encode

Time_TravelRephotography/models/encoder4editing/.gitignore

@@ -0,0 +1,133 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# Custom dataset
+pretrained_models
+results_test

Time_TravelRephotography/models/encoder4editing/LICENSE

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

Time_TravelRephotography/models/encoder4editing/README.md

@@ -0,0 +1,143 @@
+# Designing an Encoder for StyleGAN Image Manipulation (SIGGRAPH 2021)
+  <a href="https://arxiv.org/abs/2102.02766"><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>
+  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](http://colab.research.google.com/github/omertov/encoder4editing/blob/main/notebooks/inference_playground.ipynb)
+
+> Recently, there has been a surge of diverse methods for performing image editing by employing pre-trained unconditional generators. Applying these methods on real images, however, remains a challenge, as it necessarily requires the inversion of the images into their latent space. To successfully invert a real image, one needs to find a latent code that reconstructs the input image accurately, and more importantly, allows for its meaningful manipulation. In this paper, we carefully study the latent space of StyleGAN, the state-of-the-art unconditional generator. We identify and analyze the existence of a distortion-editability tradeoff and a distortion-perception tradeoff within the StyleGAN latent space. We then suggest two principles for designing encoders in a manner that allows one to control the proximity of the inversions to regions that StyleGAN was originally trained on. We present an encoder based on our two principles that is specifically designed for facilitating editing on real images by balancing these tradeoffs. By evaluating its performance qualitatively and quantitatively on numerous challenging domains, including cars and horses, we show that our inversion method, followed by common editing techniques, achieves superior real-image editing quality, with only a small reconstruction accuracy drop.
+
+<p align="center">
+<img src="docs/teaser.jpg" width="800px"/>
+</p>
+
+## Description   
+Official Implementation of "<a href="https://arxiv.org/abs/2102.02766">Designing an Encoder for StyleGAN Image Manipulation</a>" paper for both training and evaluation. 
+The e4e encoder is specifically designed to complement existing image manipulation techniques performed over StyleGAN's latent space.
+
+## Recent Updates
+`2021.08.17`: Add single style code encoder (use `--encoder_type SingleStyleCodeEncoder`). <br />
+`2021.03.25`: Add pose editing direction.
+
+## Getting Started
+### Prerequisites
+- Linux or macOS
+- NVIDIA GPU + CUDA CuDNN (CPU may be possible with some modifications, but is not inherently supported)
+- Python 3
+
+### Installation
+- Clone the repository:
+``` 
+git clone https://github.com/omertov/encoder4editing.git
+cd encoder4editing
+```
+- Dependencies:  
+We recommend running this repository using [Anaconda](https://docs.anaconda.com/anaconda/install/). 
+All dependencies for defining the environment are provided in `environment/e4e_env.yaml`.
+
+### Inference Notebook
+We provide a Jupyter notebook found in `notebooks/inference_playground.ipynb` that allows one to encode and perform several editings on real images using StyleGAN.   
+
+### Pretrained Models
+Please download the pre-trained models from the following links. Each e4e model contains the entire pSp framework architecture, including the encoder and decoder weights.
+| Path | Description
+| :--- | :----------
+|[FFHQ Inversion](https://drive.google.com/file/d/1cUv_reLE6k3604or78EranS7XzuVMWeO/view?usp=sharing)  | FFHQ e4e encoder.
+|[Cars Inversion](https://drive.google.com/file/d/17faPqBce2m1AQeLCLHUVXaDfxMRU2QcV/view?usp=sharing)  | Cars e4e encoder.
+|[Horse Inversion](https://drive.google.com/file/d/1TkLLnuX86B_BMo2ocYD0kX9kWh53rUVX/view?usp=sharing)  | Horse e4e encoder.
+|[Church Inversion](https://drive.google.com/file/d/1-L0ZdnQLwtdy6-A_Ccgq5uNJGTqE7qBa/view?usp=sharing) | Church e4e encoder.
+
+If you wish to use one of the pretrained models for training or inference, you may do so using the flag `--checkpoint_path`.
+
+In addition, we provide various auxiliary models needed for training your own e4e model from scratch.
+| Path | Description
+| :--- | :----------
+|[FFHQ StyleGAN](https://drive.google.com/file/d/1EM87UquaoQmk17Q8d5kYIAHqu0dkYqdT/view?usp=sharing) | StyleGAN model pretrained on FFHQ taken from [rosinality](https://github.com/rosinality/stylegan2-pytorch) with 1024x1024 output resolution.
+|[IR-SE50 Model](https://drive.google.com/file/d/1KW7bjndL3QG3sxBbZxreGHigcCCpsDgn/view?usp=sharing) | Pretrained IR-SE50 model taken from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch) for use in our ID loss during training.
+|[MOCOv2 Model](https://drive.google.com/file/d/18rLcNGdteX5LwT7sv_F7HWr12HpVEzVe/view?usp=sharing) | Pretrained ResNet-50 model trained using MOCOv2 for use in our simmilarity loss for domains other then human faces during training.
+
+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`. 
+
+## Training
+To train the e4e encoder, make sure the paths to the required models, as well as training and testing data is configured in `configs/path_configs.py` and `configs/data_configs.py`.
+#### **Training the e4e Encoder**
+```
+python scripts/train.py \
+--dataset_type cars_encode \
+--exp_dir new/experiment/directory \
+--start_from_latent_avg \
+--use_w_pool \
+--w_discriminator_lambda 0.1 \
+--progressive_start 20000 \
+--id_lambda 0.5 \
+--val_interval 10000 \
+--max_steps 200000 \
+--stylegan_size 512 \
+--stylegan_weights path/to/pretrained/stylegan.pt \
+--workers 8 \
+--batch_size 8 \
+--test_batch_size 4 \
+--test_workers 4 
+```
+
+#### Training on your own dataset
+In order to train the e4e encoder on a custom dataset, perform the following adjustments:
+1. Insert the paths to your train and test data into the `dataset_paths` variable defined in `configs/paths_config.py`:
+```
+dataset_paths = {
+    'my_train_data': '/path/to/train/images/directory',
+    'my_test_data': '/path/to/test/images/directory'
+}
+```
+2. Configure a new dataset under the DATASETS variable defined in `configs/data_configs.py`:
+```
+DATASETS = {
+   'my_data_encode': {
+        'transforms': transforms_config.EncodeTransforms,
+        'train_source_root': dataset_paths['my_train_data'],
+        'train_target_root': dataset_paths['my_train_data'],
+        'test_source_root': dataset_paths['my_test_data'],
+        'test_target_root': dataset_paths['my_test_data']
+    }
+}
+```
+Refer to `configs/transforms_config.py` for the transformations applied to the train and test images during training. 
+
+3. Finally, run a training session with `--dataset_type my_data_encode`.
+
+## Inference
+Having trained your model, you can use `scripts/inference.py` to apply the model on a set of images.   
+For example, 
+```
+python scripts/inference.py \
+--images_dir=/path/to/images/directory \
+--save_dir=/path/to/saving/directory \
+path/to/checkpoint.pt 
+```
+
+## Latent Editing Consistency (LEC)
+As described in the paper, we suggest a new metric, Latent Editing Consistency (LEC), for evaluating the encoder's 
+performance.
+We provide an example for calculating the metric over the FFHQ StyleGAN using the aging editing direction in 
+`metrics/LEC.py`.     
+
+To run the example: 
+```
+cd metrics
+python LEC.py \
+--images_dir=/path/to/images/directory \
+path/to/checkpoint.pt 
+```
+
+## Acknowledgments
+This code borrows heavily from [pixel2style2pixel](https://github.com/eladrich/pixel2style2pixel)
+
+## Citation
+If you use this code for your research, please cite our paper <a href="https://arxiv.org/abs/2102.02766">Designing an Encoder for StyleGAN Image Manipulation</a>:
+
+```
+@article{tov2021designing,
+  title={Designing an Encoder for StyleGAN Image Manipulation},
+  author={Tov, Omer and Alaluf, Yuval and Nitzan, Yotam and Patashnik, Or and Cohen-Or, Daniel},
+  journal={arXiv preprint arXiv:2102.02766},
+  year={2021}
+}
+```

Time_TravelRephotography/models/encoder4editing/__init__.py

@@ -0,0 +1,15 @@
+from .utils.model_utils import setup_model
+
+
+def get_latents(net, x, is_cars=False):
+    codes = net.encoder(x)
+    if net.opts.start_from_latent_avg:
+        if codes.ndim == 2:
+            codes = codes + net.latent_avg.repeat(codes.shape[0], 1, 1)[:, 0, :]
+        else:
+            codes = codes + net.latent_avg.repeat(codes.shape[0], 1, 1)
+    if codes.shape[1] == 18 and is_cars:
+        codes = codes[:, :16, :]
+    return codes
+
+

Time_TravelRephotography/models/encoder4editing/bash_scripts/inference.sh

@@ -0,0 +1,15 @@
+set -exo
+
+list="$1"
+ckpt="${2:-pretrained_models/e4e_ffhq_encode.pt}"
+
+base_dir="$REPHOTO/dataset/historically_interesting/aligned/manual_celebrity_in_19th_century/tier1/${list}/"
+save_dir="results_test/${list}/"
+
+
+TORCH_EXTENSIONS_DIR=/tmp/torch_extensions
+PYTHONPATH="" \
+python scripts/inference.py \
+    --images_dir="${base_dir}" \
+    --save_dir="${save_dir}" \
+    "${ckpt}"

Time_TravelRephotography/models/encoder4editing/configs/data_configs.py

@@ -0,0 +1,41 @@
+from configs import transforms_config
+from configs.paths_config import dataset_paths
+
+
+DATASETS = {
+	'ffhq_encode': {
+		'transforms': transforms_config.EncodeTransforms,
+		'train_source_root': dataset_paths['ffhq'],
+		'train_target_root': dataset_paths['ffhq'],
+		'test_source_root': dataset_paths['celeba_test'],
+		'test_target_root': dataset_paths['celeba_test'],
+	},
+	'cars_encode': {
+		'transforms': transforms_config.CarsEncodeTransforms,
+		'train_source_root': dataset_paths['cars_train'],
+		'train_target_root': dataset_paths['cars_train'],
+		'test_source_root': dataset_paths['cars_test'],
+		'test_target_root': dataset_paths['cars_test'],
+	},
+	'horse_encode': {
+		'transforms': transforms_config.EncodeTransforms,
+		'train_source_root': dataset_paths['horse_train'],
+		'train_target_root': dataset_paths['horse_train'],
+		'test_source_root': dataset_paths['horse_test'],
+		'test_target_root': dataset_paths['horse_test'],
+	},
+	'church_encode': {
+		'transforms': transforms_config.EncodeTransforms,
+		'train_source_root': dataset_paths['church_train'],
+		'train_target_root': dataset_paths['church_train'],
+		'test_source_root': dataset_paths['church_test'],
+		'test_target_root': dataset_paths['church_test'],
+	},
+	'cats_encode': {
+		'transforms': transforms_config.EncodeTransforms,
+		'train_source_root': dataset_paths['cats_train'],
+		'train_target_root': dataset_paths['cats_train'],
+		'test_source_root': dataset_paths['cats_test'],
+		'test_target_root': dataset_paths['cats_test'],
+	}
+}

Time_TravelRephotography/models/encoder4editing/configs/paths_config.py

@@ -0,0 +1,28 @@
+dataset_paths = {
+	#  Face Datasets (In the paper: FFHQ - train, CelebAHQ - test)
+	'ffhq': '',
+	'celeba_test': '',
+
+	#  Cars Dataset (In the paper: Stanford cars)
+	'cars_train': '',
+	'cars_test': '',
+
+	#  Horse Dataset (In the paper: LSUN Horse)
+	'horse_train': '',
+	'horse_test': '',
+
+	#  Church Dataset (In the paper: LSUN Church)
+	'church_train': '',
+	'church_test': '',
+
+	#  Cats Dataset (In the paper: LSUN Cat)
+	'cats_train': '',
+	'cats_test': ''
+}
+
+model_paths = {
+	'stylegan_ffhq': 'pretrained_models/stylegan2-ffhq-config-f.pt',
+	'ir_se50': 'pretrained_models/model_ir_se50.pth',
+	'shape_predictor': 'pretrained_models/shape_predictor_68_face_landmarks.dat',
+	'moco': 'pretrained_models/moco_v2_800ep_pretrain.pth'
+}

