        temp_file = os.path.join(cache_dir, "tmp_" + uuid.uuid4().hex + "_" + url_md5 + "_" + safe_name)
+        os.makedirs(cache_dir, exist_ok=True)
+        with open(temp_file, "wb") as f:
+            f.write(url_data)
+        os.replace(temp_file, cache_file) # atomic
+        if return_filename:
+            return cache_file
+
+    # Return data as file object.
+    assert not return_filename
+    return io.BytesIO(url_data)

PTI/models/StyleCLIP/criteria/clip_loss.py

@@ -0,0 +1,17 @@
+
+import torch
+import clip
+
+
+class CLIPLoss(torch.nn.Module):
+
+    def __init__(self, opts):
+        super(CLIPLoss, self).__init__()
+        self.model, self.preprocess = clip.load("ViT-B/32", device="cuda")
+        self.upsample = torch.nn.Upsample(scale_factor=7)
+        self.avg_pool = torch.nn.AvgPool2d(kernel_size=opts.stylegan_size // 32)
+
+    def forward(self, image, text):
+        image = self.avg_pool(self.upsample(image))
+        similarity = 1 - self.model(image, text)[0] / 100
+        return similarity

PTI/models/StyleCLIP/criteria/id_loss.py

@@ -0,0 +1,39 @@
+import torch
+from torch import nn
+
+from models.facial_recognition.model_irse import Backbone
+
+
+class IDLoss(nn.Module):
+    def __init__(self, opts):
+        super(IDLoss, self).__init__()
+        print('Loading ResNet ArcFace')
+        self.facenet = Backbone(input_size=112, num_layers=50, drop_ratio=0.6, mode='ir_se')
+        self.facenet.load_state_dict(torch.load(opts.ir_se50_weights))
+        self.pool = torch.nn.AdaptiveAvgPool2d((256, 256))
+        self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112))
+        self.facenet.eval()
+        self.opts = opts
+
+    def extract_feats(self, x):
+        if x.shape[2] != 256:
+            x = self.pool(x)
+        x = x[:, :, 35:223, 32:220]  # Crop interesting region
+        x = self.face_pool(x)
+        x_feats = self.facenet(x)
+        return x_feats
+
+    def forward(self, y_hat, y):
+        n_samples = y.shape[0]
+        y_feats = self.extract_feats(y)  # Otherwise use the feature from there
+        y_hat_feats = self.extract_feats(y_hat)
+        y_feats = y_feats.detach()
+        loss = 0
+        sim_improvement = 0
+        count = 0
+        for i in range(n_samples):
+            diff_target = y_hat_feats[i].dot(y_feats[i])
+            loss += 1 - diff_target
+            count += 1
+
+        return loss / count, sim_improvement / count

PTI/models/StyleCLIP/global_directions/GUI.py

@@ -0,0 +1,103 @@
+
+
+from tkinter import Tk,Frame ,Label,Button,messagebox,Canvas,Text,Scale
+from tkinter import  HORIZONTAL
+
+class View():
+    def __init__(self,master):
+        
+        self.width=600
+        self.height=600
+        
+        
+        self.root=master
+        self.root.geometry("600x600")
+        
+        self.left_frame=Frame(self.root,width=600)
+        self.left_frame.pack_propagate(0)
+        self.left_frame.pack(fill='both', side='left', expand='True')
+        
+        self.retrieval_frame=Frame(self.root,bg='snow3')
+        self.retrieval_frame.pack_propagate(0)
+        self.retrieval_frame.pack(fill='both', side='right', expand='True')
+        
+        self.bg_frame=Frame(self.left_frame,bg='snow3',height=600,width=600)
+        self.bg_frame.pack_propagate(0)
+        self.bg_frame.pack(fill='both', side='top', expand='True')
+        
+        self.command_frame=Frame(self.left_frame,bg='snow3')
+        self.command_frame.pack_propagate(0)
+        self.command_frame.pack(fill='both', side='bottom', expand='True')
+#        self.command_frame.grid(row=1, column=0,padx=0, pady=0)
+        
+        self.bg=Canvas(self.bg_frame,width=self.width,height=self.height, bg='gray')
+        self.bg.place(relx=0.5, rely=0.5, anchor='center')
+        
+        self.mani=Canvas(self.retrieval_frame,width=1024,height=1024, bg='gray') 
+        self.mani.grid(row=0, column=0,padx=0, pady=42)
+        
+        self.SetCommand()
+        
+        
+        
+    
+    def run(self):
+        self.root.mainloop()
+    
+    def helloCallBack(self):
+        category=self.set_category.get()
+        messagebox.showinfo( "Hello Python",category)
+    
+    def SetCommand(self):
+        
+        tmp = Label(self.command_frame, text="neutral", width=10 ,bg='snow3')
+        tmp.grid(row=1, column=0,padx=10, pady=10)
+        
+        tmp = Label(self.command_frame, text="a photo of a", width=10 ,bg='snow3')
+        tmp.grid(row=1, column=1,padx=10, pady=10)
+        
+        self.neutral = Text ( self.command_frame, height=2, width=30)
+        self.neutral.grid(row=1, column=2,padx=10, pady=10)
+        
+        
+        tmp = Label(self.command_frame, text="target", width=10 ,bg='snow3')
+        tmp.grid(row=2, column=0,padx=10, pady=10)
+        
+        tmp = Label(self.command_frame, text="a photo of a", width=10 ,bg='snow3')
+        tmp.grid(row=2, column=1,padx=10, pady=10)
+        
+        self.target = Text ( self.command_frame, height=2, width=30)
+        self.target.grid(row=2, column=2,padx=10, pady=10)
+        
+        tmp = Label(self.command_frame, text="strength", width=10 ,bg='snow3')
+        tmp.grid(row=3, column=0,padx=10, pady=10)
+        
+        self.alpha = Scale(self.command_frame, from_=-15, to=25, orient=HORIZONTAL,bg='snow3', length=250,resolution=0.01)
+        self.alpha.grid(row=3, column=2,padx=10, pady=10)
+        
+        
+        tmp = Label(self.command_frame, text="disentangle", width=10 ,bg='snow3')
+        tmp.grid(row=4, column=0,padx=10, pady=10)
+        
+        self.beta = Scale(self.command_frame, from_=0.08, to=0.4, orient=HORIZONTAL,bg='snow3', length=250,resolution=0.001)
+        self.beta.grid(row=4, column=2,padx=10, pady=10)
+        
+        self.reset = Button(self.command_frame, text='Reset') 
+        self.reset.grid(row=5, column=1,padx=10, pady=10)
+        
+        
+        self.set_init = Button(self.command_frame, text='Accept') 
+        self.set_init.grid(row=5, column=2,padx=10, pady=10)
+
+#%%
+if __name__ == "__main__":
+    master=Tk()
+    self=View(master)
+    self.run()
+    
+    
+    
+    
+    
+    
+    

PTI/models/StyleCLIP/global_directions/GenerateImg.py

@@ -0,0 +1,50 @@
+
+import os
+import numpy as np
+import argparse
+from manipulate import Manipulator
+
+from PIL import Image
+#%%
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Process some integers.')
+    
+    parser.add_argument('--dataset_name',type=str,default='ffhq',
+                    help='name of dataset, for example, ffhq')
+
+    args = parser.parse_args()
+    dataset_name=args.dataset_name
+    
+    if not os.path.isdir('./data/'+dataset_name):
+        os.system('mkdir ./data/'+dataset_name)
+    #%%
+    M=Manipulator(dataset_name=dataset_name)
+    np.set_printoptions(suppress=True)
+    print(M.dataset_name)
+    #%%
+    
+    M.img_index=0
+    M.num_images=50
+    M.alpha=[0]
+    M.step=1
+    lindex,bname=0,0
+    
+    M.manipulate_layers=[lindex]
+    codes,out=M.EditOneC(bname)
+    #%%
+    
+    for i in range(len(out)):
+        img=out[i,0]
+        img=Image.fromarray(img)
+        img.save('./data/'+dataset_name+'/'+str(i)+'.jpg')
+    #%%
+    w=np.load('./npy/'+dataset_name+'/W.npy')
+    
+    tmp=w[:M.num_images]
+    tmp=tmp[:,None,:]
+    tmp=np.tile(tmp,(1,M.Gs.components.synthesis.input_shape[1],1))
+    
+    np.save('./data/'+dataset_name+'/w_plus.npy',tmp)
+    
+    

PTI/models/StyleCLIP/global_directions/GetCode.py

@@ -0,0 +1,232 @@
+
+
+
+import os
+import pickle
+import numpy as np
+from dnnlib import tflib  
+import tensorflow as tf 
+
+import argparse
+
+def LoadModel(dataset_name):
+    # Initialize TensorFlow.
+    tflib.init_tf()
+    model_path='./model/'
+    model_name=dataset_name+'.pkl'
+    
+    tmp=os.path.join(model_path,model_name)
+    with open(tmp, 'rb') as f:
+        _, _, Gs = pickle.load(f)
+    return Gs
+
+def lerp(a,b,t):
+     return a + (b - a) * t
+
+#stylegan-ada
+def SelectName(layer_name,suffix):
+    if suffix==None:
+        tmp1='add:0' in layer_name 
+        tmp2='shape=(?,' in layer_name
+        tmp4='G_synthesis_1' in layer_name
+        tmp= tmp1 and tmp2 and tmp4  
+    else:
+        tmp1=('/Conv0_up'+suffix) in layer_name 
+        tmp2=('/Conv1'+suffix) in layer_name 
+        tmp3=('4x4/Conv'+suffix) in layer_name 
+        tmp4='G_synthesis_1' in layer_name
+        tmp5=('/ToRGB'+suffix) in layer_name
+        tmp= (tmp1 or tmp2 or tmp3 or tmp5) and tmp4 
+    return tmp
+
+
+def GetSNames(suffix):
+    #get style tensor name 
+    with tf.Session() as sess:
+        op = sess.graph.get_operations()
+    layers=[m.values() for m in op]
+    
+    
+    select_layers=[]
+    for layer in layers:
+        layer_name=str(layer)
+        if SelectName(layer_name,suffix):
+            select_layers.append(layer[0])
+    return select_layers
+
+def SelectName2(layer_name):
+    tmp1='mod_bias' in layer_name 
+    tmp2='mod_weight' in layer_name
+    tmp3='ToRGB' in layer_name 
+    
+    tmp= (tmp1 or tmp2) and (not tmp3) 
+    return tmp
+
+def GetKName(Gs):
+    
+    layers=[var for name, var in Gs.components.synthesis.vars.items()]
+    
+    select_layers=[]
+    for layer in layers:
+        layer_name=str(layer)
+        if SelectName2(layer_name):
+            select_layers.append(layer)
+    return select_layers
+
+def GetCode(Gs,random_state,num_img,num_once,dataset_name):
+    rnd = np.random.RandomState(random_state)  #5
+    
+    truncation_psi=0.7
+    truncation_cutoff=8
+    
+    dlatent_avg=Gs.get_var('dlatent_avg')
+    
+    dlatents=np.zeros((num_img,512),dtype='float32')
+    for i in range(int(num_img/num_once)):
+        src_latents =  rnd.randn(num_once, Gs.input_shape[1])
+        src_dlatents = Gs.components.mapping.run(src_latents, None) # [seed, layer, component]
+        
+        # Apply truncation trick.
+        if truncation_psi is not None and truncation_cutoff is not None:
+                layer_idx = np.arange(src_dlatents.shape[1])[np.newaxis, :, np.newaxis]
+                ones = np.ones(layer_idx.shape, dtype=np.float32)
+                coefs = np.where(layer_idx < truncation_cutoff, truncation_psi * ones, ones)
+                src_dlatents_np=lerp(dlatent_avg, src_dlatents, coefs)
+                src_dlatents=src_dlatents_np[:,0,:].astype('float32')
+                dlatents[(i*num_once):((i+1)*num_once),:]=src_dlatents
+    print('get all z and w')
+    
+    tmp='./npy/'+dataset_name+'/W'
+    np.save(tmp,dlatents)
+
+    
+def GetImg(Gs,num_img,num_once,dataset_name,save_name='images'):
+    print('Generate Image')
+    tmp='./npy/'+dataset_name+'/W.npy'
+    dlatents=np.load(tmp) 
+    fmt = dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=True)
+    
+    all_images=[]
+    for i in range(int(num_img/num_once)):
+        print(i)
+        images=[]
+        for k in range(num_once):
+            tmp=dlatents[i*num_once+k]
+            tmp=tmp[None,None,:]
+            tmp=np.tile(tmp,(1,Gs.components.synthesis.input_shape[1],1))
+            image2= Gs.components.synthesis.run(tmp, randomize_noise=False, output_transform=fmt)
+            images.append(image2)
+            
+        images=np.concatenate(images)
+        
+        all_images.append(images)
+        
+    all_images=np.concatenate(all_images)
+    
+    tmp='./npy/'+dataset_name+'/'+save_name
+    np.save(tmp,all_images)
+
+def GetS(dataset_name,num_img):
+    print('Generate S')
+    tmp='./npy/'+dataset_name+'/W.npy'
+    dlatents=np.load(tmp)[:num_img]
+    
+    with tf.Session() as sess:
+        init = tf.global_variables_initializer()
+        sess.run(init)
+        
+        Gs=LoadModel(dataset_name)
+        Gs.print_layers()  #for ada
+        select_layers1=GetSNames(suffix=None)  #None,'/mul_1:0','/mod_weight/read:0','/MatMul:0'
+        dlatents=dlatents[:,None,:]
+        dlatents=np.tile(dlatents,(1,Gs.components.synthesis.input_shape[1],1))
+        
+        all_s = sess.run(
+            select_layers1,
+            feed_dict={'G_synthesis_1/dlatents_in:0': dlatents})
+    
+    layer_names=[layer.name for layer in select_layers1]
+    save_tmp=[layer_names,all_s]
+    return save_tmp
+
+    
+
+
+def convert_images_to_uint8(images, drange=[-1,1], nchw_to_nhwc=False):
+    """Convert a minibatch of images from float32 to uint8 with configurable dynamic range.
+    Can be used as an output transformation for Network.run().
+    """
+    if nchw_to_nhwc:
+        images = np.transpose(images, [0, 2, 3, 1])
+    
+    scale = 255 / (drange[1] - drange[0])
+    images = images * scale + (0.5 - drange[0] * scale)
+    
+    np.clip(images, 0, 255, out=images)
+    images=images.astype('uint8')
+    return images
+
+
+def GetCodeMS(dlatents):
+        m=[]
+        std=[]
+        for i in range(len(dlatents)):
+            tmp= dlatents[i] 
+            tmp_mean=tmp.mean(axis=0)
+            tmp_std=tmp.std(axis=0)
+            m.append(tmp_mean)
+            std.append(tmp_std)
+        return m,std
+
+
+
+#%%
+if __name__ == "__main__":
+    
+    
+    parser = argparse.ArgumentParser(description='Process some integers.')
+    
+    parser.add_argument('--dataset_name',type=str,default='ffhq',
+                    help='name of dataset, for example, ffhq')
+    parser.add_argument('--code_type',choices=['w','s','s_mean_std'],default='w')
+    
+    args = parser.parse_args()
+    random_state=5
+    num_img=100_000 
+    num_once=1_000
+    dataset_name=args.dataset_name
+    
+    if not os.path.isfile('./model/'+dataset_name+'.pkl'):
+        url='https://nvlabs-fi-cdn.nvidia.com/stylegan2/networks/'
+        name='stylegan2-'+dataset_name+'-config-f.pkl'
+        os.system('wget ' +url+name + '  -P  ./model/')
+        os.system('mv ./model/'+name+' ./model/'+dataset_name+'.pkl')
+    
+    if not os.path.isdir('./npy/'+dataset_name):
+        os.system('mkdir ./npy/'+dataset_name)
+    
+    if args.code_type=='w':
+        Gs=LoadModel(dataset_name=dataset_name)
+        GetCode(Gs,random_state,num_img,num_once,dataset_name)
+#        GetImg(Gs,num_img=num_img,num_once=num_once,dataset_name=dataset_name,save_name='images_100K') #no need 
+    elif args.code_type=='s':
+        save_name='S'
+        save_tmp=GetS(dataset_name,num_img=2_000)
+        tmp='./npy/'+dataset_name+'/'+save_name
+        with open(tmp, "wb") as fp:
+            pickle.dump(save_tmp, fp)
+        
+    elif args.code_type=='s_mean_std':
+        save_tmp=GetS(dataset_name,num_img=num_img)
+        dlatents=save_tmp[1]
+        m,std=GetCodeMS(dlatents)
+        save_tmp=[m,std]
+        save_name='S_mean_std'
+        tmp='./npy/'+dataset_name+'/'+save_name
+        with open(tmp, "wb") as fp:
+            pickle.dump(save_tmp, fp)
+    
+    
+    
+    
+    

PTI/models/StyleCLIP/global_directions/GetGUIData.py

@@ -0,0 +1,67 @@
+
+import os
+import numpy as np
+import argparse
+from manipulate import Manipulator
+import torch
+from PIL import Image
+#%%
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Process some integers.')
+    
+    parser.add_argument('--dataset_name',type=str,default='ffhq',
+                    help='name of dataset, for example, ffhq')
+
+    parser.add_argument('--real', action='store_true')
+
+    args = parser.parse_args()
+    dataset_name=args.dataset_name
+    
+    if not os.path.isdir('./data/'+dataset_name):
+        os.system('mkdir ./data/'+dataset_name)
+    #%%
+    M=Manipulator(dataset_name=dataset_name)
+    np.set_printoptions(suppress=True)
+    print(M.dataset_name)
+    #%%
+    #remove all .jpg
+    names=os.listdir('./data/'+dataset_name+'/')
+    for name in names:
+        if '.jpg' in name:
+            os.system('rm ./data/'+dataset_name+'/'+name)
+    
+    
+    #%%
+    if args.real:
+        latents=torch.load('./data/'+dataset_name+'/latents.pt')
+        w_plus=latents.cpu().detach().numpy()
+    else:
+        w=np.load('./npy/'+dataset_name+'/W.npy')
+        tmp=w[:50] #only use 50 images
+        tmp=tmp[:,None,:]
+        w_plus=np.tile(tmp,(1,M.Gs.components.synthesis.input_shape[1],1))
+    np.save('./data/'+dataset_name+'/w_plus.npy',w_plus)
+    
+    #%%
+    tmp=M.W2S(w_plus)
+    M.dlatents=tmp
+    
+    M.img_index=0
+    M.num_images=len(w_plus)
+    M.alpha=[0]
+    M.step=1
+    lindex,bname=0,0
+    
+    M.manipulate_layers=[lindex]
+    codes,out=M.EditOneC(bname)
+    #%%
+    
+    for i in range(len(out)):
+        img=out[i,0]
+        img=Image.fromarray(img)
+        img.save('./data/'+dataset_name+'/'+str(i)+'.jpg')
+    #%%
+
+    
+    

PTI/models/StyleCLIP/global_directions/Inference.py

@@ -0,0 +1,106 @@
+
+
+from manipulate import Manipulator
+import tensorflow as tf
+import numpy as np 
+import torch
+import clip
+from MapTS import GetBoundary,GetDt
+
+class StyleCLIP():
+    
+    def __init__(self,dataset_name='ffhq'):
+        print('load clip')
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        self.model, preprocess = clip.load("ViT-B/32", device=device)
+        self.LoadData(dataset_name)
+        
+    def LoadData(self, dataset_name):
+        tf.keras.backend.clear_session()
+        M=Manipulator(dataset_name=dataset_name)
+        np.set_printoptions(suppress=True)
+        fs3=np.load('./npy/'+dataset_name+'/fs3.npy')
+        
+        self.M=M
+        self.fs3=fs3
+        
+        w_plus=np.load('./data/'+dataset_name+'/w_plus.npy')
+        self.M.dlatents=M.W2S(w_plus)
+        
+        if dataset_name=='ffhq':
+            self.c_threshold=20
+        else:
+            self.c_threshold=100
+        self.SetInitP()
+    
+    def SetInitP(self):
+        self.M.alpha=[3]
+        self.M.num_images=1
+        
+        self.target=''
+        self.neutral=''
+        self.GetDt2()
+        img_index=0
+        self.M.dlatent_tmp=[tmp[img_index:(img_index+1)] for tmp in self.M.dlatents]
+        
+        
+    def GetDt2(self):
+        classnames=[self.target,self.neutral]
+        dt=GetDt(classnames,self.model)
+        
+        self.dt=dt
+        num_cs=[]
+        betas=np.arange(0.1,0.3,0.01)
+        for i in range(len(betas)):
+            boundary_tmp2,num_c=GetBoundary(self.fs3,self.dt,self.M,threshold=betas[i])
+            print(betas[i])
+            num_cs.append(num_c)
+        
+        num_cs=np.array(num_cs)
+        select=num_cs>self.c_threshold
+        
+        if sum(select)==0:
+            self.beta=0.1
+        else:
+            self.beta=betas[select][-1]
+        
+    
+    def GetCode(self):
+        boundary_tmp2,num_c=GetBoundary(self.fs3,self.dt,self.M,threshold=self.beta)
+        codes=self.M.MSCode(self.M.dlatent_tmp,boundary_tmp2)
+        return codes
+    
+    def GetImg(self):
+        
+        codes=self.GetCode()
+        out=self.M.GenerateImg(codes)
+        img=out[0,0]
+        return img
+    
+
+
+
+#%%
+if __name__ == "__main__":
+    style_clip=StyleCLIP()
+    self=style_clip
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    

