radames commited on
Commit
d9778ff
1 Parent(s): 730589c

add inversion

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +2 -0
  2. interface/app.py +37 -4
  3. interface/examples/benedict.jpg +0 -0
  4. interface/examples/me.jpg +0 -0
  5. interface/examples/obama.jpg +0 -0
  6. interface/inversion.py +65 -0
  7. interface/model_loader.py +3 -3
  8. interface/pixel2style2pixel/LICENSE +21 -0
  9. interface/pixel2style2pixel/README.md +517 -0
  10. interface/pixel2style2pixel/cog.yaml +32 -0
  11. interface/pixel2style2pixel/configs/__init__.py +0 -0
  12. interface/pixel2style2pixel/configs/data_configs.py +41 -0
  13. interface/pixel2style2pixel/configs/paths_config.py +20 -0
  14. interface/pixel2style2pixel/configs/transforms_config.py +154 -0
  15. interface/pixel2style2pixel/criteria/__init__.py +0 -0
  16. interface/pixel2style2pixel/criteria/id_loss.py +44 -0
  17. interface/pixel2style2pixel/criteria/lpips/__init__.py +0 -0
  18. interface/pixel2style2pixel/criteria/lpips/lpips.py +35 -0
  19. interface/pixel2style2pixel/criteria/lpips/networks.py +96 -0
  20. interface/pixel2style2pixel/criteria/lpips/utils.py +30 -0
  21. interface/pixel2style2pixel/criteria/moco_loss.py +69 -0
  22. interface/pixel2style2pixel/criteria/w_norm.py +14 -0
  23. interface/pixel2style2pixel/datasets/__init__.py +0 -0
  24. interface/pixel2style2pixel/datasets/augmentations.py +110 -0
  25. interface/pixel2style2pixel/datasets/gt_res_dataset.py +32 -0
  26. interface/pixel2style2pixel/datasets/images_dataset.py +33 -0
  27. interface/pixel2style2pixel/datasets/inference_dataset.py +22 -0
  28. interface/pixel2style2pixel/docs/encoding_inputs.jpg +0 -0
  29. interface/pixel2style2pixel/docs/encoding_outputs.jpg +0 -0
  30. interface/pixel2style2pixel/docs/frontalization_inputs.jpg +0 -0
  31. interface/pixel2style2pixel/docs/frontalization_outputs.jpg +0 -0
  32. interface/pixel2style2pixel/docs/seg2image.png +3 -0
  33. interface/pixel2style2pixel/docs/sketch2image.png +3 -0
  34. interface/pixel2style2pixel/docs/super_res_32.jpg +0 -0
  35. interface/pixel2style2pixel/docs/super_res_style_mixing.jpg +0 -0
  36. interface/pixel2style2pixel/docs/teaser.png +3 -0
  37. interface/pixel2style2pixel/docs/toonify_input.jpg +0 -0
  38. interface/pixel2style2pixel/docs/toonify_output.jpg +0 -0
  39. interface/pixel2style2pixel/download-weights.sh +12 -0
  40. interface/pixel2style2pixel/environment/psp_env.yaml +37 -0
  41. interface/pixel2style2pixel/licenses/LICENSE_HuangYG123 +21 -0
  42. interface/pixel2style2pixel/licenses/LICENSE_S-aiueo32 +25 -0
  43. interface/pixel2style2pixel/licenses/LICENSE_TreB1eN +21 -0
  44. interface/pixel2style2pixel/licenses/LICENSE_lessw2020 +201 -0
  45. interface/pixel2style2pixel/licenses/LICENSE_rosinality +21 -0
  46. interface/pixel2style2pixel/models/__init__.py +0 -0
  47. interface/pixel2style2pixel/models/encoders/__init__.py +0 -0
  48. interface/pixel2style2pixel/models/encoders/helpers.py +119 -0
  49. interface/pixel2style2pixel/models/encoders/model_irse.py +84 -0
  50. interface/pixel2style2pixel/models/encoders/psp_encoders.py +186 -0
.gitattributes ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ *.png filter=lfs diff=lfs merge=lfs -text
2
+ *.npy filter=lfs diff=lfs merge=lfs -text
interface/app.py CHANGED
@@ -4,6 +4,7 @@ import sys
4
  sys.path.append(".")
5
  sys.path.append("..")
6
  from model_loader import Model
 
7
  from PIL import Image
8
  import cv2
9
  from huggingface_hub import snapshot_download
@@ -25,7 +26,10 @@ models_files = {
25
  }
26
 
27
  models = {name: Model(models_path + "/" + path) for name, path in models_files.items()}
28
-
 
 
 
29
 
30
  canvas_html = """<draggan-canvas id="canvas-root" style='display:flex;max-width: 500px;margin: 0 auto;'></draggan-canvas>"""
31
  load_js = """
@@ -68,6 +72,13 @@ def random_sample(model_name: str):
68
  return img_pil, model_name, latents
69
 
70
 
 
 
 
 
 
 
 
71
  def transform(model_state, latents_state, dxdysxsy=default_dxdysxsy, dz=0):
72
  if "w1" not in latents_state or "w1_initial" not in latents_state:
73
  raise gr.Error("Generate a random sample first")
@@ -107,7 +118,7 @@ def image_click(evt: gr.SelectData):
107
 
108
 
109
  with gr.Blocks() as block:
110
- model_state = gr.State(value="cat")
111
  latents_state = gr.State({})
112
  gr.Markdown(
113
  """# UserControllableLT: User Controllable Latent Transformer
@@ -128,7 +139,7 @@ Double click to add or remove stop points.
128
  model_name = gr.Dropdown(
129
  choices=list(models_files.keys()),
130
  label="Select Pretrained Model",
131
- value="cat",
132
  )
133
  with gr.Row():
134
  button = gr.Button("Random sample")
@@ -144,7 +155,23 @@ Double click to add or remove stop points.
144
  minimum=-15, maximum=15, step_size=0.01, label="zoom", value=0.0
145
  )
146
  image = gr.Image(type="pil", visible=False, preprocess=False)
147
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
148
  with gr.Column():
149
  html = gr.HTML(canvas_html, label="output")
150
 
@@ -176,6 +203,12 @@ Double click to add or remove stop points.
176
  show_progress=False,
177
  )
178
  image.change(None, inputs=[image], outputs=None, _js=image_change)
 
 
 
 
 
 
179
  block.load(None, None, None, _js=load_js)
180
  block.load(
181
  random_sample, inputs=[model_name], outputs=[image, model_state, latents_state]
 
4
  sys.path.append(".")
5
  sys.path.append("..")
6
  from model_loader import Model
7
+ from inversion import InversionModel
8
  from PIL import Image
9
  import cv2
10
  from huggingface_hub import snapshot_download
 
26
  }
27
 
28
  models = {name: Model(models_path + "/" + path) for name, path in models_files.items()}
29
+ inversion_model = InversionModel(
30
+ models_path + "/psp_ffhq_encode.pt",
31
+ models_path + "/shape_predictor_68_face_landmarks.dat",
32
+ )
33
 
34
  canvas_html = """<draggan-canvas id="canvas-root" style='display:flex;max-width: 500px;margin: 0 auto;'></draggan-canvas>"""
35
  load_js = """
 
72
  return img_pil, model_name, latents
73
 
74
 
75
+ def load_from_img_file(image_path: str):
76
+ img_pil, latents = inversion_model.inference(image_path)
77
+ if RESIZE:
78
+ img_pil = img_pil.resize((128, 128))
79
+ return img_pil, "ffhq", latents
80
+
81
+
82
  def transform(model_state, latents_state, dxdysxsy=default_dxdysxsy, dz=0):
83
  if "w1" not in latents_state or "w1_initial" not in latents_state:
84
  raise gr.Error("Generate a random sample first")
 
118
 
119
 
120
  with gr.Blocks() as block:
121
+ model_state = gr.State(value="ffhq")
122
  latents_state = gr.State({})
123
  gr.Markdown(
124
  """# UserControllableLT: User Controllable Latent Transformer
 
139
  model_name = gr.Dropdown(
140
  choices=list(models_files.keys()),
141
  label="Select Pretrained Model",
142
+ value="ffhq",
143
  )
144
  with gr.Row():
145
  button = gr.Button("Random sample")
 