Time_TravelRephotography/models/encoder4editing/configs/transforms_config.py

@@ -0,0 +1,62 @@
+from abc import abstractmethod
+import torchvision.transforms as transforms
+
+
+class TransformsConfig(object):
+
+	def __init__(self, opts):
+		self.opts = opts
+
+	@abstractmethod
+	def get_transforms(self):
+		pass
+
+
+class EncodeTransforms(TransformsConfig):
+
+	def __init__(self, opts):
+		super(EncodeTransforms, self).__init__(opts)
+
+	def get_transforms(self):
+		transforms_dict = {
+			'transform_gt_train': transforms.Compose([
+				transforms.Resize((256, 256)),
+				transforms.RandomHorizontalFlip(0.5),
+				transforms.ToTensor(),
+				transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
+			'transform_source': None,
+			'transform_test': transforms.Compose([
+				transforms.Resize((256, 256)),
+				transforms.ToTensor(),
+				transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
+			'transform_inference': transforms.Compose([
+				transforms.Resize((256, 256)),
+				transforms.ToTensor(),
+				transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
+		}
+		return transforms_dict
+
+
+class CarsEncodeTransforms(TransformsConfig):
+
+	def __init__(self, opts):
+		super(CarsEncodeTransforms, self).__init__(opts)
+
+	def get_transforms(self):
+		transforms_dict = {
+			'transform_gt_train': transforms.Compose([
+				transforms.Resize((192, 256)),
+				transforms.RandomHorizontalFlip(0.5),
+				transforms.ToTensor(),
+				transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
+			'transform_source': None,
+			'transform_test': transforms.Compose([
+				transforms.Resize((192, 256)),
+				transforms.ToTensor(),
+				transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
+			'transform_inference': transforms.Compose([
+				transforms.Resize((192, 256)),
+				transforms.ToTensor(),
+				transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
+		}
+		return transforms_dict

Time_TravelRephotography/models/encoder4editing/criteria/id_loss.py

@@ -0,0 +1,47 @@
+import torch
+from torch import nn
+from configs.paths_config import model_paths
+from models.encoders.model_irse import Backbone
+
+
+class IDLoss(nn.Module):
+    def __init__(self):
+        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(model_paths['ir_se50']))
+        self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112))
+        self.facenet.eval()
+        for module in [self.facenet, self.face_pool]:
+            for param in module.parameters():
+                param.requires_grad = False
+
+    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, y_hat, y, x):
+        n_samples = x.shape[0]
+        x_feats = self.extract_feats(x)
+        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
+        id_logs = []
+        count = 0
+        for i in range(n_samples):
+            diff_target = y_hat_feats[i].dot(y_feats[i])
+            diff_input = y_hat_feats[i].dot(x_feats[i])
+            diff_views = y_feats[i].dot(x_feats[i])
+            id_logs.append({'diff_target': float(diff_target),
+                            'diff_input': float(diff_input),
+                            'diff_views': float(diff_views)})
+            loss += 1 - diff_target
+            id_diff = float(diff_target) - float(diff_views)
+            sim_improvement += id_diff
+            count += 1
+
+        return loss / count, sim_improvement / count, id_logs

Time_TravelRephotography/models/encoder4editing/criteria/lpips/lpips.py

@@ -0,0 +1,35 @@
+import torch
+import torch.nn as nn
+
+from criteria.lpips.networks import get_network, LinLayers
+from criteria.lpips.utils import get_state_dict
+
+
+class LPIPS(nn.Module):
+    r"""Creates a criterion that measures
+    Learned Perceptual Image Patch Similarity (LPIPS).
+    Arguments:
+        net_type (str): the network type to compare the features:
+                        'alex' | 'squeeze' | 'vgg'. Default: 'alex'.
+        version (str): the version of LPIPS. Default: 0.1.
+    """
+    def __init__(self, net_type: str = 'alex', version: str = '0.1'):
+
+        assert version in ['0.1'], 'v0.1 is only supported now'
+
+        super(LPIPS, self).__init__()
+
+        # pretrained network
+        self.net = get_network(net_type).to("cuda")
+
+        # linear layers
+        self.lin = LinLayers(self.net.n_channels_list).to("cuda")
+        self.lin.load_state_dict(get_state_dict(net_type, version))
+
+    def forward(self, x: torch.Tensor, y: torch.Tensor):
+        feat_x, feat_y = self.net(x), self.net(y)
+
+        diff = [(fx - fy) ** 2 for fx, fy in zip(feat_x, feat_y)]
+        res = [l(d).mean((2, 3), True) for d, l in zip(diff, self.lin)]
+
+        return torch.sum(torch.cat(res, 0)) / x.shape[0]

Time_TravelRephotography/models/encoder4editing/criteria/lpips/networks.py

@@ -0,0 +1,96 @@
+from typing import Sequence
+
+from itertools import chain
+
+import torch
+import torch.nn as nn
+from torchvision import models
+
+from criteria.lpips.utils import normalize_activation
+
+
+def get_network(net_type: str):
+    if net_type == 'alex':
+        return AlexNet()
+    elif net_type == 'squeeze':
+        return SqueezeNet()
+    elif net_type == 'vgg':
+        return VGG16()
+    else:
+        raise NotImplementedError('choose net_type from [alex, squeeze, vgg].')
+
+
+class LinLayers(nn.ModuleList):
+    def __init__(self, n_channels_list: Sequence[int]):
+        super(LinLayers, self).__init__([
+            nn.Sequential(
+                nn.Identity(),
+                nn.Conv2d(nc, 1, 1, 1, 0, bias=False)
+            ) for nc in n_channels_list
+        ])
+
+        for param in self.parameters():
+            param.requires_grad = False
+
+
+class BaseNet(nn.Module):
+    def __init__(self):
+        super(BaseNet, self).__init__()
+
+        # register buffer
+        self.register_buffer(
+            'mean', torch.Tensor([-.030, -.088, -.188])[None, :, None, None])
+        self.register_buffer(
+            'std', torch.Tensor([.458, .448, .450])[None, :, None, None])
+
+    def set_requires_grad(self, state: bool):
+        for param in chain(self.parameters(), self.buffers()):
+            param.requires_grad = state
+
+    def z_score(self, x: torch.Tensor):
+        return (x - self.mean) / self.std
+
+    def forward(self, x: torch.Tensor):
+        x = self.z_score(x)
+
+        output = []
+        for i, (_, layer) in enumerate(self.layers._modules.items(), 1):
+            x = layer(x)
+            if i in self.target_layers:
+                output.append(normalize_activation(x))
+            if len(output) == len(self.target_layers):
+                break
+        return output
+
+
+class SqueezeNet(BaseNet):
+    def __init__(self):
+        super(SqueezeNet, self).__init__()
+
+        self.layers = models.squeezenet1_1(True).features
+        self.target_layers = [2, 5, 8, 10, 11, 12, 13]
+        self.n_channels_list = [64, 128, 256, 384, 384, 512, 512]
+
+        self.set_requires_grad(False)
+
+
+class AlexNet(BaseNet):
+    def __init__(self):
+        super(AlexNet, self).__init__()
+
+        self.layers = models.alexnet(True).features
+        self.target_layers = [2, 5, 8, 10, 12]
+        self.n_channels_list = [64, 192, 384, 256, 256]
+
+        self.set_requires_grad(False)
+
+
+class VGG16(BaseNet):
+    def __init__(self):
+        super(VGG16, self).__init__()
+
+        self.layers = models.vgg16(True).features
+        self.target_layers = [4, 9, 16, 23, 30]
+        self.n_channels_list = [64, 128, 256, 512, 512]
+
+        self.set_requires_grad(False)

Time_TravelRephotography/models/encoder4editing/criteria/lpips/utils.py

@@ -0,0 +1,30 @@
+from collections import OrderedDict
+
+import torch
+
+
+def normalize_activation(x, eps=1e-10):
+    norm_factor = torch.sqrt(torch.sum(x ** 2, dim=1, keepdim=True))
+    return x / (norm_factor + eps)
+
+
+def get_state_dict(net_type: str = 'alex', version: str = '0.1'):
+    # build url
+    url = 'https://raw.githubusercontent.com/richzhang/PerceptualSimilarity/' \
+        + f'master/lpips/weights/v{version}/{net_type}.pth'
+
+    # download
+    old_state_dict = torch.hub.load_state_dict_from_url(
+        url, progress=True,
+        map_location=None if torch.cuda.is_available() else torch.device('cpu')
+    )
+
+    # rename keys
+    new_state_dict = OrderedDict()
+    for key, val in old_state_dict.items():
+        new_key = key
+        new_key = new_key.replace('lin', '')
+        new_key = new_key.replace('model.', '')
+        new_state_dict[new_key] = val
+
+    return new_state_dict

Time_TravelRephotography/models/encoder4editing/criteria/moco_loss.py

@@ -0,0 +1,71 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+from configs.paths_config import model_paths
+
+
+class MocoLoss(nn.Module):
+
+    def __init__(self, opts):
+        super(MocoLoss, self).__init__()
+        print("Loading MOCO model from path: {}".format(model_paths["moco"]))
+        self.model = self.__load_model()
+        self.model.eval()
+        for param in self.model.parameters():
+            param.requires_grad = False
+
+    @staticmethod
+    def __load_model():
+        import torchvision.models as models
+        model = models.__dict__["resnet50"]()
+        # freeze all layers but the last fc
+        for name, param in model.named_parameters():
+            if name not in ['fc.weight', 'fc.bias']:
+                param.requires_grad = False
+        checkpoint = torch.load(model_paths['moco'], map_location="cpu")
+        state_dict = checkpoint['state_dict']
+        # rename moco pre-trained keys
+        for k in list(state_dict.keys()):
+            # retain only encoder_q up to before the embedding layer
+            if k.startswith('module.encoder_q') and not k.startswith('module.encoder_q.fc'):
+                # remove prefix
+                state_dict[k[len("module.encoder_q."):]] = state_dict[k]
+            # delete renamed or unused k
+            del state_dict[k]
+        msg = model.load_state_dict(state_dict, strict=False)
+        assert set(msg.missing_keys) == {"fc.weight", "fc.bias"}
+        # remove output layer
+        model = nn.Sequential(*list(model.children())[:-1]).cuda()
+        return model
+
+    def extract_feats(self, x):
+        x = F.interpolate(x, size=224)
+        x_feats = self.model(x)
+        x_feats = nn.functional.normalize(x_feats, dim=1)
+        x_feats = x_feats.squeeze()
+        return x_feats
+
+    def forward(self, y_hat, y, x):
+        n_samples = x.shape[0]
+        x_feats = self.extract_feats(x)
+        y_feats = self.extract_feats(y)
+        y_hat_feats = self.extract_feats(y_hat)
+        y_feats = y_feats.detach()
+        loss = 0
+        sim_improvement = 0
+        sim_logs = []
+        count = 0
+        for i in range(n_samples):
+            diff_target = y_hat_feats[i].dot(y_feats[i])
+            diff_input = y_hat_feats[i].dot(x_feats[i])
+            diff_views = y_feats[i].dot(x_feats[i])
+            sim_logs.append({'diff_target': float(diff_target),
+                             'diff_input': float(diff_input),
+                             'diff_views': float(diff_views)})
+            loss += 1 - diff_target
+            sim_diff = float(diff_target) - float(diff_views)
+            sim_improvement += sim_diff
+            count += 1
+
+        return loss / count, sim_improvement / count, sim_logs

Time_TravelRephotography/models/encoder4editing/criteria/w_norm.py

@@ -0,0 +1,14 @@
+import torch
+from torch import nn
+
+
+class WNormLoss(nn.Module):
+
+	def __init__(self, start_from_latent_avg=True):
+		super(WNormLoss, self).__init__()
+		self.start_from_latent_avg = start_from_latent_avg
+
+	def forward(self, latent, latent_avg=None):
+		if self.start_from_latent_avg:
+			latent = latent - latent_avg
+		return torch.sum(latent.norm(2, dim=(1, 2))) / latent.shape[0]

Time_TravelRephotography/models/encoder4editing/datasets/gt_res_dataset.py

@@ -0,0 +1,32 @@
+#!/usr/bin/python
+# encoding: utf-8
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+import torch
+
+class GTResDataset(Dataset):
+
+	def __init__(self, root_path, gt_dir=None, transform=None, transform_train=None):
+		self.pairs = []
+		for f in os.listdir(root_path):
+			image_path = os.path.join(root_path, f)
+			gt_path = os.path.join(gt_dir, f)
+			if f.endswith(".jpg") or f.endswith(".png"):
+				self.pairs.append([image_path, gt_path.replace('.png', '.jpg'), None])
+		self.transform = transform
+		self.transform_train = transform_train
+
+	def __len__(self):
+		return len(self.pairs)
+
+	def __getitem__(self, index):
+		from_path, to_path, _ = self.pairs[index]
+		from_im = Image.open(from_path).convert('RGB')
+		to_im = Image.open(to_path).convert('RGB')
+
+		if self.transform:
+			to_im = self.transform(to_im)
+			from_im = self.transform(from_im)
+
+		return from_im, to_im