PTI/models/StyleCLIP/global_directions/MapTS.py

@@ -0,0 +1,394 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Thu Feb  4 17:36:31 2021
+
+@author: wuzongze
+"""
+
+import os
+#os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+#os.environ["CUDA_VISIBLE_DEVICES"] = "1" #(or "1" or "2")
+
+import sys 
+
+#sys.path=['', '/usr/local/tensorflow/avx-avx2-gpu/1.14.0/python3.7/site-packages', '/usr/local/matlab/2018b/lib/python3.7/site-packages', '/cs/labs/danix/wuzongze/pythonV/venv3.7/lib/python37.zip', '/cs/labs/danix/wuzongze/pythonV/venv3.7/lib/python3.7', '/cs/labs/danix/wuzongze/pythonV/venv3.7/lib/python3.7/lib-dynload', '/usr/lib/python3.7', '/cs/labs/danix/wuzongze/pythonV/venv3.7/lib/python3.7/site-packages', '/cs/labs/danix/wuzongze/pythonV/venv3.7/lib/python3.7/site-packages/copkmeans-1.5-py3.7.egg', '/cs/labs/danix/wuzongze/pythonV/venv3.7/lib/python3.7/site-packages/spherecluster-0.1.7-py3.7.egg', '/usr/lib/python3/dist-packages', '/usr/local/lib/python3.7/dist-packages', '/usr/lib/python3/dist-packages/IPython/extensions']
+
+import tensorflow as tf
+
+import numpy as np 
+import torch
+import clip
+from PIL import Image
+import pickle
+import copy
+import matplotlib.pyplot as plt
+
+def GetAlign(out,dt,model,preprocess):
+    imgs=out
+    imgs1=imgs.reshape([-1]+list(imgs.shape[2:]))
+    
+    tmp=[]
+    for i in range(len(imgs1)):
+        
+        img=Image.fromarray(imgs1[i])
+        image = preprocess(img).unsqueeze(0).to(device)
+        tmp.append(image)
+    
+    image=torch.cat(tmp)
+    
+    with torch.no_grad():
+        image_features = model.encode_image(image)
+        image_features = image_features / image_features.norm(dim=-1, keepdim=True)
+    
+    image_features1=image_features.cpu().numpy()
+    
+    image_features1=image_features1.reshape(list(imgs.shape[:2])+[512])
+    
+    fd=image_features1[:,1:,:]-image_features1[:,:-1,:]
+    
+    fd1=fd.reshape([-1,512])
+    fd2=fd1/np.linalg.norm(fd1,axis=1)[:,None]
+    
+    tmp=np.dot(fd2,dt)
+    m=tmp.mean()
+    acc=np.sum(tmp>0)/len(tmp)
+    print(m,acc)
+    return m,acc
+
+
+def SplitS(ds_p,M,if_std):
+    all_ds=[]
+    start=0
+    for i in M.mindexs:
+        tmp=M.dlatents[i].shape[1]
+        end=start+tmp
+        tmp=ds_p[start:end]
+#        tmp=tmp*M.code_std[i]
+        
+        all_ds.append(tmp)
+        start=end
+    
+    all_ds2=[]
+    tmp_index=0
+    for i in range(len(M.s_names)):
+        if (not 'RGB' in M.s_names[i]) and (not len(all_ds[tmp_index])==0):
+            
+#            tmp=np.abs(all_ds[tmp_index]/M.code_std[i])
+#            print(i,tmp.mean())
+#            tmp=np.dot(M.latent_codes[i],all_ds[tmp_index])
+#            print(tmp)
+            if if_std:
+                tmp=all_ds[tmp_index]*M.code_std[i]
+            else:
+                tmp=all_ds[tmp_index]
+            
+            all_ds2.append(tmp)
+            tmp_index+=1
+        else:
+            tmp=np.zeros(len(M.dlatents[i][0]))
+            all_ds2.append(tmp)
+    return all_ds2
+
+
+imagenet_templates = [
+    'a bad photo of a {}.',
+#    'a photo of many {}.',
+    'a sculpture of a {}.',
+    'a photo of the hard to see {}.',
+    'a low resolution photo of the {}.',
+    'a rendering of a {}.',
+    'graffiti of a {}.',
+    'a bad photo of the {}.',
+    'a cropped photo of the {}.',
+    'a tattoo of a {}.',
+    'the embroidered {}.',
+    'a photo of a hard to see {}.',
+    'a bright photo of a {}.',
+    'a photo of a clean {}.',
+    'a photo of a dirty {}.',
+    'a dark photo of the {}.',
+    'a drawing of a {}.',
+    'a photo of my {}.',
+    'the plastic {}.',
+    'a photo of the cool {}.',
+    'a close-up photo of a {}.',
+    'a black and white photo of the {}.',
+    'a painting of the {}.',
+    'a painting of a {}.',
+    'a pixelated photo of the {}.',
+    'a sculpture of the {}.',
+    'a bright photo of the {}.',
+    'a cropped photo of a {}.',
+    'a plastic {}.',
+    'a photo of the dirty {}.',
+    'a jpeg corrupted photo of a {}.',
+    'a blurry photo of the {}.',
+    'a photo of the {}.',
+    'a good photo of the {}.',
+    'a rendering of the {}.',
+    'a {} in a video game.',
+    'a photo of one {}.',
+    'a doodle of a {}.',
+    'a close-up photo of the {}.',
+    'a photo of a {}.',
+    'the origami {}.',
+    'the {} in a video game.',
+    'a sketch of a {}.',
+    'a doodle of the {}.',
+    'a origami {}.',
+    'a low resolution photo of a {}.',
+    'the toy {}.',
+    'a rendition of the {}.',
+    'a photo of the clean {}.',
+    'a photo of a large {}.',
+    'a rendition of a {}.',
+    'a photo of a nice {}.',
+    'a photo of a weird {}.',
+    'a blurry photo of a {}.',
+    'a cartoon {}.',
+    'art of a {}.',
+    'a sketch of the {}.',
+    'a embroidered {}.',
+    'a pixelated photo of a {}.',
+    'itap of the {}.',
+    'a jpeg corrupted photo of the {}.',
+    'a good photo of a {}.',
+    'a plushie {}.',
+    'a photo of the nice {}.',
+    'a photo of the small {}.',
+    'a photo of the weird {}.',
+    'the cartoon {}.',
+    'art of the {}.',
+    'a drawing of the {}.',
+    'a photo of the large {}.',
+    'a black and white photo of a {}.',
+    'the plushie {}.',
+    'a dark photo of a {}.',
+    'itap of a {}.',
+    'graffiti of the {}.',
+    'a toy {}.',
+    'itap of my {}.',
+    'a photo of a cool {}.',
+    'a photo of a small {}.',
+    'a tattoo of the {}.',
+]
+
+
+def zeroshot_classifier(classnames, templates,model):
+    with torch.no_grad():
+        zeroshot_weights = []
+        for classname in classnames:
+            texts = [template.format(classname) for template in templates] #format with class
+            texts = clip.tokenize(texts).cuda() #tokenize
+            class_embeddings = model.encode_text(texts) #embed with text encoder
+            class_embeddings /= class_embeddings.norm(dim=-1, keepdim=True)
+            class_embedding = class_embeddings.mean(dim=0)
+            class_embedding /= class_embedding.norm()
+            zeroshot_weights.append(class_embedding)
+        zeroshot_weights = torch.stack(zeroshot_weights, dim=1).cuda()
+    return zeroshot_weights
+
+
+def GetDt(classnames,model):
+    text_features=zeroshot_classifier(classnames, imagenet_templates,model).t()
+    
+    dt=text_features[0]-text_features[1]
+    dt=dt.cpu().numpy()
+    
+#    t_m1=t_m/np.linalg.norm(t_m)
+#    dt=text_features.cpu().numpy()[0]-t_m1
+    print(np.linalg.norm(dt))
+    dt=dt/np.linalg.norm(dt)
+    return dt
+
+
+def GetBoundary(fs3,dt,M,threshold):
+    tmp=np.dot(fs3,dt)
+    
+    ds_imp=copy.copy(tmp)
+    select=np.abs(tmp)<threshold
+    num_c=np.sum(~select)
+
+
+    ds_imp[select]=0
+    tmp=np.abs(ds_imp).max()
+    ds_imp/=tmp
+    
+    boundary_tmp2=SplitS(ds_imp,M,if_std=True)
+    print('num of channels being manipulated:',num_c)
+    return boundary_tmp2,num_c
+
+def GetFs(file_path):
+    fs=np.load(file_path+'single_channel.npy')
+    tmp=np.linalg.norm(fs,axis=-1)
+    fs1=fs/tmp[:,:,:,None]
+    fs2=fs1[:,:,1,:]-fs1[:,:,0,:]  # 5*sigma - (-5)* sigma
+    fs3=fs2/np.linalg.norm(fs2,axis=-1)[:,:,None]
+    fs3=fs3.mean(axis=1)
+    fs3=fs3/np.linalg.norm(fs3,axis=-1)[:,None]
+    return fs3
+#%%
+
+if __name__ == "__main__":
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model, preprocess = clip.load("ViT-B/32", device=device)
+    #%%
+    sys.path.append('/cs/labs/danix/wuzongze/Gan_Manipulation/play')
+    from example_try import Manipulator4
+    
+    M=Manipulator4(dataset_name='ffhq',code_type='S')
+    np.set_printoptions(suppress=True)
+
+    #%%
+    
+    
+    file_path='/cs/labs/danix/wuzongze/Tansformer_Manipulation/CLIP/results/'+M.dataset_name+'/'
+    fs3=GetFs(file_path)
+    
+
+    
+    #%%
+    '''
+    text_features=zeroshot_classifier2(classnames, imagenet_templates) #.t()
+        
+    tmp=np.linalg.norm(text_features,axis=2)
+    text_features/=tmp[:,:,None]
+    dt=text_features[0]-text_features[1]
+    
+    tmp=np.linalg.norm(dt,axis=1)
+    dt/=tmp[:,None]
+    dt=dt.mean(axis=0)
+    '''
+    
+    #%%
+    '''
+    all_tmp=[]
+    tmp=torch.load('/cs/labs/danix/wuzongze/downloads/harris_latent.pt')
+    tmp=tmp.cpu().detach().numpy() #[:,:14,:]
+    all_tmp.append(tmp)
+    
+    tmp=torch.load('/cs/labs/danix/wuzongze/downloads/ariana_latent.pt')
+    tmp=tmp.cpu().detach().numpy() #[:,:14,:]
+    all_tmp.append(tmp)
+    
+    tmp=torch.load('/cs/labs/danix/wuzongze/downloads/federer.pt')
+    tmp=tmp.cpu().detach().numpy() #[:,:14,:]
+    all_tmp.append(tmp)
+    
+    all_tmp=np.array(all_tmp)[:,0]
+    
+    dlatent_tmp=M.W2S(all_tmp)
+    '''
+    '''
+    tmp=torch.load('/cs/labs/danix/wuzongze/downloads/all_cars.pt')
+    tmp=tmp.cpu().detach().numpy()[:300]
+    dlatent_tmp=M.W2S(tmp)
+    '''
+    '''
+    tmp=torch.load('/cs/labs/danix/wuzongze/downloads/faces.pt')
+    tmp=tmp.cpu().detach().numpy()[:100]
+    dlatent_tmp=M.W2S(tmp)
+    '''
+    #%%
+#    M.viz_size=1024
+    M.img_index=0
+    M.num_images=30
+    dlatent_tmp=[tmp[M.img_index:(M.img_index+M.num_images)] for tmp in M.dlatents]
+    #%%
+    
+    classnames=['face','face with glasses']
+    
+#    classnames=['car','classic car']
+#    classnames=['dog','happy dog']
+#    classnames=['bedroom','modern bedroom']
+    
+#    classnames=['church','church without watermark']
+#    classnames=['natural scene','natural scene without grass']
+    dt=GetDt(classnames,model)
+#    tmp=np.dot(fs3,dt)
+#    
+#    ds_imp=copy.copy(tmp)
+#    select=np.abs(tmp)<0.1
+#    num_c=np.sum(~select)
+#
+#
+#    ds_imp[select]=0
+#    tmp=np.abs(ds_imp).max()
+#    ds_imp/=tmp
+#    
+#    boundary_tmp2=SplitS(ds_imp,M,if_std=True)
+#    print('num of channels being manipulated:',num_c)
+    
+    boundary_tmp2=GetBoundary(fs3,dt,M,threshold=0.13)
+    
+    #%%
+    M.start_distance=-20
+    M.end_distance=20
+    M.step=7
+#    M.num_images=100
+    codes=M.MSCode(dlatent_tmp,boundary_tmp2)
+    out=M.GenerateImg(codes)
+    M.Vis2(str('tmp'),'filter2',out)
+    
+#    full=GetAlign(out,dt,model,preprocess)
+    
+    
+    #%%
+    boundary_tmp3=copy.copy(boundary_tmp2) #primary
+    boundary_tmp4=copy.copy(boundary_tmp2) #condition
+    #%%
+    boundary_tmp2=copy.copy(boundary_tmp3)
+    for i in range(len(boundary_tmp3)):
+        select=boundary_tmp4[i]==0
+        boundary_tmp2[i][~select]=0
+    
+    
+
+    
+    
+    
+    
+    #%%1
+    
+    
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

PTI/models/StyleCLIP/global_directions/PlayInteractively.py

@@ -0,0 +1,197 @@
+
+
+
+from tkinter import Tk 
+from PIL import Image, ImageTk
+from tkinter.filedialog import askopenfilename
+from GUI import View
+from Inference import StyleCLIP
+import argparse
+#%%
+
+
+class PlayInteractively():  #Controller
+    '''
+    followed Model View Controller Design Pattern
+    
+    controller, model, view
+    '''
+    def __init__(self,dataset_name='ffhq'):
+        
+        self.root = Tk()
+        self.view=View(self.root)
+        self.img_ratio=2
+        self.style_clip=StyleCLIP(dataset_name)
+        
+        self.view.neutral.bind("<Return>", self.text_n)
+        self.view.target.bind("<Return>", self.text_t)
+        self.view.alpha.bind('<ButtonRelease-1>', self.ChangeAlpha)
+        self.view.beta.bind('<ButtonRelease-1>', self.ChangeBeta)
+        self.view.set_init.bind('<ButtonPress-1>', self.SetInit) 
+        self.view.reset.bind('<ButtonPress-1>', self.Reset) 
+        self.view.bg.bind('<Double-1>', self.open_img)
+        
+        
+        self.drawn  = None
+        
+        self.view.target.delete(1.0, "end")
+        self.view.target.insert("end", self.style_clip.target)
+#        
+        self.view.neutral.delete(1.0, "end")
+        self.view.neutral.insert("end", self.style_clip.neutral)
+        
+    
+    def Reset(self,event):
+        self.style_clip.GetDt2()
+        self.style_clip.M.alpha=[0]
+        
+        self.view.beta.set(self.style_clip.beta)
+        self.view.alpha.set(0)
+        
+        img=self.style_clip.GetImg()
+        img=Image.fromarray(img)
+        img = ImageTk.PhotoImage(img)
+        self.addImage_m(img)
+        
+    
+    def SetInit(self,event):
+        codes=self.style_clip.GetCode()
+        self.style_clip.M.dlatent_tmp=[tmp[:,0] for tmp in codes]
+        print('set init')
+    
+    def ChangeAlpha(self,event):
+        tmp=self.view.alpha.get()
+        self.style_clip.M.alpha=[float(tmp)]
+        
+        img=self.style_clip.GetImg()
+        print('manipulate one')
+        img=Image.fromarray(img)
+        img = ImageTk.PhotoImage(img)
+        self.addImage_m(img)
+        
+    def ChangeBeta(self,event):
+        tmp=self.view.beta.get()
+        self.style_clip.beta=float(tmp)
+        
+        img=self.style_clip.GetImg()
+        print('manipulate one')
+        img=Image.fromarray(img)
+        img = ImageTk.PhotoImage(img)
+        self.addImage_m(img)
+
+    def ChangeDataset(self,event):
+        
+        dataset_name=self.view.set_category.get()
+        
+        self.style_clip.LoadData(dataset_name)
+        
+        self.view.target.delete(1.0, "end")
+        self.view.target.insert("end", self.style_clip.target)
+        
+        self.view.neutral.delete(1.0, "end")
+        self.view.neutral.insert("end", self.style_clip.neutral)
+    
+    def text_t(self,event):
+        tmp=self.view.target.get("1.0",'end')
+        tmp=tmp.replace('\n','')
+        
+        self.view.target.delete(1.0, "end")
+        self.view.target.insert("end", tmp)
+        
+        print('target',tmp,'###')
+        self.style_clip.target=tmp
+        self.style_clip.GetDt2()
+        self.view.beta.set(self.style_clip.beta)
+        self.view.alpha.set(3)
+        self.style_clip.M.alpha=[3]
+        
+        img=self.style_clip.GetImg()
+        print('manipulate one')
+        img=Image.fromarray(img)
+        img = ImageTk.PhotoImage(img)
+        self.addImage_m(img)
+        
+        
+    def text_n(self,event):
+        tmp=self.view.neutral.get("1.0",'end')
+        tmp=tmp.replace('\n','')
+        
+        self.view.neutral.delete(1.0, "end")
+        self.view.neutral.insert("end", tmp)
+        
+        print('neutral',tmp,'###')
+        self.style_clip.neutral=tmp
+        self.view.target.delete(1.0, "end")
+        self.view.target.insert("end", tmp)
+        
+        
+    def run(self):
+        self.root.mainloop()
+    
+    def addImage(self,img):
+        self.view.bg.create_image(self.view.width/2, self.view.height/2, image=img, anchor='center')
+        self.image=img #save a copy of image. if not the image will disappear
+        
+    def addImage_m(self,img):
+        self.view.mani.create_image(512, 512, image=img, anchor='center')
+        self.image2=img
+        
+    
+    def openfn(self):
+        filename = askopenfilename(title='open',initialdir='./data/'+self.style_clip.M.dataset_name+'/',filetypes=[("all image format", ".jpg"),("all image format", ".png")])
+        return filename
+    
+    def open_img(self,event):
+        x = self.openfn()
+        print(x)
+        
+        
+        img = Image.open(x)
+        img2 = img.resize(( 512,512), Image.ANTIALIAS)
+        img2 = ImageTk.PhotoImage(img2)
+        self.addImage(img2)
+        
+        img = ImageTk.PhotoImage(img)
+        self.addImage_m(img)
+        
+        img_index=x.split('/')[-1].split('.')[0]
+        img_index=int(img_index)
+        print(img_index)
+        self.style_clip.M.img_index=img_index
+        self.style_clip.M.dlatent_tmp=[tmp[img_index:(img_index+1)] for tmp in self.style_clip.M.dlatents]
+        
+        
+        self.style_clip.GetDt2()
+        self.view.beta.set(self.style_clip.beta)
+        self.view.alpha.set(3)
+        
+    #%%
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Process some integers.')
+    
+    parser.add_argument('--dataset_name',type=str,default='ffhq',
+                    help='name of dataset, for example, ffhq')
+    
+    args = parser.parse_args()
+    dataset_name=args.dataset_name
+    
+    self=PlayInteractively(dataset_name)
+    self.run()
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    

PTI/models/StyleCLIP/global_directions/SingleChannel.py

@@ -0,0 +1,109 @@
+
+
+
+import numpy as np 
+import torch
+import clip
+from PIL import Image
+import copy
+from manipulate import Manipulator
+import argparse
+
+def GetImgF(out,model,preprocess):
+    imgs=out
+    imgs1=imgs.reshape([-1]+list(imgs.shape[2:]))
+    
+    tmp=[]
+    for i in range(len(imgs1)):
+        
+        img=Image.fromarray(imgs1[i])
+        image = preprocess(img).unsqueeze(0).to(device)
+        tmp.append(image)
+    
+    image=torch.cat(tmp)
+    with torch.no_grad():
+        image_features = model.encode_image(image)
+    
+    image_features1=image_features.cpu().numpy()
+    image_features1=image_features1.reshape(list(imgs.shape[:2])+[512])
+    
+    return image_features1
+
+def GetFs(fs):
+    tmp=np.linalg.norm(fs,axis=-1)
+    fs1=fs/tmp[:,:,:,None]
+    fs2=fs1[:,:,1,:]-fs1[:,:,0,:]  # 5*sigma - (-5)* sigma
+    fs3=fs2/np.linalg.norm(fs2,axis=-1)[:,:,None]
+    fs3=fs3.mean(axis=1)
+    fs3=fs3/np.linalg.norm(fs3,axis=-1)[:,None]
+    return fs3
+
+#%%
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Process some integers.')
+    
+    parser.add_argument('--dataset_name',type=str,default='cat',
+                    help='name of dataset, for example, ffhq')
+    args = parser.parse_args()
+    dataset_name=args.dataset_name
+    
+    #%%
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model, preprocess = clip.load("ViT-B/32", device=device)
+    #%%
+    M=Manipulator(dataset_name=dataset_name)
+    np.set_printoptions(suppress=True)
+    print(M.dataset_name)
+    #%%
+    img_sindex=0
+    num_images=100
+    dlatents_o=[]
+    tmp=img_sindex*num_images
+    for i in range(len(M.dlatents)):
+        tmp1=M.dlatents[i][tmp:(tmp+num_images)]
+        dlatents_o.append(tmp1)
+    #%%
+    
+    all_f=[]
+    M.alpha=[-5,5] #ffhq 5
+    M.step=2
+    M.num_images=num_images
+    select=np.array(M.mindexs)<=16 #below or equal to 128 resolution 
+    mindexs2=np.array(M.mindexs)[select]
+    for lindex in mindexs2: #ignore ToRGB layers
+        print(lindex)
+        num_c=M.dlatents[lindex].shape[1]
+        for cindex in range(num_c):
+            
+            M.dlatents=copy.copy(dlatents_o)
+            M.dlatents[lindex][:,cindex]=M.code_mean[lindex][cindex]
+            
+            M.manipulate_layers=[lindex]
+            codes,out=M.EditOneC(cindex) 
+            image_features1=GetImgF(out,model,preprocess)
+            all_f.append(image_features1)
+    
+    all_f=np.array(all_f)
+    
+    fs3=GetFs(all_f)
+    
+    #%%
+    file_path='./npy/'+M.dataset_name+'/'
+    np.save(file_path+'fs3',fs3)
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    

PTI/models/StyleCLIP/global_directions/data/ffhq/w_plus.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:394f0f166305654f49cd1b0cd3d4f2b7a51e740a449a1ebfa1c69f79d01399fa
+size 2506880

PTI/models/StyleCLIP/global_directions/dnnlib/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from .util import EasyDict, make_cache_dir_path

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/__init__.py

@@ -0,0 +1,20 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from . import autosummary
+from . import network
+from . import optimizer
+from . import tfutil
+from . import custom_ops
+
+from .tfutil import *
+from .network import Network
+
+from .optimizer import Optimizer
+
+from .custom_ops import get_plugin