155
  minimum=-15, maximum=15, step_size=0.01, label="zoom", value=0.0
156
  )
157
  image = gr.Image(type="pil", visible=False, preprocess=False)
158
+ with gr.Accordion(label="Upload your face image", open=False):
159
+ gr.Markdown("<small> This only works on FFHQ model </small>")
160
+ with gr.Row():
161
+ image_path = gr.Image(
162
+ type="filepath", label="input image", interactive=True
163
+ )
164
+ examples = gr.Examples(
165
+ examples=[
166
+ "interface/examples/benedict.jpg",
167
+ "interface/examples/obama.jpg",
168
+ "interface/examples/me.jpg",
169
+ ],
170
+ fn=load_from_img_file,
171
+ run_on_click=True,
172
+ inputs=[image_path],
173
+ outputs=[image, model_state, latents_state],
174
+ )
175
  with gr.Column():
176
  html = gr.HTML(canvas_html, label="output")
177
 
 
203
  show_progress=False,
204
  )
205
  image.change(None, inputs=[image], outputs=None, _js=image_change)
206
+ image_path.upload(
207
+ load_from_img_file,
208
+ inputs=[image_path],
209
+ outputs=[image, model_state, latents_state],
210
+ )
211
+
212
  block.load(None, None, None, _js=load_js)
213
  block.load(
214
  random_sample, inputs=[model_name], outputs=[image, model_state, latents_state]
interface/examples/benedict.jpg ADDED
interface/examples/me.jpg ADDED
interface/examples/obama.jpg ADDED
interface/inversion.py ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from argparse import Namespace
2
+ import time
3
+ import torch
4
+ import torchvision.transforms as transforms
5
+ import dlib
6
+ import numpy as np
7
+ from PIL import Image
8
+
9
+ from pixel2style2pixel.utils.common import tensor2im
10
+ from pixel2style2pixel.models.psp import pSp
11
+ from pixel2style2pixel.scripts.align_all_parallel import align_face
12
+
13
+
14
+ class InversionModel:
15
+ def __init__(self, checkpoint_path: str, dlib_path: str) -> None:
16
+ self.dlib_path = dlib_path
17
+ self.dlib_predictor = dlib.shape_predictor(dlib_path)
18
+
19
+ self.tranform_image = transforms.Compose(
20
+ [
21
+ transforms.Resize((256, 256)),
22
+ transforms.ToTensor(),
23
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
24
+ ]
25
+ )
26
+ ckpt = torch.load(checkpoint_path, map_location="cpu")
27
+ opts = ckpt["opts"]
28
+ opts["checkpoint_path"] = checkpoint_path
29
+ opts["learn_in_w"] = False
30
+ opts["output_size"] = 1024
31
+
32
+ self.opts = Namespace(**opts)
33
+ self.net = pSp(self.opts)
34
+ self.net.eval()
35
+ self.net.cuda()
36
+ print("Model successfully loaded!")
37
+
38
+ def run_alignment(self, image_path: str):
39
+ aligned_image = align_face(filepath=image_path, predictor=self.dlib_predictor)
40
+ print("Aligned image has shape: {}".format(aligned_image.size))
41
+ return aligned_image
42
+
43
+ def inference(self, image_path: str):
44
+ input_image = self.run_alignment(image_path)
45
+ input_image = input_image.resize((256, 256))
46
+ transformed_image = self.tranform_image(input_image)
47
+
48
+ with torch.no_grad():
49
+ tic = time.time()
50
+ result_image, latents = self.net(
51
+ transformed_image.unsqueeze(0).to("cuda").float(),
52
+ return_latents=True,
53
+ randomize_noise=False,
54
+ )
55
+ toc = time.time()
56
+ print("Inference took {:.4f} seconds.".format(toc - tic))
57
+
58
+ res_image = tensor2im(result_image[0])
59
+ return (
60
+ res_image,
61
+ {
62
+ "w1": latents.cpu().detach().numpy(),
63
+ "w1_initial": latents.cpu().detach().numpy(),
64
+ },
65
+ )
interface/model_loader.py CHANGED
@@ -12,7 +12,7 @@ class Model:
12
  ):
13
  self.truncation = truncation
14
  self.use_average_code_as_input = use_average_code_as_input
15
- ckpt = torch.load(checkpoint_path, map_location="cuda")
16
  opts = ckpt["opts"]
17
  opts["checkpoint_path"] = checkpoint_path
18
  self.opts = Namespace(**ckpt["opts"])
@@ -84,7 +84,7 @@ class Model:
84
 
85
  dxyz = np.array([dxy[0], dxy[1], dz], dtype=np.float32)
86
  dxy_norm = np.linalg.norm(dxyz[:2], ord=2)
87
- epsilon = 1e-8
88
  dxy_norm = dxy_norm + epsilon
89
  dxyz[:2] = dxyz[:2] / dxy_norm
90
  vec_num = dxy_norm / 10
@@ -166,7 +166,7 @@ class Model:
166
  result,
167
  {
168
  "w1": w1_new.cpu().detach().numpy(),
169
- "w1_initial": w1_new.cpu().detach().numpy(),
170
  },
171
  )
172
 
 
12
  ):
13
  self.truncation = truncation
14
  self.use_average_code_as_input = use_average_code_as_input
15
+ ckpt = torch.load(checkpoint_path, map_location="cpu")
16
  opts = ckpt["opts"]
17
  opts["checkpoint_path"] = checkpoint_path
18
  self.opts = Namespace(**ckpt["opts"])
 
84
 
85
  dxyz = np.array([dxy[0], dxy[1], dz], dtype=np.float32)
86
  dxy_norm = np.linalg.norm(dxyz[:2], ord=2)
87
+ epsilon = 1e-8
88
  dxy_norm = dxy_norm + epsilon
89
  dxyz[:2] = dxyz[:2] / dxy_norm
90
  vec_num = dxy_norm / 10
 
166
  result,
167
  {
168
  "w1": w1_new.cpu().detach().numpy(),
169
+ "w1_initial": w1_initial.cpu().detach().numpy(),
170
  },
171
  )
172
 