Time_TravelRephotography/models/encoder4editing/datasets/images_dataset.py

@@ -0,0 +1,33 @@
+from torch.utils.data import Dataset
+from PIL import Image
+from utils import data_utils
+
+
+class ImagesDataset(Dataset):
+
+	def __init__(self, source_root, target_root, opts, target_transform=None, source_transform=None):
+		self.source_paths = sorted(data_utils.make_dataset(source_root))
+		self.target_paths = sorted(data_utils.make_dataset(target_root))
+		self.source_transform = source_transform
+		self.target_transform = target_transform
+		self.opts = opts
+
+	def __len__(self):
+		return len(self.source_paths)
+
+	def __getitem__(self, index):
+		from_path = self.source_paths[index]
+		from_im = Image.open(from_path)
+		from_im = from_im.convert('RGB')
+
+		to_path = self.target_paths[index]
+		to_im = Image.open(to_path).convert('RGB')
+		if self.target_transform:
+			to_im = self.target_transform(to_im)
+
+		if self.source_transform:
+			from_im = self.source_transform(from_im)
+		else:
+			from_im = to_im
+
+		return from_im, to_im

Time_TravelRephotography/models/encoder4editing/datasets/inference_dataset.py

@@ -0,0 +1,25 @@
+from torch.utils.data import Dataset
+from PIL import Image
+from utils import data_utils
+
+
+class InferenceDataset(Dataset):
+
+	def __init__(self, root, opts, transform=None, preprocess=None):
+		self.paths = sorted(data_utils.make_dataset(root))
+		self.transform = transform
+		self.preprocess = preprocess
+		self.opts = opts
+
+	def __len__(self):
+		return len(self.paths)
+
+	def __getitem__(self, index):
+		from_path = self.paths[index]
+		if self.preprocess is not None:
+			from_im = self.preprocess(from_path)
+		else:
+			from_im = Image.open(from_path).convert('RGB')
+		if self.transform:
+			from_im = self.transform(from_im)
+		return from_im

Time_TravelRephotography/models/encoder4editing/editings/ganspace.py

@@ -0,0 +1,22 @@
+import torch
+
+
+def edit(latents, pca, edit_directions):
+    edit_latents = []
+    for latent in latents:
+        for pca_idx, start, end, strength in edit_directions:
+            delta = get_delta(pca, latent, pca_idx, strength)
+            delta_padded = torch.zeros(latent.shape).to('cuda')
+            delta_padded[start:end] += delta.repeat(end - start, 1)
+            edit_latents.append(latent + delta_padded)
+    return torch.stack(edit_latents)
+
+
+def get_delta(pca, latent, idx, strength):
+    # pca: ganspace checkpoint. latent: (16, 512) w+
+    w_centered = latent - pca['mean'].to('cuda')
+    lat_comp = pca['comp'].to('cuda')
+    lat_std = pca['std'].to('cuda')
+    w_coord = torch.sum(w_centered[0].reshape(-1)*lat_comp[idx].reshape(-1)) / lat_std[idx]
+    delta = (strength - w_coord)*lat_comp[idx]*lat_std[idx]
+    return delta

Time_TravelRephotography/models/encoder4editing/editings/ganspace_pca/cars_pca.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5c3bae61ecd85de077fbbf103f5f30cf4b7676fe23a8508166eaf2ce73c8392
+size 167562

Time_TravelRephotography/models/encoder4editing/editings/ganspace_pca/ffhq_pca.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d7f9df1c96180d9026b9cb8d04753579fbf385f321a9d0e263641601c5e5d36
+size 167562

Time_TravelRephotography/models/encoder4editing/editings/interfacegan_directions/age.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50074516b1629707d89b5e43d6b8abd1792212fa3b961a87a11323d6a5222ae0
+size 2808

Time_TravelRephotography/models/encoder4editing/editings/interfacegan_directions/pose.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:736e0eacc8488fa0b020a2e7bd256b957284c364191dfea693705e5d06d43e7d
+size 37624

Time_TravelRephotography/models/encoder4editing/editings/interfacegan_directions/smile.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:817a7e732b59dee9eba862bec8bd7e8373568443bc9f9731a21cf9b0356f0653
+size 2808

Time_TravelRephotography/models/encoder4editing/editings/latent_editor.py

@@ -0,0 +1,45 @@
+import torch
+import sys
+sys.path.append(".")
+sys.path.append("..")
+from editings import ganspace, sefa
+from utils.common import tensor2im
+
+
+class LatentEditor(object):
+    def __init__(self, stylegan_generator, is_cars=False):
+        self.generator = stylegan_generator
+        self.is_cars = is_cars  # Since the cars StyleGAN output is 384x512, there is a need to crop the 512x512 output.
+
+    def apply_ganspace(self, latent, ganspace_pca, edit_directions):
+        edit_latents = ganspace.edit(latent, ganspace_pca, edit_directions)
+        return self._latents_to_image(edit_latents)
+
+    def apply_interfacegan(self, latent, direction, factor=1, factor_range=None):
+        edit_latents = []
+        if factor_range is not None:  # Apply a range of editing factors. for example, (-5, 5)
+            for f in range(*factor_range):
+                edit_latent = latent + f * direction
+                edit_latents.append(edit_latent)
+            edit_latents = torch.cat(edit_latents)
+        else:
+            edit_latents = latent + factor * direction
+        return self._latents_to_image(edit_latents)
+
+    def apply_sefa(self, latent, indices=[2, 3, 4, 5], **kwargs):
+        edit_latents = sefa.edit(self.generator, latent, indices, **kwargs)
+        return self._latents_to_image(edit_latents)
+
+    # Currently, in order to apply StyleFlow editings, one should run inference,
+    # save the latent codes and load them form the official StyleFlow repository.
+    # def apply_styleflow(self):
+    #     pass
+
+    def _latents_to_image(self, latents):
+        with torch.no_grad():
+            images, _ = self.generator([latents], randomize_noise=False, input_is_latent=True)
+            if self.is_cars:
+                images = images[:, :, 64:448, :]  # 512x512 -> 384x512
+        horizontal_concat_image = torch.cat(list(images), 2)
+        final_image = tensor2im(horizontal_concat_image)
+        return final_image

Time_TravelRephotography/models/encoder4editing/editings/sefa.py

@@ -0,0 +1,46 @@
+import torch
+import numpy as np
+from tqdm import tqdm
+
+
+def edit(generator, latents, indices, semantics=1, start_distance=-15.0, end_distance=15.0, num_samples=1, step=11):
+
+    layers, boundaries, values = factorize_weight(generator, indices)
+    codes = latents.detach().cpu().numpy()  # (1,18,512)
+
+    # Generate visualization pages.
+    distances = np.linspace(start_distance, end_distance, step)
+    num_sam = num_samples
+    num_sem = semantics
+
+    edited_latents = []
+    for sem_id in tqdm(range(num_sem), desc='Semantic ', leave=False):
+        boundary = boundaries[sem_id:sem_id + 1]
+        for sam_id in tqdm(range(num_sam), desc='Sample ', leave=False):
+            code = codes[sam_id:sam_id + 1]
+            for col_id, d in enumerate(distances, start=1):
+                temp_code = code.copy()
+                temp_code[:, layers, :] += boundary * d
+                edited_latents.append(torch.from_numpy(temp_code).float().cuda())
+    return torch.cat(edited_latents)
+
+
+def factorize_weight(g_ema, layers='all'):
+
+    weights = []
+    if layers == 'all' or 0 in layers:
+        weight = g_ema.conv1.conv.modulation.weight.T
+        weights.append(weight.cpu().detach().numpy())
+
+    if layers == 'all':
+        layers = list(range(g_ema.num_layers - 1))
+    else:
+        layers = [l - 1 for l in layers if l != 0]
+
+    for idx in layers:
+        weight = g_ema.convs[idx].conv.modulation.weight.T
+        weights.append(weight.cpu().detach().numpy())
+    weight = np.concatenate(weights, axis=1).astype(np.float32)
+    weight = weight / np.linalg.norm(weight, axis=0, keepdims=True)
+    eigen_values, eigen_vectors = np.linalg.eig(weight.dot(weight.T))
+    return layers, eigen_vectors.T, eigen_values

Time_TravelRephotography/models/encoder4editing/environment/e4e_env.yaml

@@ -0,0 +1,73 @@
+name: e4e_env
+channels:
+  - conda-forge
+  - defaults
+dependencies:
+  - _libgcc_mutex=0.1=main
+  - ca-certificates=2020.4.5.1=hecc5488_0
+  - certifi=2020.4.5.1=py36h9f0ad1d_0
+  - libedit=3.1.20181209=hc058e9b_0
+  - libffi=3.2.1=hd88cf55_4
+  - libgcc-ng=9.1.0=hdf63c60_0
+  - libstdcxx-ng=9.1.0=hdf63c60_0
+  - ncurses=6.2=he6710b0_1
+  - ninja=1.10.0=hc9558a2_0
+  - openssl=1.1.1g=h516909a_0
+  - pip=20.0.2=py36_3
+  - python=3.6.7=h0371630_0
+  - python_abi=3.6=1_cp36m
+  - readline=7.0=h7b6447c_5
+  - setuptools=46.4.0=py36_0
+  - sqlite=3.31.1=h62c20be_1
+  - tk=8.6.8=hbc83047_0
+  - wheel=0.34.2=py36_0
+  - xz=5.2.5=h7b6447c_0
+  - zlib=1.2.11=h7b6447c_3
+  - pip:
+    - absl-py==0.9.0
+    - cachetools==4.1.0
+    - chardet==3.0.4
+    - cycler==0.10.0
+    - decorator==4.4.2
+    - future==0.18.2
+    - google-auth==1.15.0
+    - google-auth-oauthlib==0.4.1
+    - grpcio==1.29.0
+    - idna==2.9
+    - imageio==2.8.0
+    - importlib-metadata==1.6.0
+    - kiwisolver==1.2.0
+    - markdown==3.2.2
+    - matplotlib==3.2.1
+    - mxnet==1.6.0
+    - networkx==2.4
+    - numpy==1.18.4
+    - oauthlib==3.1.0
+    - opencv-python==4.2.0.34
+    - pillow==7.1.2
+    - protobuf==3.12.1
+    - pyasn1==0.4.8
+    - pyasn1-modules==0.2.8
+    - pyparsing==2.4.7
+    - python-dateutil==2.8.1
+    - pytorch-lightning==0.7.1
+    - pywavelets==1.1.1
+    - requests==2.23.0
+    - requests-oauthlib==1.3.0
+    - rsa==4.0
+    - scikit-image==0.17.2
+    - scipy==1.4.1
+    - six==1.15.0
+    - tensorboard==2.2.1
+    - tensorboard-plugin-wit==1.6.0.post3
+    - tensorboardx==1.9
+    - tifffile==2020.5.25
+    - torch==1.6.0
+    - torchvision==0.7.1
+    - tqdm==4.46.0
+    - urllib3==1.25.9
+    - werkzeug==1.0.1
+    - zipp==3.1.0
+    - pyaml
+prefix: ~/anaconda3/envs/e4e_env
+