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/autosummary.py

@@ -0,0 +1,193 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Helper for adding automatically tracked values to Tensorboard.
+
+Autosummary creates an identity op that internally keeps track of the input
+values and automatically shows up in TensorBoard. The reported value
+represents an average over input components. The average is accumulated
+constantly over time and flushed when save_summaries() is called.
+
+Notes:
+- The output tensor must be used as an input for something else in the
+  graph. Otherwise, the autosummary op will not get executed, and the average
+  value will not get accumulated.
+- It is perfectly fine to include autosummaries with the same name in
+  several places throughout the graph, even if they are executed concurrently.
+- It is ok to also pass in a python scalar or numpy array. In this case, it
+  is added to the average immediately.
+"""
+
+from collections import OrderedDict
+import numpy as np
+import tensorflow as tf
+from tensorboard import summary as summary_lib
+from tensorboard.plugins.custom_scalar import layout_pb2
+
+from . import tfutil
+from .tfutil import TfExpression
+from .tfutil import TfExpressionEx
+
+# Enable "Custom scalars" tab in TensorBoard for advanced formatting.
+# Disabled by default to reduce tfevents file size.
+enable_custom_scalars = False
+
+_dtype = tf.float64
+_vars = OrderedDict()  # name => [var, ...]
+_immediate = OrderedDict()  # name => update_op, update_value
+_finalized = False
+_merge_op = None
+
+
+def _create_var(name: str, value_expr: TfExpression) -> TfExpression:
+    """Internal helper for creating autosummary accumulators."""
+    assert not _finalized
+    name_id = name.replace("/", "_")
+    v = tf.cast(value_expr, _dtype)
+
+    if v.shape.is_fully_defined():
+        size = np.prod(v.shape.as_list())
+        size_expr = tf.constant(size, dtype=_dtype)
+    else:
+        size = None
+        size_expr = tf.reduce_prod(tf.cast(tf.shape(v), _dtype))
+
+    if size == 1:
+        if v.shape.ndims != 0:
+            v = tf.reshape(v, [])
+        v = [size_expr, v, tf.square(v)]
+    else:
+        v = [size_expr, tf.reduce_sum(v), tf.reduce_sum(tf.square(v))]
+    v = tf.cond(tf.is_finite(v[1]), lambda: tf.stack(v), lambda: tf.zeros(3, dtype=_dtype))
+
+    with tfutil.absolute_name_scope("Autosummary/" + name_id), tf.control_dependencies(None):
+        var = tf.Variable(tf.zeros(3, dtype=_dtype), trainable=False)  # [sum(1), sum(x), sum(x**2)]
+    update_op = tf.cond(tf.is_variable_initialized(var), lambda: tf.assign_add(var, v), lambda: tf.assign(var, v))
+
+    if name in _vars:
+        _vars[name].append(var)
+    else:
+        _vars[name] = [var]
+    return update_op
+
+
+def autosummary(name: str, value: TfExpressionEx, passthru: TfExpressionEx = None, condition: TfExpressionEx = True) -> TfExpressionEx:
+    """Create a new autosummary.
+
+    Args:
+        name:     Name to use in TensorBoard
+        value:    TensorFlow expression or python value to track
+        passthru: Optionally return this TF node without modifications but tack an autosummary update side-effect to this node.
+
+    Example use of the passthru mechanism:
+
+    n = autosummary('l2loss', loss, passthru=n)
+
+    This is a shorthand for the following code:
+
+    with tf.control_dependencies([autosummary('l2loss', loss)]):
+        n = tf.identity(n)
+    """
+    tfutil.assert_tf_initialized()
+    name_id = name.replace("/", "_")
+
+    if tfutil.is_tf_expression(value):
+        with tf.name_scope("summary_" + name_id), tf.device(value.device):
+            condition = tf.convert_to_tensor(condition, name='condition')
+            update_op = tf.cond(condition, lambda: tf.group(_create_var(name, value)), tf.no_op)
+            with tf.control_dependencies([update_op]):
+                return tf.identity(value if passthru is None else passthru)
+
+    else:  # python scalar or numpy array
+        assert not tfutil.is_tf_expression(passthru)
+        assert not tfutil.is_tf_expression(condition)
+        if condition:
+            if name not in _immediate:
+                with tfutil.absolute_name_scope("Autosummary/" + name_id), tf.device(None), tf.control_dependencies(None):
+                    update_value = tf.placeholder(_dtype)
+                    update_op = _create_var(name, update_value)
+                    _immediate[name] = update_op, update_value
+            update_op, update_value = _immediate[name]
+            tfutil.run(update_op, {update_value: value})
+        return value if passthru is None else passthru
+
+
+def finalize_autosummaries() -> None:
+    """Create the necessary ops to include autosummaries in TensorBoard report.
+    Note: This should be done only once per graph.
+    """
+    global _finalized
+    tfutil.assert_tf_initialized()
+
+    if _finalized:
+        return None
+
+    _finalized = True
+    tfutil.init_uninitialized_vars([var for vars_list in _vars.values() for var in vars_list])
+
+    # Create summary ops.
+    with tf.device(None), tf.control_dependencies(None):
+        for name, vars_list in _vars.items():
+            name_id = name.replace("/", "_")
+            with tfutil.absolute_name_scope("Autosummary/" + name_id):
+                moments = tf.add_n(vars_list)
+                moments /= moments[0]
+                with tf.control_dependencies([moments]):  # read before resetting
+                    reset_ops = [tf.assign(var, tf.zeros(3, dtype=_dtype)) for var in vars_list]
+                    with tf.name_scope(None), tf.control_dependencies(reset_ops):  # reset before reporting
+                        mean = moments[1]
+                        std = tf.sqrt(moments[2] - tf.square(moments[1]))
+                        tf.summary.scalar(name, mean)
+                        if enable_custom_scalars:
+                            tf.summary.scalar("xCustomScalars/" + name + "/margin_lo", mean - std)
+                            tf.summary.scalar("xCustomScalars/" + name + "/margin_hi", mean + std)
+
+    # Setup layout for custom scalars.
+    layout = None
+    if enable_custom_scalars:
+        cat_dict = OrderedDict()
+        for series_name in sorted(_vars.keys()):
+            p = series_name.split("/")
+            cat = p[0] if len(p) >= 2 else ""
+            chart = "/".join(p[1:-1]) if len(p) >= 3 else p[-1]
+            if cat not in cat_dict:
+                cat_dict[cat] = OrderedDict()
+            if chart not in cat_dict[cat]:
+                cat_dict[cat][chart] = []
+            cat_dict[cat][chart].append(series_name)
+        categories = []
+        for cat_name, chart_dict in cat_dict.items():
+            charts = []
+            for chart_name, series_names in chart_dict.items():
+                series = []
+                for series_name in series_names:
+                    series.append(layout_pb2.MarginChartContent.Series(
+                        value=series_name,
+                        lower="xCustomScalars/" + series_name + "/margin_lo",
+                        upper="xCustomScalars/" + series_name + "/margin_hi"))
+                margin = layout_pb2.MarginChartContent(series=series)
+                charts.append(layout_pb2.Chart(title=chart_name, margin=margin))
+            categories.append(layout_pb2.Category(title=cat_name, chart=charts))
+        layout = summary_lib.custom_scalar_pb(layout_pb2.Layout(category=categories))
+    return layout
+
+def save_summaries(file_writer, global_step=None):
+    """Call FileWriter.add_summary() with all summaries in the default graph,
+    automatically finalizing and merging them on the first call.
+    """
+    global _merge_op
+    tfutil.assert_tf_initialized()
+
+    if _merge_op is None:
+        layout = finalize_autosummaries()
+        if layout is not None:
+            file_writer.add_summary(layout)
+        with tf.device(None), tf.control_dependencies(None):
+            _merge_op = tf.summary.merge_all()
+
+    file_writer.add_summary(_merge_op.eval(), global_step)

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/custom_ops.py

@@ -0,0 +1,181 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""TensorFlow custom ops builder.
+"""
+
+import glob
+import os
+import re
+import uuid
+import hashlib
+import tempfile
+import shutil
+import tensorflow as tf
+from tensorflow.python.client import device_lib # pylint: disable=no-name-in-module
+
+from .. import util
+
+#----------------------------------------------------------------------------
+# Global configs.
+
+cuda_cache_path = None
+cuda_cache_version_tag = 'v1'
+do_not_hash_included_headers = True # Speed up compilation by assuming that headers included by the CUDA code never change.
+verbose = True # Print status messages to stdout.
+
+#----------------------------------------------------------------------------
+# Internal helper funcs.
+
+def _find_compiler_bindir():
+    hostx64_paths = sorted(glob.glob('C:/Program Files (x86)/Microsoft Visual Studio/*/Professional/VC/Tools/MSVC/*/bin/Hostx64/x64'), reverse=True)
+    if hostx64_paths != []:
+        return hostx64_paths[0]
+    hostx64_paths = sorted(glob.glob('C:/Program Files (x86)/Microsoft Visual Studio/*/BuildTools/VC/Tools/MSVC/*/bin/Hostx64/x64'), reverse=True)
+    if hostx64_paths != []:
+        return hostx64_paths[0]
+    hostx64_paths = sorted(glob.glob('C:/Program Files (x86)/Microsoft Visual Studio/*/Community/VC/Tools/MSVC/*/bin/Hostx64/x64'), reverse=True)
+    if hostx64_paths != []:
+        return hostx64_paths[0]
+    vc_bin_dir = 'C:/Program Files (x86)/Microsoft Visual Studio 14.0/vc/bin'
+    if os.path.isdir(vc_bin_dir):
+        return vc_bin_dir
+    return None
+
+def _get_compute_cap(device):
+    caps_str = device.physical_device_desc
+    m = re.search('compute capability: (\\d+).(\\d+)', caps_str)
+    major = m.group(1)
+    minor = m.group(2)
+    return (major, minor)
+
+def _get_cuda_gpu_arch_string():
+    gpus = [x for x in device_lib.list_local_devices() if x.device_type == 'GPU']
+    if len(gpus) == 0:
+        raise RuntimeError('No GPU devices found')
+    (major, minor) = _get_compute_cap(gpus[0])
+    return 'sm_%s%s' % (major, minor)
+
+def _run_cmd(cmd):
+    with os.popen(cmd) as pipe:
+        output = pipe.read()
+        status = pipe.close()
+    if status is not None:
+        raise RuntimeError('NVCC returned an error. See below for full command line and output log:\n\n%s\n\n%s' % (cmd, output))
+
+def _prepare_nvcc_cli(opts):
+    cmd = 'nvcc ' + opts.strip()
+    cmd += ' --disable-warnings'
+    cmd += ' --include-path "%s"' % tf.sysconfig.get_include()
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'protobuf_archive', 'src')
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'com_google_absl')
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'eigen_archive')
+
+    compiler_bindir = _find_compiler_bindir()
+    if compiler_bindir is None:
+        # Require that _find_compiler_bindir succeeds on Windows.  Allow
+        # nvcc to use whatever is the default on Linux.
+        if os.name == 'nt':
+            raise RuntimeError('Could not find MSVC/GCC/CLANG installation on this computer. Check compiler_bindir_search_path list in "%s".' % __file__)
+    else:
+        cmd += ' --compiler-bindir "%s"' % compiler_bindir
+    cmd += ' 2>&1'
+    return cmd
+
+#----------------------------------------------------------------------------
+# Main entry point.
+
+_plugin_cache = dict()
+
+def get_plugin(cuda_file, extra_nvcc_options=[]):
+    cuda_file_base = os.path.basename(cuda_file)
+    cuda_file_name, cuda_file_ext = os.path.splitext(cuda_file_base)
+
+    # Already in cache?
+    if cuda_file in _plugin_cache:
+        return _plugin_cache[cuda_file]
+
+    # Setup plugin.
+    if verbose:
+        print('Setting up TensorFlow plugin "%s": ' % cuda_file_base, end='', flush=True)
+    try:
+        # Hash CUDA source.
+        md5 = hashlib.md5()
+        with open(cuda_file, 'rb') as f:
+            md5.update(f.read())
+        md5.update(b'\n')
+
+        # Hash headers included by the CUDA code by running it through the preprocessor.
+        if not do_not_hash_included_headers:
+            if verbose:
+                print('Preprocessing... ', end='', flush=True)
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                tmp_file = os.path.join(tmp_dir, cuda_file_name + '_tmp' + cuda_file_ext)
+                _run_cmd(_prepare_nvcc_cli('"%s" --preprocess -o "%s" --keep --keep-dir "%s"' % (cuda_file, tmp_file, tmp_dir)))
+                with open(tmp_file, 'rb') as f:
+                    bad_file_str = ('"' + cuda_file.replace('\\', '/') + '"').encode('utf-8') # __FILE__ in error check macros
+                    good_file_str = ('"' + cuda_file_base + '"').encode('utf-8')
+                    for ln in f:
+                        if not ln.startswith(b'# ') and not ln.startswith(b'#line '): # ignore line number pragmas
+                            ln = ln.replace(bad_file_str, good_file_str)
+                            md5.update(ln)
+                    md5.update(b'\n')
+
+        # Select compiler configs.
+        compile_opts = ''
+        if os.name == 'nt':
+            compile_opts += '"%s"' % os.path.join(tf.sysconfig.get_lib(), 'python', '_pywrap_tensorflow_internal.lib')
+        elif os.name == 'posix':
+            compile_opts += f' --compiler-options \'-fPIC\''
+            compile_opts += f' --compiler-options \'{" ".join(tf.sysconfig.get_compile_flags())}\''
+            compile_opts += f' --linker-options \'{" ".join(tf.sysconfig.get_link_flags())}\''
+        else:
+            assert False # not Windows or Linux, w00t?
+        compile_opts += f' --gpu-architecture={_get_cuda_gpu_arch_string()}'
+        compile_opts += ' --use_fast_math'
+        for opt in extra_nvcc_options:
+            compile_opts += ' ' + opt
+        nvcc_cmd = _prepare_nvcc_cli(compile_opts)
+
+        # Hash build configuration.
+        md5.update(('nvcc_cmd: ' + nvcc_cmd).encode('utf-8') + b'\n')
+        md5.update(('tf.VERSION: ' + tf.VERSION).encode('utf-8') + b'\n')
+        md5.update(('cuda_cache_version_tag: ' + cuda_cache_version_tag).encode('utf-8') + b'\n')
+
+        # Compile if not already compiled.
+        cache_dir = util.make_cache_dir_path('tflib-cudacache') if cuda_cache_path is None else cuda_cache_path
+        bin_file_ext = '.dll' if os.name == 'nt' else '.so'
+        bin_file = os.path.join(cache_dir, cuda_file_name + '_' + md5.hexdigest() + bin_file_ext)
+        if not os.path.isfile(bin_file):
+            if verbose:
+                print('Compiling... ', end='', flush=True)
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                tmp_file = os.path.join(tmp_dir, cuda_file_name + '_tmp' + bin_file_ext)
+                _run_cmd(nvcc_cmd + ' "%s" --shared -o "%s" --keep --keep-dir "%s"' % (cuda_file, tmp_file, tmp_dir))
+                os.makedirs(cache_dir, exist_ok=True)
+                intermediate_file = os.path.join(cache_dir, cuda_file_name + '_' + uuid.uuid4().hex + '_tmp' + bin_file_ext)
+                shutil.copyfile(tmp_file, intermediate_file)
+                os.rename(intermediate_file, bin_file) # atomic
+
+        # Load.
+        if verbose:
+            print('Loading... ', end='', flush=True)
+        plugin = tf.load_op_library(bin_file)
+
+        # Add to cache.
+        _plugin_cache[cuda_file] = plugin
+        if verbose:
+            print('Done.', flush=True)
+        return plugin
+
+    except:
+        if verbose:
+            print('Failed!', flush=True)
+        raise
+
+#----------------------------------------------------------------------------