interface/pixel2style2pixel/LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2020 Elad Richardson, Yuval Alaluf
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
interface/pixel2style2pixel/README.md ADDED
@@ -0,0 +1,517 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation
2
+ <a href="https://arxiv.org/abs/2008.00951"><img src="https://img.shields.io/badge/arXiv-2008.00951-b31b1b.svg" height=22.5></a>
3
+ <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-yellow.svg" height=22.5></a>
4
+
5
+ <a href="https://www.youtube.com/watch?v=bfvSwhqsTgM"><img src="https://img.shields.io/static/v1?label=CVPR 2021&message=5 Minute Video&color=red" height=22.5></a>
6
+ <a href="https://replicate.ai/eladrich/pixel2style2pixel"><img src="https://img.shields.io/static/v1?label=Replicate&message=Demo and Docker Image&color=darkgreen" height=22.5></a>
7
+
8
+ <a href="http://colab.research.google.com/github/eladrich/pixel2style2pixel/blob/master/notebooks/inference_playground.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" height=22.5></a>
9
+
10
+ > We present a generic image-to-image translation framework, pixel2style2pixel (pSp).
11
+ Our pSp framework is based on a novel encoder network that directly generates a series of style vectors which are fed into a pretrained StyleGAN generator,
12
+ forming the extended W+ latent space. We first show that our encoder can directly embed real images into W+, with no additional optimization.
13
+ Next, we propose utilizing our encoder to directly solve image-to-image translation tasks, defining them as encoding problems from some input domain into the
14
+ latent domain. By deviating from the standard "invert first, edit later" methodology used with previous StyleGAN encoders, our approach can handle a variety of
15
+ tasks even when the input image is not represented in the StyleGAN domain. We show that solving translation tasks through StyleGAN significantly simplifies the training process, as no adversary is required, has better support
16
+ >for solving tasks without pixel-to-pixel correspondence, and inherently supports multi-modal synthesis via the resampling of styles.
17
+ Finally, we demonstrate the potential of our framework on a variety of facial image-to-image translation tasks, even when compared to state-of-the-art solutions designed specifically for a single task, and further show that it can be extended beyond the human facial domain.
18
+
19
+ <p align="center">
20
+ <img src="docs/teaser.png" width="800px"/>
21
+ <br>
22
+ The proposed pixel2style2pixel framework can be used to solve a wide variety of image-to-image translation tasks. Here we show results of pSp on StyleGAN inversion, multi-modal conditional image synthesis, facial frontalization, inpainting and super-resolution.
23
+ </p>
24
+
25
+ ## Description
26
+ Official Implementation of our pSp paper for both training and evaluation. The pSp method extends the StyleGAN model to
27
+ allow solving different image-to-image translation problems using its encoder.
28
+
29
+ ## Table of Contents
30
+ * [Description](#description)
31
+ * [Table of Contents](#table-of-contents)
32
+ * [Recent Updates](#recent-updates)
33
+ * [Applications](#applications)
34
+ + [StyleGAN Encoding](#stylegan-encoding)
35
+ + [Face Frontalization](#face-frontalization)
36
+ + [Conditional Image Synthesis](#conditional-image-synthesis)
37
+ + [Super Resolution](#super-resolution)
38
+ * [Getting Started](#getting-started)
39
+ + [Prerequisites](#prerequisites)
40
+ + [Installation](#installation)
41
+ + [Inference Notebook](#inference-notebook)
42
+ + [Pretrained Models](#pretrained-models)
43
+ * [Training](#training)
44
+ + [Preparing your Data](#preparing-your-data)
45
+ + [Training pSp](#training-psp)
46
+ - [Training the pSp Encoder](#training-the-psp-encoder)
47
+ - [Frontalization](#frontalization)
48
+ - [Sketch to Face](#sketch-to-face)
49
+ - [Segmentation Map to Face](#segmentation-map-to-face)
50
+ - [Super Resolution](#super-resolution-1)
51
+ + [Additional Notes](#additional-notes)
52
+ + [Weights & Biases Integration](#weights--biases-integration)
53
+ * [Testing](#testing)
54
+ + [Inference](#inference)
55
+ + [Multi-Modal Synthesis with Style-Mixing](#multi-modal-synthesis-with-style-mixing)
56
+ + [Computing Metrics](#computing-metrics)
57
+ * [Additional Applications](#additional-applications)
58
+ + [Toonify](#toonify)
59
+ * [Repository structure](#repository-structure)
60
+ * [TODOs](#todos)
61
+ * [Credits](#credits)
62
+ * [Inspired by pSp](#inspired-by-psp)
63
+ * [pSp in the Media](#psp-in-the-media)
64
+ * [Citation](#citation)
65
+
66
+ ## Recent Updates
67
+ **`2020.10.04`**: Initial code release
68
+ **`2020.10.06`**: Add pSp toonify model (Thanks to the great work from [Doron Adler](https://linktr.ee/Norod78) and [Justin Pinkney](https://www.justinpinkney.com/))!
69
+ **`2021.04.23`**: Added several new features:
70
+ - Added supported for StyleGANs of different resolutions (e.g., 256, 512, 1024). This can be set using the flag `--output_size`, which is set to 1024 by default.
71
+ - Added support for the MoCo-Based similarity loss introduced in [encoder4editing (Tov et al. 2021)](https://github.com/omertov/encoder4editing). More details are provided [below](https://github.com/eladrich/pixel2style2pixel#training-psp).
72
+
73
+ **`2021.07.06`**: Added support for training with Weights & Biases. [See below for details](https://github.com/eladrich/pixel2style2pixel#weights--biases-integration).
74
+
75
+ ## Applications
76
+ ### StyleGAN Encoding
77
+ Here, we use pSp to find the latent code of real images in the latent domain of a pretrained StyleGAN generator.
78
+ <p align="center">
79
+ <img src="docs/encoding_inputs.jpg" width="800px"/>
80
+ <img src="docs/encoding_outputs.jpg" width="800px"/>
81
+ </p>
82
+
83
+
84
+ ### Face Frontalization
85
+ In this application we want to generate a front-facing face from a given input image.
86
+ <p align="center">
87
+ <img src="docs/frontalization_inputs.jpg" width="800px"/>
88
+ <img src="docs/frontalization_outputs.jpg" width="800px"/>
89
+ </p>
90
+
91
+ ### Conditional Image Synthesis
92
+ Here we wish to generate photo-realistic face images from ambiguous sketch images or segmentation maps. Using style-mixing, we inherently support multi-modal synthesis for a single input.
93
+ <p align="center">
94
+ <img src="docs/seg2image.png" width="800px"/>
95
+ <img src="docs/sketch2image.png" width="800px"/>
96
+ </p>
97
+
98
+ ### Super Resolution
99
+ Given a low-resolution input image, we generate a corresponding high-resolution image. As this too is an ambiguous task, we can use style-mixing to produce several plausible results.
100
+ <p align="center">
101
+ <img src="docs/super_res_32.jpg" width="800px"/>
102
+ <img src="docs/super_res_style_mixing.jpg" width="800px"/>
103
+ </p>
104
+
105
+
106
+ ## Getting Started
107
+ ### Prerequisites
108
+ - Linux or macOS
109
+ - NVIDIA GPU + CUDA CuDNN (CPU may be possible with some modifications, but is not inherently supported)
110
+ - Python 2 or 3
111
+
112
+ ### Installation
113
+ - Clone this repo:
114
+ ```
115
+ git clone https://github.com/eladrich/pixel2style2pixel.git
116
+ cd pixel2style2pixel
117
+ ```
118
+ - Dependencies:
119
+ We recommend running this repository using [Anaconda](https://docs.anaconda.com/anaconda/install/).
120
+ All dependencies for defining the environment are provided in `environment/psp_env.yaml`.
121
+
122
+ ### Inference Notebook
123
+ To help visualize the pSp framework on multiple tasks and to help you get started, we provide a Jupyter notebook found in `notebooks/inference_playground.ipynb` that allows one to visualize the various applications of pSp.
124
+ The notebook will download the necessary pretrained models and run inference on the images found in `notebooks/images`.
125