Time_TravelRephotography/models/encoder4editing/metrics/LEC.py

@@ -0,0 +1,134 @@
+import sys
+import argparse
+import torch
+import numpy as np
+from torch.utils.data import DataLoader
+
+sys.path.append(".")
+sys.path.append("..")
+
+from configs import data_configs
+from datasets.images_dataset import ImagesDataset
+from utils.model_utils import setup_model
+
+
+class LEC:
+    def __init__(self, net, is_cars=False):
+        """
+        Latent Editing Consistency metric as proposed in the main paper.
+        :param net: e4e model loaded over the pSp framework.
+        :param is_cars: An indication as to whether or not to crop the middle of the StyleGAN's output images.
+        """
+        self.net = net
+        self.is_cars = is_cars
+
+    def _encode(self, images):
+        """
+        Encodes the given images into StyleGAN's latent space.
+        :param images: Tensor of shape NxCxHxW representing the images to be encoded.
+        :return: Tensor of shape NxKx512 representing the latent space embeddings of the given image (in W(K, *) space).
+        """
+        codes = self.net.encoder(images)
+        assert codes.ndim == 3, f"Invalid latent codes shape, should be NxKx512 but is {codes.shape}"
+        # normalize with respect to the center of an average face
+        if self.net.opts.start_from_latent_avg:
+            codes = codes + self.net.latent_avg.repeat(codes.shape[0], 1, 1)
+        return codes
+
+    def _generate(self, codes):
+        """
+        Generate the StyleGAN2 images of the given codes
+        :param codes: Tensor of shape NxKx512 representing the StyleGAN's latent codes (in W(K, *) space).
+        :return: Tensor of shape  NxCxHxW representing the generated images.
+        """
+        images, _ = self.net.decoder([codes], input_is_latent=True, randomize_noise=False, return_latents=True)
+        images = self.net.face_pool(images)
+        if self.is_cars:
+            images = images[:, :, 32:224, :]
+        return images
+
+    @staticmethod
+    def _filter_outliers(arr):
+        arr = np.array(arr)
+
+        lo = np.percentile(arr, 1, interpolation="lower")
+        hi = np.percentile(arr, 99, interpolation="higher")
+        return np.extract(
+            np.logical_and(lo <= arr, arr <= hi), arr
+        )
+
+    def calculate_metric(self, data_loader, edit_function, inverse_edit_function):
+        """
+        Calculate the LEC metric score.
+        :param data_loader: An iterable that returns a tuple of (images, _), similar to the training data loader.
+        :param edit_function: A function that receives latent codes and performs a semantically meaningful edit in the
+                              latent space.
+        :param inverse_edit_function: A function that receives latent codes and performs the inverse edit of the
+                                      `edit_function` parameter.
+        :return: The LEC metric score.
+        """
+        distances = []
+        with torch.no_grad():
+            for batch in data_loader:
+                x, _ = batch
+                inputs = x.to(device).float()
+
+                codes = self._encode(inputs)
+                edited_codes = edit_function(codes)
+                edited_image = self._generate(edited_codes)
+                edited_image_inversion_codes = self._encode(edited_image)
+                inverse_edit_codes = inverse_edit_function(edited_image_inversion_codes)
+
+                dist = (codes - inverse_edit_codes).norm(2, dim=(1, 2)).mean()
+                distances.append(dist.to("cpu").numpy())
+
+        distances = self._filter_outliers(distances)
+        return distances.mean()
+
+
+if __name__ == "__main__":
+    device = "cuda"
+
+    parser = argparse.ArgumentParser(description="LEC metric calculator")
+
+    parser.add_argument("--batch", type=int, default=8, help="batch size for the models")
+    parser.add_argument("--images_dir", type=str, default=None,
+                        help="Path to the images directory on which we calculate the LEC score")
+    parser.add_argument("ckpt", metavar="CHECKPOINT", help="path to the model checkpoints")
+
+    args = parser.parse_args()
+    print(args)
+
+    net, opts = setup_model(args.ckpt, device)
+    dataset_args = data_configs.DATASETS[opts.dataset_type]
+    transforms_dict = dataset_args['transforms'](opts).get_transforms()
+
+    images_directory = dataset_args['test_source_root'] if args.images_dir is None else args.images_dir
+    test_dataset = ImagesDataset(source_root=images_directory,
+                                 target_root=images_directory,
+                                 source_transform=transforms_dict['transform_source'],
+                                 target_transform=transforms_dict['transform_test'],
+                                 opts=opts)
+
+    data_loader = DataLoader(test_dataset,
+                             batch_size=args.batch,
+                             shuffle=False,
+                             num_workers=2,
+                             drop_last=True)
+
+    print(f'dataset length: {len(test_dataset)}')
+
+    # In the following example, we are using an InterfaceGAN based editing to calculate the LEC metric.
+    # Change the provided example according to your domain and needs.
+    direction = torch.load('../editings/interfacegan_directions/age.pt').to(device)
+
+    def edit_func_example(codes):
+        return codes + 3 * direction
+
+
+    def inverse_edit_func_example(codes):
+        return codes - 3 * direction
+
+    lec = LEC(net, is_cars='car' in opts.dataset_type)
+    result = lec.calculate_metric(data_loader, edit_func_example, inverse_edit_func_example)
+    print(f"LEC: {result}")

Time_TravelRephotography/models/encoder4editing/models/discriminator.py

@@ -0,0 +1,20 @@
+from torch import nn
+
+
+class LatentCodesDiscriminator(nn.Module):
+    def __init__(self, style_dim, n_mlp):
+        super().__init__()
+
+        self.style_dim = style_dim
+
+        layers = []
+        for i in range(n_mlp-1):
+            layers.append(
+                nn.Linear(style_dim, style_dim)
+            )
+            layers.append(nn.LeakyReLU(0.2))
+        layers.append(nn.Linear(512, 1))
+        self.mlp = nn.Sequential(*layers)
+
+    def forward(self, w):
+        return self.mlp(w)

Time_TravelRephotography/models/encoder4editing/models/encoders/helpers.py

@@ -0,0 +1,140 @@
+from collections import namedtuple
+import torch
+import torch.nn.functional as F
+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
+
+
+def _upsample_add(x, y):
+    """Upsample and add two feature maps.
+    Args:
+      x: (Variable) top feature map to be upsampled.
+      y: (Variable) lateral feature map.
+    Returns:
+      (Variable) added feature map.
+    Note in PyTorch, when input size is odd, the upsampled feature map
+    with `F.upsample(..., scale_factor=2, mode='nearest')`
+    maybe not equal to the lateral feature map size.
+    e.g.
+    original input size: [N,_,15,15] ->
+    conv2d feature map size: [N,_,8,8] ->
+    upsampled feature map size: [N,_,16,16]
+    So we choose bilinear upsample which supports arbitrary output sizes.
+    """
+    _, _, H, W = y.size()
+    return F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True) + y

Time_TravelRephotography/models/encoder4editing/models/encoders/model_irse.py

@@ -0,0 +1,84 @@
+from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, Dropout, Sequential, Module
+from models.encoders.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

Time_TravelRephotography/models/encoder4editing/models/encoders/psp_encoders.py

@@ -0,0 +1,235 @@
+from enum import Enum
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import Conv2d, BatchNorm2d, PReLU, Sequential, Module
+
+from .helpers import get_blocks, bottleneck_IR, bottleneck_IR_SE, _upsample_add
+from ..stylegan2.model import EqualLinear
+
+
+class ProgressiveStage(Enum):
+    WTraining = 0
+    Delta1Training = 1
+    Delta2Training = 2
+    Delta3Training = 3
+    Delta4Training = 4
+    Delta5Training = 5
+    Delta6Training = 6
+    Delta7Training = 7
+    Delta8Training = 8
+    Delta9Training = 9
+    Delta10Training = 10
+    Delta11Training = 11
+    Delta12Training = 12
+    Delta13Training = 13
+    Delta14Training = 14
+    Delta15Training = 15
+    Delta16Training = 16
+    Delta17Training = 17
+    Inference = 18
+
+
+class GradualStyleBlock(Module):
+    def __init__(self, in_c, out_c, spatial):
+        super(GradualStyleBlock, self).__init__()
+        self.out_c = out_c
+        self.spatial = spatial
+        num_pools = int(np.log2(spatial))
+        modules = []
+        modules += [Conv2d(in_c, out_c, kernel_size=3, stride=2, padding=1),
+                    nn.LeakyReLU()]
+        for i in range(num_pools - 1):
+            modules += [
+                Conv2d(out_c, out_c, kernel_size=3, stride=2, padding=1),
+                nn.LeakyReLU()
+            ]
+        self.convs = nn.Sequential(*modules)
+        self.linear = EqualLinear(out_c, out_c, lr_mul=1)
+
+    def forward(self, x):
+        x = self.convs(x)
+        x = x.view(-1, self.out_c)
+        x = self.linear(x)
+        return x
+
+
+class GradualStyleEncoder(Module):
+    def __init__(self, num_layers, mode='ir', opts=None):
+        super(GradualStyleEncoder, self).__init__()
+        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))
+        modules = []
+        for block in blocks:
+            for bottleneck in block:
+                modules.append(unit_module(bottleneck.in_channel,
+                                           bottleneck.depth,
+                                           bottleneck.stride))
+        self.body = Sequential(*modules)
+
+        self.styles = nn.ModuleList()
+        log_size = int(math.log(opts.stylegan_size, 2))
+        self.style_count = 2 * log_size - 2
+        self.coarse_ind = 3
+        self.middle_ind = 7
+        for i in range(self.style_count):
+            if i < self.coarse_ind:
+                style = GradualStyleBlock(512, 512, 16)
+            elif i < self.middle_ind:
+                style = GradualStyleBlock(512, 512, 32)
+            else:
+                style = GradualStyleBlock(512, 512, 64)
+            self.styles.append(style)
+        self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0)
+        self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        x = self.input_layer(x)
+
+        latents = []
+        modulelist = list(self.body._modules.values())
+        for i, l in enumerate(modulelist):
+            x = l(x)
+            if i == 6:
+                c1 = x
+            elif i == 20:
+                c2 = x
+            elif i == 23:
+                c3 = x
+
+        for j in range(self.coarse_ind):
+            latents.append(self.styles[j](c3))
+
+        p2 = _upsample_add(c3, self.latlayer1(c2))
+        for j in range(self.coarse_ind, self.middle_ind):
+            latents.append(self.styles[j](p2))
+
+        p1 = _upsample_add(p2, self.latlayer2(c1))
+        for j in range(self.middle_ind, self.style_count):
+            latents.append(self.styles[j](p1))
+
+        out = torch.stack(latents, dim=1)
+        return out
+
+
+class Encoder4Editing(Module):
+    def __init__(self, num_layers, mode='ir', opts=None):
+        super(Encoder4Editing, self).__init__()
+        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))
+        modules = []
+        for block in blocks:
+            for bottleneck in block:
+                modules.append(unit_module(bottleneck.in_channel,
+                                           bottleneck.depth,
+                                           bottleneck.stride))
+        self.body = Sequential(*modules)
+
+        self.styles = nn.ModuleList()
+        log_size = int(math.log(opts.stylegan_size, 2))
+        self.style_count = 2 * log_size - 2
+        self.coarse_ind = 3
+        self.middle_ind = 7
+
+        for i in range(self.style_count):
+            if i < self.coarse_ind:
+                style = GradualStyleBlock(512, 512, 16)
+            elif i < self.middle_ind:
+                style = GradualStyleBlock(512, 512, 32)
+            else:
+                style = GradualStyleBlock(512, 512, 64)
+            self.styles.append(style)
+
+        self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0)
+        self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0)
+
+        self.progressive_stage = ProgressiveStage.Inference
+
+    def get_deltas_starting_dimensions(self):
+        ''' Get a list of the initial dimension of every delta from which it is applied '''
+        return list(range(self.style_count))  # Each dimension has a delta applied to it
+
+    def set_progressive_stage(self, new_stage: ProgressiveStage):
+        self.progressive_stage = new_stage
+        print('Changed progressive stage to: ', new_stage)
+
+    def forward(self, x):
+        x = self.input_layer(x)
+
+        modulelist = list(self.body._modules.values())
+        for i, l in enumerate(modulelist):
+            x = l(x)
+            if i == 6:
+                c1 = x
+            elif i == 20:
+                c2 = x
+            elif i == 23:
+                c3 = x
+
+        # Infer main W and duplicate it
+        w0 = self.styles[0](c3)
+        w = w0.repeat(self.style_count, 1, 1).permute(1, 0, 2)
+        stage = self.progressive_stage.value
+        features = c3
+        for i in range(1, min(stage + 1, self.style_count)):  # Infer additional deltas
+            if i == self.coarse_ind:
+                p2 = _upsample_add(c3, self.latlayer1(c2))  # FPN's middle features
+                features = p2
+            elif i == self.middle_ind:
+                p1 = _upsample_add(p2, self.latlayer2(c1))  # FPN's fine features
+                features = p1
+            delta_i = self.styles[i](features)
+            w[:, i] += delta_i
+        return w
+
+
+class BackboneEncoderUsingLastLayerIntoW(Module):
+    def __init__(self, num_layers, mode='ir', opts=None):
+        super(BackboneEncoderUsingLastLayerIntoW, self).__init__()
+        print('Using BackboneEncoderUsingLastLayerIntoW')
+        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))
+        self.output_pool = torch.nn.AdaptiveAvgPool2d((1, 1))
+        self.linear = EqualLinear(512, 512, lr_mul=1)
+        modules = []
+        for block in blocks:
+            for bottleneck in block:
+                modules.append(unit_module(bottleneck.in_channel,
+                                           bottleneck.depth,
+                                           bottleneck.stride))
+        self.body = Sequential(*modules)
+        log_size = int(math.log(opts.stylegan_size, 2))
+        self.style_count = 2 * log_size - 2
+
+    def forward(self, x):
+        x = self.input_layer(x)
+        x = self.body(x)
+        x = self.output_pool(x)
+        x = x.view(-1, 512)
+        x = self.linear(x)
+        return x.repeat(self.style_count, 1, 1).permute(1, 0, 2)