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/network.py

@@ -0,0 +1,781 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Helper for managing networks."""
+
+import types
+import inspect
+import re
+import uuid
+import sys
+import copy
+import numpy as np
+import tensorflow as tf
+
+from collections import OrderedDict
+from typing import Any, List, Tuple, Union, Callable
+
+from . import tfutil
+from .. import util
+
+from .tfutil import TfExpression, TfExpressionEx
+
+# pylint: disable=protected-access
+# pylint: disable=attribute-defined-outside-init
+# pylint: disable=too-many-public-methods
+
+_import_handlers = []  # Custom import handlers for dealing with legacy data in pickle import.
+_import_module_src = dict()  # Source code for temporary modules created during pickle import.
+
+
+def import_handler(handler_func):
+    """Function decorator for declaring custom import handlers."""
+    _import_handlers.append(handler_func)
+    return handler_func
+
+
+class Network:
+    """Generic network abstraction.
+
+    Acts as a convenience wrapper for a parameterized network construction
+    function, providing several utility methods and convenient access to
+    the inputs/outputs/weights.
+
+    Network objects can be safely pickled and unpickled for long-term
+    archival purposes. The pickling works reliably as long as the underlying
+    network construction function is defined in a standalone Python module
+    that has no side effects or application-specific imports.
+
+    Args:
+        name: Network name. Used to select TensorFlow name and variable scopes. Defaults to build func name if None.
+        func_name: Fully qualified name of the underlying network construction function, or a top-level function object.
+        static_kwargs: Keyword arguments to be passed in to the network construction function.
+    """
+
+    def __init__(self, name: str = None, func_name: Any = None, **static_kwargs):
+        # Locate the user-specified build function.
+        assert isinstance(func_name, str) or util.is_top_level_function(func_name)
+        if util.is_top_level_function(func_name):
+            func_name = util.get_top_level_function_name(func_name)
+        module, func_name = util.get_module_from_obj_name(func_name)
+        func = util.get_obj_from_module(module, func_name)
+
+        # Dig up source code for the module containing the build function.
+        module_src = _import_module_src.get(module, None)
+        if module_src is None:
+            module_src = inspect.getsource(module)
+
+        # Initialize fields.
+        self._init_fields(name=(name or func_name), static_kwargs=static_kwargs, build_func=func, build_func_name=func_name, build_module_src=module_src)
+
+    def _init_fields(self, name: str, static_kwargs: dict, build_func: Callable, build_func_name: str, build_module_src: str) -> None:
+        tfutil.assert_tf_initialized()
+        assert isinstance(name, str)
+        assert len(name) >= 1
+        assert re.fullmatch(r"[A-Za-z0-9_.\\-]*", name)
+        assert isinstance(static_kwargs, dict)
+        assert util.is_pickleable(static_kwargs)
+        assert callable(build_func)
+        assert isinstance(build_func_name, str)
+        assert isinstance(build_module_src, str)
+
+        # Choose TensorFlow name scope.
+        with tf.name_scope(None):
+            scope = tf.get_default_graph().unique_name(name, mark_as_used=True)
+
+        # Query current TensorFlow device.
+        with tfutil.absolute_name_scope(scope), tf.control_dependencies(None):
+            device = tf.no_op(name="_QueryDevice").device
+
+        # Immutable state.
+        self._name                  = name
+        self._scope                 = scope
+        self._device                = device
+        self._static_kwargs         = util.EasyDict(copy.deepcopy(static_kwargs))
+        self._build_func            = build_func
+        self._build_func_name       = build_func_name
+        self._build_module_src      = build_module_src
+
+        # State before _init_graph().
+        self._var_inits             = dict()    # var_name => initial_value, set to None by _init_graph()
+        self._all_inits_known       = False     # Do we know for sure that _var_inits covers all the variables?
+        self._components            = None      # subnet_name => Network, None if the components are not known yet
+
+        # Initialized by _init_graph().
+        self._input_templates       = None
+        self._output_templates      = None
+        self._own_vars              = None
+
+        # Cached values initialized the respective methods.
+        self._input_shapes          = None
+        self._output_shapes         = None
+        self._input_names           = None
+        self._output_names          = None
+        self._vars                  = None
+        self._trainables            = None
+        self._var_global_to_local   = None
+        self._run_cache             = dict()
+
+    def _init_graph(self) -> None:
+        assert self._var_inits is not None
+        assert self._input_templates is None
+        assert self._output_templates is None
+        assert self._own_vars is None
+
+        # Initialize components.
+        if self._components is None:
+            self._components = util.EasyDict()
+
+        # Choose build func kwargs.
+        build_kwargs = dict(self.static_kwargs)
+        build_kwargs["is_template_graph"] = True
+        build_kwargs["components"] = self._components
+
+        # Override scope and device, and ignore surrounding control dependencies.
+        with tfutil.absolute_variable_scope(self.scope, reuse=False), tfutil.absolute_name_scope(self.scope), tf.device(self.device), tf.control_dependencies(None):
+            assert tf.get_variable_scope().name == self.scope
+            assert tf.get_default_graph().get_name_scope() == self.scope
+
+            # Create input templates.
+            self._input_templates = []
+            for param in inspect.signature(self._build_func).parameters.values():
+                if param.kind == param.POSITIONAL_OR_KEYWORD and param.default is param.empty:
+                    self._input_templates.append(tf.placeholder(tf.float32, name=param.name))
+
+            # Call build func.
+            out_expr = self._build_func(*self._input_templates, **build_kwargs)
+
+        # Collect output templates and variables.
+        assert tfutil.is_tf_expression(out_expr) or isinstance(out_expr, tuple)
+        self._output_templates = [out_expr] if tfutil.is_tf_expression(out_expr) else list(out_expr)
+        self._own_vars = OrderedDict((var.name[len(self.scope) + 1:].split(":")[0], var) for var in tf.global_variables(self.scope + "/"))
+
+        # Check for errors.
+        if len(self._input_templates) == 0:
+            raise ValueError("Network build func did not list any inputs.")
+        if len(self._output_templates) == 0:
+            raise ValueError("Network build func did not return any outputs.")
+        if any(not tfutil.is_tf_expression(t) for t in self._output_templates):
+            raise ValueError("Network outputs must be TensorFlow expressions.")
+        if any(t.shape.ndims is None for t in self._input_templates):
+            raise ValueError("Network input shapes not defined. Please call x.set_shape() for each input.")
+        if any(t.shape.ndims is None for t in self._output_templates):
+            raise ValueError("Network output shapes not defined. Please call x.set_shape() where applicable.")
+        if any(not isinstance(comp, Network) for comp in self._components.values()):
+            raise ValueError("Components of a Network must be Networks themselves.")
+        if len(self._components) != len(set(comp.name for comp in self._components.values())):
+            raise ValueError("Components of a Network must have unique names.")
+
+        # Initialize variables.
+        if len(self._var_inits):
+            tfutil.set_vars({self._get_vars()[name]: value for name, value in self._var_inits.items() if name in self._get_vars()})
+        remaining_inits = [var.initializer for name, var in self._own_vars.items() if name not in self._var_inits]
+        if self._all_inits_known:
+            assert len(remaining_inits) == 0
+        else:
+            tfutil.run(remaining_inits)
+        self._var_inits = None
+
+    @property
+    def name(self):
+        """User-specified name string."""
+        return self._name
+
+    @property
+    def scope(self):
+        """Unique TensorFlow scope containing template graph and variables, derived from the user-specified name."""
+        return self._scope
+
+    @property
+    def device(self):
+        """Name of the TensorFlow device that the weights of this network reside on. Determined by the current device at construction time."""
+        return self._device
+
+    @property
+    def static_kwargs(self):
+        """EasyDict of arguments passed to the user-supplied build func."""
+        return copy.deepcopy(self._static_kwargs)
+
+    @property
+    def components(self):
+        """EasyDict of sub-networks created by the build func."""
+        return copy.copy(self._get_components())
+
+    def _get_components(self):
+        if self._components is None:
+            self._init_graph()
+            assert self._components is not None
+        return self._components
+
+    @property
+    def input_shapes(self):
+        """List of input tensor shapes, including minibatch dimension."""
+        if self._input_shapes is None:
+            self._input_shapes = [t.shape.as_list() for t in self.input_templates]
+        return copy.deepcopy(self._input_shapes)
+
+    @property
+    def output_shapes(self):
+        """List of output tensor shapes, including minibatch dimension."""
+        if self._output_shapes is None:
+            self._output_shapes = [t.shape.as_list() for t in self.output_templates]
+        return copy.deepcopy(self._output_shapes)
+
+    @property
+    def input_shape(self):
+        """Short-hand for input_shapes[0]."""
+        return self.input_shapes[0]
+
+    @property
+    def output_shape(self):
+        """Short-hand for output_shapes[0]."""
+        return self.output_shapes[0]
+
+    @property
+    def num_inputs(self):
+        """Number of input tensors."""
+        return len(self.input_shapes)
+
+    @property
+    def num_outputs(self):
+        """Number of output tensors."""
+        return len(self.output_shapes)
+
+    @property
+    def input_names(self):
+        """Name string for each input."""
+        if self._input_names is None:
+            self._input_names = [t.name.split("/")[-1].split(":")[0] for t in self.input_templates]
+        return copy.copy(self._input_names)
+
+    @property
+    def output_names(self):
+        """Name string for each output."""
+        if self._output_names is None:
+            self._output_names = [t.name.split("/")[-1].split(":")[0] for t in self.output_templates]
+        return copy.copy(self._output_names)
+
+    @property
+    def input_templates(self):
+        """Input placeholders in the template graph."""
+        if self._input_templates is None:
+            self._init_graph()
+            assert self._input_templates is not None
+        return copy.copy(self._input_templates)
+
+    @property
+    def output_templates(self):
+        """Output tensors in the template graph."""
+        if self._output_templates is None:
+            self._init_graph()
+            assert self._output_templates is not None
+        return copy.copy(self._output_templates)
+
+    @property
+    def own_vars(self):
+        """Variables defined by this network (local_name => var), excluding sub-networks."""
+        return copy.copy(self._get_own_vars())
+
+    def _get_own_vars(self):
+        if self._own_vars is None:
+            self._init_graph()
+            assert self._own_vars is not None
+        return self._own_vars
+
+    @property
+    def vars(self):
+        """All variables (local_name => var)."""
+        return copy.copy(self._get_vars())
+
+    def _get_vars(self):
+        if self._vars is None:
+            self._vars = OrderedDict(self._get_own_vars())
+            for comp in self._get_components().values():
+                self._vars.update((comp.name + "/" + name, var) for name, var in comp._get_vars().items())
+        return self._vars
+
+    @property
+    def trainables(self):
+        """All trainable variables (local_name => var)."""
+        return copy.copy(self._get_trainables())
+
+    def _get_trainables(self):
+        if self._trainables is None:
+            self._trainables = OrderedDict((name, var) for name, var in self.vars.items() if var.trainable)
+        return self._trainables
+
+    @property
+    def var_global_to_local(self):
+        """Mapping from variable global names to local names."""
+        return copy.copy(self._get_var_global_to_local())
+
+    def _get_var_global_to_local(self):
+        if self._var_global_to_local is None:
+            self._var_global_to_local = OrderedDict((var.name.split(":")[0], name) for name, var in self.vars.items())
+        return self._var_global_to_local
+
+    def reset_own_vars(self) -> None:
+        """Re-initialize all variables of this network, excluding sub-networks."""
+        if self._var_inits is None or self._components is None:
+            tfutil.run([var.initializer for var in self._get_own_vars().values()])
+        else:
+            self._var_inits.clear()
+            self._all_inits_known = False
+
+    def reset_vars(self) -> None:
+        """Re-initialize all variables of this network, including sub-networks."""
+        if self._var_inits is None:
+            tfutil.run([var.initializer for var in self._get_vars().values()])
+        else:
+            self._var_inits.clear()
+            self._all_inits_known = False
+            if self._components is not None:
+                for comp in self._components.values():
+                    comp.reset_vars()
+
+    def reset_trainables(self) -> None:
+        """Re-initialize all trainable variables of this network, including sub-networks."""
+        tfutil.run([var.initializer for var in self._get_trainables().values()])
+
+    def get_output_for(self, *in_expr: TfExpression, return_as_list: bool = False, **dynamic_kwargs) -> Union[TfExpression, List[TfExpression]]:
+        """Construct TensorFlow expression(s) for the output(s) of this network, given the input expression(s).
+        The graph is placed on the current TensorFlow device."""
+        assert len(in_expr) == self.num_inputs
+        assert not all(expr is None for expr in in_expr)
+        self._get_vars()  # ensure that all variables have been created
+
+        # Choose build func kwargs.
+        build_kwargs = dict(self.static_kwargs)
+        build_kwargs.update(dynamic_kwargs)
+        build_kwargs["is_template_graph"] = False
+        build_kwargs["components"] = self._components
+
+        # Build TensorFlow graph to evaluate the network.
+        with tfutil.absolute_variable_scope(self.scope, reuse=True), tf.name_scope(self.name):
+            assert tf.get_variable_scope().name == self.scope
+            valid_inputs = [expr for expr in in_expr if expr is not None]
+            final_inputs = []
+            for expr, name, shape in zip(in_expr, self.input_names, self.input_shapes):
+                if expr is not None:
+                    expr = tf.identity(expr, name=name)
+                else:
+                    expr = tf.zeros([tf.shape(valid_inputs[0])[0]] + shape[1:], name=name)
+                final_inputs.append(expr)
+            out_expr = self._build_func(*final_inputs, **build_kwargs)
+
+        # Propagate input shapes back to the user-specified expressions.
+        for expr, final in zip(in_expr, final_inputs):
+            if isinstance(expr, tf.Tensor):
+                expr.set_shape(final.shape)
+
+        # Express outputs in the desired format.
+        assert tfutil.is_tf_expression(out_expr) or isinstance(out_expr, tuple)
+        if return_as_list:
+            out_expr = [out_expr] if tfutil.is_tf_expression(out_expr) else list(out_expr)
+        return out_expr
+
+    def get_var_local_name(self, var_or_global_name: Union[TfExpression, str]) -> str:
+        """Get the local name of a given variable, without any surrounding name scopes."""
+        assert tfutil.is_tf_expression(var_or_global_name) or isinstance(var_or_global_name, str)
+        global_name = var_or_global_name if isinstance(var_or_global_name, str) else var_or_global_name.name
+        return self._get_var_global_to_local()[global_name]
+
+    def find_var(self, var_or_local_name: Union[TfExpression, str]) -> TfExpression:
+        """Find variable by local or global name."""
+        assert tfutil.is_tf_expression(var_or_local_name) or isinstance(var_or_local_name, str)
+        return self._get_vars()[var_or_local_name] if isinstance(var_or_local_name, str) else var_or_local_name
+
+    def get_var(self, var_or_local_name: Union[TfExpression, str]) -> np.ndarray:
+        """Get the value of a given variable as NumPy array.
+        Note: This method is very inefficient -- prefer to use tflib.run(list_of_vars) whenever possible."""
+        return self.find_var(var_or_local_name).eval()
+
+    def set_var(self, var_or_local_name: Union[TfExpression, str], new_value: Union[int, float, np.ndarray]) -> None:
+        """Set the value of a given variable based on the given NumPy array.
+        Note: This method is very inefficient -- prefer to use tflib.set_vars() whenever possible."""
+        tfutil.set_vars({self.find_var(var_or_local_name): new_value})
+
+    def __getstate__(self) -> dict:
+        """Pickle export."""
+        state = dict()
+        state["version"]            = 5
+        state["name"]               = self.name
+        state["static_kwargs"]      = dict(self.static_kwargs)
+        state["components"]         = dict(self.components)
+        state["build_module_src"]   = self._build_module_src
+        state["build_func_name"]    = self._build_func_name
+        state["variables"]          = list(zip(self._get_own_vars().keys(), tfutil.run(list(self._get_own_vars().values()))))
+        state["input_shapes"]       = self.input_shapes
+        state["output_shapes"]      = self.output_shapes
+        state["input_names"]        = self.input_names
+        state["output_names"]       = self.output_names
+        return state
+
+    def __setstate__(self, state: dict) -> None:
+        """Pickle import."""
+
+        # Execute custom import handlers.
+        for handler in _import_handlers:
+            state = handler(state)
+
+        # Get basic fields.
+        assert state["version"] in [2, 3, 4, 5]
+        name = state["name"]
+        static_kwargs = state["static_kwargs"]
+        build_module_src = state["build_module_src"]
+        build_func_name = state["build_func_name"]
+
+        # Create temporary module from the imported source code.
+        module_name = "_tflib_network_import_" + uuid.uuid4().hex
+        module = types.ModuleType(module_name)
+        sys.modules[module_name] = module
+        _import_module_src[module] = build_module_src
+        exec(build_module_src, module.__dict__) # pylint: disable=exec-used
+        build_func = util.get_obj_from_module(module, build_func_name)
+
+        # Initialize fields.
+        self._init_fields(name=name, static_kwargs=static_kwargs, build_func=build_func, build_func_name=build_func_name, build_module_src=build_module_src)
+        self._var_inits.update(copy.deepcopy(state["variables"]))
+        self._all_inits_known   = True
+        self._components        = util.EasyDict(state.get("components", {}))
+        self._input_shapes      = copy.deepcopy(state.get("input_shapes", None))
+        self._output_shapes     = copy.deepcopy(state.get("output_shapes", None))
+        self._input_names       = copy.deepcopy(state.get("input_names", None))
+        self._output_names      = copy.deepcopy(state.get("output_names", None))
+
+    def clone(self, name: str = None, **new_static_kwargs) -> "Network":
+        """Create a clone of this network with its own copy of the variables."""
+        static_kwargs = dict(self.static_kwargs)
+        static_kwargs.update(new_static_kwargs)
+        net = object.__new__(Network)
+        net._init_fields(name=(name or self.name), static_kwargs=static_kwargs, build_func=self._build_func, build_func_name=self._build_func_name, build_module_src=self._build_module_src)
+        net.copy_vars_from(self)
+        return net
+
+    def copy_own_vars_from(self, src_net: "Network") -> None:
+        """Copy the values of all variables from the given network, excluding sub-networks."""
+
+        # Source has unknown variables or unknown components => init now.
+        if (src_net._var_inits is not None and not src_net._all_inits_known) or src_net._components is None:
+            src_net._get_vars()
+
+       # Both networks are inited => copy directly.
+        if src_net._var_inits is None and self._var_inits is None:
+            names = [name for name in self._get_own_vars().keys() if name in src_net._get_own_vars()]
+            tfutil.set_vars(tfutil.run({self._get_vars()[name]: src_net._get_vars()[name] for name in names}))
+            return
+
+        # Read from source.
+        if src_net._var_inits is None:
+            value_dict = tfutil.run(src_net._get_own_vars())
+        else:
+            value_dict = src_net._var_inits
+
+        # Write to destination.
+        if self._var_inits is None:
+            tfutil.set_vars({self._get_vars()[name]: value for name, value in value_dict.items() if name in self._get_vars()})
+        else:
+            self._var_inits.update(value_dict)
+
+    def copy_vars_from(self, src_net: "Network") -> None:
+        """Copy the values of all variables from the given network, including sub-networks."""
+
+        # Source has unknown variables or unknown components => init now.
+        if (src_net._var_inits is not None and not src_net._all_inits_known) or src_net._components is None:
+            src_net._get_vars()
+
+        # Source is inited, but destination components have not been created yet => set as initial values.
+        if src_net._var_inits is None and self._components is None:
+            self._var_inits.update(tfutil.run(src_net._get_vars()))
+            return
+
+        # Destination has unknown components => init now.
+        if self._components is None:
+            self._get_vars()
+
+        # Both networks are inited => copy directly.
+        if src_net._var_inits is None and self._var_inits is None:
+            names = [name for name in self._get_vars().keys() if name in src_net._get_vars()]
+            tfutil.set_vars(tfutil.run({self._get_vars()[name]: src_net._get_vars()[name] for name in names}))
+            return
+
+        # Copy recursively, component by component.
+        self.copy_own_vars_from(src_net)
+        for name, src_comp in src_net._components.items():
+            if name in self._components:
+                self._components[name].copy_vars_from(src_comp)
+
+    def copy_trainables_from(self, src_net: "Network") -> None:
+        """Copy the values of all trainable variables from the given network, including sub-networks."""
+        names = [name for name in self._get_trainables().keys() if name in src_net._get_trainables()]
+        tfutil.set_vars(tfutil.run({self._get_vars()[name]: src_net._get_vars()[name] for name in names}))
+
+    def convert(self, new_func_name: str, new_name: str = None, **new_static_kwargs) -> "Network":
+        """Create new network with the given parameters, and copy all variables from this network."""
+        if new_name is None:
+            new_name = self.name
+        static_kwargs = dict(self.static_kwargs)
+        static_kwargs.update(new_static_kwargs)
+        net = Network(name=new_name, func_name=new_func_name, **static_kwargs)
+        net.copy_vars_from(self)
+        return net
+
+    def setup_as_moving_average_of(self, src_net: "Network", beta: TfExpressionEx = 0.99, beta_nontrainable: TfExpressionEx = 0.0) -> tf.Operation:
+        """Construct a TensorFlow op that updates the variables of this network
+        to be slightly closer to those of the given network."""
+        with tfutil.absolute_name_scope(self.scope + "/_MovingAvg"):
+            ops = []
+            for name, var in self._get_vars().items():
+                if name in src_net._get_vars():
+                    cur_beta = beta if var.trainable else beta_nontrainable
+                    new_value = tfutil.lerp(src_net._get_vars()[name], var, cur_beta)
+                    ops.append(var.assign(new_value))
+            return tf.group(*ops)
+
+    def run(self,
+            *in_arrays: Tuple[Union[np.ndarray, None], ...],
+            input_transform: dict = None,
+            output_transform: dict = None,
+            return_as_list: bool = False,
+            print_progress: bool = False,
+            minibatch_size: int = None,
+            num_gpus: int = 1,
+            assume_frozen: bool = False,
+            **dynamic_kwargs) -> Union[np.ndarray, Tuple[np.ndarray, ...], List[np.ndarray]]:
+        """Run this network for the given NumPy array(s), and return the output(s) as NumPy array(s).
+
+        Args:
+            input_transform:    A dict specifying a custom transformation to be applied to the input tensor(s) before evaluating the network.
+                                The dict must contain a 'func' field that points to a top-level function. The function is called with the input
+                                TensorFlow expression(s) as positional arguments. Any remaining fields of the dict will be passed in as kwargs.
+            output_transform:   A dict specifying a custom transformation to be applied to the output tensor(s) after evaluating the network.
+                                The dict must contain a 'func' field that points to a top-level function. The function is called with the output
+                                TensorFlow expression(s) as positional arguments. Any remaining fields of the dict will be passed in as kwargs.
+            return_as_list:     True = return a list of NumPy arrays, False = return a single NumPy array, or a tuple if there are multiple outputs.
+            print_progress:     Print progress to the console? Useful for very large input arrays.
+            minibatch_size:     Maximum minibatch size to use, None = disable batching.
+            num_gpus:           Number of GPUs to use.
+            assume_frozen:      Improve multi-GPU performance by assuming that the trainable parameters will remain changed between calls.
+            dynamic_kwargs:     Additional keyword arguments to be passed into the network build function.
+        """
+        assert len(in_arrays) == self.num_inputs
+        assert not all(arr is None for arr in in_arrays)
+        assert input_transform is None or util.is_top_level_function(input_transform["func"])
+        assert output_transform is None or util.is_top_level_function(output_transform["func"])
+        output_transform, dynamic_kwargs = _handle_legacy_output_transforms(output_transform, dynamic_kwargs)
+        num_items = in_arrays[0].shape[0]
+        if minibatch_size is None:
+            minibatch_size = num_items
+
+        # Construct unique hash key from all arguments that affect the TensorFlow graph.
+        key = dict(input_transform=input_transform, output_transform=output_transform, num_gpus=num_gpus, assume_frozen=assume_frozen, dynamic_kwargs=dynamic_kwargs)
+        def unwind_key(obj):
+            if isinstance(obj, dict):
+                return [(key, unwind_key(value)) for key, value in sorted(obj.items())]
+            if callable(obj):
+                return util.get_top_level_function_name(obj)
+            return obj
+        key = repr(unwind_key(key))
+
+        # Build graph.
+        if key not in self._run_cache:
+            with tfutil.absolute_name_scope(self.scope + "/_Run"), tf.control_dependencies(None):
+                with tf.device("/cpu:0"):
+                    in_expr = [tf.placeholder(tf.float32, name=name) for name in self.input_names]
+                    in_split = list(zip(*[tf.split(x, num_gpus) for x in in_expr]))
+
+                out_split = []
+                for gpu in range(num_gpus):
+                    with tf.device(self.device if num_gpus == 1 else "/gpu:%d" % gpu):
+                        net_gpu = self.clone() if assume_frozen else self
+                        in_gpu = in_split[gpu]
+
+                        if input_transform is not None:
+                            in_kwargs = dict(input_transform)
+                            in_gpu = in_kwargs.pop("func")(*in_gpu, **in_kwargs)
+                            in_gpu = [in_gpu] if tfutil.is_tf_expression(in_gpu) else list(in_gpu)
+
+                        assert len(in_gpu) == self.num_inputs
+                        out_gpu = net_gpu.get_output_for(*in_gpu, return_as_list=True, **dynamic_kwargs)
+
+                        if output_transform is not None:
+                            out_kwargs = dict(output_transform)
+                            out_gpu = out_kwargs.pop("func")(*out_gpu, **out_kwargs)
+                            out_gpu = [out_gpu] if tfutil.is_tf_expression(out_gpu) else list(out_gpu)
+
+                        assert len(out_gpu) == self.num_outputs
+                        out_split.append(out_gpu)
+
+                with tf.device("/cpu:0"):
+                    out_expr = [tf.concat(outputs, axis=0) for outputs in zip(*out_split)]
+                    self._run_cache[key] = in_expr, out_expr
+
+        # Run minibatches.
+        in_expr, out_expr = self._run_cache[key]
+        out_arrays = [np.empty([num_items] + expr.shape.as_list()[1:], expr.dtype.name) for expr in out_expr]
+
+        for mb_begin in range(0, num_items, minibatch_size):
+            if print_progress:
+                print("\r%d / %d" % (mb_begin, num_items), end="")
+
+            mb_end = min(mb_begin + minibatch_size, num_items)
+            mb_num = mb_end - mb_begin
+            mb_in = [src[mb_begin : mb_end] if src is not None else np.zeros([mb_num] + shape[1:]) for src, shape in zip(in_arrays, self.input_shapes)]
+            mb_out = tf.get_default_session().run(out_expr, dict(zip(in_expr, mb_in)))
+
+            for dst, src in zip(out_arrays, mb_out):
+                dst[mb_begin: mb_end] = src
+
+        # Done.
+        if print_progress:
+            print("\r%d / %d" % (num_items, num_items))
+
+        if not return_as_list:
+            out_arrays = out_arrays[0] if len(out_arrays) == 1 else tuple(out_arrays)
+        return out_arrays
+
+    def list_ops(self) -> List[TfExpression]:
+        _ = self.output_templates  # ensure that the template graph has been created
+        include_prefix = self.scope + "/"
+        exclude_prefix = include_prefix + "_"
+        ops = tf.get_default_graph().get_operations()
+        ops = [op for op in ops if op.name.startswith(include_prefix)]
+        ops = [op for op in ops if not op.name.startswith(exclude_prefix)]
+        return ops
+
+    def list_layers(self) -> List[Tuple[str, TfExpression, List[TfExpression]]]:
+        """Returns a list of (layer_name, output_expr, trainable_vars) tuples corresponding to
+        individual layers of the network. Mainly intended to be used for reporting."""
+        layers = []
+
+        def recurse(scope, parent_ops, parent_vars, level):
+            if len(parent_ops) == 0 and len(parent_vars) == 0:
+                return
+
+            # Ignore specific patterns.
+            if any(p in scope for p in ["/Shape", "/strided_slice", "/Cast", "/concat", "/Assign"]):
+                return
+
+            # Filter ops and vars by scope.
+            global_prefix = scope + "/"
+            local_prefix = global_prefix[len(self.scope) + 1:]
+            cur_ops = [op for op in parent_ops if op.name.startswith(global_prefix) or op.name == global_prefix[:-1]]
+            cur_vars = [(name, var) for name, var in parent_vars if name.startswith(local_prefix) or name == local_prefix[:-1]]
+            if not cur_ops and not cur_vars:
+                return
+
+            # Filter out all ops related to variables.
+            for var in [op for op in cur_ops if op.type.startswith("Variable")]:
+                var_prefix = var.name + "/"
+                cur_ops = [op for op in cur_ops if not op.name.startswith(var_prefix)]
+
+            # Scope does not contain ops as immediate children => recurse deeper.
+            contains_direct_ops = any("/" not in op.name[len(global_prefix):] and op.type not in ["Identity", "Cast", "Transpose"] for op in cur_ops)
+            if (level == 0 or not contains_direct_ops) and (len(cur_ops) != 0 or len(cur_vars) != 0):
+                visited = set()
+                for rel_name in [op.name[len(global_prefix):] for op in cur_ops] + [name[len(local_prefix):] for name, _var in cur_vars]:
+                    token = rel_name.split("/")[0]
+                    if token not in visited:
+                        recurse(global_prefix + token, cur_ops, cur_vars, level + 1)
+                        visited.add(token)
+                return
+
+            # Report layer.
+            layer_name = scope[len(self.scope) + 1:]
+            layer_output = cur_ops[-1].outputs[0] if cur_ops else cur_vars[-1][1]
+            layer_trainables = [var for _name, var in cur_vars if var.trainable]
+            layers.append((layer_name, layer_output, layer_trainables))
+
+        recurse(self.scope, self.list_ops(), list(self._get_vars().items()), 0)
+        return layers
+
+    def print_layers(self, title: str = None, hide_layers_with_no_params: bool = False) -> None:
+        """Print a summary table of the network structure."""
+        rows = [[title if title is not None else self.name, "Params", "OutputShape", "WeightShape"]]
+        rows += [["---"] * 4]
+        total_params = 0
+
+        for layer_name, layer_output, layer_trainables in self.list_layers():
+            num_params = sum(int(np.prod(var.shape.as_list())) for var in layer_trainables)
+            weights = [var for var in layer_trainables if var.name.endswith("/weight:0")]
+            weights.sort(key=lambda x: len(x.name))
+            if len(weights) == 0 and len(layer_trainables) == 1:
+                weights = layer_trainables
+            total_params += num_params
+
+            if not hide_layers_with_no_params or num_params != 0:
+                num_params_str = str(num_params) if num_params > 0 else "-"
+                output_shape_str = str(layer_output.shape)
+                weight_shape_str = str(weights[0].shape) if len(weights) >= 1 else "-"
+                rows += [[layer_name, num_params_str, output_shape_str, weight_shape_str]]
+
+        rows += [["---"] * 4]
+        rows += [["Total", str(total_params), "", ""]]
+
+        widths = [max(len(cell) for cell in column) for column in zip(*rows)]
+        print()
+        for row in rows:
+            print("  ".join(cell + " " * (width - len(cell)) for cell, width in zip(row, widths)))
+        print()
+
+    def setup_weight_histograms(self, title: str = None) -> None:
+        """Construct summary ops to include histograms of all trainable parameters in TensorBoard."""
+        if title is None:
+            title = self.name
+
+        with tf.name_scope(None), tf.device(None), tf.control_dependencies(None):
+            for local_name, var in self._get_trainables().items():
+                if "/" in local_name:
+                    p = local_name.split("/")
+                    name = title + "_" + p[-1] + "/" + "_".join(p[:-1])
+                else:
+                    name = title + "_toplevel/" + local_name
+
+                tf.summary.histogram(name, var)
+
+#----------------------------------------------------------------------------
+# Backwards-compatible emulation of legacy output transformation in Network.run().
+
+_print_legacy_warning = True
+
+def _handle_legacy_output_transforms(output_transform, dynamic_kwargs):
+    global _print_legacy_warning
+    legacy_kwargs = ["out_mul", "out_add", "out_shrink", "out_dtype"]
+    if not any(kwarg in dynamic_kwargs for kwarg in legacy_kwargs):
+        return output_transform, dynamic_kwargs
+
+    if _print_legacy_warning:
+        _print_legacy_warning = False
+        print()
+        print("WARNING: Old-style output transformations in Network.run() are deprecated.")
+        print("Consider using 'output_transform=dict(func=tflib.convert_images_to_uint8)'")
+        print("instead of 'out_mul=127.5, out_add=127.5, out_dtype=np.uint8'.")
+        print()
+    assert output_transform is None
+
+    new_kwargs = dict(dynamic_kwargs)
+    new_transform = {kwarg: new_kwargs.pop(kwarg) for kwarg in legacy_kwargs if kwarg in dynamic_kwargs}
+    new_transform["func"] = _legacy_output_transform_func
+    return new_transform, new_kwargs
+
+def _legacy_output_transform_func(*expr, out_mul=1.0, out_add=0.0, out_shrink=1, out_dtype=None):
+    if out_mul != 1.0:
+        expr = [x * out_mul for x in expr]
+
+    if out_add != 0.0:
+        expr = [x + out_add for x in expr]
+
+    if out_shrink > 1:
+        ksize = [1, 1, out_shrink, out_shrink]
+        expr = [tf.nn.avg_pool(x, ksize=ksize, strides=ksize, padding="VALID", data_format="NCHW") for x in expr]
+
+    if out_dtype is not None:
+        if tf.as_dtype(out_dtype).is_integer:
+            expr = [tf.round(x) for x in expr]
+        expr = [tf.saturate_cast(x, out_dtype) for x in expr]
+    return expr