+ For the tasks of conditional image synthesis and super resolution, the notebook also demonstrates pSp's ability to perform multi-modal synthesis using
126
+ style-mixing.
127
+
128
+ ### Pretrained Models
129
+ Please download the pre-trained models from the following links. Each pSp model contains the entire pSp architecture, including the encoder and decoder weights.
130
+ | Path | Description
131
+ | :--- | :----------
132
+ |[StyleGAN Inversion](https://drive.google.com/file/d/1bMTNWkh5LArlaWSc_wa8VKyq2V42T2z0/view?usp=sharing) | pSp trained with the FFHQ dataset for StyleGAN inversion.
133
+ |[Face Frontalization](https://drive.google.com/file/d/1_S4THAzXb-97DbpXmanjHtXRyKxqjARv/view?usp=sharing) | pSp trained with the FFHQ dataset for face frontalization.
134
+ |[Sketch to Image](https://drive.google.com/file/d/1lB7wk7MwtdxL-LL4Z_T76DuCfk00aSXA/view?usp=sharing) | pSp trained with the CelebA-HQ dataset for image synthesis from sketches.
135
+ |[Segmentation to Image](https://drive.google.com/file/d/1VpEKc6E6yG3xhYuZ0cq8D2_1CbT0Dstz/view?usp=sharing) | pSp trained with the CelebAMask-HQ dataset for image synthesis from segmentation maps.
136
+ |[Super Resolution](https://drive.google.com/file/d/1ZpmSXBpJ9pFEov6-jjQstAlfYbkebECu/view?usp=sharing) | pSp trained with the CelebA-HQ dataset for super resolution (up to x32 down-sampling).
137
+ |[Toonify](https://drive.google.com/file/d/1YKoiVuFaqdvzDP5CZaqa3k5phL-VDmyz/view) | pSp trained with the FFHQ dataset for toonification using StyleGAN generator from [Doron Adler](https://linktr.ee/Norod78) and [Justin Pinkney](https://www.justinpinkney.com/).
138
+
139
+ If you wish to use one of the pretrained models for training or inference, you may do so using the flag `--checkpoint_path`.
140
+
141
+ In addition, we provide various auxiliary models needed for training your own pSp model from scratch as well as pretrained models needed for computing the ID metrics reported in the paper.
142
+ | Path | Description
143
+ | :--- | :----------
144
+ |[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.
145
+ |[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 pSp training.
146
+ |[MoCo ResNet-50](https://drive.google.com/file/d/18rLcNGdteX5LwT7sv_F7HWr12HpVEzVe/view?usp=sharing) | Pretrained ResNet-50 model trained using MOCOv2 for computing MoCo-based similarity loss on non-facial domains. The model is taken from the [official implementation](https://github.com/facebookresearch/moco).
147
+ |[CurricularFace Backbone](https://drive.google.com/file/d/1f4IwVa2-Bn9vWLwB-bUwm53U_MlvinAj/view?usp=sharing) | Pretrained CurricularFace model taken from [HuangYG123](https://github.com/HuangYG123/CurricularFace) for use in ID similarity metric computation.
148
+ |[MTCNN](https://drive.google.com/file/d/1tJ7ih-wbCO6zc3JhI_1ZGjmwXKKaPlja/view?usp=sharing) | Weights for MTCNN model taken from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch) for use in ID similarity metric computation. (Unpack the tar.gz to extract the 3 model weights.)
149
+
150
+ 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`.
151
+
152
+ ## Training
153
+ ### Preparing your Data
154
+ - Currently, we provide support for numerous datasets and experiments (encoding, frontalization, etc.).
155
+ - Refer to `configs/paths_config.py` to define the necessary data paths and model paths for training and evaluation.
156
+ - Refer to `configs/transforms_config.py` for the transforms defined for each dataset/experiment.
157
+ - Finally, refer to `configs/data_configs.py` for the source/target data paths for the train and test sets
158
+ as well as the transforms.
159
+ - If you wish to experiment with your own dataset, you can simply make the necessary adjustments in
160
+ 1. `data_configs.py` to define your data paths.
161
+ 2. `transforms_configs.py` to define your own data transforms.
162
+
163
+ As an example, assume we wish to run encoding using ffhq (`dataset_type=ffhq_encode`).
164
+ We first go to `configs/paths_config.py` and define:
165
+ ```
166
+ dataset_paths = {
167
+ 'ffhq': '/path/to/ffhq/images256x256'
168
+ 'celeba_test': '/path/to/CelebAMask-HQ/test_img',
169
+ }
170
+ ```
171
+ The transforms for the experiment are defined in the class `EncodeTransforms` in `configs/transforms_config.py`.
172
+ Finally, in `configs/data_configs.py`, we define:
173
+ ```
174
+ DATASETS = {
175
+ 'ffhq_encode': {
176
+ 'transforms': transforms_config.EncodeTransforms,
177
+ 'train_source_root': dataset_paths['ffhq'],
178
+ 'train_target_root': dataset_paths['ffhq'],
179
+ 'test_source_root': dataset_paths['celeba_test'],
180
+ 'test_target_root': dataset_paths['celeba_test'],
181
+ },
182
+ }
183
+ ```
184
+ When defining our datasets, we will take the values in the above dictionary.
185
+
186
+
187
+ ### Training pSp
188
+ The main training script can be found in `scripts/train.py`.
189
+ Intermediate training results are saved to `opts.exp_dir`. This includes checkpoints, train outputs, and test outputs.
190
+ Additionally, if you have tensorboard installed, you can visualize tensorboard logs in `opts.exp_dir/logs`.
191
+
192
+ #### Training the pSp Encoder
193
+ ```
194
+ python scripts/train.py \
195
+ --dataset_type=ffhq_encode \
196
+ --exp_dir=/path/to/experiment \
197
+ --workers=8 \
198
+ --batch_size=8 \
199
+ --test_batch_size=8 \
200
+ --test_workers=8 \
201
+ --val_interval=2500 \
202
+ --save_interval=5000 \
203
+ --encoder_type=GradualStyleEncoder \
204
+ --start_from_latent_avg \
205
+ --lpips_lambda=0.8 \
206
+ --l2_lambda=1 \
207
+ --id_lambda=0.1
208
+ ```
209
+
210
+ #### Frontalization
211
+ ```
212
+ python scripts/train.py \
213
+ --dataset_type=ffhq_frontalize \
214
+ --exp_dir=/path/to/experiment \
215
+ --workers=8 \
216
+ --batch_size=8 \
217
+ --test_batch_size=8 \
218
+ --test_workers=8 \
219
+ --val_interval=2500 \
220
+ --save_interval=5000 \
221
+ --encoder_type=GradualStyleEncoder \
222
+ --start_from_latent_avg \
223
+ --lpips_lambda=0.08 \
224
+ --l2_lambda=0.001 \
225
+ --lpips_lambda_crop=0.8 \
226
+ --l2_lambda_crop=0.01 \
227
+ --id_lambda=1 \
228
+ --w_norm_lambda=0.005
229
+ ```
230
+
231
+ #### Sketch to Face
232
+ ```
233
+ python scripts/train.py \
234
+ --dataset_type=celebs_sketch_to_face \
235
+ --exp_dir=/path/to/experiment \
236
+ --workers=8 \
237
+ --batch_size=8 \
238
+ --test_batch_size=8 \
239
+ --test_workers=8 \
240
+ --val_interval=2500 \
241
+ --save_interval=5000 \
242
+ --encoder_type=GradualStyleEncoder \
243
+ --start_from_latent_avg \
244
+ --lpips_lambda=0.8 \
245
+ --l2_lambda=1 \
246
+ --id_lambda=0 \
247
+ --w_norm_lambda=0.005 \
248
+ --label_nc=1 \
249
+ --input_nc=1
250
+ ```
251
+
252
+ #### Segmentation Map to Face
253
+ ```
254
+ python scripts/train.py \
255
+ --dataset_type=celebs_seg_to_face \
256
+ --exp_dir=/path/to/experiment \
257
+ --workers=8 \
258
+ --batch_size=8 \
259
+ --test_batch_size=8 \
260
+ --test_workers=8 \
261
+ --val_interval=2500 \
262
+ --save_interval=5000 \
263
+ --encoder_type=GradualStyleEncoder \
264
+ --start_from_latent_avg \
265
+ --lpips_lambda=0.8 \
266
+ --l2_lambda=1 \
267
+ --id_lambda=0 \
268
+ --w_norm_lambda=0.005 \
269
+ --label_nc=19 \
270
+ --input_nc=19
271
+ ```
272
+ Notice with conditional image synthesis no identity loss is utilized (i.e. `--id_lambda=0`)
273
+
274
+ #### Super Resolution
275
+ ```
276
+ python scripts/train.py \
277
+ --dataset_type=celebs_super_resolution \
278
+ --exp_dir=/path/to/experiment \
279
+ --workers=8 \
280
+ --batch_size=8 \
281
+ --test_batch_size=8 \
282
+ --test_workers=8 \
283
+ --val_interval=2500 \
284
+ --save_interval=5000 \
285
+ --encoder_type=GradualStyleEncoder \
286
+ --start_from_latent_avg \
287
+ --lpips_lambda=0.8 \
288
+ --l2_lambda=1 \
289
+ --id_lambda=0.1 \
290
+ --w_norm_lambda=0.005 \
291
+ --resize_factors=1,2,4,8,16,32
292
+ ```
293
+
294
+ ### Additional Notes