Time_TravelRephotography/models/encoder4editing/models/latent_codes_pool.py

@@ -0,0 +1,55 @@
+import random
+import torch
+
+
+class LatentCodesPool:
+    """This class implements latent codes buffer that stores previously generated w latent codes.
+    This buffer enables us to update discriminators using a history of generated w's
+    rather than the ones produced by the latest encoder.
+    """
+
+    def __init__(self, pool_size):
+        """Initialize the ImagePool class
+        Parameters:
+            pool_size (int) -- the size of image buffer, if pool_size=0, no buffer will be created
+        """
+        self.pool_size = pool_size
+        if self.pool_size > 0:  # create an empty pool
+            self.num_ws = 0
+            self.ws = []
+
+    def query(self, ws):
+        """Return w's from the pool.
+        Parameters:
+            ws: the latest generated w's from the generator
+        Returns w's from the buffer.
+        By 50/100, the buffer will return input w's.
+        By 50/100, the buffer will return w's previously stored in the buffer,
+        and insert the current w's to the buffer.
+        """
+        if self.pool_size == 0:  # if the buffer size is 0, do nothing
+            return ws
+        return_ws = []
+        for w in ws:  # ws.shape: (batch, 512) or (batch, n_latent, 512)
+            # w = torch.unsqueeze(image.data, 0)
+            if w.ndim == 2:
+                i = random.randint(0, len(w) - 1)  # apply a random latent index as a candidate
+                w = w[i]
+            self.handle_w(w, return_ws)
+        return_ws = torch.stack(return_ws, 0)   # collect all the images and return
+        return return_ws
+
+    def handle_w(self, w, return_ws):
+        if self.num_ws < self.pool_size:  # if the buffer is not full; keep inserting current codes to the buffer
+            self.num_ws = self.num_ws + 1
+            self.ws.append(w)
+            return_ws.append(w)
+        else:
+            p = random.uniform(0, 1)
+            if p > 0.5:  # by 50% chance, the buffer will return a previously stored latent code, and insert the current code into the buffer
+                random_id = random.randint(0, self.pool_size - 1)  # randint is inclusive
+                tmp = self.ws[random_id].clone()
+                self.ws[random_id] = w
+                return_ws.append(tmp)
+            else:  # by another 50% chance, the buffer will return the current image
+                return_ws.append(w)

Time_TravelRephotography/models/encoder4editing/models/psp.py

@@ -0,0 +1,100 @@
+import matplotlib
+
+matplotlib.use('Agg')
+import torch
+from torch import nn
+from .encoders import psp_encoders
+from .stylegan2.model import Generator
+from ..configs.paths_config import model_paths
+
+
+def get_keys(d, name):
+    if 'state_dict' in d:
+        d = d['state_dict']
+    d_filt = {k[len(name) + 1:]: v for k, v in d.items() if k[:len(name)] == name}
+    return d_filt
+
+
+class pSp(nn.Module):
+
+    def __init__(self, opts):
+        super(pSp, self).__init__()
+        self.opts = opts
+        # Define architecture
+        self.encoder = self.set_encoder()
+        self.decoder = Generator(opts.stylegan_size, 512, 8, channel_multiplier=2)
+        self.face_pool = torch.nn.AdaptiveAvgPool2d((256, 256))
+        # Load weights if needed
+        self.load_weights()
+
+    def set_encoder(self):
+        if self.opts.encoder_type == 'GradualStyleEncoder':
+            encoder = psp_encoders.GradualStyleEncoder(50, 'ir_se', self.opts)
+        elif self.opts.encoder_type == 'Encoder4Editing':
+            encoder = psp_encoders.Encoder4Editing(50, 'ir_se', self.opts)
+        elif self.opts.encoder_type == 'SingleStyleCodeEncoder':
+            encoder = psp_encoders.BackboneEncoderUsingLastLayerIntoW(50, 'ir_se', self.opts)
+        else:
+            raise Exception('{} is not a valid encoders'.format(self.opts.encoder_type))
+        return encoder
+
+    def load_weights(self):
+        if self.opts.checkpoint_path is not None:
+            print('Loading e4e over the pSp framework from checkpoint: {}'.format(self.opts.checkpoint_path))
+            ckpt = torch.load(self.opts.checkpoint_path, map_location='cpu')
+            self.encoder.load_state_dict(get_keys(ckpt, 'encoder'), strict=True)
+            self.decoder.load_state_dict(get_keys(ckpt, 'decoder'), strict=True)
+            self.__load_latent_avg(ckpt)
+        else:
+            print('Loading encoders weights from irse50!')
+            encoder_ckpt = torch.load(model_paths['ir_se50'])
+            self.encoder.load_state_dict(encoder_ckpt, strict=False)
+            print('Loading decoder weights from pretrained!')
+            ckpt = torch.load(self.opts.stylegan_weights)
+            self.decoder.load_state_dict(ckpt['g_ema'], strict=False)
+            self.__load_latent_avg(ckpt, repeat=self.encoder.style_count)
+
+    def forward(self, x, resize=True, latent_mask=None, input_code=False, randomize_noise=True,
+                inject_latent=None, return_latents=False, alpha=None):
+        if input_code:
+            codes = x
+        else:
+            codes = self.encoder(x)
+            # normalize with respect to the center of an average face
+            if self.opts.start_from_latent_avg:
+                if codes.ndim == 2:
+                    codes = codes + self.latent_avg.repeat(codes.shape[0], 1, 1)[:, 0, :]
+                else:
+                    codes = codes + self.latent_avg.repeat(codes.shape[0], 1, 1)
+
+        if latent_mask is not None:
+            for i in latent_mask:
+                if inject_latent is not None:
+                    if alpha is not None:
+                        codes[:, i] = alpha * inject_latent[:, i] + (1 - alpha) * codes[:, i]
+                    else:
+                        codes[:, i] = inject_latent[:, i]
+                else:
+                    codes[:, i] = 0
+
+        input_is_latent = not input_code
+        images, result_latent = self.decoder([codes],
+                                             input_is_latent=input_is_latent,
+                                             randomize_noise=randomize_noise,
+                                             return_latents=return_latents)
+
+        if resize:
+            images = self.face_pool(images)
+
+        if return_latents:
+            return images, result_latent
+        else:
+            return images
+
+    def __load_latent_avg(self, ckpt, repeat=None):
+        if 'latent_avg' in ckpt:
+            self.latent_avg = ckpt['latent_avg'].to(self.opts.device)
+            if repeat is not None:
+                self.latent_avg = self.latent_avg.repeat(repeat, 1)
+        else:
+            self.latent_avg = None