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/ops/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+# empty

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/ops/fused_bias_act.cu

@@ -0,0 +1,220 @@
+// Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#define EIGEN_USE_GPU
+#define __CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/framework/shape_inference.h"
+#include <stdio.h>
+
+using namespace tensorflow;
+using namespace tensorflow::shape_inference;
+
+#define OP_CHECK_CUDA_ERROR(CTX, CUDA_CALL) do { cudaError_t err = CUDA_CALL; OP_REQUIRES(CTX, err == cudaSuccess, errors::Internal(cudaGetErrorName(err))); } while (false)
+
+//------------------------------------------------------------------------
+// CUDA kernel.
+
+template <class T>
+struct FusedBiasActKernelParams
+{
+    const T*    x;      // [sizeX]
+    const T*    b;      // [sizeB] or NULL
+    const T*    xref;   // [sizeX] or NULL
+    const T*    yref;   // [sizeX] or NULL
+    T*          y;      // [sizeX]
+
+    int         grad;
+    int         axis;
+    int         act;
+    float       alpha;
+    float       gain;
+    float       clamp;
+
+    int         sizeX;
+    int         sizeB;
+    int         stepB;
+    int         loopX;
+};
+
+template <class T>
+static __global__ void FusedBiasActKernel(const FusedBiasActKernelParams<T> p)
+{
+    const float expRange        = 80.0f;
+    const float halfExpRange    = 40.0f;
+    const float seluScale       = 1.0507009873554804934193349852946f;
+    const float seluAlpha       = 1.6732632423543772848170429916717f;
+
+    // Loop over elements.
+    int xi = blockIdx.x * p.loopX * blockDim.x + threadIdx.x;
+    for (int loopIdx = 0; loopIdx < p.loopX && xi < p.sizeX; loopIdx++, xi += blockDim.x)
+    {
+        // Load and apply bias.
+        float x = (float)p.x[xi];
+        if (p.b)
+            x += (float)p.b[(xi / p.stepB) % p.sizeB];
+        float xref = (p.xref) ? (float)p.xref[xi] : 0.0f;
+        float yref = (p.yref) ? (float)p.yref[xi] : 0.0f;
+        float yy = (p.gain != 0.0f) ? yref / p.gain : 0.0f;
+
+        // Evaluate activation func.
+        float y;
+        switch (p.act * 10 + p.grad)
+        {
+            // linear
+            default:
+            case 10: y = x; break;
+            case 11: y = x; break;
+            case 12: y = 0.0f; break;
+
+            // relu
+            case 20: y = (x > 0.0f) ? x : 0.0f; break;
+            case 21: y = (yy > 0.0f) ? x : 0.0f; break;
+            case 22: y = 0.0f; break;
+
+            // lrelu
+            case 30: y = (x > 0.0f) ? x : x * p.alpha; break;
+            case 31: y = (yy > 0.0f) ? x : x * p.alpha; break;
+            case 32: y = 0.0f; break;
+
+            // tanh
+            case 40: { float c = expf(x); float d = 1.0f / c; y = (x < -expRange) ? -1.0f : (x > expRange) ? 1.0f : (c - d) / (c + d); } break;
+            case 41: y = x * (1.0f - yy * yy); break;
+            case 42: y = x * (1.0f - yy * yy) * (-2.0f * yy); break;
+
+            // sigmoid
+            case 50: y = (x < -expRange) ? 0.0f : 1.0f / (expf(-x) + 1.0f); break;
+            case 51: y = x * yy * (1.0f - yy); break;
+            case 52: y = x * yy * (1.0f - yy) * (1.0f - 2.0f * yy); break;
+
+            // elu
+            case 60: y = (x >= 0.0f) ? x : expf(x) - 1.0f; break;
+            case 61: y = (yy >= 0.0f) ? x : x * (yy + 1.0f); break;
+            case 62: y = (yy >= 0.0f) ? 0.0f : x * (yy + 1.0f); break;
+
+            // selu
+            case 70: y = (x >= 0.0f) ? seluScale * x : (seluScale * seluAlpha) * (expf(x) - 1.0f); break;
+            case 71: y = (yy >= 0.0f) ? x * seluScale : x * (yy + seluScale * seluAlpha); break;
+            case 72: y = (yy >= 0.0f) ? 0.0f : x * (yy + seluScale * seluAlpha); break;
+
+            // softplus
+            case 80: y = (x > expRange) ? x : logf(expf(x) + 1.0f); break;
+            case 81: y = x * (1.0f - expf(-yy)); break;
+            case 82: { float c = expf(-yy); y = x * c * (1.0f - c); } break;
+
+            // swish
+            case 90: y = (x < -expRange) ? 0.0f : x / (expf(-x) + 1.0f); break;
+            case 91:
+            case 92:
+                {
+                    float c = expf(xref);
+                    float d = c + 1.0f;
+                    if (p.grad == 1)
+                        y = (xref > halfExpRange) ? x : x * c * (xref + d) / (d * d);
+                    else
+                        y = (xref > halfExpRange) ? 0.0f : x * c * (xref * (2.0f - d) + 2.0f * d) / (d * d * d);
+                    yref = (xref < -expRange) ? 0.0f : xref / (expf(-xref) + 1.0f) * p.gain;
+                }
+                break;
+        }
+
+        // Apply gain.
+        y *= p.gain;
+
+        // Clamp.
+        if (p.clamp >= 0.0f)
+        {
+            if (p.grad == 0)
+                y = (fabsf(y) < p.clamp) ? y : (y >= 0.0f) ? p.clamp : -p.clamp;
+            else
+                y = (fabsf(yref) < p.clamp) ? y : 0.0f;
+        }
+
+        // Store.
+        p.y[xi] = (T)y;
+    }
+}
+
+//------------------------------------------------------------------------
+// TensorFlow op.
+
+template <class T>
+struct FusedBiasActOp : public OpKernel
+{
+    FusedBiasActKernelParams<T> m_attribs;
+
+    FusedBiasActOp(OpKernelConstruction* ctx) : OpKernel(ctx)
+    {
+        memset(&m_attribs, 0, sizeof(m_attribs));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("grad",    &m_attribs.grad));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("axis",    &m_attribs.axis));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("act",     &m_attribs.act));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("alpha",   &m_attribs.alpha));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("gain",    &m_attribs.gain));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("clamp",   &m_attribs.clamp));
+        OP_REQUIRES(ctx, m_attribs.grad >= 0, errors::InvalidArgument("grad must be non-negative"));
+        OP_REQUIRES(ctx, m_attribs.axis >= 0, errors::InvalidArgument("axis must be non-negative"));
+        OP_REQUIRES(ctx, m_attribs.act >= 0, errors::InvalidArgument("act must be non-negative"));
+    }
+
+    void Compute(OpKernelContext* ctx)
+    {
+        FusedBiasActKernelParams<T> p = m_attribs;
+        cudaStream_t stream = ctx->eigen_device<Eigen::GpuDevice>().stream();
+
+        const Tensor& x     = ctx->input(0); // [...]
+        const Tensor& b     = ctx->input(1); // [sizeB] or [0]
+        const Tensor& xref  = ctx->input(2); // x.shape or [0]
+        const Tensor& yref  = ctx->input(3); // x.shape or [0]
+        p.x = x.flat<T>().data();
+        p.b = (b.NumElements()) ? b.flat<T>().data() : NULL;
+        p.xref = (xref.NumElements()) ? xref.flat<T>().data() : NULL;
+        p.yref = (yref.NumElements()) ? yref.flat<T>().data() : NULL;
+        OP_REQUIRES(ctx, b.NumElements() == 0 || m_attribs.axis < x.dims(), errors::InvalidArgument("axis out of bounds"));
+        OP_REQUIRES(ctx, b.dims() == 1, errors::InvalidArgument("b must have rank 1"));
+        OP_REQUIRES(ctx, b.NumElements() == 0 || b.NumElements() == x.dim_size(m_attribs.axis), errors::InvalidArgument("b has wrong number of elements"));
+        OP_REQUIRES(ctx, xref.NumElements() == 0 || xref.NumElements() == x.NumElements(), errors::InvalidArgument("xref has wrong number of elements"));
+        OP_REQUIRES(ctx, yref.NumElements() == 0 || yref.NumElements() == x.NumElements(), errors::InvalidArgument("yref has wrong number of elements"));
+        OP_REQUIRES(ctx, x.NumElements() <= kint32max, errors::InvalidArgument("x is too large"));
+
+        p.sizeX = (int)x.NumElements();
+        p.sizeB = (int)b.NumElements();
+        p.stepB = 1;
+        for (int i = m_attribs.axis + 1; i < x.dims(); i++)
+            p.stepB *= (int)x.dim_size(i);
+
+        Tensor* y = NULL; // x.shape
+        OP_REQUIRES_OK(ctx, ctx->allocate_output(0, x.shape(), &y));
+        p.y = y->flat<T>().data();
+
+        p.loopX = 4;
+        int blockSize = 4 * 32;
+        int gridSize = (p.sizeX - 1) / (p.loopX * blockSize) + 1;
+        void* args[] = {&p};
+        OP_CHECK_CUDA_ERROR(ctx, cudaLaunchKernel((void*)FusedBiasActKernel<T>, gridSize, blockSize, args, 0, stream));
+    }
+};
+
+REGISTER_OP("FusedBiasAct")
+    .Input      ("x: T")
+    .Input      ("b: T")
+    .Input      ("xref: T")
+    .Input      ("yref: T")
+    .Output     ("y: T")
+    .Attr       ("T: {float, half}")
+    .Attr       ("grad: int = 0")
+    .Attr       ("axis: int = 1")
+    .Attr       ("act: int = 0")
+    .Attr       ("alpha: float = 0.0")
+    .Attr       ("gain: float = 1.0")
+    .Attr       ("clamp: float = -1.0");
+REGISTER_KERNEL_BUILDER(Name("FusedBiasAct").Device(DEVICE_GPU).TypeConstraint<float>("T"), FusedBiasActOp<float>);
+REGISTER_KERNEL_BUILDER(Name("FusedBiasAct").Device(DEVICE_GPU).TypeConstraint<Eigen::half>("T"), FusedBiasActOp<Eigen::half>);
+
+//------------------------------------------------------------------------

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/ops/fused_bias_act.py

@@ -0,0 +1,211 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom TensorFlow ops for efficient bias and activation."""
+
+import os
+import numpy as np
+import tensorflow as tf
+from .. import custom_ops
+from ...util import EasyDict
+
+def _get_plugin():
+    return custom_ops.get_plugin(os.path.splitext(__file__)[0] + '.cu')
+
+#----------------------------------------------------------------------------
+
+activation_funcs = {
+    'linear':   EasyDict(func=lambda x, **_:        x,                          def_alpha=None, def_gain=1.0,           cuda_idx=1, ref='y', zero_2nd_grad=True),
+    'relu':     EasyDict(func=lambda x, **_:        tf.nn.relu(x),              def_alpha=None, def_gain=np.sqrt(2),    cuda_idx=2, ref='y', zero_2nd_grad=True),
+    'lrelu':    EasyDict(func=lambda x, alpha, **_: tf.nn.leaky_relu(x, alpha), def_alpha=0.2,  def_gain=np.sqrt(2),    cuda_idx=3, ref='y', zero_2nd_grad=True),
+    'tanh':     EasyDict(func=lambda x, **_:        tf.nn.tanh(x),              def_alpha=None, def_gain=1.0,           cuda_idx=4, ref='y', zero_2nd_grad=False),
+    'sigmoid':  EasyDict(func=lambda x, **_:        tf.nn.sigmoid(x),           def_alpha=None, def_gain=1.0,           cuda_idx=5, ref='y', zero_2nd_grad=False),
+    'elu':      EasyDict(func=lambda x, **_:        tf.nn.elu(x),               def_alpha=None, def_gain=1.0,           cuda_idx=6, ref='y', zero_2nd_grad=False),
+    'selu':     EasyDict(func=lambda x, **_:        tf.nn.selu(x),              def_alpha=None, def_gain=1.0,           cuda_idx=7, ref='y', zero_2nd_grad=False),
+    'softplus': EasyDict(func=lambda x, **_:        tf.nn.softplus(x),          def_alpha=None, def_gain=1.0,           cuda_idx=8, ref='y', zero_2nd_grad=False),
+    'swish':    EasyDict(func=lambda x, **_:        tf.nn.sigmoid(x) * x,       def_alpha=None, def_gain=np.sqrt(2),    cuda_idx=9, ref='x', zero_2nd_grad=False),
+}
+
+#----------------------------------------------------------------------------
+
+def fused_bias_act(x, b=None, axis=1, act='linear', alpha=None, gain=None, clamp=None, impl='cuda'):
+    r"""Fused bias and activation function.
+
+    Adds bias `b` to activation tensor `x`, evaluates activation function `act`,
+    and scales the result by `gain`. Each of the steps is optional. In most cases,
+    the fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports first and second order gradients,
+    but not third order gradients.
+
+    Args:
+        x:      Input activation tensor. Can have any shape, but if `b` is defined, the
+                dimension corresponding to `axis`, as well as the rank, must be known.
+        b:      Bias vector, or `None` to disable. Must be a 1D tensor of the same type
+                as `x`. The shape must be known, and it must match the dimension of `x`
+                corresponding to `axis`.
+        axis:   The dimension in `x` corresponding to the elements of `b`.
+                The value of `axis` is ignored if `b` is not specified.
+        act:    Name of the activation function to evaluate, or `"linear"` to disable.
+                Can be e.g. `"relu"`, `"lrelu"`, `"tanh"`, `"sigmoid"`, `"swish"`, etc.
+                See `activation_funcs` for a full list. `None` is not allowed.
+        alpha:  Shape parameter for the activation function, or `None` to use the default.
+        gain:   Scaling factor for the output tensor, or `None` to use default.
+                See `activation_funcs` for the default scaling of each activation function.
+                If unsure, consider specifying `1.0`.
+        clamp:  Clamp the output values to `[-clamp, +clamp]`, or `None` to disable
+                the clamping (default).
+        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the same shape and datatype as `x`.
+    """
+
+    impl_dict = {
+        'ref':  _fused_bias_act_ref,
+        'cuda': _fused_bias_act_cuda,
+    }
+    return impl_dict[impl](x=x, b=b, axis=axis, act=act, alpha=alpha, gain=gain, clamp=clamp)
+
+#----------------------------------------------------------------------------
+
+def _fused_bias_act_ref(x, b, axis, act, alpha, gain, clamp):
+    """Slow reference implementation of `fused_bias_act()` using standard TensorFlow ops."""
+
+    # Validate arguments.
+    x = tf.convert_to_tensor(x)
+    b = tf.convert_to_tensor(b) if b is not None else tf.constant([], dtype=x.dtype)
+    act_spec = activation_funcs[act]
+    assert b.shape.rank == 1 and (b.shape[0] == 0 or b.shape[0] == x.shape[axis])
+    assert b.shape[0] == 0 or 0 <= axis < x.shape.rank
+    if alpha is None:
+        alpha = act_spec.def_alpha
+    if gain is None:
+        gain = act_spec.def_gain
+
+    # Add bias.
+    if b.shape[0] != 0:
+        x += tf.reshape(b, [-1 if i == axis else 1 for i in range(x.shape.rank)])
+
+    # Evaluate activation function.
+    x = act_spec.func(x, alpha=alpha)
+
+    # Scale by gain.
+    if gain != 1:
+        x *= gain
+
+    # Clamp.
+    if clamp is not None:
+        clamp = np.asarray(clamp, dtype=x.dtype.name)
+        assert clamp.shape == () and clamp >= 0
+        x = tf.clip_by_value(x, -clamp, clamp)
+    return x
+
+#----------------------------------------------------------------------------
+
+def _fused_bias_act_cuda(x, b, axis, act, alpha, gain, clamp):
+    """Fast CUDA implementation of `fused_bias_act()` using custom ops."""
+
+    # Validate arguments.
+    x = tf.convert_to_tensor(x)
+    empty_tensor = tf.constant([], dtype=x.dtype)
+    b = tf.convert_to_tensor(b) if b is not None else empty_tensor
+    act_spec = activation_funcs[act]
+    assert b.shape.rank == 1 and (b.shape[0] == 0 or b.shape[0] == x.shape[axis])
+    assert b.shape[0] == 0 or 0 <= axis < x.shape.rank
+    if alpha is None:
+        alpha = act_spec.def_alpha
+    if gain is None:
+        gain = act_spec.def_gain
+
+    # Special cases.
+    if act == 'linear' and b is None and gain == 1.0:
+        return x
+    if act_spec.cuda_idx is None:
+        return _fused_bias_act_ref(x=x, b=b, axis=axis, act=act, alpha=alpha, gain=gain, clamp=clamp)
+
+    # CUDA op.
+    cuda_op = _get_plugin().fused_bias_act
+    cuda_kwargs = dict(axis=int(axis), act=int(act_spec.cuda_idx), gain=float(gain))
+    if alpha is not None:
+        cuda_kwargs['alpha'] = float(alpha)
+    if clamp is not None:
+        clamp = np.asarray(clamp, dtype=x.dtype.name)
+        assert clamp.shape == () and clamp >= 0
+        cuda_kwargs['clamp'] = float(clamp.astype(np.float32))
+    def ref(tensor, name):
+        return tensor if act_spec.ref == name else empty_tensor
+
+    # Forward pass: y = func(x, b).
+    def func_y(x, b):
+        y = cuda_op(x=x, b=b, xref=empty_tensor, yref=empty_tensor, grad=0, **cuda_kwargs)
+        y.set_shape(x.shape)
+        return y
+
+    # Backward pass: dx, db = grad(dy, x, y)
+    def grad_dx(dy, x, y):
+        dx = cuda_op(x=dy, b=empty_tensor, xref=ref(x,'x'), yref=ref(y,'y'), grad=1, **cuda_kwargs)
+        dx.set_shape(x.shape)
+        return dx
+    def grad_db(dx):
+        if b.shape[0] == 0:
+            return empty_tensor
+        db = dx
+        if axis < x.shape.rank - 1:
+            db = tf.reduce_sum(db, list(range(axis + 1, x.shape.rank)))
+        if axis > 0:
+            db = tf.reduce_sum(db, list(range(axis)))
+        db.set_shape(b.shape)
+        return db
+
+    # Second order gradients: d_dy, d_x = grad2(d_dx, d_db, x, y)
+    def grad2_d_dy(d_dx, d_db, x, y):
+        d_dy = cuda_op(x=d_dx, b=d_db, xref=ref(x,'x'), yref=ref(y,'y'), grad=1, **cuda_kwargs)
+        d_dy.set_shape(x.shape)
+        return d_dy
+    def grad2_d_x(d_dx, d_db, x, y):
+        d_x = cuda_op(x=d_dx, b=d_db, xref=ref(x,'x'), yref=ref(y,'y'), grad=2, **cuda_kwargs)
+        d_x.set_shape(x.shape)
+        return d_x
+
+    # Fast version for piecewise-linear activation funcs.
+    @tf.custom_gradient
+    def func_zero_2nd_grad(x, b):
+        y = func_y(x, b)
+        @tf.custom_gradient
+        def grad(dy):
+            dx = grad_dx(dy, x, y)
+            db = grad_db(dx)
+            def grad2(d_dx, d_db):
+                d_dy = grad2_d_dy(d_dx, d_db, x, y)
+                return d_dy
+            return (dx, db), grad2
+        return y, grad
+
+    # Slow version for general activation funcs.
+    @tf.custom_gradient
+    def func_nonzero_2nd_grad(x, b):
+        y = func_y(x, b)
+        def grad_wrap(dy):
+            @tf.custom_gradient
+            def grad_impl(dy, x):
+                dx = grad_dx(dy, x, y)
+                db = grad_db(dx)
+                def grad2(d_dx, d_db):
+                    d_dy = grad2_d_dy(d_dx, d_db, x, y)
+                    d_x = grad2_d_x(d_dx, d_db, x, y)
+                    return d_dy, d_x
+                return (dx, db), grad2
+            return grad_impl(dy, x)
+        return y, grad_wrap
+
+    # Which version to use?
+    if act_spec.zero_2nd_grad:
+        return func_zero_2nd_grad(x, b)
+    return func_nonzero_2nd_grad(x, b)
+
+#----------------------------------------------------------------------------