295
+ - See `options/train_options.py` for all training-specific flags.
296
+ - See `options/test_options.py` for all test-specific flags.
297
+ - If you wish to resume from a specific checkpoint (e.g. a pretrained pSp model), you may do so using `--checkpoint_path`.
298
+ - By default, we assume that the StyleGAN used outputs images at resolution `1024x1024`. If you wish to use a StyleGAN at a smaller resolution, you can do so by using the flag `--output_size` (e.g., `--output_size=256`).
299
+ - If you wish to generate images from segmentation maps, please specify `--label_nc=N` and `--input_nc=N` where `N`
300
+ is the number of semantic categories.
301
+ - Similarly, for generating images from sketches, please specify `--label_nc=1` and `--input_nc=1`.
302
+ - Specifying `--label_nc=0` (the default value), will directly use the RGB colors as input.
303
+
304
+ ** Identity/Similarity Losses **
305
+ In pSp, we introduce a facial identity loss using a pre-trained ArcFace network for facial recognition. When operating on the human facial domain, we
306
+ highly recommend employing this loss objective by using the flag `--id_lambda`.
307
+ In a more recent paper, [encoder4editing](https://github.com/omertov/encoder4editing), the authors generalize this identity loss to other domains by
308
+ using a MoCo-based ResNet to extract features instead of an ArcFace network.
309
+ Applying this MoCo-based similarity loss can be done by using the flag `--moco_lambda`. We recommend setting `--moco_lambda=0.5` in your experiments.
310
+ Please note, you <ins>cannot</ins> set both `id_lambda` and `moco_lambda` to be active simultaneously (e.g., to use the MoCo-based loss, you should specify,
311
+ `--moco_lambda=0.5 --id_lambda=0`).
312
+
313
+ ### Weights & Biases Integration
314
+ To help track your experiments, we've integrated [Weights & Biases](https://wandb.ai/home) into our training process.
315
+ To enable Weights & Biases (`wandb`), first make an account on the platform's webpage and install `wandb` using
316
+ `pip install wandb`. Then, to train pSp using `wandb`, simply add the flag `--use_wandb`.
317
+
318
+ Note that when running for the first time, you will be asked to provide your access key which can be accessed via the
319
+ Weights & Biases platform.
320
+
321
+ Using Weights & Biases will allow you to visualize the training and testing loss curves as well as
322
+ intermediate training results.
323
+
324
+
325
+ ## Testing
326
+ ### Inference
327
+ Having trained your model, you can use `scripts/inference.py` to apply the model on a set of images.
328
+ For example,
329
+ ```
330
+ python scripts/inference.py \
331
+ --exp_dir=/path/to/experiment \
332
+ --checkpoint_path=experiment/checkpoints/best_model.pt \
333
+ --data_path=/path/to/test_data \
334
+ --test_batch_size=4 \
335
+ --test_workers=4 \
336
+ --couple_outputs
337
+ ```
338
+ Additional notes to consider:
339
+ - During inference, the options used during training are loaded from the saved checkpoint and are then updated using the
340
+ test options passed to the inference script. For example, there is no need to pass `--dataset_type` or `--label_nc` to the
341
+ inference script, as they are taken from the loaded `opts`.
342
+ - When running inference for segmentation-to-image or sketch-to-image, it is highly recommend to do so with a style-mixing,
343
+ as is done in the paper. This can simply be done by adding `--latent_mask=8,9,10,11,12,13,14,15,16,17` when calling the
344
+ script.
345
+ - When running inference for super-resolution, please provide a single down-sampling value using `--resize_factors`.
346
+ - Adding the flag `--couple_outputs` will save an additional image containing the input and output images side-by-side in the sub-directory
347
+ `inference_coupled`. Otherwise, only the output image is saved to the sub-directory `inference_results`.
348
+ - By default, the images will be saved at resolutiosn of 1024x1024, the original output size of StyleGAN. If you wish to save
349
+ outputs resized to resolutions of 256x256, you can do so by adding the flag `--resize_outputs`.
350
+
351
+
352
+ ### Multi-Modal Synthesis with Style-Mixing
353
+ Given a trained model for conditional image synthesis or super-resolution, we can easily generate multiple outputs
354
+ for a given input image. This can be done using the script `scripts/style_mixing.py`.
355
+ For example, running the following command will perform style-mixing for a segmentation-to-image experiment:
356
+ ```
357
+ python scripts/style_mixing.py \
358
+ --exp_dir=/path/to/experiment \
359
+ --checkpoint_path=/path/to/experiment/checkpoints/best_model.pt \
360
+ --data_path=/path/to/test_data/ \
361
+ --test_batch_size=4 \
362
+ --test_workers=4 \
363
+ --n_images=25 \
364
+ --n_outputs_to_generate=5 \
365
+ --latent_mask=8,9,10,11,12,13,14,15,16,17
366
+ ```
367
+ Here, we inject `5` randomly drawn vectors and perform style-mixing on the latents `[8,9,10,11,12,13,14,15,16,17]`.
368
+
369
+ Additional notes to consider:
370
+ - To perform style-mixing on a subset of images, you may use the flag `--n_images`. The default value of `None` will perform
371
+ style mixing on every image in the given `data_path`.
372
+ - You may also include the argument `--mix_alpha=m` where `m` is a float defining the mixing coefficient between the
373
+ input latent and the randomly drawn latent.
374
+ - When performing style-mixing for super-resolution, please provide a single down-sampling value using `--resize_factors`.
375
+ - By default, the images will be saved at resolutiosn of 1024x1024, the original output size of StyleGAN. If you wish to save
376
+ outputs resized to resolutions of 256x256, you can do so by adding the flag `--resize_outputs`.
377
+
378
+
379
+ ### Computing Metrics
380
+ Similarly, given a trained model and generated outputs, we can compute the loss metrics on a given dataset.
381
+ These scripts receive the inference output directory and ground truth directory.
382
+ - Calculating the identity loss:
383
+ ```
384
+ python scripts/calc_id_loss_parallel.py \
385
+ --data_path=/path/to/experiment/inference_outputs \
386
+ --gt_path=/path/to/test_images \
387
+ ```
388
+ - Calculating LPIPS loss:
389
+ ```
390
+ python scripts/calc_losses_on_images.py \
391
+ --mode lpips
392
+ --data_path=/path/to/experiment/inference_outputs \
393
+ --gt_path=/path/to/test_images \
394
+ ```
395
+ - Calculating L2 loss:
396
+ ```
397
+ python scripts/calc_losses_on_images.py \
398
+ --mode l2
399
+ --data_path=/path/to/experiment/inference_outputs \
400
+ --gt_path=/path/to/test_images \
401
+ ```
402
+
403
+ ## Additional Applications
404
+ To better show the flexibility of our pSp framework we present additional applications below.
405
+
406
+ As with our main applications, you may download the pretrained models here:
407
+ | Path | Description
408
+ | :--- | :----------
409
+ |[Toonify](https://drive.google.com/file/d/1YKoiVuFaqdvzDP5CZaqa3k5phL-VDmyz/view) | pSp trained with the FFHQ dataset for toonification using StyleGAN generator from [Doron Adler](https://linktr.ee/Norod78) and [Justin Pinkney](https://www.justinpinkney.com/).
410
+
411
+ ### Toonify
412
+ Using the toonify StyleGAN built by [Doron Adler](https://linktr.ee/Norod78) and [Justin Pinkney](https://www.justinpinkney.com/),
413
+ we take a real face image and generate a toonified version of the given image. We train the pSp encoder to directly reconstruct real
414
+ face images inside the toons latent space resulting in a projection of each image to the closest toon. We do so without requiring any labeled pairs
415
+ or distillation!
416
+ <p align="center">
417
+ <img src="docs/toonify_input.jpg" width="800px"/>
418
+ <img src="docs/toonify_output.jpg" width="800px"/>
419
+ </p>
420
+
421
+ This is trained exactly like the StyleGAN inversion task with several changes:
422
+ - Change from FFHQ StyleGAN to toonifed StyleGAN (can be set using `--stylegan_weights`)
423
+ - The toonify generator is taken from [Doron Adler](https://linktr.ee/Norod78) and [Justin Pinkney](https://www.justinpinkney.com/)
424
+ and converted to Pytorch using [rosinality's](https://github.com/rosinality/stylegan2-pytorch) conversion script.
425
+ - For convenience, the converted generator Pytorch model may be downloaded [here](https://drive.google.com/file/d/1r3XVCt_WYUKFZFxhNH-xO2dTtF6B5szu/view?usp=sharing).
426
+ - Increase `id_lambda` from `0.1` to `1`
427
+ - Increase `w_norm_lambda` from `0.005` to `0.025`
428
+
429
+ We obtain the best results after around `6000` iterations of training (can be set using `--max_steps`)
430
+
431
+
432
+ ## Repository structure
433
+ | Path | Description <img width=200>
434
+ | :--- | :---
435
+ | pixel2style2pixel | Repository root folder
436
+ | &boxvr;&nbsp; configs | Folder containing configs defining model/data paths and data transforms
437
+ | &boxvr;&nbsp; criteria | Folder containing various loss criterias for training
438
+ | &boxvr;&nbsp; datasets | Folder with various dataset objects and augmentations
439
+ | &boxvr;&nbsp; environment | Folder containing Anaconda environment used in our experiments
440
+ | &boxvr; models | Folder containting all the models and training objects
441
+ | &boxv;&nbsp; &boxvr;&nbsp; encoders | Folder containing our pSp encoder architecture implementation and ArcFace encoder implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
442
+ | &boxv;&nbsp; &boxvr;&nbsp; mtcnn | MTCNN implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
443
+ | &boxv;&nbsp; &boxvr;&nbsp; stylegan2 | StyleGAN2 model from [rosinality](https://github.com/rosinality/stylegan2-pytorch)
444
+ | &boxv;&nbsp; &boxur;&nbsp; psp.py | Implementation of our pSp framework
445
+ | &boxvr;&nbsp; notebook | Folder with jupyter notebook containing pSp inference playground
446
+ | &boxvr;&nbsp; options | Folder with training and test command-line options
447
+ | &boxvr;&nbsp; scripts | Folder with running scripts for training and inference
448
+ | &boxvr;&nbsp; training | Folder with main training logic and Ranger implementation from [lessw2020](https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer)
449
+ | &boxvr;&nbsp; utils | Folder with various utility functions
450
+ | <img width=300> | <img>
451
+
452
+ ## TODOs
453
+ - [ ] Add multi-gpu support
454
+
455
+ ## Credits
456
+ **StyleGAN2 implementation:**
457
+ https://github.com/rosinality/stylegan2-pytorch
458
+ Copyright (c) 2019 Kim Seonghyeon
459
+ License (MIT) https://github.com/rosinality/stylegan2-pytorch/blob/master/LICENSE
460
+
461
+ **MTCNN, IR-SE50, and ArcFace models and implementations:**
462
+ https://github.com/TreB1eN/InsightFace_Pytorch
463
+ Copyright (c) 2018 TreB1eN
464
+ License (MIT) https://github.com/TreB1eN/InsightFace_Pytorch/blob/master/LICENSE
465
+
466
+ **CurricularFace model and implementation:**
467
+ https://github.com/HuangYG123/CurricularFace
468
+ Copyright (c) 2020 HuangYG123
469
+ License (MIT) https://github.com/HuangYG123/CurricularFace/blob/master/LICENSE
470
+
471
+ **Ranger optimizer implementation:**
472
+ https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
473
+ License (Apache License 2.0) https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer/blob/master/LICENSE
474
+
475
+ **LPIPS implementation:**
476
+ https://github.com/S-aiueo32/lpips-pytorch
477
+ Copyright (c) 2020, Sou Uchida
478
+ License (BSD 2-Clause) https://github.com/S-aiueo32/lpips-pytorch/blob/master/LICENSE
479
+
480
+ **Please Note**: The CUDA files under the [StyleGAN2 ops directory](https://github.com/eladrich/pixel2style2pixel/tree/master/models/stylegan2/op) are made available under the [Nvidia Source Code License-NC](https://nvlabs.github.io/stylegan2/license.html)
481
+
482
+ ## Inspired by pSp
483
+ Below are several works inspired by pSp that we found particularly interesting:
484
+
485
+ **Reverse Toonification**
486
+ Using our pSp encoder, artist [Nathan Shipley](https://linktr.ee/nathan_shipley) transformed animated figures and paintings into real life. Check out his amazing work on his [twitter page](https://twitter.com/citizenplain?lang=en) and [website](http://www.nathanshipley.com/gan).
487
+
488
+ **Deploying pSp with StyleSpace for Editing**
489
+ Awesome work from [Justin Pinkney](https://www.justinpinkney.com/) who deployed our pSp model on Runway and provided support for editing the resulting inversions using the [StyleSpace Analysis paper](https://arxiv.org/abs/2011.12799). Check out his repository [here](https://github.com/justinpinkney/pixel2style2pixel).
490
+
491
+ **Encoder4Editing (e4e)**
492
+ Building on the work of pSp, Tov et al. design an encoder to enable high quality edits on real images. Check out their [paper](https://arxiv.org/abs/2102.02766) and [code](https://github.com/omertov/encoder4editing).
493
+
494
+ **Style-based Age Manipulation (SAM)**
495
+ Leveraging pSp and the rich semantics of StyleGAN, SAM learns non-linear latent space paths for modeling the age transformation of real face images. Check out the project page [here](https://yuval-alaluf.github.io/SAM/).
496
+
497
+ **ReStyle**
498
+ ReStyle builds on recent encoders such as pSp and e4e by introducing an iterative refinment mechanism to gradually improve the inversion of real images. Check out the project page [here](https://yuval-alaluf.github.io/restyle-encoder/).
499
+
500
+ ## pSp in the Media
501
+ * bycloud: [AI Generates Cartoon Characters In Real Life Pixel2Style2Pixel](https://www.youtube.com/watch?v=g-N8lfceclI&ab_channel=bycloud)
502
+ * Synced: [Pixel2Style2Pixel: Novel Encoder Architecture Boosts Facial Image-To-Image Translation](https://syncedreview.com/2020/08/07/pixel2style2pixel-novel-encoder-architecture-boosts-facial-image-to-image-translation/)
503
+ * Cartoon Brew: [An Artist Has Used Machine Learning To Turn Animated Characters Into Creepy Photorealistic Figures](https://www.cartoonbrew.com/tech/an-artist-has-used-machine-learning-to-turn-animated-characters-into-creepy-photorealistic-figures-197975.html)
504
+
505
+
506
+ ## Citation
507
+ If you use this code for your research, please cite our paper <a href="https://arxiv.org/abs/2008.00951">Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation</a>:
508
+
509
+ ```
510
+ @InProceedings{richardson2021encoding,
511
+ author = {Richardson, Elad and Alaluf, Yuval and Patashnik, Or and Nitzan, Yotam and Azar, Yaniv and Shapiro, Stav and Cohen-Or, Daniel},
512
+ title = {Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation},
513
+ booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
514
+ month = {June},
515
+ year = {2021}
516
+ }
517
+ ```
interface/pixel2style2pixel/cog.yaml ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ build:
2
+ gpu: true
3
+ python_version: "3.8"
4
+ system_packages:
5
+ - "libgl1-mesa-glx"
6
+ - "libglib2.0-0"
7
+ - "ninja-build"
8
+ python_packages:
9
+ - "cmake==3.21.2"
10
+ - "torch==1.8.0"
11
+ - "torchvision==0.9.0"
12
+ - "numpy==1.21.1"
13
+ - "ipython==7.21.0"
14
+ - "tensorboard==2.6.0"
15
+ - "tqdm==4.43.0"
16
+ - "torch-optimizer==0.1.0"
17
+ - "opencv-python==4.5.3.56"
18
+ - "Pillow==8.3.2"
19
+ - "matplotlib==3.2.1"
20
+ - "scipy==1.7.1"
21
+ run:
22
+ - pip install dlib
23
+
24
+ predict: "predict.py:Predictor"
25
+
26
+
27
+
28
+
29
+
30
+
31
+
32
+
interface/pixel2style2pixel/configs/__init__.py ADDED
File without changes
interface/pixel2style2pixel/configs/data_configs.py ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from configs import transforms_config
2
+ from configs.paths_config import dataset_paths
3
+
4
+
5
+ DATASETS = {
6
+ 'ffhq_encode': {
7
+ 'transforms': transforms_config.EncodeTransforms,
8
+ 'train_source_root': dataset_paths['ffhq'],