Time_TravelRephotography/models/encoder4editing/models/stylegan2/model.py

@@ -0,0 +1,673 @@
+import math
+import random
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from .op import FusedLeakyReLU, fused_leaky_relu, upfirdn2d
+
+
+class PixelNorm(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, input):
+        return input * torch.rsqrt(torch.mean(input ** 2, dim=1, keepdim=True) + 1e-8)
+
+
+def make_kernel(k):
+    k = torch.tensor(k, dtype=torch.float32)
+
+    if k.ndim == 1:
+        k = k[None, :] * k[:, None]
+
+    k /= k.sum()
+
+    return k
+
+
+class Upsample(nn.Module):
+    def __init__(self, kernel, factor=2):
+        super().__init__()
+
+        self.factor = factor
+        kernel = make_kernel(kernel) * (factor ** 2)
+        self.register_buffer('kernel', kernel)
+
+        p = kernel.shape[0] - factor
+
+        pad0 = (p + 1) // 2 + factor - 1
+        pad1 = p // 2
+
+        self.pad = (pad0, pad1)
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, up=self.factor, down=1, pad=self.pad)
+
+        return out
+
+
+class Downsample(nn.Module):
+    def __init__(self, kernel, factor=2):
+        super().__init__()
+
+        self.factor = factor
+        kernel = make_kernel(kernel)
+        self.register_buffer('kernel', kernel)
+
+        p = kernel.shape[0] - factor
+
+        pad0 = (p + 1) // 2
+        pad1 = p // 2
+
+        self.pad = (pad0, pad1)
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, up=1, down=self.factor, pad=self.pad)
+
+        return out
+
+
+class Blur(nn.Module):
+    def __init__(self, kernel, pad, upsample_factor=1):
+        super().__init__()
+
+        kernel = make_kernel(kernel)
+
+        if upsample_factor > 1:
+            kernel = kernel * (upsample_factor ** 2)
+
+        self.register_buffer('kernel', kernel)
+
+        self.pad = pad
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, pad=self.pad)
+
+        return out
+
+
+class EqualConv2d(nn.Module):
+    def __init__(
+            self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=True
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(
+            torch.randn(out_channel, in_channel, kernel_size, kernel_size)
+        )
+        self.scale = 1 / math.sqrt(in_channel * kernel_size ** 2)
+
+        self.stride = stride
+        self.padding = padding
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channel))
+
+        else:
+            self.bias = None
+
+    def forward(self, input):
+        out = F.conv2d(
+            input,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]},'
+            f' {self.weight.shape[2]}, stride={self.stride}, padding={self.padding})'
+        )
+
+
+class EqualLinear(nn.Module):
+    def __init__(
+            self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
+
+        else:
+            self.bias = None
+
+        self.activation = activation
+
+        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
+        self.lr_mul = lr_mul
+
+    def forward(self, input):
+        if self.activation:
+            out = F.linear(input, self.weight * self.scale)
+            out = fused_leaky_relu(out, self.bias * self.lr_mul)
+
+        else:
+            out = F.linear(
+                input, self.weight * self.scale, bias=self.bias * self.lr_mul
+            )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]})'
+        )
+
+
+class ScaledLeakyReLU(nn.Module):
+    def __init__(self, negative_slope=0.2):
+        super().__init__()
+
+        self.negative_slope = negative_slope
+
+    def forward(self, input):
+        out = F.leaky_relu(input, negative_slope=self.negative_slope)
+
+        return out * math.sqrt(2)
+
+
+class ModulatedConv2d(nn.Module):
+    def __init__(
+            self,
+            in_channel,
+            out_channel,
+            kernel_size,
+            style_dim,
+            demodulate=True,
+            upsample=False,
+            downsample=False,
+            blur_kernel=[1, 3, 3, 1],
+    ):
+        super().__init__()
+
+        self.eps = 1e-8
+        self.kernel_size = kernel_size
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+        self.upsample = upsample
+        self.downsample = downsample
+
+        if upsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) - (kernel_size - 1)
+            pad0 = (p + 1) // 2 + factor - 1
+            pad1 = p // 2 + 1
+
+            self.blur = Blur(blur_kernel, pad=(pad0, pad1), upsample_factor=factor)
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            pad0 = (p + 1) // 2
+            pad1 = p // 2
+
+            self.blur = Blur(blur_kernel, pad=(pad0, pad1))
+
+        fan_in = in_channel * kernel_size ** 2
+        self.scale = 1 / math.sqrt(fan_in)
+        self.padding = kernel_size // 2
+
+        self.weight = nn.Parameter(
+            torch.randn(1, out_channel, in_channel, kernel_size, kernel_size)
+        )
+
+        self.modulation = EqualLinear(style_dim, in_channel, bias_init=1)
+
+        self.demodulate = demodulate
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.in_channel}, {self.out_channel}, {self.kernel_size}, '
+            f'upsample={self.upsample}, downsample={self.downsample})'
+        )
+
+    def forward(self, input, style):
+        batch, in_channel, height, width = input.shape
+
+        style = self.modulation(style).view(batch, 1, in_channel, 1, 1)
+        weight = self.scale * self.weight * style
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + 1e-8)
+            weight = weight * demod.view(batch, self.out_channel, 1, 1, 1)
+
+        weight = weight.view(
+            batch * self.out_channel, in_channel, self.kernel_size, self.kernel_size
+        )
+
+        if self.upsample:
+            input = input.view(1, batch * in_channel, height, width)
+            weight = weight.view(
+                batch, self.out_channel, in_channel, self.kernel_size, self.kernel_size
+            )
+            weight = weight.transpose(1, 2).reshape(
+                batch * in_channel, self.out_channel, self.kernel_size, self.kernel_size
+            )
+            out = F.conv_transpose2d(input, weight, padding=0, stride=2, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+            out = self.blur(out)
+
+        elif self.downsample:
+            input = self.blur(input)
+            _, _, height, width = input.shape
+            input = input.view(1, batch * in_channel, height, width)
+            out = F.conv2d(input, weight, padding=0, stride=2, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+
+        else:
+            input = input.view(1, batch * in_channel, height, width)
+            out = F.conv2d(input, weight, padding=self.padding, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+
+        return out
+
+
+class NoiseInjection(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.zeros(1))
+
+    def forward(self, image, noise=None):
+        if noise is None:
+            batch, _, height, width = image.shape
+            noise = image.new_empty(batch, 1, height, width).normal_()
+
+        return image + self.weight * noise
+
+
+class ConstantInput(nn.Module):
+    def __init__(self, channel, size=4):
+        super().__init__()
+
+        self.input = nn.Parameter(torch.randn(1, channel, size, size))
+
+    def forward(self, input):
+        batch = input.shape[0]
+        out = self.input.repeat(batch, 1, 1, 1)
+
+        return out
+
+
+class StyledConv(nn.Module):
+    def __init__(
+            self,
+            in_channel,
+            out_channel,
+            kernel_size,
+            style_dim,
+            upsample=False,
+            blur_kernel=[1, 3, 3, 1],
+            demodulate=True,
+    ):
+        super().__init__()
+
+        self.conv = ModulatedConv2d(
+            in_channel,
+            out_channel,
+            kernel_size,
+            style_dim,
+            upsample=upsample,
+            blur_kernel=blur_kernel,
+            demodulate=demodulate,
+        )
+
+        self.noise = NoiseInjection()
+        # self.bias = nn.Parameter(torch.zeros(1, out_channel, 1, 1))
+        # self.activate = ScaledLeakyReLU(0.2)
+        self.activate = FusedLeakyReLU(out_channel)
+
+    def forward(self, input, style, noise=None):
+        out = self.conv(input, style)
+        out = self.noise(out, noise=noise)
+        # out = out + self.bias
+        out = self.activate(out)
+
+        return out
+
+
+class ToRGB(nn.Module):
+    def __init__(self, in_channel, style_dim, upsample=True, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        if upsample:
+            self.upsample = Upsample(blur_kernel)
+
+        self.conv = ModulatedConv2d(in_channel, 3, 1, style_dim, demodulate=False)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, input, style, skip=None):
+        out = self.conv(input, style)
+        out = out + self.bias
+
+        if skip is not None:
+            skip = self.upsample(skip)
+
+            out = out + skip
+
+        return out
+
+
+class Generator(nn.Module):
+    def __init__(
+            self,
+            size,
+            style_dim,
+            n_mlp,
+            channel_multiplier=2,
+            blur_kernel=[1, 3, 3, 1],
+            lr_mlp=0.01,
+    ):
+        super().__init__()
+
+        self.size = size
+
+        self.style_dim = style_dim
+
+        layers = [PixelNorm()]
+
+        for i in range(n_mlp):
+            layers.append(
+                EqualLinear(
+                    style_dim, style_dim, lr_mul=lr_mlp, activation='fused_lrelu'
+                )
+            )
+
+        self.style = nn.Sequential(*layers)
+
+        self.channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        self.input = ConstantInput(self.channels[4])
+        self.conv1 = StyledConv(
+            self.channels[4], self.channels[4], 3, style_dim, blur_kernel=blur_kernel
+        )
+        self.to_rgb1 = ToRGB(self.channels[4], style_dim, upsample=False)
+
+        self.log_size = int(math.log(size, 2))
+        self.num_layers = (self.log_size - 2) * 2 + 1
+
+        self.convs = nn.ModuleList()
+        self.upsamples = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+
+        in_channel = self.channels[4]
+
+        for layer_idx in range(self.num_layers):
+            res = (layer_idx + 5) // 2
+            shape = [1, 1, 2 ** res, 2 ** res]
+            self.noises.register_buffer(f'noise_{layer_idx}', torch.randn(*shape))
+
+        for i in range(3, self.log_size + 1):
+            out_channel = self.channels[2 ** i]
+
+            self.convs.append(
+                StyledConv(
+                    in_channel,
+                    out_channel,
+                    3,
+                    style_dim,
+                    upsample=True,
+                    blur_kernel=blur_kernel,
+                )
+            )
+
+            self.convs.append(
+                StyledConv(
+                    out_channel, out_channel, 3, style_dim, blur_kernel=blur_kernel
+                )
+            )
+
+            self.to_rgbs.append(ToRGB(out_channel, style_dim))
+
+            in_channel = out_channel
+
+        self.n_latent = self.log_size * 2 - 2
+
+    def make_noise(self):
+        device = self.input.input.device
+
+        noises = [torch.randn(1, 1, 2 ** 2, 2 ** 2, device=device)]
+
+        for i in range(3, self.log_size + 1):
+            for _ in range(2):
+                noises.append(torch.randn(1, 1, 2 ** i, 2 ** i, device=device))
+
+        return noises
+
+    def mean_latent(self, n_latent):
+        latent_in = torch.randn(
+            n_latent, self.style_dim, device=self.input.input.device
+        )
+        latent = self.style(latent_in).mean(0, keepdim=True)
+
+        return latent
+
+    def get_latent(self, input):
+        return self.style(input)
+
+    def forward(
+            self,
+            styles,
+            return_latents=False,
+            return_features=False,
+            inject_index=None,
+            truncation=1,
+            truncation_latent=None,
+            input_is_latent=False,
+            noise=None,
+            randomize_noise=True,
+    ):
+        if not input_is_latent:
+            styles = [self.style(s) for s in styles]
+
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers
+            else:
+                noise = [
+                    getattr(self.noises, f'noise_{i}') for i in range(self.num_layers)
+                ]
+
+        if truncation < 1:
+            style_t = []
+
+            for style in styles:
+                style_t.append(
+                    truncation_latent + truncation * (style - truncation_latent)
+                )
+
+            styles = style_t
+
+        if len(styles) < 2:
+            inject_index = self.n_latent
+
+            if styles[0].ndim < 3:
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:
+                latent = styles[0]
+
+        else:
+            if inject_index is None:
+                inject_index = random.randint(1, self.n_latent - 1)
+
+            latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.n_latent - inject_index, 1)
+
+            latent = torch.cat([latent, latent2], 1)
+
+        out = self.input(latent)
+        out = self.conv1(out, latent[:, 0], noise=noise[0])
+
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(
+                self.convs[::2], self.convs[1::2], noise[1::2], noise[2::2], self.to_rgbs
+        ):
+            out = conv1(out, latent[:, i], noise=noise1)
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)
+
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        elif return_features:
+            return image, out
+        else:
+            return image, None
+
+
+class ConvLayer(nn.Sequential):
+    def __init__(
+            self,
+            in_channel,
+            out_channel,
+            kernel_size,
+            downsample=False,
+            blur_kernel=[1, 3, 3, 1],
+            bias=True,
+            activate=True,
+    ):
+        layers = []
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            pad0 = (p + 1) // 2
+            pad1 = p // 2
+
+            layers.append(Blur(blur_kernel, pad=(pad0, pad1)))
+
+            stride = 2
+            self.padding = 0
+
+        else:
+            stride = 1
+            self.padding = kernel_size // 2
+
+        layers.append(
+            EqualConv2d(
+                in_channel,
+                out_channel,
+                kernel_size,
+                padding=self.padding,
+                stride=stride,
+                bias=bias and not activate,
+            )
+        )
+
+        if activate:
+            if bias:
+                layers.append(FusedLeakyReLU(out_channel))
+
+            else:
+                layers.append(ScaledLeakyReLU(0.2))
+
+        super().__init__(*layers)
+
+
+class ResBlock(nn.Module):
+    def __init__(self, in_channel, out_channel, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        self.conv1 = ConvLayer(in_channel, in_channel, 3)
+        self.conv2 = ConvLayer(in_channel, out_channel, 3, downsample=True)
+
+        self.skip = ConvLayer(
+            in_channel, out_channel, 1, downsample=True, activate=False, bias=False
+        )
+
+    def forward(self, input):
+        out = self.conv1(input)
+        out = self.conv2(out)
+
+        skip = self.skip(input)
+        out = (out + skip) / math.sqrt(2)
+
+        return out
+
+
+class Discriminator(nn.Module):
+    def __init__(self, size, channel_multiplier=2, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        convs = [ConvLayer(3, channels[size], 1)]
+
+        log_size = int(math.log(size, 2))
+
+        in_channel = channels[size]
+
+        for i in range(log_size, 2, -1):
+            out_channel = channels[2 ** (i - 1)]
+
+            convs.append(ResBlock(in_channel, out_channel, blur_kernel))
+
+            in_channel = out_channel
+
+        self.convs = nn.Sequential(*convs)
+
+        self.stddev_group = 4
+        self.stddev_feat = 1
+
+        self.final_conv = ConvLayer(in_channel + 1, channels[4], 3)
+        self.final_linear = nn.Sequential(
+            EqualLinear(channels[4] * 4 * 4, channels[4], activation='fused_lrelu'),
+            EqualLinear(channels[4], 1),
+        )
+
+    def forward(self, input):
+        out = self.convs(input)
+
+        batch, channel, height, width = out.shape
+        group = min(batch, self.stddev_group)
+        stddev = out.view(
+            group, -1, self.stddev_feat, channel // self.stddev_feat, height, width
+        )
+        stddev = torch.sqrt(stddev.var(0, unbiased=False) + 1e-8)
+        stddev = stddev.mean([2, 3, 4], keepdims=True).squeeze(2)
+        stddev = stddev.repeat(group, 1, height, width)
+        out = torch.cat([out, stddev], 1)
+
+        out = self.final_conv(out)
+
+        out = out.view(batch, -1)
+        out = self.final_linear(out)
+
+        return out

Time_TravelRephotography/models/encoder4editing/models/stylegan2/op/__init__.py

@@ -0,0 +1,2 @@
+from .fused_act import FusedLeakyReLU, fused_leaky_relu
+from .upfirdn2d import upfirdn2d

Time_TravelRephotography/models/encoder4editing/models/stylegan2/op/fused_act.py

@@ -0,0 +1,85 @@
+import os
+
+import torch
+from torch import nn
+from torch.autograd import Function
+from torch.utils.cpp_extension import load
+
+module_path = os.path.dirname(__file__)
+fused = load(
+    'fused',
+    sources=[
+        os.path.join(module_path, 'fused_bias_act.cpp'),
+        os.path.join(module_path, 'fused_bias_act_kernel.cu'),
+    ],
+)
+
+
+class FusedLeakyReLUFunctionBackward(Function):
+    @staticmethod
+    def forward(ctx, grad_output, out, negative_slope, scale):
+        ctx.save_for_backward(out)
+        ctx.negative_slope = negative_slope
+        ctx.scale = scale
+
+        empty = grad_output.new_empty(0)
+
+        grad_input = fused.fused_bias_act(
+            grad_output, empty, out, 3, 1, negative_slope, scale
+        )
+
+        dim = [0]
+
+        if grad_input.ndim > 2:
+            dim += list(range(2, grad_input.ndim))
+
+        grad_bias = grad_input.sum(dim).detach()
+
+        return grad_input, grad_bias
+
+    @staticmethod
+    def backward(ctx, gradgrad_input, gradgrad_bias):
+        out, = ctx.saved_tensors
+        gradgrad_out = fused.fused_bias_act(
+            gradgrad_input, gradgrad_bias, out, 3, 1, ctx.negative_slope, ctx.scale
+        )
+
+        return gradgrad_out, None, None, None
+
+
+class FusedLeakyReLUFunction(Function):
+    @staticmethod
+    def forward(ctx, input, bias, negative_slope, scale):
+        empty = input.new_empty(0)
+        out = fused.fused_bias_act(input, bias, empty, 3, 0, negative_slope, scale)
+        ctx.save_for_backward(out)
+        ctx.negative_slope = negative_slope
+        ctx.scale = scale
+
+        return out
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        out, = ctx.saved_tensors
+
+        grad_input, grad_bias = FusedLeakyReLUFunctionBackward.apply(
+            grad_output, out, ctx.negative_slope, ctx.scale
+        )
+
+        return grad_input, grad_bias, None, None
+
+
+class FusedLeakyReLU(nn.Module):
+    def __init__(self, channel, negative_slope=0.2, scale=2 ** 0.5):
+        super().__init__()
+
+        self.bias = nn.Parameter(torch.zeros(channel))
+        self.negative_slope = negative_slope
+        self.scale = scale
+
+    def forward(self, input):
+        return fused_leaky_relu(input, self.bias, self.negative_slope, self.scale)
+
+
+def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2 ** 0.5):
+    return FusedLeakyReLUFunction.apply(input, bias, negative_slope, scale)

Time_TravelRephotography/models/encoder4editing/models/stylegan2/op/fused_bias_act.cpp

@@ -0,0 +1,21 @@
+#include <torch/extension.h>
+
+
+torch::Tensor fused_bias_act_op(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer,
+    int act, int grad, float alpha, float scale);
+
+#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+
+torch::Tensor fused_bias_act(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer,
+    int act, int grad, float alpha, float scale) {
+    CHECK_CUDA(input);
+    CHECK_CUDA(bias);
+
+    return fused_bias_act_op(input, bias, refer, act, grad, alpha, scale);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("fused_bias_act", &fused_bias_act, "fused bias act (CUDA)");
+}

Time_TravelRephotography/models/encoder4editing/models/stylegan2/op/fused_bias_act_kernel.cu