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/ops/upfirdn_2d.cu

@@ -0,0 +1,359 @@
+// Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#define EIGEN_USE_GPU
+#define __CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/framework/shape_inference.h"
+#include <stdio.h>
+
+using namespace tensorflow;
+using namespace tensorflow::shape_inference;
+
+//------------------------------------------------------------------------
+// Helpers.
+
+#define OP_CHECK_CUDA_ERROR(CTX, CUDA_CALL) do { cudaError_t err = CUDA_CALL; OP_REQUIRES(CTX, err == cudaSuccess, errors::Internal(cudaGetErrorName(err))); } while (false)
+
+static __host__ __device__ __forceinline__ int floorDiv(int a, int b)
+{
+    int t = 1 - a / b;
+    return (a + t * b) / b - t;
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel params.
+
+template <class T>
+struct UpFirDn2DKernelParams
+{
+    const T*    x;          // [majorDim, inH, inW, minorDim]
+    const T*    k;          // [kernelH, kernelW]
+    T*          y;          // [majorDim, outH, outW, minorDim]
+
+    int         upx;
+    int         upy;
+    int         downx;
+    int         downy;
+    int         padx0;
+    int         padx1;
+    int         pady0;
+    int         pady1;
+
+    int         majorDim;
+    int         inH;
+    int         inW;
+    int         minorDim;
+    int         kernelH;
+    int         kernelW;
+    int         outH;
+    int         outW;
+    int         loopMajor;
+    int         loopX;
+};
+
+//------------------------------------------------------------------------
+// General CUDA implementation for large filter kernels.
+
+template <class T>
+static __global__ void UpFirDn2DKernel_large(const UpFirDn2DKernelParams<T> p)
+{
+    // Calculate thread index.
+    int minorIdx = blockIdx.x * blockDim.x + threadIdx.x;
+    int outY = minorIdx / p.minorDim;
+    minorIdx -= outY * p.minorDim;
+    int outXBase = blockIdx.y * p.loopX * blockDim.y + threadIdx.y;
+    int majorIdxBase = blockIdx.z * p.loopMajor;
+    if (outXBase >= p.outW || outY >= p.outH || majorIdxBase >= p.majorDim)
+        return;
+
+    // Setup Y receptive field.
+    int midY = outY * p.downy + p.upy - 1 - p.pady0;
+    int inY = min(max(floorDiv(midY, p.upy), 0), p.inH);
+    int h = min(max(floorDiv(midY + p.kernelH, p.upy), 0), p.inH) - inY;
+    int kernelY = midY + p.kernelH - (inY + 1) * p.upy;
+
+    // Loop over majorDim and outX.
+    for (int loopMajor = 0, majorIdx = majorIdxBase; loopMajor < p.loopMajor && majorIdx < p.majorDim; loopMajor++, majorIdx++)
+    for (int loopX = 0, outX = outXBase; loopX < p.loopX && outX < p.outW; loopX++, outX += blockDim.y)
+    {
+        // Setup X receptive field.
+        int midX = outX * p.downx + p.upx - 1 - p.padx0;
+        int inX = min(max(floorDiv(midX, p.upx), 0), p.inW);
+        int w = min(max(floorDiv(midX + p.kernelW, p.upx), 0), p.inW) - inX;
+        int kernelX = midX + p.kernelW - (inX + 1) * p.upx;
+
+        // Initialize pointers.
+        const T* xp = &p.x[((majorIdx * p.inH + inY) * p.inW + inX) * p.minorDim + minorIdx];
+        const T* kp = &p.k[kernelY * p.kernelW + kernelX];
+        int xpx = p.minorDim;
+        int kpx = -p.upx;
+        int xpy = p.inW * p.minorDim;
+        int kpy = -p.upy * p.kernelW;
+
+        // Inner loop.
+        float v = 0.0f;
+        for (int y = 0; y < h; y++)
+        {
+            for (int x = 0; x < w; x++)
+            {
+                v += (float)(*xp) * (float)(*kp);
+                xp += xpx;
+                kp += kpx;
+            }
+            xp += xpy - w * xpx;
+            kp += kpy - w * kpx;
+        }
+
+        // Store result.
+        p.y[((majorIdx * p.outH + outY) * p.outW + outX) * p.minorDim + minorIdx] = (T)v;
+    }
+}
+
+//------------------------------------------------------------------------
+// Specialized CUDA implementation for small filter kernels.
+
+template <class T, int upx, int upy, int downx, int downy, int kernelW, int kernelH, int tileOutW, int tileOutH>
+static __global__ void UpFirDn2DKernel_small(const UpFirDn2DKernelParams<T> p)
+{
+    //assert(kernelW % upx == 0);
+    //assert(kernelH % upy == 0);
+    const int tileInW = ((tileOutW - 1) * downx + kernelW - 1) / upx + 1;
+    const int tileInH = ((tileOutH - 1) * downy + kernelH - 1) / upy + 1;
+    __shared__ volatile float sk[kernelH][kernelW];
+    __shared__ volatile float sx[tileInH][tileInW];
+
+    // Calculate tile index.
+    int minorIdx = blockIdx.x;
+    int tileOutY = minorIdx / p.minorDim;
+    minorIdx -= tileOutY * p.minorDim;
+    tileOutY *= tileOutH;
+    int tileOutXBase = blockIdx.y * p.loopX * tileOutW;
+    int majorIdxBase = blockIdx.z * p.loopMajor;
+    if (tileOutXBase >= p.outW | tileOutY >= p.outH | majorIdxBase >= p.majorDim)
+        return;
+
+    // Load filter kernel (flipped).
+    for (int tapIdx = threadIdx.x; tapIdx < kernelH * kernelW; tapIdx += blockDim.x)
+    {
+        int ky = tapIdx / kernelW;
+        int kx = tapIdx - ky * kernelW;
+        float v = 0.0f;
+        if (kx < p.kernelW & ky < p.kernelH)
+            v = (float)p.k[(p.kernelH - 1 - ky) * p.kernelW + (p.kernelW - 1 - kx)];
+        sk[ky][kx] = v;
+    }
+
+    // Loop over majorDim and outX.
+    for (int loopMajor = 0, majorIdx = majorIdxBase; loopMajor < p.loopMajor & majorIdx < p.majorDim; loopMajor++, majorIdx++)
+    for (int loopX = 0, tileOutX = tileOutXBase; loopX < p.loopX & tileOutX < p.outW; loopX++, tileOutX += tileOutW)
+    {
+        // Load input pixels.
+        int tileMidX = tileOutX * downx + upx - 1 - p.padx0;
+        int tileMidY = tileOutY * downy + upy - 1 - p.pady0;
+        int tileInX = floorDiv(tileMidX, upx);
+        int tileInY = floorDiv(tileMidY, upy);
+        __syncthreads();
+        for (int inIdx = threadIdx.x; inIdx < tileInH * tileInW; inIdx += blockDim.x)
+        {
+            int relInY = inIdx / tileInW;
+            int relInX = inIdx - relInY * tileInW;
+            int inX = relInX + tileInX;
+            int inY = relInY + tileInY;
+            float v = 0.0f;
+            if (inX >= 0 & inY >= 0 & inX < p.inW & inY < p.inH)
+                v = (float)p.x[((majorIdx * p.inH + inY) * p.inW + inX) * p.minorDim + minorIdx];
+            sx[relInY][relInX] = v;
+        }
+
+        // Loop over output pixels.
+        __syncthreads();
+        for (int outIdx = threadIdx.x; outIdx < tileOutH * tileOutW; outIdx += blockDim.x)
+        {
+            int relOutY = outIdx / tileOutW;
+            int relOutX = outIdx - relOutY * tileOutW;
+            int outX = relOutX + tileOutX;
+            int outY = relOutY + tileOutY;
+
+            // Setup receptive field.
+            int midX = tileMidX + relOutX * downx;
+            int midY = tileMidY + relOutY * downy;
+            int inX = floorDiv(midX, upx);
+            int inY = floorDiv(midY, upy);
+            int relInX = inX - tileInX;
+            int relInY = inY - tileInY;
+            int kernelX = (inX + 1) * upx - midX - 1; // flipped
+            int kernelY = (inY + 1) * upy - midY - 1; // flipped
+
+            // Inner loop.
+            float v = 0.0f;
+            #pragma unroll
+            for (int y = 0; y < kernelH / upy; y++)
+                #pragma unroll
+                for (int x = 0; x < kernelW / upx; x++)
+                    v += sx[relInY + y][relInX + x] * sk[kernelY + y * upy][kernelX + x * upx];
+
+            // Store result.
+            if (outX < p.outW & outY < p.outH)
+                p.y[((majorIdx * p.outH + outY) * p.outW + outX) * p.minorDim + minorIdx] = (T)v;
+        }
+    }
+}
+
+//------------------------------------------------------------------------
+// TensorFlow op.
+
+template <class T>
+struct UpFirDn2DOp : public OpKernel
+{
+    UpFirDn2DKernelParams<T> m_attribs;
+
+    UpFirDn2DOp(OpKernelConstruction* ctx) : OpKernel(ctx)
+    {
+        memset(&m_attribs, 0, sizeof(m_attribs));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("upx", &m_attribs.upx));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("upy", &m_attribs.upy));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("downx", &m_attribs.downx));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("downy", &m_attribs.downy));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("padx0", &m_attribs.padx0));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("padx1", &m_attribs.padx1));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("pady0", &m_attribs.pady0));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("pady1", &m_attribs.pady1));
+        OP_REQUIRES(ctx, m_attribs.upx >= 1 && m_attribs.upy >= 1, errors::InvalidArgument("upx and upy must be at least 1x1"));
+        OP_REQUIRES(ctx, m_attribs.downx >= 1 && m_attribs.downy >= 1, errors::InvalidArgument("downx and downy must be at least 1x1"));
+    }
+
+    void Compute(OpKernelContext* ctx)
+    {
+        UpFirDn2DKernelParams<T> p = m_attribs;
+        cudaStream_t stream = ctx->eigen_device<Eigen::GpuDevice>().stream();
+
+        const Tensor& x = ctx->input(0); // [majorDim, inH, inW, minorDim]
+        const Tensor& k = ctx->input(1); // [kernelH, kernelW]
+        p.x = x.flat<T>().data();
+        p.k = k.flat<T>().data();
+        OP_REQUIRES(ctx, x.dims() == 4, errors::InvalidArgument("input must have rank 4"));
+        OP_REQUIRES(ctx, k.dims() == 2, errors::InvalidArgument("kernel must have rank 2"));
+        OP_REQUIRES(ctx, x.NumElements() <= kint32max, errors::InvalidArgument("input too large"));
+        OP_REQUIRES(ctx, k.NumElements() <= kint32max, errors::InvalidArgument("kernel too large"));
+
+        p.majorDim  = (int)x.dim_size(0);
+        p.inH       = (int)x.dim_size(1);
+        p.inW       = (int)x.dim_size(2);
+        p.minorDim  = (int)x.dim_size(3);
+        p.kernelH   = (int)k.dim_size(0);
+        p.kernelW   = (int)k.dim_size(1);
+        OP_REQUIRES(ctx, p.kernelW >= 1 && p.kernelH >= 1, errors::InvalidArgument("kernel must be at least 1x1"));
+
+        p.outW = (p.inW * p.upx + p.padx0 + p.padx1 - p.kernelW + p.downx) / p.downx;
+        p.outH = (p.inH * p.upy + p.pady0 + p.pady1 - p.kernelH + p.downy) / p.downy;
+        OP_REQUIRES(ctx, p.outW >= 1 && p.outH >= 1, errors::InvalidArgument("output must be at least 1x1"));
+
+        Tensor* y = NULL; // [majorDim, outH, outW, minorDim]
+        TensorShape ys;
+        ys.AddDim(p.majorDim);
+        ys.AddDim(p.outH);
+        ys.AddDim(p.outW);
+        ys.AddDim(p.minorDim);
+        OP_REQUIRES_OK(ctx, ctx->allocate_output(0, ys, &y));
+        p.y = y->flat<T>().data();
+        OP_REQUIRES(ctx, y->NumElements() <= kint32max, errors::InvalidArgument("output too large"));
+
+        // Choose CUDA kernel to use.
+        void* cudaKernel = (void*)UpFirDn2DKernel_large<T>;
+        int tileOutW = -1;
+        int tileOutH = -1;
+
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 7  && p.kernelH <= 7 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 7,7,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 6  && p.kernelH <= 6 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 6,6,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 5  && p.kernelH <= 5 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 5,5,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 4  && p.kernelH <= 4 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 4,4,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 3  && p.kernelH <= 3 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 3,3,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 24 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 24,1, 128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 20 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 20,1, 128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 16 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 16,1, 128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 12 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 12,1, 128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 8  && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 8,1,  128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 24) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 1,24, 32,32>; tileOutW = 32;  tileOutH = 32; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 20) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 1,20, 32,32>; tileOutW = 32;  tileOutH = 32; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 16) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 1,16, 32,32>; tileOutW = 32;  tileOutH = 32; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 12) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 1,12, 32,32>; tileOutW = 32;  tileOutH = 32; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 8 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 1,8,  32,32>; tileOutW = 32;  tileOutH = 32; }
+
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 8  && p.kernelH <= 8 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 8,8,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 6  && p.kernelH <= 6 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 6,6,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 4  && p.kernelH <= 4 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 4,4,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 2  && p.kernelH <= 2 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 2,2,  64,16>; tileOutW = 64;  tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 24 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,1, 1,1, 24,1, 128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 2 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 20 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,1, 1,1, 20,1, 128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 2 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 16 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,1, 1,1, 16,1, 128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 2 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 12 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,1, 1,1, 12,1, 128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 2 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 8  && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,1, 1,1, 8,1,  128,8>; tileOutW = 128; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 24) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,2, 1,1, 1,24, 32,32>; tileOutW = 32;  tileOutH = 32; }
+        if (p.upx == 1 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 20) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,2, 1,1, 1,20, 32,32>; tileOutW = 32;  tileOutH = 32; }
+        if (p.upx == 1 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 16) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,2, 1,1, 1,16, 32,32>; tileOutW = 32;  tileOutH = 32; }
+        if (p.upx == 1 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 12) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,2, 1,1, 1,12, 32,32>; tileOutW = 32;  tileOutH = 32; }
+        if (p.upx == 1 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 1  && p.kernelH <= 8 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,2, 1,1, 1,8,  32,32>; tileOutW = 32;  tileOutH = 32; }
+
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 8  && p.kernelH <= 8 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 8,8,  32,8 >; tileOutW = 32;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 6  && p.kernelH <= 6 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 6,6,  32,8 >; tileOutW = 32;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 4  && p.kernelH <= 4 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 4,4,  32,8 >; tileOutW = 32;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 2  && p.kernelH <= 2 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 2,2,  32,8 >; tileOutW = 32;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 1 && p.kernelW <= 24 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,1, 24,1, 64,8 >; tileOutW = 64;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 1 && p.kernelW <= 20 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,1, 20,1, 64,8 >; tileOutW = 64;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 1 && p.kernelW <= 16 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,1, 16,1, 64,8 >; tileOutW = 64;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 1 && p.kernelW <= 12 && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,1, 12,1, 64,8 >; tileOutW = 64;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 1 && p.kernelW <= 8  && p.kernelH <= 1 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,1, 8,1,  64,8 >; tileOutW = 64;  tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 2 && p.kernelW <= 1  && p.kernelH <= 24) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,2, 1,24, 32,16>; tileOutW = 32;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 2 && p.kernelW <= 1  && p.kernelH <= 20) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,2, 1,20, 32,16>; tileOutW = 32;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 2 && p.kernelW <= 1  && p.kernelH <= 16) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,2, 1,16, 32,16>; tileOutW = 32;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 2 && p.kernelW <= 1  && p.kernelH <= 12) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,2, 1,12, 32,16>; tileOutW = 32;  tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 2 && p.kernelW <= 1  && p.kernelH <= 8 ) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,2, 1,8,  32,16>; tileOutW = 32;  tileOutH = 16; }
+
+        // Choose launch params.
+        dim3 blockSize;
+        dim3 gridSize;
+        if (tileOutW > 0 && tileOutH > 0) // small
+        {
+            p.loopMajor = (p.majorDim - 1) / 16384 + 1;
+            p.loopX = 1;
+            blockSize = dim3(32 * 8, 1, 1);
+            gridSize = dim3(((p.outH - 1) / tileOutH + 1) * p.minorDim, (p.outW - 1) / (p.loopX * tileOutW) + 1, (p.majorDim - 1) / p.loopMajor + 1);
+        }
+        else // large
+        {
+            p.loopMajor = (p.majorDim - 1) / 16384 + 1;
+            p.loopX = 4;
+            blockSize = dim3(4, 32, 1);
+            gridSize = dim3((p.outH * p.minorDim - 1) / blockSize.x + 1, (p.outW - 1) / (p.loopX * blockSize.y) + 1, (p.majorDim - 1) / p.loopMajor + 1);
+        }
+
+        // Launch CUDA kernel.
+        void* args[] = {&p};
+        OP_CHECK_CUDA_ERROR(ctx, cudaLaunchKernel(cudaKernel, gridSize, blockSize, args, 0, stream));
+    }
+};
+
+REGISTER_OP("UpFirDn2D")
+    .Input      ("x: T")
+    .Input      ("k: T")
+    .Output     ("y: T")
+    .Attr       ("T: {float, half}")
+    .Attr       ("upx: int = 1")
+    .Attr       ("upy: int = 1")
+    .Attr       ("downx: int = 1")
+    .Attr       ("downy: int = 1")
+    .Attr       ("padx0: int = 0")
+    .Attr       ("padx1: int = 0")
+    .Attr       ("pady0: int = 0")
+    .Attr       ("pady1: int = 0");
+REGISTER_KERNEL_BUILDER(Name("UpFirDn2D").Device(DEVICE_GPU).TypeConstraint<float>("T"), UpFirDn2DOp<float>);
+REGISTER_KERNEL_BUILDER(Name("UpFirDn2D").Device(DEVICE_GPU).TypeConstraint<Eigen::half>("T"), UpFirDn2DOp<Eigen::half>);
+
+//------------------------------------------------------------------------