9
+ 'train_target_root': dataset_paths['ffhq'],
10
+ 'test_source_root': dataset_paths['celeba_test'],
11
+ 'test_target_root': dataset_paths['celeba_test'],
12
+ },
13
+ 'ffhq_frontalize': {
14
+ 'transforms': transforms_config.FrontalizationTransforms,
15
+ 'train_source_root': dataset_paths['ffhq'],
16
+ 'train_target_root': dataset_paths['ffhq'],
17
+ 'test_source_root': dataset_paths['celeba_test'],
18
+ 'test_target_root': dataset_paths['celeba_test'],
19
+ },
20
+ 'celebs_sketch_to_face': {
21
+ 'transforms': transforms_config.SketchToImageTransforms,
22
+ 'train_source_root': dataset_paths['celeba_train_sketch'],
23
+ 'train_target_root': dataset_paths['celeba_train'],
24
+ 'test_source_root': dataset_paths['celeba_test_sketch'],
25
+ 'test_target_root': dataset_paths['celeba_test'],
26
+ },
27
+ 'celebs_seg_to_face': {
28
+ 'transforms': transforms_config.SegToImageTransforms,
29
+ 'train_source_root': dataset_paths['celeba_train_segmentation'],
30
+ 'train_target_root': dataset_paths['celeba_train'],
31
+ 'test_source_root': dataset_paths['celeba_test_segmentation'],
32
+ 'test_target_root': dataset_paths['celeba_test'],
33
+ },
34
+ 'celebs_super_resolution': {
35
+ 'transforms': transforms_config.SuperResTransforms,
36
+ 'train_source_root': dataset_paths['celeba_train'],
37
+ 'train_target_root': dataset_paths['celeba_train'],
38
+ 'test_source_root': dataset_paths['celeba_test'],
39
+ 'test_target_root': dataset_paths['celeba_test'],
40
+ },
41
+ }
interface/pixel2style2pixel/configs/paths_config.py ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dataset_paths = {
2
+ 'celeba_train': '',
3
+ 'celeba_test': '',
4
+ 'celeba_train_sketch': '',
5
+ 'celeba_test_sketch': '',
6
+ 'celeba_train_segmentation': '',
7
+ 'celeba_test_segmentation': '',
8
+ 'ffhq': '',
9
+ }
10
+
11
+ model_paths = {
12
+ 'stylegan_ffhq': 'pretrained_models/stylegan2-ffhq-config-f.pt',
13
+ 'ir_se50': 'pretrained_models/model_ir_se50.pth',
14
+ 'circular_face': 'pretrained_models/CurricularFace_Backbone.pth',
15
+ 'mtcnn_pnet': 'pretrained_models/mtcnn/pnet.npy',
16
+ 'mtcnn_rnet': 'pretrained_models/mtcnn/rnet.npy',
17
+ 'mtcnn_onet': 'pretrained_models/mtcnn/onet.npy',
18
+ 'shape_predictor': 'shape_predictor_68_face_landmarks.dat',
19
+ 'moco': 'pretrained_models/moco_v2_800ep_pretrain.pth.tar'
20
+ }
interface/pixel2style2pixel/configs/transforms_config.py ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from abc import abstractmethod
2
+ import torchvision.transforms as transforms
3
+ from datasets import augmentations
4
+
5
+
6
+ class TransformsConfig(object):
7
+
8
+ def __init__(self, opts):
9
+ self.opts = opts
10
+
11
+ @abstractmethod
12
+ def get_transforms(self):
13
+ pass
14
+
15
+
16
+ class EncodeTransforms(TransformsConfig):
17
+
18
+ def __init__(self, opts):
19
+ super(EncodeTransforms, self).__init__(opts)
20
+
21
+ def get_transforms(self):
22
+ transforms_dict = {
23
+ 'transform_gt_train': transforms.Compose([
24
+ transforms.Resize((256, 256)),
25
+ transforms.RandomHorizontalFlip(0.5),
26
+ transforms.ToTensor(),
27
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
28
+ 'transform_source': None,
29
+ 'transform_test': transforms.Compose([
30
+ transforms.Resize((256, 256)),
31
+ transforms.ToTensor(),
32
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
33
+ 'transform_inference': transforms.Compose([
34
+ transforms.Resize((256, 256)),
35
+ transforms.ToTensor(),
36
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
37
+ }
38
+ return transforms_dict
39
+
40
+
41
+ class FrontalizationTransforms(TransformsConfig):
42
+
43
+ def __init__(self, opts):
44
+ super(FrontalizationTransforms, self).__init__(opts)
45
+
46
+ def get_transforms(self):
47
+ transforms_dict = {
48
+ 'transform_gt_train': transforms.Compose([
49
+ transforms.Resize((256, 256)),
50
+ transforms.RandomHorizontalFlip(0.5),
51
+ transforms.ToTensor(),
52
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
53
+ 'transform_source': transforms.Compose([
54
+ transforms.Resize((256, 256)),
55
+ transforms.RandomHorizontalFlip(0.5),
56
+ transforms.ToTensor(),
57
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
58
+ 'transform_test': transforms.Compose([
59
+ transforms.Resize((256, 256)),
60
+ transforms.ToTensor(),
61
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
62
+ 'transform_inference': transforms.Compose([
63
+ transforms.Resize((256, 256)),
64
+ transforms.ToTensor(),
65
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
66
+ }
67
+ return transforms_dict
68
+
69
+
70
+ class SketchToImageTransforms(TransformsConfig):
71
+
72
+ def __init__(self, opts):
73
+ super(SketchToImageTransforms, self).__init__(opts)
74
+
75
+ def get_transforms(self):
76
+ transforms_dict = {
77
+ 'transform_gt_train': transforms.Compose([
78
+ transforms.Resize((256, 256)),
79
+ transforms.ToTensor(),
80
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
81
+ 'transform_source': transforms.Compose([
82
+ transforms.Resize((256, 256)),
83
+ transforms.ToTensor()]),
84
+ 'transform_test': transforms.Compose([
85
+ transforms.Resize((256, 256)),
86
+ transforms.ToTensor(),
87
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
88
+ 'transform_inference': transforms.Compose([
89
+ transforms.Resize((256, 256)),
90
+ transforms.ToTensor()]),
91
+ }
92
+ return transforms_dict
93
+
94
+
95
+ class SegToImageTransforms(TransformsConfig):
96
+
97
+ def __init__(self, opts):
98
+ super(SegToImageTransforms, self).__init__(opts)
99
+
100
+ def get_transforms(self):
101
+ transforms_dict = {
102
+ 'transform_gt_train': transforms.Compose([
103
+ transforms.Resize((256, 256)),
104
+ transforms.ToTensor(),
105
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
106
+ 'transform_source': transforms.Compose([
107
+ transforms.Resize((256, 256)),
108
+ augmentations.ToOneHot(self.opts.label_nc),
109
+ transforms.ToTensor()]),
110
+ 'transform_test': transforms.Compose([
111
+ transforms.Resize((256, 256)),
112
+ transforms.ToTensor(),
113
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
114
+ 'transform_inference': transforms.Compose([
115
+ transforms.Resize((256, 256)),
116
+ augmentations.ToOneHot(self.opts.label_nc),
117
+ transforms.ToTensor()])
118
+ }
119
+ return transforms_dict
120
+
121
+
122
+ class SuperResTransforms(TransformsConfig):
123
+
124
+ def __init__(self, opts):
125
+ super(SuperResTransforms, self).__init__(opts)
126
+
127
+ def get_transforms(self):
128
+ if self.opts.resize_factors is None:
129
+ self.opts.resize_factors = '1,2,4,8,16,32'
130
+ factors = [int(f) for f in self.opts.resize_factors.split(",")]
131
+ print("Performing down-sampling with factors: {}".format(factors))
132
+ transforms_dict = {
133
+ 'transform_gt_train': transforms.Compose([
134
+ transforms.Resize((256, 256)),
135
+ transforms.ToTensor(),
136
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
137
+ 'transform_source': transforms.Compose([
138
+ transforms.Resize((256, 256)),
139
+ augmentations.BilinearResize(factors=factors),
140
+ transforms.Resize((256, 256)),
141
+ transforms.ToTensor(),
142
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
143
+ 'transform_test': transforms.Compose([
144
+ transforms.Resize((256, 256)),
145
+ transforms.ToTensor(),
146
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]),
147
+ 'transform_inference': transforms.Compose([
148
+ transforms.Resize((256, 256)),
149
+ augmentations.BilinearResize(factors=factors),
150
+ transforms.Resize((256, 256)),
151
+ transforms.ToTensor(),
152
+ transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
153
+ }
154
+ return transforms_dict
interface/pixel2style2pixel/criteria/__init__.py ADDED
File without changes
interface/pixel2style2pixel/criteria/id_loss.py ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+ from configs.paths_config import model_paths