@@ -0,0 +1,99 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#include <torch/types.h>
+
+#include <ATen/ATen.h>
+#include <ATen/AccumulateType.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+
+template <typename scalar_t>
+static __global__ void fused_bias_act_kernel(scalar_t* out, const scalar_t* p_x, const scalar_t* p_b, const scalar_t* p_ref,
+    int act, int grad, scalar_t alpha, scalar_t scale, int loop_x, int size_x, int step_b, int size_b, int use_bias, int use_ref) {
+    int xi = blockIdx.x * loop_x * blockDim.x + threadIdx.x;
+
+    scalar_t zero = 0.0;
+
+    for (int loop_idx = 0; loop_idx < loop_x && xi < size_x; loop_idx++, xi += blockDim.x) {
+        scalar_t x = p_x[xi];
+
+        if (use_bias) {
+            x += p_b[(xi / step_b) % size_b];
+        }
+
+        scalar_t ref = use_ref ? p_ref[xi] : zero;
+
+        scalar_t y;
+
+        switch (act * 10 + grad) {
+            default:
+            case 10: y = x; break;
+            case 11: y = x; break;
+            case 12: y = 0.0; break;
+
+            case 30: y = (x > 0.0) ? x : x * alpha; break;
+            case 31: y = (ref > 0.0) ? x : x * alpha; break;
+            case 32: y = 0.0; break;
+        }
+
+        out[xi] = y * scale;
+    }
+}
+
+
+torch::Tensor fused_bias_act_op(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer,
+    int act, int grad, float alpha, float scale) {
+    int curDevice = -1;
+    cudaGetDevice(&curDevice);
+    cudaStream_t stream = at::cuda::getCurrentCUDAStream(curDevice);
+
+    auto x = input.contiguous();
+    auto b = bias.contiguous();
+    auto ref = refer.contiguous();
+
+    int use_bias = b.numel() ? 1 : 0;
+    int use_ref = ref.numel() ? 1 : 0;
+
+    int size_x = x.numel();
+    int size_b = b.numel();
+    int step_b = 1;
+
+    for (int i = 1 + 1; i < x.dim(); i++) {
+        step_b *= x.size(i);
+    }
+
+    int loop_x = 4;
+    int block_size = 4 * 32;
+    int grid_size = (size_x - 1) / (loop_x * block_size) + 1;
+
+    auto y = torch::empty_like(x);
+
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "fused_bias_act_kernel", [&] {
+        fused_bias_act_kernel<scalar_t><<<grid_size, block_size, 0, stream>>>(
+            y.data_ptr<scalar_t>(),
+            x.data_ptr<scalar_t>(),
+            b.data_ptr<scalar_t>(),
+            ref.data_ptr<scalar_t>(),
+            act,
+            grad,
+            alpha,
+            scale,
+            loop_x,
+            size_x,
+            step_b,
+            size_b,
+            use_bias,
+            use_ref
+        );
+    });
+
+    return y;
+}

Time_TravelRephotography/models/encoder4editing/models/stylegan2/op/upfirdn2d.cpp

@@ -0,0 +1,23 @@
+#include <torch/extension.h>
+
+
+torch::Tensor upfirdn2d_op(const torch::Tensor& input, const torch::Tensor& kernel,
+                            int up_x, int up_y, int down_x, int down_y,
+                            int pad_x0, int pad_x1, int pad_y0, int pad_y1);
+
+#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+
+torch::Tensor upfirdn2d(const torch::Tensor& input, const torch::Tensor& kernel,
+                        int up_x, int up_y, int down_x, int down_y,
+                        int pad_x0, int pad_x1, int pad_y0, int pad_y1) {
+    CHECK_CUDA(input);
+    CHECK_CUDA(kernel);
+
+    return upfirdn2d_op(input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("upfirdn2d", &upfirdn2d, "upfirdn2d (CUDA)");
+}

Time_TravelRephotography/models/encoder4editing/models/stylegan2/op/upfirdn2d.py

@@ -0,0 +1,184 @@
+import os
+
+import torch
+from torch.autograd import Function
+from torch.utils.cpp_extension import load
+
+module_path = os.path.dirname(__file__)
+upfirdn2d_op = load(
+    'upfirdn2d',
+    sources=[
+        os.path.join(module_path, 'upfirdn2d.cpp'),
+        os.path.join(module_path, 'upfirdn2d_kernel.cu'),
+    ],
+)
+
+
+class UpFirDn2dBackward(Function):
+    @staticmethod
+    def forward(
+            ctx, grad_output, kernel, grad_kernel, up, down, pad, g_pad, in_size, out_size
+    ):
+        up_x, up_y = up
+        down_x, down_y = down
+        g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1 = g_pad
+
+        grad_output = grad_output.reshape(-1, out_size[0], out_size[1], 1)
+
+        grad_input = upfirdn2d_op.upfirdn2d(
+            grad_output,
+            grad_kernel,
+            down_x,
+            down_y,
+            up_x,
+            up_y,
+            g_pad_x0,
+            g_pad_x1,
+            g_pad_y0,
+            g_pad_y1,
+        )
+        grad_input = grad_input.view(in_size[0], in_size[1], in_size[2], in_size[3])
+
+        ctx.save_for_backward(kernel)
+
+        pad_x0, pad_x1, pad_y0, pad_y1 = pad
+
+        ctx.up_x = up_x
+        ctx.up_y = up_y
+        ctx.down_x = down_x
+        ctx.down_y = down_y
+        ctx.pad_x0 = pad_x0
+        ctx.pad_x1 = pad_x1
+        ctx.pad_y0 = pad_y0
+        ctx.pad_y1 = pad_y1
+        ctx.in_size = in_size
+        ctx.out_size = out_size
+
+        return grad_input
+
+    @staticmethod
+    def backward(ctx, gradgrad_input):
+        kernel, = ctx.saved_tensors
+
+        gradgrad_input = gradgrad_input.reshape(-1, ctx.in_size[2], ctx.in_size[3], 1)
+
+        gradgrad_out = upfirdn2d_op.upfirdn2d(
+            gradgrad_input,
+            kernel,
+            ctx.up_x,
+            ctx.up_y,
+            ctx.down_x,
+            ctx.down_y,
+            ctx.pad_x0,
+            ctx.pad_x1,
+            ctx.pad_y0,
+            ctx.pad_y1,
+        )
+        # gradgrad_out = gradgrad_out.view(ctx.in_size[0], ctx.out_size[0], ctx.out_size[1], ctx.in_size[3])
+        gradgrad_out = gradgrad_out.view(
+            ctx.in_size[0], ctx.in_size[1], ctx.out_size[0], ctx.out_size[1]
+        )
+
+        return gradgrad_out, None, None, None, None, None, None, None, None
+
+
+class UpFirDn2d(Function):
+    @staticmethod
+    def forward(ctx, input, kernel, up, down, pad):
+        up_x, up_y = up
+        down_x, down_y = down
+        pad_x0, pad_x1, pad_y0, pad_y1 = pad
+
+        kernel_h, kernel_w = kernel.shape
+        batch, channel, in_h, in_w = input.shape
+        ctx.in_size = input.shape
+
+        input = input.reshape(-1, in_h, in_w, 1)
+
+        ctx.save_for_backward(kernel, torch.flip(kernel, [0, 1]))
+
+        out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
+        out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1
+        ctx.out_size = (out_h, out_w)
+
+        ctx.up = (up_x, up_y)
+        ctx.down = (down_x, down_y)
+        ctx.pad = (pad_x0, pad_x1, pad_y0, pad_y1)
+
+        g_pad_x0 = kernel_w - pad_x0 - 1
+        g_pad_y0 = kernel_h - pad_y0 - 1
+        g_pad_x1 = in_w * up_x - out_w * down_x + pad_x0 - up_x + 1
+        g_pad_y1 = in_h * up_y - out_h * down_y + pad_y0 - up_y + 1
+
+        ctx.g_pad = (g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1)
+
+        out = upfirdn2d_op.upfirdn2d(
+            input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1
+        )
+        # out = out.view(major, out_h, out_w, minor)
+        out = out.view(-1, channel, out_h, out_w)
+
+        return out
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        kernel, grad_kernel = ctx.saved_tensors
+
+        grad_input = UpFirDn2dBackward.apply(
+            grad_output,
+            kernel,
+            grad_kernel,
+            ctx.up,
+            ctx.down,
+            ctx.pad,
+            ctx.g_pad,
+            ctx.in_size,
+            ctx.out_size,
+        )
+
+        return grad_input, None, None, None, None
+
+
+def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)):
+    out = UpFirDn2d.apply(
+        input, kernel, (up, up), (down, down), (pad[0], pad[1], pad[0], pad[1])
+    )
+
+    return out
+
+
+def upfirdn2d_native(
+        input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1
+):
+    _, in_h, in_w, minor = input.shape
+    kernel_h, kernel_w = kernel.shape
+
+    out = input.view(-1, in_h, 1, in_w, 1, minor)
+    out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
+    out = out.view(-1, in_h * up_y, in_w * up_x, minor)
+
+    out = F.pad(
+        out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)]
+    )
+    out = out[
+          :,
+          max(-pad_y0, 0): out.shape[1] - max(-pad_y1, 0),
+          max(-pad_x0, 0): out.shape[2] - max(-pad_x1, 0),
+          :,
+          ]
+
+    out = out.permute(0, 3, 1, 2)
+    out = out.reshape(
+        [-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1]
+    )
+    w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
+    out = F.conv2d(out, w)
+    out = out.reshape(
+        -1,
+        minor,
+        in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
+        in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
+    )
+    out = out.permute(0, 2, 3, 1)
+
+    return out[:, ::down_y, ::down_x, :]

Time_TravelRephotography/models/encoder4editing/models/stylegan2/op/upfirdn2d_kernel.cu