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/ops/upfirdn_2d.py

@@ -0,0 +1,418 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom TensorFlow ops for efficient resampling of 2D images."""
+
+import os
+import numpy as np
+import tensorflow as tf
+from .. import custom_ops
+
+def _get_plugin():
+    return custom_ops.get_plugin(os.path.splitext(__file__)[0] + '.cu')
+
+#----------------------------------------------------------------------------
+
+def upfirdn_2d(x, k, upx=1, upy=1, downx=1, downy=1, padx0=0, padx1=0, pady0=0, pady1=0, impl='cuda'):
+    r"""Pad, upsample, FIR filter, and downsample a batch of 2D images.
+
+    Accepts a batch of 2D images of the shape `[majorDim, inH, inW, minorDim]`
+    and performs the following operations for each image, batched across
+    `majorDim` and `minorDim`:
+
+    1. Upsample the image by inserting the zeros after each pixel (`upx`, `upy`).
+
+    2. Pad the image with zeros by the specified number of pixels on each side
+       (`padx0`, `padx1`, `pady0`, `pady1`). Specifying a negative value
+       corresponds to cropping the image.
+
+    3. Convolve the image with the specified 2D FIR filter (`k`), shrinking the
+       image so that the footprint of all output pixels lies within the input image.
+
+    4. Downsample the image by throwing away pixels (`downx`, `downy`).
+
+    This sequence of operations bears close resemblance to scipy.signal.upfirdn().
+    The fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:      Input tensor of the shape `[majorDim, inH, inW, minorDim]`.
+        k:      2D FIR filter of the shape `[firH, firW]`.
+        upx:    Integer upsampling factor along the X-axis (default: 1).
+        upy:    Integer upsampling factor along the Y-axis (default: 1).
+        downx:  Integer downsampling factor along the X-axis (default: 1).
+        downy:  Integer downsampling factor along the Y-axis (default: 1).
+        padx0:  Number of pixels to pad on the left side (default: 0).
+        padx1:  Number of pixels to pad on the right side (default: 0).
+        pady0:  Number of pixels to pad on the top side (default: 0).
+        pady1:  Number of pixels to pad on the bottom side (default: 0).
+        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[majorDim, outH, outW, minorDim]`, and same datatype as `x`.
+    """
+
+    impl_dict = {
+        'ref':  _upfirdn_2d_ref,
+        'cuda': _upfirdn_2d_cuda,
+    }
+    return impl_dict[impl](x=x, k=k, upx=upx, upy=upy, downx=downx, downy=downy, padx0=padx0, padx1=padx1, pady0=pady0, pady1=pady1)
+
+#----------------------------------------------------------------------------
+
+def _upfirdn_2d_ref(x, k, upx, upy, downx, downy, padx0, padx1, pady0, pady1):
+    """Slow reference implementation of `upfirdn_2d()` using standard TensorFlow ops."""
+
+    x = tf.convert_to_tensor(x)
+    k = np.asarray(k, dtype=np.float32)
+    assert x.shape.rank == 4
+    inH = x.shape[1].value
+    inW = x.shape[2].value
+    minorDim = _shape(x, 3)
+    kernelH, kernelW = k.shape
+    assert inW >= 1 and inH >= 1
+    assert kernelW >= 1 and kernelH >= 1
+    assert isinstance(upx, int) and isinstance(upy, int)
+    assert isinstance(downx, int) and isinstance(downy, int)
+    assert isinstance(padx0, int) and isinstance(padx1, int)
+    assert isinstance(pady0, int) and isinstance(pady1, int)
+
+    # Upsample (insert zeros).
+    x = tf.reshape(x, [-1, inH, 1, inW, 1, minorDim])
+    x = tf.pad(x, [[0, 0], [0, 0], [0, upy - 1], [0, 0], [0, upx - 1], [0, 0]])
+    x = tf.reshape(x, [-1, inH * upy, inW * upx, minorDim])
+
+    # Pad (crop if negative).
+    x = tf.pad(x, [[0, 0], [max(pady0, 0), max(pady1, 0)], [max(padx0, 0), max(padx1, 0)], [0, 0]])
+    x = x[:, max(-pady0, 0) : x.shape[1].value - max(-pady1, 0), max(-padx0, 0) : x.shape[2].value - max(-padx1, 0), :]
+
+    # Convolve with filter.
+    x = tf.transpose(x, [0, 3, 1, 2])
+    x = tf.reshape(x, [-1, 1, inH * upy + pady0 + pady1, inW * upx + padx0 + padx1])
+    w = tf.constant(k[::-1, ::-1, np.newaxis, np.newaxis], dtype=x.dtype)
+    x = tf.nn.conv2d(x, w, strides=[1,1,1,1], padding='VALID', data_format='NCHW')
+    x = tf.reshape(x, [-1, minorDim, inH * upy + pady0 + pady1 - kernelH + 1, inW * upx + padx0 + padx1 - kernelW + 1])
+    x = tf.transpose(x, [0, 2, 3, 1])
+
+    # Downsample (throw away pixels).
+    return x[:, ::downy, ::downx, :]
+
+#----------------------------------------------------------------------------
+
+def _upfirdn_2d_cuda(x, k, upx, upy, downx, downy, padx0, padx1, pady0, pady1):
+    """Fast CUDA implementation of `upfirdn_2d()` using custom ops."""
+
+    x = tf.convert_to_tensor(x)
+    k = np.asarray(k, dtype=np.float32)
+    majorDim, inH, inW, minorDim = x.shape.as_list()
+    kernelH, kernelW = k.shape
+    assert inW >= 1 and inH >= 1
+    assert kernelW >= 1 and kernelH >= 1
+    assert isinstance(upx, int) and isinstance(upy, int)
+    assert isinstance(downx, int) and isinstance(downy, int)
+    assert isinstance(padx0, int) and isinstance(padx1, int)
+    assert isinstance(pady0, int) and isinstance(pady1, int)
+
+    outW = (inW * upx + padx0 + padx1 - kernelW) // downx + 1
+    outH = (inH * upy + pady0 + pady1 - kernelH) // downy + 1
+    assert outW >= 1 and outH >= 1
+
+    cuda_op = _get_plugin().up_fir_dn2d
+    kc = tf.constant(k, dtype=x.dtype)
+    gkc = tf.constant(k[::-1, ::-1], dtype=x.dtype)
+    gpadx0 = kernelW - padx0 - 1
+    gpady0 = kernelH - pady0 - 1
+    gpadx1 = inW * upx - outW * downx + padx0 - upx + 1
+    gpady1 = inH * upy - outH * downy + pady0 - upy + 1
+
+    @tf.custom_gradient
+    def func(x):
+        y = cuda_op(x=x, k=kc, upx=int(upx), upy=int(upy), downx=int(downx), downy=int(downy), padx0=int(padx0), padx1=int(padx1), pady0=int(pady0), pady1=int(pady1))
+        y.set_shape([majorDim, outH, outW, minorDim])
+        @tf.custom_gradient
+        def grad(dy):
+            dx = cuda_op(x=dy, k=gkc, upx=int(downx), upy=int(downy), downx=int(upx), downy=int(upy), padx0=int(gpadx0), padx1=int(gpadx1), pady0=int(gpady0), pady1=int(gpady1))
+            dx.set_shape([majorDim, inH, inW, minorDim])
+            return dx, func
+        return y, grad
+    return func(x)
+
+#----------------------------------------------------------------------------
+
+def filter_2d(x, k, gain=1, padding=0, data_format='NCHW', impl='cuda'):
+    r"""Filter a batch of 2D images with the given FIR filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and filters each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        padding:      Number of pixels to pad or crop the output on each side (default: 0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the same shape and datatype as `x`.
+    """
+
+    assert isinstance(padding, int)
+    k = _FilterKernel(k=k, gain=gain)
+    assert k.w == k.h
+    pad0 = k.w // 2 + padding
+    pad1 = (k.w - 1) // 2 + padding
+    return _simple_upfirdn_2d(x, k, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def upsample_2d(x, k=None, factor=2, gain=1, padding=0, data_format='NCHW', impl='cuda'):
+    r"""Upsample a batch of 2D images with the given filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and upsamples each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero, and the filter is padded with
+    zeros so that its shape is a multiple of the upsampling factor.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to nearest-neighbor
+                      upsampling.
+        factor:       Integer upsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        padding:      Number of pixels to pad or crop the output on each side (default: 0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H * factor, W * factor]` or
+        `[N, H * factor, W * factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    assert isinstance(padding, int)
+    k = _FilterKernel(k if k is not None else [1] * factor, gain * (factor ** 2))
+    assert k.w == k.h
+    pad0 = (k.w + factor - 1) // 2 + padding
+    pad1 = (k.w - factor) // 2 + padding
+    return _simple_upfirdn_2d(x, k, up=factor, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def downsample_2d(x, k=None, factor=2, gain=1, padding=0, data_format='NCHW', impl='cuda'):
+    r"""Downsample a batch of 2D images with the given filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and downsamples each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero, and the filter is padded with
+    zeros so that its shape is a multiple of the downsampling factor.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to average pooling.
+        factor:       Integer downsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        padding:      Number of pixels to pad or crop the output on each side (default: 0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H // factor, W // factor]` or
+        `[N, H // factor, W // factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    assert isinstance(padding, int)
+    k = _FilterKernel(k if k is not None else [1] * factor, gain)
+    assert k.w == k.h
+    pad0 = (k.w - factor + 1) // 2 + padding * factor
+    pad1 = (k.w - factor) // 2 + padding * factor
+    return _simple_upfirdn_2d(x, k, down=factor, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def upsample_conv_2d(x, w, k=None, factor=2, gain=1, padding=0, data_format='NCHW', impl='cuda'):
+    r"""Fused `upsample_2d()` followed by `tf.nn.conv2d()`.
+
+    Padding is performed only once at the beginning, not between the operations.
+    The fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        w:            Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`.
+                      Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to nearest-neighbor
+                      upsampling.
+        factor:       Integer upsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        padding:      Number of pixels to pad or crop the output on each side (default: 0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H * factor, W * factor]` or
+        `[N, H * factor, W * factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    assert isinstance(padding, int)
+
+    # Check weight shape.
+    w = tf.convert_to_tensor(w)
+    ch, cw, _inC, _outC = w.shape.as_list()
+    inC = _shape(w, 2)
+    outC = _shape(w, 3)
+    assert cw == ch
+
+    # Fast path for 1x1 convolution.
+    if cw == 1 and ch == 1:
+        x = tf.nn.conv2d(x, w, data_format=data_format, strides=[1,1,1,1], padding='VALID')
+        x = upsample_2d(x, k, factor=factor, gain=gain, padding=padding, data_format=data_format, impl=impl)
+        return x
+
+    # Setup filter kernel.
+    k = _FilterKernel(k if k is not None else [1] * factor, gain * (factor ** 2))
+    assert k.w == k.h
+
+    # Determine data dimensions.
+    if data_format == 'NCHW':
+        stride = [1, 1, factor, factor]
+        output_shape = [_shape(x, 0), outC, (_shape(x, 2) - 1) * factor + ch, (_shape(x, 3) - 1) * factor + cw]
+        num_groups = _shape(x, 1) // inC
+    else:
+        stride = [1, factor, factor, 1]
+        output_shape = [_shape(x, 0), (_shape(x, 1) - 1) * factor + ch, (_shape(x, 2) - 1) * factor + cw, outC]
+        num_groups = _shape(x, 3) // inC
+
+    # Transpose weights.
+    w = tf.reshape(w, [ch, cw, inC, num_groups, -1])
+    w = tf.transpose(w[::-1, ::-1], [0, 1, 4, 3, 2])
+    w = tf.reshape(w, [ch, cw, -1, num_groups * inC])
+
+    # Execute.
+    x = tf.nn.conv2d_transpose(x, w, output_shape=output_shape, strides=stride, padding='VALID', data_format=data_format)
+    pad0 = (k.w + factor - cw) // 2 + padding
+    pad1 = (k.w - factor - cw + 3) // 2 + padding
+    return _simple_upfirdn_2d(x, k, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def conv_downsample_2d(x, w, k=None, factor=2, gain=1, padding=0, data_format='NCHW', impl='cuda'):
+    r"""Fused `tf.nn.conv2d()` followed by `downsample_2d()`.
+
+    Padding is performed only once at the beginning, not between the operations.
+    The fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        w:            Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`.
+                      Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to average pooling.
+        factor:       Integer downsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        padding:      Number of pixels to pad or crop the output on each side (default: 0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H // factor, W // factor]` or
+        `[N, H // factor, W // factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    assert isinstance(padding, int)
+
+    # Check weight shape.
+    w = tf.convert_to_tensor(w)
+    ch, cw, _inC, _outC = w.shape.as_list()
+    assert cw == ch
+
+    # Fast path for 1x1 convolution.
+    if cw == 1 and ch == 1:
+        x = downsample_2d(x, k, factor=factor, gain=gain, padding=padding, data_format=data_format, impl=impl)
+        x = tf.nn.conv2d(x, w, data_format=data_format, strides=[1,1,1,1], padding='VALID')
+        return x
+
+    # Setup filter kernel.
+    k = _FilterKernel(k if k is not None else [1] * factor, gain)
+    assert k.w == k.h
+
+    # Determine stride.
+    if data_format == 'NCHW':
+        s = [1, 1, factor, factor]
+    else:
+        s = [1, factor, factor, 1]
+
+    # Execute.
+    pad0 = (k.w - factor + cw) // 2 + padding * factor
+    pad1 = (k.w - factor + cw - 1) // 2 + padding * factor
+    x = _simple_upfirdn_2d(x, k, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
+    return tf.nn.conv2d(x, w, strides=s, padding='VALID', data_format=data_format)
+
+#----------------------------------------------------------------------------
+# Internal helpers.
+
+class _FilterKernel:
+    def __init__(self, k, gain=1):
+        k = np.asarray(k, dtype=np.float32)
+        k /= np.sum(k)
+
+        # Separable.
+        if k.ndim == 1 and k.size >= 8:
+            self.w = k.size
+            self.h = k.size
+            self.kx = k[np.newaxis, :]
+            self.ky = k[:, np.newaxis] * gain
+            self.kxy = None
+
+        # Non-separable.
+        else:
+            if k.ndim == 1:
+                k = np.outer(k, k)
+            assert k.ndim == 2
+            self.w = k.shape[1]
+            self.h = k.shape[0]
+            self.kx = None
+            self.ky = None
+            self.kxy = k * gain
+
+def _simple_upfirdn_2d(x, k, up=1, down=1, pad0=0, pad1=0, data_format='NCHW', impl='cuda'):
+    assert isinstance(k, _FilterKernel)
+    assert data_format in ['NCHW', 'NHWC']
+    assert x.shape.rank == 4
+    y = x
+    if data_format == 'NCHW':
+        y = tf.reshape(y, [-1, _shape(y, 2), _shape(y, 3), 1])
+    if k.kx is not None:
+        y = upfirdn_2d(y, k.kx, upx=up, downx=down, padx0=pad0, padx1=pad1, impl=impl)
+    if k.ky is not None:
+        y = upfirdn_2d(y, k.ky, upy=up, downy=down, pady0=pad0, pady1=pad1, impl=impl)
+    if k.kxy is not None:
+        y = upfirdn_2d(y, k.kxy, upx=up, upy=up, downx=down, downy=down, padx0=pad0, padx1=pad1, pady0=pad0, pady1=pad1, impl=impl)
+    if data_format == 'NCHW':
+        y = tf.reshape(y, [-1, _shape(x, 1), _shape(y, 1), _shape(y, 2)])
+    return y
+
+def _shape(tf_expr, dim_idx):
+    if tf_expr.shape.rank is not None:
+        dim = tf_expr.shape[dim_idx].value
+        if dim is not None:
+            return dim
+    return tf.shape(tf_expr)[dim_idx]
+
+#----------------------------------------------------------------------------