@@ -0,0 +1,272 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#include <torch/types.h>
+
+#include <ATen/ATen.h>
+#include <ATen/AccumulateType.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+
+static __host__ __device__ __forceinline__ int floor_div(int a, int b) {
+    int c = a / b;
+
+    if (c * b > a) {
+        c--;
+    }
+
+    return c;
+}
+
+
+struct UpFirDn2DKernelParams {
+    int up_x;
+    int up_y;
+    int down_x;
+    int down_y;
+    int pad_x0;
+    int pad_x1;
+    int pad_y0;
+    int pad_y1;
+
+    int major_dim;
+    int in_h;
+    int in_w;
+    int minor_dim;
+    int kernel_h;
+    int kernel_w;
+    int out_h;
+    int out_w;
+    int loop_major;
+    int loop_x;
+};
+
+
+template <typename scalar_t, int up_x, int up_y, int down_x, int down_y, int kernel_h, int kernel_w, int tile_out_h, int tile_out_w>
+__global__ void upfirdn2d_kernel(scalar_t* out, const scalar_t* input, const scalar_t* kernel, const UpFirDn2DKernelParams p) {
+    const int tile_in_h = ((tile_out_h - 1) * down_y + kernel_h - 1) / up_y + 1;
+    const int tile_in_w = ((tile_out_w - 1) * down_x + kernel_w - 1) / up_x + 1;
+
+    __shared__ volatile float sk[kernel_h][kernel_w];
+    __shared__ volatile float sx[tile_in_h][tile_in_w];
+
+    int minor_idx = blockIdx.x;
+    int tile_out_y = minor_idx / p.minor_dim;
+    minor_idx -= tile_out_y * p.minor_dim;
+    tile_out_y *= tile_out_h;
+    int tile_out_x_base = blockIdx.y * p.loop_x * tile_out_w;
+    int major_idx_base = blockIdx.z * p.loop_major;
+
+    if (tile_out_x_base >= p.out_w | tile_out_y >= p.out_h | major_idx_base >= p.major_dim) {
+        return;
+    }
+
+    for (int tap_idx = threadIdx.x; tap_idx < kernel_h * kernel_w; tap_idx += blockDim.x) {
+        int ky = tap_idx / kernel_w;
+        int kx = tap_idx - ky * kernel_w;
+        scalar_t v = 0.0;
+
+        if (kx < p.kernel_w & ky < p.kernel_h) {
+            v = kernel[(p.kernel_h - 1 - ky) * p.kernel_w + (p.kernel_w - 1 - kx)];
+        }
+
+        sk[ky][kx] = v;
+    }
+
+    for (int loop_major = 0, major_idx = major_idx_base; loop_major < p.loop_major & major_idx < p.major_dim; loop_major++, major_idx++) {
+        for (int loop_x = 0, tile_out_x = tile_out_x_base; loop_x < p.loop_x & tile_out_x < p.out_w; loop_x++, tile_out_x += tile_out_w) {
+            int tile_mid_x = tile_out_x * down_x + up_x - 1 - p.pad_x0;
+            int tile_mid_y = tile_out_y * down_y + up_y - 1 - p.pad_y0;
+            int tile_in_x = floor_div(tile_mid_x, up_x);
+            int tile_in_y = floor_div(tile_mid_y, up_y);
+
+            __syncthreads();
+
+            for (int in_idx = threadIdx.x; in_idx < tile_in_h * tile_in_w; in_idx += blockDim.x) {
+                int rel_in_y = in_idx / tile_in_w;
+                int rel_in_x = in_idx - rel_in_y * tile_in_w;
+                int in_x = rel_in_x + tile_in_x;
+                int in_y = rel_in_y + tile_in_y;
+
+                scalar_t v = 0.0;
+
+                if (in_x >= 0 & in_y >= 0 & in_x < p.in_w & in_y < p.in_h) {
+                    v = input[((major_idx * p.in_h + in_y) * p.in_w + in_x) * p.minor_dim + minor_idx];
+                }
+
+                sx[rel_in_y][rel_in_x] = v;
+            }
+
+            __syncthreads();
+            for (int out_idx = threadIdx.x; out_idx < tile_out_h * tile_out_w; out_idx += blockDim.x) {
+                int rel_out_y = out_idx / tile_out_w;
+                int rel_out_x = out_idx - rel_out_y * tile_out_w;
+                int out_x = rel_out_x + tile_out_x;
+                int out_y = rel_out_y + tile_out_y;
+
+                int mid_x = tile_mid_x + rel_out_x * down_x;
+                int mid_y = tile_mid_y + rel_out_y * down_y;
+                int in_x = floor_div(mid_x, up_x);
+                int in_y = floor_div(mid_y, up_y);
+                int rel_in_x = in_x - tile_in_x;
+                int rel_in_y = in_y - tile_in_y;
+                int kernel_x = (in_x + 1) * up_x - mid_x - 1;
+                int kernel_y = (in_y + 1) * up_y - mid_y - 1;
+
+                scalar_t v = 0.0;
+
+                #pragma unroll
+                for (int y = 0; y < kernel_h / up_y; y++)
+                    #pragma unroll
+                    for (int x = 0; x < kernel_w / up_x; x++)
+                        v += sx[rel_in_y + y][rel_in_x + x] * sk[kernel_y + y * up_y][kernel_x + x * up_x];
+
+                if (out_x < p.out_w & out_y < p.out_h) {
+                    out[((major_idx * p.out_h + out_y) * p.out_w + out_x) * p.minor_dim + minor_idx] = v;
+                }
+            }
+        }
+    }
+}
+
+
+torch::Tensor upfirdn2d_op(const torch::Tensor& input, const torch::Tensor& kernel,
+    int up_x, int up_y, int down_x, int down_y,
+    int pad_x0, int pad_x1, int pad_y0, int pad_y1) {
+    int curDevice = -1;
+    cudaGetDevice(&curDevice);
+    cudaStream_t stream = at::cuda::getCurrentCUDAStream(curDevice);
+
+    UpFirDn2DKernelParams p;
+
+    auto x = input.contiguous();
+    auto k = kernel.contiguous();
+
+    p.major_dim = x.size(0);
+    p.in_h = x.size(1);
+    p.in_w = x.size(2);
+    p.minor_dim = x.size(3);
+    p.kernel_h = k.size(0);
+    p.kernel_w = k.size(1);
+    p.up_x = up_x;
+    p.up_y = up_y;
+    p.down_x = down_x;
+    p.down_y = down_y;
+    p.pad_x0 = pad_x0;
+    p.pad_x1 = pad_x1;
+    p.pad_y0 = pad_y0;
+    p.pad_y1 = pad_y1;
+
+    p.out_h = (p.in_h * p.up_y + p.pad_y0 + p.pad_y1 - p.kernel_h + p.down_y) / p.down_y;
+    p.out_w = (p.in_w * p.up_x + p.pad_x0 + p.pad_x1 - p.kernel_w + p.down_x) / p.down_x;
+
+    auto out = at::empty({p.major_dim, p.out_h, p.out_w, p.minor_dim}, x.options());
+
+    int mode = -1;
+
+    int tile_out_h;
+    int tile_out_w;
+
+    if (p.up_x == 1 && p.up_y == 1 && p.down_x == 1 && p.down_y == 1 && p.kernel_h <= 4 && p.kernel_w <= 4) {
+        mode = 1;
+        tile_out_h = 16;
+        tile_out_w = 64;
+    }
+
+    if (p.up_x == 1 && p.up_y == 1 && p.down_x == 1 && p.down_y == 1 && p.kernel_h <= 3 && p.kernel_w <= 3) {
+        mode = 2;
+        tile_out_h = 16;
+        tile_out_w = 64;
+    }
+
+    if (p.up_x == 2 && p.up_y == 2 && p.down_x == 1 && p.down_y == 1 && p.kernel_h <= 4 && p.kernel_w <= 4) {
+        mode = 3;
+        tile_out_h = 16;
+        tile_out_w = 64;
+    }
+
+    if (p.up_x == 2 && p.up_y == 2 && p.down_x == 1 && p.down_y == 1 && p.kernel_h <= 2 && p.kernel_w <= 2) {
+        mode = 4;
+        tile_out_h = 16;
+        tile_out_w = 64;
+    }
+
+    if (p.up_x == 1 && p.up_y == 1 && p.down_x == 2 && p.down_y == 2 && p.kernel_h <= 4 && p.kernel_w <= 4) {
+        mode = 5;
+        tile_out_h = 8;
+        tile_out_w = 32;
+    }
+
+    if (p.up_x == 1 && p.up_y == 1 && p.down_x == 2 && p.down_y == 2 && p.kernel_h <= 2 && p.kernel_w <= 2) {
+        mode = 6;
+        tile_out_h = 8;
+        tile_out_w = 32;
+    }
+
+    dim3 block_size;
+    dim3 grid_size;
+
+    if (tile_out_h > 0 && tile_out_w) {
+        p.loop_major = (p.major_dim - 1) / 16384 + 1;
+        p.loop_x = 1;
+        block_size = dim3(32 * 8, 1, 1);
+        grid_size = dim3(((p.out_h - 1) / tile_out_h + 1) * p.minor_dim,
+                         (p.out_w - 1) / (p.loop_x * tile_out_w) + 1,
+                         (p.major_dim - 1) / p.loop_major + 1);
+    }
+
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "upfirdn2d_cuda", [&] {
+        switch (mode) {
+        case 1:
+            upfirdn2d_kernel<scalar_t, 1, 1, 1, 1, 4, 4, 16, 64><<<grid_size, block_size, 0, stream>>>(
+                out.data_ptr<scalar_t>(), x.data_ptr<scalar_t>(), k.data_ptr<scalar_t>(), p
+            );
+
+            break;
+
+        case 2:
+            upfirdn2d_kernel<scalar_t, 1, 1, 1, 1, 3, 3, 16, 64><<<grid_size, block_size, 0, stream>>>(
+                out.data_ptr<scalar_t>(), x.data_ptr<scalar_t>(), k.data_ptr<scalar_t>(), p
+            );
+
+            break;
+
+        case 3:
+            upfirdn2d_kernel<scalar_t, 2, 2, 1, 1, 4, 4, 16, 64><<<grid_size, block_size, 0, stream>>>(
+                out.data_ptr<scalar_t>(), x.data_ptr<scalar_t>(), k.data_ptr<scalar_t>(), p
+            );
+
+            break;
+
+        case 4:
+            upfirdn2d_kernel<scalar_t, 2, 2, 1, 1, 2, 2, 16, 64><<<grid_size, block_size, 0, stream>>>(
+                out.data_ptr<scalar_t>(), x.data_ptr<scalar_t>(), k.data_ptr<scalar_t>(), p
+            );
+
+            break;
+
+        case 5:
+            upfirdn2d_kernel<scalar_t, 1, 1, 2, 2, 4, 4, 8, 32><<<grid_size, block_size, 0, stream>>>(
+                out.data_ptr<scalar_t>(), x.data_ptr<scalar_t>(), k.data_ptr<scalar_t>(), p
+            );
+
+            break;
+
+        case 6:
+            upfirdn2d_kernel<scalar_t, 1, 1, 2, 2, 4, 4, 8, 32><<<grid_size, block_size, 0, stream>>>(
+                out.data_ptr<scalar_t>(), x.data_ptr<scalar_t>(), k.data_ptr<scalar_t>(), p
+            );
+
+            break;
+        }
+    });
+
+    return out;
+}

Time_TravelRephotography/models/encoder4editing/options/train_options.py

@@ -0,0 +1,84 @@
+from argparse import ArgumentParser
+from configs.paths_config import model_paths
+
+
+class TrainOptions:
+
+    def __init__(self):
+        self.parser = ArgumentParser()
+        self.initialize()
+
+    def initialize(self):
+        self.parser.add_argument('--exp_dir', type=str, help='Path to experiment output directory')
+        self.parser.add_argument('--dataset_type', default='ffhq_encode', type=str,
+                                 help='Type of dataset/experiment to run')
+        self.parser.add_argument('--encoder_type', default='Encoder4Editing', type=str, help='Which encoder to use')
+
+        self.parser.add_argument('--batch_size', default=4, type=int, help='Batch size for training')
+        self.parser.add_argument('--test_batch_size', default=2, type=int, help='Batch size for testing and inference')
+        self.parser.add_argument('--workers', default=4, type=int, help='Number of train dataloader workers')
+        self.parser.add_argument('--test_workers', default=2, type=int,
+                                 help='Number of test/inference dataloader workers')
+
+        self.parser.add_argument('--learning_rate', default=0.0001, type=float, help='Optimizer learning rate')
+        self.parser.add_argument('--optim_name', default='ranger', type=str, help='Which optimizer to use')
+        self.parser.add_argument('--train_decoder', default=False, type=bool, help='Whether to train the decoder model')
+        self.parser.add_argument('--start_from_latent_avg', action='store_true',
+                                 help='Whether to add average latent vector to generate codes from encoder.')
+        self.parser.add_argument('--lpips_type', default='alex', type=str, help='LPIPS backbone')
+
+        self.parser.add_argument('--lpips_lambda', default=0.8, type=float, help='LPIPS loss multiplier factor')
+        self.parser.add_argument('--id_lambda', default=0.1, type=float, help='ID loss multiplier factor')
+        self.parser.add_argument('--l2_lambda', default=1.0, type=float, help='L2 loss multiplier factor')
+
+        self.parser.add_argument('--stylegan_weights', default=model_paths['stylegan_ffhq'], type=str,
+                                 help='Path to StyleGAN model weights')
+        self.parser.add_argument('--stylegan_size', default=1024, type=int,
+                                 help='size of pretrained StyleGAN Generator')
+        self.parser.add_argument('--checkpoint_path', default=None, type=str, help='Path to pSp model checkpoint')
+
+        self.parser.add_argument('--max_steps', default=500000, type=int, help='Maximum number of training steps')
+        self.parser.add_argument('--image_interval', default=100, type=int,
+                                 help='Interval for logging train images during training')
+        self.parser.add_argument('--board_interval', default=50, type=int,
+                                 help='Interval for logging metrics to tensorboard')
+        self.parser.add_argument('--val_interval', default=1000, type=int, help='Validation interval')
+        self.parser.add_argument('--save_interval', default=None, type=int, help='Model checkpoint interval')
+
+        # Discriminator flags
+        self.parser.add_argument('--w_discriminator_lambda', default=0, type=float, help='Dw loss multiplier')
+        self.parser.add_argument('--w_discriminator_lr', default=2e-5, type=float, help='Dw learning rate')
+        self.parser.add_argument("--r1", type=float, default=10, help="weight of the r1 regularization")
+        self.parser.add_argument("--d_reg_every", type=int, default=16,
+                                 help="interval for applying r1 regularization")
+        self.parser.add_argument('--use_w_pool', action='store_true',
+                                 help='Whether to store a latnet codes pool for the discriminator\'s training')
+        self.parser.add_argument("--w_pool_size", type=int, default=50,
+                                 help="W\'s pool size, depends on --use_w_pool")
+
+        # e4e specific
+        self.parser.add_argument('--delta_norm', type=int, default=2, help="norm type of the deltas")
+        self.parser.add_argument('--delta_norm_lambda', type=float, default=2e-4, help="lambda for delta norm loss")
+
+        # Progressive training
+        self.parser.add_argument('--progressive_steps', nargs='+', type=int, default=None,
+                                 help="The training steps of training new deltas. steps[i] starts the delta_i training")
+        self.parser.add_argument('--progressive_start', type=int, default=None,
+                                 help="The training step to start training the deltas, overrides progressive_steps")
+        self.parser.add_argument('--progressive_step_every', type=int, default=2_000,
+                                 help="Amount of training steps for each progressive step")
+
+        # Save additional training info to enable future training continuation from produced checkpoints
+        self.parser.add_argument('--save_training_data', action='store_true',
+                                 help='Save intermediate training data to resume training from the checkpoint')
+        self.parser.add_argument('--sub_exp_dir', default=None, type=str, help='Name of sub experiment directory')
+        self.parser.add_argument('--keep_optimizer', action='store_true',
+                                 help='Whether to continue from the checkpoint\'s optimizer')
+        self.parser.add_argument('--resume_training_from_ckpt', default=None, type=str,
+                                 help='Path to training checkpoint, works when --save_training_data was set to True')
+        self.parser.add_argument('--update_param_list', nargs='+', type=str, default=None,
+                                 help="Name of training parameters to update the loaded training checkpoint")
+
+    def parse(self):
+        opts = self.parser.parse_args()
+        return opts