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/optimizer.py

@@ -0,0 +1,372 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Helper wrapper for a Tensorflow optimizer."""
+
+import platform
+import numpy as np
+import tensorflow as tf
+
+from collections import OrderedDict
+from typing import List, Union
+
+from . import autosummary
+from . import tfutil
+from .. import util
+
+from .tfutil import TfExpression, TfExpressionEx
+
+_collective_ops_warning_printed = False
+_collective_ops_group_key       = 831766147
+_collective_ops_instance_key    = 436340067
+
+class Optimizer:
+    """A Wrapper for tf.train.Optimizer.
+
+    Automatically takes care of:
+    - Gradient averaging for multi-GPU training.
+    - Gradient accumulation for arbitrarily large minibatches.
+    - Dynamic loss scaling and typecasts for FP16 training.
+    - Ignoring corrupted gradients that contain NaNs/Infs.
+    - Reporting statistics.
+    - Well-chosen default settings.
+    """
+
+    def __init__(self,
+        name:                   str             = "Train",                  # Name string that will appear in TensorFlow graph.
+        tf_optimizer:           str             = "tf.train.AdamOptimizer", # Underlying optimizer class.
+        learning_rate:          TfExpressionEx  = 0.001,                    # Learning rate. Can vary over time.
+        minibatch_multiplier:   TfExpressionEx  = None,                     # Treat N consecutive minibatches as one by accumulating gradients.
+        share:                  "Optimizer"     = None,                     # Share internal state with a previously created optimizer?
+        use_loss_scaling:       bool            = False,                    # Enable dynamic loss scaling for robust mixed-precision training?
+        loss_scaling_init:      float           = 64.0,                     # Log2 of initial loss scaling factor.
+        loss_scaling_inc:       float           = 0.0005,                   # Log2 of per-minibatch loss scaling increment when there is no overflow.
+        loss_scaling_dec:       float           = 1.0,                      # Log2 of per-minibatch loss scaling decrement when there is an overflow.
+        report_mem_usage:       bool            = False,                    # Report fine-grained memory usage statistics in TensorBoard?
+        **kwargs):
+
+        # Public fields.
+        self.name                   = name
+        self.learning_rate          = learning_rate
+        self.minibatch_multiplier   = minibatch_multiplier
+        self.id                     = self.name.replace("/", ".")
+        self.scope                  = tf.get_default_graph().unique_name(self.id)
+        self.optimizer_class        = util.get_obj_by_name(tf_optimizer)
+        self.optimizer_kwargs       = dict(kwargs)
+        self.use_loss_scaling       = use_loss_scaling
+        self.loss_scaling_init      = loss_scaling_init
+        self.loss_scaling_inc       = loss_scaling_inc
+        self.loss_scaling_dec       = loss_scaling_dec
+
+        # Private fields.
+        self._updates_applied       = False
+        self._devices               = OrderedDict() # device_name => EasyDict()
+        self._shared_optimizers     = OrderedDict() # device_name => optimizer_class
+        self._gradient_shapes       = None          # [shape, ...]
+        self._report_mem_usage      = report_mem_usage
+
+        # Validate arguments.
+        assert callable(self.optimizer_class)
+
+        # Share internal state if requested.
+        if share is not None:
+            assert isinstance(share, Optimizer)
+            assert self.optimizer_class is share.optimizer_class
+            assert self.learning_rate is share.learning_rate
+            assert self.optimizer_kwargs == share.optimizer_kwargs
+            self._shared_optimizers = share._shared_optimizers # pylint: disable=protected-access
+
+    def _get_device(self, device_name: str):
+        """Get internal state for the given TensorFlow device."""
+        tfutil.assert_tf_initialized()
+        if device_name in self._devices:
+            return self._devices[device_name]
+
+        # Initialize fields.
+        device = util.EasyDict()
+        device.name             = device_name
+        device.optimizer        = None          # Underlying optimizer:     optimizer_class
+        device.loss_scaling_var = None          # Log2 of loss scaling:     tf.Variable
+        device.grad_raw         = OrderedDict() # Raw gradients:            var => [grad, ...]
+        device.grad_clean       = OrderedDict() # Clean gradients:          var => grad
+        device.grad_acc_vars    = OrderedDict() # Accumulation sums:        var => tf.Variable
+        device.grad_acc_count   = None          # Accumulation counter:     tf.Variable
+        device.grad_acc         = OrderedDict() # Accumulated gradients:    var => grad
+
+        # Setup TensorFlow objects.
+        with tfutil.absolute_name_scope(self.scope + "/Devices"), tf.device(device_name), tf.control_dependencies(None):
+            if device_name not in self._shared_optimizers:
+                optimizer_name = self.scope.replace("/", "_") + "_opt%d" % len(self._shared_optimizers)
+                self._shared_optimizers[device_name] = self.optimizer_class(name=optimizer_name, learning_rate=self.learning_rate, **self.optimizer_kwargs)
+            device.optimizer = self._shared_optimizers[device_name]
+            if self.use_loss_scaling:
+                device.loss_scaling_var = tf.Variable(np.float32(self.loss_scaling_init), trainable=False, name="loss_scaling_var")
+
+        # Register device.
+        self._devices[device_name] = device
+        return device
+
+    def register_gradients(self, loss: TfExpression, trainable_vars: Union[List, dict]) -> None:
+        """Register the gradients of the given loss function with respect to the given variables.
+        Intended to be called once per GPU."""
+        tfutil.assert_tf_initialized()
+        assert not self._updates_applied
+        device = self._get_device(loss.device)
+
+        # Validate trainables.
+        if isinstance(trainable_vars, dict):
+            trainable_vars = list(trainable_vars.values())  # allow passing in Network.trainables as vars
+        assert isinstance(trainable_vars, list) and len(trainable_vars) >= 1
+        assert all(tfutil.is_tf_expression(expr) for expr in trainable_vars + [loss])
+        assert all(var.device == device.name for var in trainable_vars)
+
+        # Validate shapes.
+        if self._gradient_shapes is None:
+            self._gradient_shapes = [var.shape.as_list() for var in trainable_vars]
+        assert len(trainable_vars) == len(self._gradient_shapes)
+        assert all(var.shape.as_list() == var_shape for var, var_shape in zip(trainable_vars, self._gradient_shapes))
+
+        # Report memory usage if requested.
+        deps = [loss]
+        if self._report_mem_usage:
+            self._report_mem_usage = False
+            try:
+                with tf.name_scope(self.id + '_mem'), tf.device(device.name), tf.control_dependencies([loss]):
+                    deps.append(autosummary.autosummary(self.id + "/mem_usage_gb", tf.contrib.memory_stats.BytesInUse() / 2**30))
+            except tf.errors.NotFoundError:
+                pass
+
+        # Compute gradients.
+        with tf.name_scope(self.id + "_grad"), tf.device(device.name), tf.control_dependencies(deps):
+            loss = self.apply_loss_scaling(tf.cast(loss, tf.float32))
+            gate = tf.train.Optimizer.GATE_NONE  # disable gating to reduce memory usage
+            grad_list = device.optimizer.compute_gradients(loss=loss, var_list=trainable_vars, gate_gradients=gate)
+
+        # Register gradients.
+        for grad, var in grad_list:
+            if var not in device.grad_raw:
+                device.grad_raw[var] = []
+            device.grad_raw[var].append(grad)
+
+    def apply_updates(self, allow_no_op: bool = False) -> tf.Operation:
+        """Construct training op to update the registered variables based on their gradients."""
+        tfutil.assert_tf_initialized()
+        assert not self._updates_applied
+        self._updates_applied = True
+        all_ops = []
+
+        # Check for no-op.
+        if allow_no_op and len(self._devices) == 0:
+            with tfutil.absolute_name_scope(self.scope):
+                return tf.no_op(name='TrainingOp')
+
+        # Clean up gradients.
+        for device_idx, device in enumerate(self._devices.values()):
+            with tfutil.absolute_name_scope(self.scope + "/Clean%d" % device_idx), tf.device(device.name):
+                for var, grad in device.grad_raw.items():
+
+                    # Filter out disconnected gradients and convert to float32.
+                    grad = [g for g in grad if g is not None]
+                    grad = [tf.cast(g, tf.float32) for g in grad]
+
+                    # Sum within the device.
+                    if len(grad) == 0:
+                        grad = tf.zeros(var.shape)  # No gradients => zero.
+                    elif len(grad) == 1:
+                        grad = grad[0]              # Single gradient => use as is.
+                    else:
+                        grad = tf.add_n(grad)       # Multiple gradients => sum.
+
+                    # Scale as needed.
+                    scale = 1.0 / len(device.grad_raw[var]) / len(self._devices)
+                    scale = tf.constant(scale, dtype=tf.float32, name="scale")
+                    if self.minibatch_multiplier is not None:
+                        scale /= tf.cast(self.minibatch_multiplier, tf.float32)
+                    scale = self.undo_loss_scaling(scale)
+                    device.grad_clean[var] = grad * scale
+
+        # Sum gradients across devices.
+        if len(self._devices) > 1:
+            with tfutil.absolute_name_scope(self.scope + "/Broadcast"), tf.device(None):
+                if platform.system() == "Windows":    # Windows => NCCL ops are not available.
+                    self._broadcast_fallback()
+                elif tf.VERSION.startswith("1.15."):  # TF 1.15 => NCCL ops are broken: https://github.com/tensorflow/tensorflow/issues/41539
+                    self._broadcast_fallback()
+                else:                                 # Otherwise => NCCL ops are safe to use.
+                    self._broadcast_nccl()
+
+        # Apply updates separately on each device.
+        for device_idx, device in enumerate(self._devices.values()):
+            with tfutil.absolute_name_scope(self.scope + "/Apply%d" % device_idx), tf.device(device.name):
+                # pylint: disable=cell-var-from-loop
+
+                # Accumulate gradients over time.
+                if self.minibatch_multiplier is None:
+                    acc_ok = tf.constant(True, name='acc_ok')
+                    device.grad_acc = OrderedDict(device.grad_clean)
+                else:
+                    # Create variables.
+                    with tf.control_dependencies(None):
+                        for var in device.grad_clean.keys():
+                            device.grad_acc_vars[var] = tf.Variable(tf.zeros(var.shape), trainable=False, name="grad_acc_var")
+                        device.grad_acc_count = tf.Variable(tf.zeros([]), trainable=False, name="grad_acc_count")
+
+                    # Track counter.
+                    count_cur = device.grad_acc_count + 1.0
+                    count_inc_op = lambda: tf.assign(device.grad_acc_count, count_cur)
+                    count_reset_op = lambda: tf.assign(device.grad_acc_count, tf.zeros([]))
+                    acc_ok = (count_cur >= tf.cast(self.minibatch_multiplier, tf.float32))
+                    all_ops.append(tf.cond(acc_ok, count_reset_op, count_inc_op))
+
+                    # Track gradients.
+                    for var, grad in device.grad_clean.items():
+                        acc_var = device.grad_acc_vars[var]
+                        acc_cur = acc_var + grad
+                        device.grad_acc[var] = acc_cur
+                        with tf.control_dependencies([acc_cur]):
+                            acc_inc_op = lambda: tf.assign(acc_var, acc_cur)
+                            acc_reset_op = lambda: tf.assign(acc_var, tf.zeros(var.shape))
+                            all_ops.append(tf.cond(acc_ok, acc_reset_op, acc_inc_op))
+
+                # No overflow => apply gradients.
+                all_ok = tf.reduce_all(tf.stack([acc_ok] + [tf.reduce_all(tf.is_finite(g)) for g in device.grad_acc.values()]))
+                apply_op = lambda: device.optimizer.apply_gradients([(tf.cast(grad, var.dtype), var) for var, grad in device.grad_acc.items()])
+                all_ops.append(tf.cond(all_ok, apply_op, tf.no_op))
+
+                # Adjust loss scaling.
+                if self.use_loss_scaling:
+                    ls_inc_op = lambda: tf.assign_add(device.loss_scaling_var, self.loss_scaling_inc)
+                    ls_dec_op = lambda: tf.assign_sub(device.loss_scaling_var, self.loss_scaling_dec)
+                    ls_update_op = lambda: tf.group(tf.cond(all_ok, ls_inc_op, ls_dec_op))
+                    all_ops.append(tf.cond(acc_ok, ls_update_op, tf.no_op))
+
+                # Last device => report statistics.
+                if device_idx == len(self._devices) - 1:
+                    all_ops.append(autosummary.autosummary(self.id + "/learning_rate", tf.convert_to_tensor(self.learning_rate)))
+                    all_ops.append(autosummary.autosummary(self.id + "/overflow_frequency", tf.where(all_ok, 0, 1), condition=acc_ok))
+                    if self.use_loss_scaling:
+                        all_ops.append(autosummary.autosummary(self.id + "/loss_scaling_log2", device.loss_scaling_var))
+
+        # Initialize variables.
+        self.reset_optimizer_state()
+        if self.use_loss_scaling:
+            tfutil.init_uninitialized_vars([device.loss_scaling_var for device in self._devices.values()])
+        if self.minibatch_multiplier is not None:
+            tfutil.run([var.initializer for device in self._devices.values() for var in list(device.grad_acc_vars.values()) + [device.grad_acc_count]])
+
+        # Group everything into a single op.
+        with tfutil.absolute_name_scope(self.scope):
+            return tf.group(*all_ops, name="TrainingOp")
+
+    def reset_optimizer_state(self) -> None:
+        """Reset internal state of the underlying optimizer."""
+        tfutil.assert_tf_initialized()
+        tfutil.run([var.initializer for device in self._devices.values() for var in device.optimizer.variables()])
+
+    def get_loss_scaling_var(self, device: str) -> Union[tf.Variable, None]:
+        """Get or create variable representing log2 of the current dynamic loss scaling factor."""
+        return self._get_device(device).loss_scaling_var
+
+    def apply_loss_scaling(self, value: TfExpression) -> TfExpression:
+        """Apply dynamic loss scaling for the given expression."""
+        assert tfutil.is_tf_expression(value)
+        if not self.use_loss_scaling:
+            return value
+        return value * tfutil.exp2(self.get_loss_scaling_var(value.device))
+
+    def undo_loss_scaling(self, value: TfExpression) -> TfExpression:
+        """Undo the effect of dynamic loss scaling for the given expression."""
+        assert tfutil.is_tf_expression(value)
+        if not self.use_loss_scaling:
+            return value
+        return value * tfutil.exp2(-self.get_loss_scaling_var(value.device)) # pylint: disable=invalid-unary-operand-type
+
+    def _broadcast_nccl(self):
+        """Sum gradients across devices using NCCL ops (fast path)."""
+        from tensorflow.python.ops import nccl_ops # pylint: disable=no-name-in-module
+        for all_vars in zip(*[device.grad_clean.keys() for device in self._devices.values()]):
+            if any(x.shape.num_elements() > 0 for x in all_vars):
+                all_grads = [device.grad_clean[var] for device, var in zip(self._devices.values(), all_vars)]
+                all_grads = nccl_ops.all_sum(all_grads)
+                for device, var, grad in zip(self._devices.values(), all_vars, all_grads):
+                    device.grad_clean[var] = grad
+
+    def _broadcast_fallback(self):
+        """Sum gradients across devices using TensorFlow collective ops (slow fallback path)."""
+        from tensorflow.python.ops import collective_ops # pylint: disable=no-name-in-module
+        global _collective_ops_warning_printed, _collective_ops_group_key, _collective_ops_instance_key
+        if all(x.shape.num_elements() == 0 for device in self._devices.values() for x in device.grad_clean.values()):
+            return
+        if not _collective_ops_warning_printed:
+            print("------------------------------------------------------------------------")
+            print("WARNING: Using slow fallback implementation for inter-GPU communication.")
+            print("Please use TensorFlow 1.14 on Linux for optimal training performance.")
+            print("------------------------------------------------------------------------")
+            _collective_ops_warning_printed = True
+        for device in self._devices.values():
+            with tf.device(device.name):
+                combo = [tf.reshape(x, [x.shape.num_elements()]) for x in device.grad_clean.values()]
+                combo = tf.concat(combo, axis=0)
+                combo = collective_ops.all_reduce(combo, merge_op='Add', final_op='Id',
+                    group_size=len(self._devices), group_key=_collective_ops_group_key,
+                    instance_key=_collective_ops_instance_key)
+                cur_ofs = 0
+                for var, grad_old in device.grad_clean.items():
+                    grad_new = tf.reshape(combo[cur_ofs : cur_ofs + grad_old.shape.num_elements()], grad_old.shape)
+                    cur_ofs += grad_old.shape.num_elements()
+                    device.grad_clean[var] = grad_new
+        _collective_ops_instance_key += 1
+
+
+class SimpleAdam:
+    """Simplified version of tf.train.AdamOptimizer that behaves identically when used with dnnlib.tflib.Optimizer."""
+
+    def __init__(self, name="Adam", learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-8):
+        self.name = name
+        self.learning_rate = learning_rate
+        self.beta1 = beta1
+        self.beta2 = beta2
+        self.epsilon = epsilon
+        self.all_state_vars = []
+
+    def variables(self):
+        return self.all_state_vars
+
+    def compute_gradients(self, loss, var_list, gate_gradients=tf.train.Optimizer.GATE_NONE):
+        assert gate_gradients == tf.train.Optimizer.GATE_NONE
+        return list(zip(tf.gradients(loss, var_list), var_list))
+
+    def apply_gradients(self, grads_and_vars):
+        with tf.name_scope(self.name):
+            state_vars = []
+            update_ops = []
+
+            # Adjust learning rate to deal with startup bias.
+            with tf.control_dependencies(None):
+                b1pow_var = tf.Variable(dtype=tf.float32, initial_value=1, trainable=False)
+                b2pow_var = tf.Variable(dtype=tf.float32, initial_value=1, trainable=False)
+                state_vars += [b1pow_var, b2pow_var]
+            b1pow_new = b1pow_var * self.beta1
+            b2pow_new = b2pow_var * self.beta2
+            update_ops += [tf.assign(b1pow_var, b1pow_new), tf.assign(b2pow_var, b2pow_new)]
+            lr_new = self.learning_rate * tf.sqrt(1 - b2pow_new) / (1 - b1pow_new)
+
+            # Construct ops to update each variable.
+            for grad, var in grads_and_vars:
+                with tf.control_dependencies(None):
+                    m_var = tf.Variable(dtype=tf.float32, initial_value=tf.zeros_like(var), trainable=False)
+                    v_var = tf.Variable(dtype=tf.float32, initial_value=tf.zeros_like(var), trainable=False)
+                    state_vars += [m_var, v_var]
+                m_new = self.beta1 * m_var + (1 - self.beta1) * grad
+                v_new = self.beta2 * v_var + (1 - self.beta2) * tf.square(grad)
+                var_delta = lr_new * m_new / (tf.sqrt(v_new) + self.epsilon)
+                update_ops += [tf.assign(m_var, m_new), tf.assign(v_var, v_new), tf.assign_sub(var, var_delta)]
+
+            # Group everything together.
+            self.all_state_vars += state_vars
+            return tf.group(*update_ops)

PTI/models/StyleCLIP/global_directions/dnnlib/tflib/tfutil.py

@@ -0,0 +1,262 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Miscellaneous helper utils for Tensorflow."""
+
+import os
+import numpy as np
+import tensorflow as tf
+
+# Silence deprecation warnings from TensorFlow 1.13 onwards
+import logging
+logging.getLogger('tensorflow').setLevel(logging.ERROR)
+import tensorflow.contrib   # requires TensorFlow 1.x!
+tf.contrib = tensorflow.contrib
+
+from typing import Any, Iterable, List, Union
+
+TfExpression = Union[tf.Tensor, tf.Variable, tf.Operation]
+"""A type that represents a valid Tensorflow expression."""
+
+TfExpressionEx = Union[TfExpression, int, float, np.ndarray]
+"""A type that can be converted to a valid Tensorflow expression."""
+
+
+def run(*args, **kwargs) -> Any:
+    """Run the specified ops in the default session."""
+    assert_tf_initialized()
+    return tf.get_default_session().run(*args, **kwargs)
+
+
+def is_tf_expression(x: Any) -> bool:
+    """Check whether the input is a valid Tensorflow expression, i.e., Tensorflow Tensor, Variable, or Operation."""
+    return isinstance(x, (tf.Tensor, tf.Variable, tf.Operation))
+
+
+def shape_to_list(shape: Iterable[tf.Dimension]) -> List[Union[int, None]]:
+    """Convert a Tensorflow shape to a list of ints. Retained for backwards compatibility -- use TensorShape.as_list() in new code."""
+    return [dim.value for dim in shape]
+
+
+def flatten(x: TfExpressionEx) -> TfExpression:
+    """Shortcut function for flattening a tensor."""
+    with tf.name_scope("Flatten"):
+        return tf.reshape(x, [-1])
+
+
+def log2(x: TfExpressionEx) -> TfExpression:
+    """Logarithm in base 2."""
+    with tf.name_scope("Log2"):
+        return tf.log(x) * np.float32(1.0 / np.log(2.0))
+
+
+def exp2(x: TfExpressionEx) -> TfExpression:
+    """Exponent in base 2."""
+    with tf.name_scope("Exp2"):
+        return tf.exp(x * np.float32(np.log(2.0)))
+
+
+def erfinv(y: TfExpressionEx) -> TfExpression:
+    """Inverse of the error function."""
+    # pylint: disable=no-name-in-module
+    from tensorflow.python.ops.distributions import special_math
+    return special_math.erfinv(y)
+
+
+def lerp(a: TfExpressionEx, b: TfExpressionEx, t: TfExpressionEx) -> TfExpressionEx:
+    """Linear interpolation."""
+    with tf.name_scope("Lerp"):
+        return a + (b - a) * t
+
+
+def lerp_clip(a: TfExpressionEx, b: TfExpressionEx, t: TfExpressionEx) -> TfExpression:
+    """Linear interpolation with clip."""
+    with tf.name_scope("LerpClip"):
+        return a + (b - a) * tf.clip_by_value(t, 0.0, 1.0)
+
+
+def absolute_name_scope(scope: str) -> tf.name_scope:
+    """Forcefully enter the specified name scope, ignoring any surrounding scopes."""
+    return tf.name_scope(scope + "/")
+
+
+def absolute_variable_scope(scope: str, **kwargs) -> tf.variable_scope:
+    """Forcefully enter the specified variable scope, ignoring any surrounding scopes."""
+    return tf.variable_scope(tf.VariableScope(name=scope, **kwargs), auxiliary_name_scope=False)
+
+
+def _sanitize_tf_config(config_dict: dict = None) -> dict:
+    # Defaults.
+    cfg = dict()
+    cfg["rnd.np_random_seed"]               = None      # Random seed for NumPy. None = keep as is.
+    cfg["rnd.tf_random_seed"]               = "auto"    # Random seed for TensorFlow. 'auto' = derive from NumPy random state. None = keep as is.
+    cfg["env.TF_CPP_MIN_LOG_LEVEL"]         = "1"       # 0 = Print all available debug info from TensorFlow. 1 = Print warnings and errors, but disable debug info.
+    cfg["env.HDF5_USE_FILE_LOCKING"]        = "FALSE"   # Disable HDF5 file locking to avoid concurrency issues with network shares.
+    cfg["graph_options.place_pruned_graph"] = True      # False = Check that all ops are available on the designated device. True = Skip the check for ops that are not used.
+    cfg["gpu_options.allow_growth"]         = True      # False = Allocate all GPU memory at the beginning. True = Allocate only as much GPU memory as needed.
+
+    # Remove defaults for environment variables that are already set.
+    for key in list(cfg):
+        fields = key.split(".")
+        if fields[0] == "env":
+            assert len(fields) == 2
+            if fields[1] in os.environ:
+                del cfg[key]
+
+    # User overrides.
+    if config_dict is not None:
+        cfg.update(config_dict)
+    return cfg
+
+
+def init_tf(config_dict: dict = None) -> None:
+    """Initialize TensorFlow session using good default settings."""
+    # Skip if already initialized.
+    if tf.get_default_session() is not None:
+        return
+
+    # Setup config dict and random seeds.
+    cfg = _sanitize_tf_config(config_dict)
+    np_random_seed = cfg["rnd.np_random_seed"]
+    if np_random_seed is not None:
+        np.random.seed(np_random_seed)
+    tf_random_seed = cfg["rnd.tf_random_seed"]
+    if tf_random_seed == "auto":
+        tf_random_seed = np.random.randint(1 << 31)
+    if tf_random_seed is not None:
+        tf.set_random_seed(tf_random_seed)
+
+    # Setup environment variables.
+    for key, value in cfg.items():
+        fields = key.split(".")
+        if fields[0] == "env":
+            assert len(fields) == 2
+            os.environ[fields[1]] = str(value)
+
+    # Create default TensorFlow session.
+    create_session(cfg, force_as_default=True)
+
+
+def assert_tf_initialized():
+    """Check that TensorFlow session has been initialized."""
+    if tf.get_default_session() is None:
+        raise RuntimeError("No default TensorFlow session found. Please call dnnlib.tflib.init_tf().")
+
+
+def create_session(config_dict: dict = None, force_as_default: bool = False) -> tf.Session:
+    """Create tf.Session based on config dict."""
+    # Setup TensorFlow config proto.
+    cfg = _sanitize_tf_config(config_dict)
+    config_proto = tf.ConfigProto()
+    for key, value in cfg.items():
+        fields = key.split(".")
+        if fields[0] not in ["rnd", "env"]:
+            obj = config_proto
+            for field in fields[:-1]:
+                obj = getattr(obj, field)
+            setattr(obj, fields[-1], value)
+
+    # Create session.
+    session = tf.Session(config=config_proto)
+    if force_as_default:
+        # pylint: disable=protected-access
+        session._default_session = session.as_default()
+        session._default_session.enforce_nesting = False
+        session._default_session.__enter__()
+    return session
+
+
+def init_uninitialized_vars(target_vars: List[tf.Variable] = None) -> None:
+    """Initialize all tf.Variables that have not already been initialized.
+
+    Equivalent to the following, but more efficient and does not bloat the tf graph:
+    tf.variables_initializer(tf.report_uninitialized_variables()).run()
+    """
+    assert_tf_initialized()
+    if target_vars is None:
+        target_vars = tf.global_variables()
+
+    test_vars = []
+    test_ops = []
+
+    with tf.control_dependencies(None):  # ignore surrounding control_dependencies
+        for var in target_vars:
+            assert is_tf_expression(var)
+
+            try:
+                tf.get_default_graph().get_tensor_by_name(var.name.replace(":0", "/IsVariableInitialized:0"))
+            except KeyError:
+                # Op does not exist => variable may be uninitialized.
+                test_vars.append(var)
+
+                with absolute_name_scope(var.name.split(":")[0]):
+                    test_ops.append(tf.is_variable_initialized(var))
+
+    init_vars = [var for var, inited in zip(test_vars, run(test_ops)) if not inited]
+    run([var.initializer for var in init_vars])
+
+
+def set_vars(var_to_value_dict: dict) -> None:
+    """Set the values of given tf.Variables.
+
+    Equivalent to the following, but more efficient and does not bloat the tf graph:
+    tflib.run([tf.assign(var, value) for var, value in var_to_value_dict.items()]
+    """
+    assert_tf_initialized()
+    ops = []
+    feed_dict = {}
+
+    for var, value in var_to_value_dict.items():
+        assert is_tf_expression(var)
+
+        try:
+            setter = tf.get_default_graph().get_tensor_by_name(var.name.replace(":0", "/setter:0"))  # look for existing op
+        except KeyError:
+            with absolute_name_scope(var.name.split(":")[0]):
+                with tf.control_dependencies(None):  # ignore surrounding control_dependencies
+                    setter = tf.assign(var, tf.placeholder(var.dtype, var.shape, "new_value"), name="setter")  # create new setter
+
+        ops.append(setter)
+        feed_dict[setter.op.inputs[1]] = value
+
+    run(ops, feed_dict)
+
+
+def create_var_with_large_initial_value(initial_value: np.ndarray, *args, **kwargs):
+    """Create tf.Variable with large initial value without bloating the tf graph."""
+    assert_tf_initialized()
+    assert isinstance(initial_value, np.ndarray)
+    zeros = tf.zeros(initial_value.shape, initial_value.dtype)
+    var = tf.Variable(zeros, *args, **kwargs)
+    set_vars({var: initial_value})
+    return var
+
+
+def convert_images_from_uint8(images, drange=[-1,1], nhwc_to_nchw=False):
+    """Convert a minibatch of images from uint8 to float32 with configurable dynamic range.
+    Can be used as an input transformation for Network.run().
+    """
+    images = tf.cast(images, tf.float32)
+    if nhwc_to_nchw:
+        images = tf.transpose(images, [0, 3, 1, 2])
+    return images * ((drange[1] - drange[0]) / 255) + drange[0]
+
+
+def convert_images_to_uint8(images, drange=[-1,1], nchw_to_nhwc=False, shrink=1):
+    """Convert a minibatch of images from float32 to uint8 with configurable dynamic range.
+    Can be used as an output transformation for Network.run().
+    """
+    images = tf.cast(images, tf.float32)
+    if shrink > 1:
+        ksize = [1, 1, shrink, shrink]
+        images = tf.nn.avg_pool(images, ksize=ksize, strides=ksize, padding="VALID", data_format="NCHW")
+    if nchw_to_nhwc:
+        images = tf.transpose(images, [0, 2, 3, 1])
+    scale = 255 / (drange[1] - drange[0])
+    images = images * scale + (0.5 - drange[0] * scale)
+    return tf.saturate_cast(images, tf.uint8)