4
+ from models.encoders.model_irse import Backbone
5
+
6
+
7
+ class IDLoss(nn.Module):
8
+ def __init__(self):
9
+ super(IDLoss, self).__init__()
10
+ print('Loading ResNet ArcFace')
11
+ self.facenet = Backbone(input_size=112, num_layers=50, drop_ratio=0.6, mode='ir_se')
12
+ self.facenet.load_state_dict(torch.load(model_paths['ir_se50']))
13
+ self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112))
14
+ self.facenet.eval()
15
+
16
+ def extract_feats(self, x):
17
+ x = x[:, :, 35:223, 32:220] # Crop interesting region
18
+ x = self.face_pool(x)
19
+ x_feats = self.facenet(x)
20
+ return x_feats
21
+
22
+ def forward(self, y_hat, y, x):
23
+ n_samples = x.shape[0]
24
+ x_feats = self.extract_feats(x)
25
+ y_feats = self.extract_feats(y) # Otherwise use the feature from there
26
+ y_hat_feats = self.extract_feats(y_hat)
27
+ y_feats = y_feats.detach()
28
+ loss = 0
29
+ sim_improvement = 0
30
+ id_logs = []
31
+ count = 0
32
+ for i in range(n_samples):
33
+ diff_target = y_hat_feats[i].dot(y_feats[i])
34
+ diff_input = y_hat_feats[i].dot(x_feats[i])
35
+ diff_views = y_feats[i].dot(x_feats[i])
36
+ id_logs.append({'diff_target': float(diff_target),
37
+ 'diff_input': float(diff_input),
38
+ 'diff_views': float(diff_views)})
39
+ loss += 1 - diff_target
40
+ id_diff = float(diff_target) - float(diff_views)
41
+ sim_improvement += id_diff
42
+ count += 1
43
+
44
+ return loss / count, sim_improvement / count, id_logs
interface/pixel2style2pixel/criteria/lpips/__init__.py ADDED
File without changes
interface/pixel2style2pixel/criteria/lpips/lpips.py ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+
4
+ from criteria.lpips.networks import get_network, LinLayers
5
+ from criteria.lpips.utils import get_state_dict
6
+
7
+
8
+ class LPIPS(nn.Module):
9
+ r"""Creates a criterion that measures
10
+ Learned Perceptual Image Patch Similarity (LPIPS).
11
+ Arguments:
12
+ net_type (str): the network type to compare the features:
13
+ 'alex' | 'squeeze' | 'vgg'. Default: 'alex'.
14
+ version (str): the version of LPIPS. Default: 0.1.
15
+ """
16
+ def __init__(self, net_type: str = 'alex', version: str = '0.1'):
17
+
18
+ assert version in ['0.1'], 'v0.1 is only supported now'
19
+
20
+ super(LPIPS, self).__init__()
21
+
22
+ # pretrained network
23
+ self.net = get_network(net_type).to("cuda")
24
+
25
+ # linear layers
26
+ self.lin = LinLayers(self.net.n_channels_list).to("cuda")
27
+ self.lin.load_state_dict(get_state_dict(net_type, version))
28
+
29
+ def forward(self, x: torch.Tensor, y: torch.Tensor):
30
+ feat_x, feat_y = self.net(x), self.net(y)
31
+
32
+ diff = [(fx - fy) ** 2 for fx, fy in zip(feat_x, feat_y)]
33
+ res = [l(d).mean((2, 3), True) for d, l in zip(diff, self.lin)]
34
+
35
+ return torch.sum(torch.cat(res, 0)) / x.shape[0]
interface/pixel2style2pixel/criteria/lpips/networks.py ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Sequence
2
+
3
+ from itertools import chain
4
+
5
+ import torch
6
+ import torch.nn as nn
7
+ from torchvision import models
8
+
9
+ from criteria.lpips.utils import normalize_activation
10
+
11
+
12
+ def get_network(net_type: str):
13
+ if net_type == 'alex':
14
+ return AlexNet()
15
+ elif net_type == 'squeeze':
16
+ return SqueezeNet()
17
+ elif net_type == 'vgg':
18
+ return VGG16()
19
+ else:
20
+ raise NotImplementedError('choose net_type from [alex, squeeze, vgg].')
21
+
22
+
23
+ class LinLayers(nn.ModuleList):
24
+ def __init__(self, n_channels_list: Sequence[int]):
25
+ super(LinLayers, self).__init__([
26
+ nn.Sequential(
27
+ nn.Identity(),
28
+ nn.Conv2d(nc, 1, 1, 1, 0, bias=False)
29
+ ) for nc in n_channels_list
30
+ ])
31
+
32
+ for param in self.parameters():
33
+ param.requires_grad = False
34
+
35
+
36
+ class BaseNet(nn.Module):
37
+ def __init__(self):
38
+ super(BaseNet, self).__init__()
39
+
40
+ # register buffer
41
+ self.register_buffer(
42
+ 'mean', torch.Tensor([-.030, -.088, -.188])[None, :, None, None])
43
+ self.register_buffer(
44
+ 'std', torch.Tensor([.458, .448, .450])[None, :, None, None])
45
+
46
+ def set_requires_grad(self, state: bool):
47
+ for param in chain(self.parameters(), self.buffers()):
48
+ param.requires_grad = state
49
+
50
+ def z_score(self, x: torch.Tensor):
51
+ return (x - self.mean) / self.std
52
+
53
+ def forward(self, x: torch.Tensor):
54
+ x = self.z_score(x)
55
+
56
+ output = []
57
+ for i, (_, layer) in enumerate(self.layers._modules.items(), 1):
58
+ x = layer(x)
59
+ if i in self.target_layers:
60
+ output.append(normalize_activation(x))
61
+ if len(output) == len(self.target_layers):
62
+ break
63
+ return output
64
+
65
+
66
+ class SqueezeNet(BaseNet):
67
+ def __init__(self):
68
+ super(SqueezeNet, self).__init__()
69
+
70
+ self.layers = models.squeezenet1_1(True).features
71
+ self.target_layers = [2, 5, 8, 10, 11, 12, 13]
72
+ self.n_channels_list = [64, 128, 256, 384, 384, 512, 512]
73
+
74
+ self.set_requires_grad(False)
75
+
76
+
77
+ class AlexNet(BaseNet):
78
+ def __init__(self):
79
+ super(AlexNet, self).__init__()
80
+
81
+ self.layers = models.alexnet(True).features
82
+ self.target_layers = [2, 5, 8, 10, 12]
83
+ self.n_channels_list = [64, 192, 384, 256, 256]
84
+
85
+ self.set_requires_grad(False)
86
+
87
+
88
+ class VGG16(BaseNet):
89
+ def __init__(self):
90
+ super(VGG16, self).__init__()
91
+
92
+ self.layers = models.vgg16(True).features
93
+ self.target_layers = [4, 9, 16, 23, 30]
94
+ self.n_channels_list = [64, 128, 256, 512, 512]
95
+
96
+ self.set_requires_grad(False)
interface/pixel2style2pixel/criteria/lpips/utils.py ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from collections import OrderedDict
2
+
3
+ import torch
4
+
5
+
6
+ def normalize_activation(x, eps=1e-10):
7
+ norm_factor = torch.sqrt(torch.sum(x ** 2, dim=1, keepdim=True))
8
+ return x / (norm_factor + eps)
9
+
10
+
11
+ def get_state_dict(net_type: str = 'alex', version: str = '0.1'):
12
+ # build url
13
+ url = 'https://raw.githubusercontent.com/richzhang/PerceptualSimilarity/' \
14
+ + f'master/lpips/weights/v{version}/{net_type}.pth'
15
+
16
+ # download
17
+ old_state_dict = torch.hub.load_state_dict_from_url(
18
+ url, progress=True,
19
+ map_location=None if torch.cuda.is_available() else torch.device('cpu')
20
+ )
21
+
22
+ # rename keys
23
+ new_state_dict = OrderedDict()
24
+ for key, val in old_state_dict.items():
25
+ new_key = key
26
+ new_key = new_key.replace('lin', '')
27
+ new_key = new_key.replace('model.', '')
28
+ new_state_dict[new_key] = val
29
+
30
+ return new_state_dict
interface/pixel2style2pixel/criteria/moco_loss.py ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+ import torch.nn.functional as F
4
+ from configs.paths_config import model_paths
5
+
6
+
7
+ class MocoLoss(nn.Module):
8
+
9
+ def __init__(self):
10
+ super(MocoLoss, self).__init__()
11
+ print("Loading MOCO model from path: {}".format(model_paths["moco"]))
12
+ self.model = self.__load_model()
13
+ self.model.cuda()
14
+ self.model.eval()
15
+
16
+ @staticmethod
17
+ def __load_model():
18
+ import torchvision.models as models
19
+ model = models.__dict__["resnet50"]()
20
+ # freeze all layers but the last fc
21
+ for name, param in model.named_parameters():
22
+ if name not in ['fc.weight', 'fc.bias']:
23
+ param.requires_grad = False
24
+ checkpoint = torch.load(model_paths['moco'], map_location="cpu")
25
+ state_dict = checkpoint['state_dict']
26
+ # rename moco pre-trained keys
27
+ for k in list(state_dict.keys()):
28
+ # retain only encoder_q up to before the embedding layer
29
+ if k.startswith('module.encoder_q') and not k.startswith('module.encoder_q.fc'):
30
+ # remove prefix
31
+ state_dict[k[len("module.encoder_q."):]] = state_dict[k]
32
+ # delete renamed or unused k
33
+ del state_dict[k]
34
+ msg = model.load_state_dict(state_dict, strict=False)
35
+ assert set(msg.missing_keys) == {"fc.weight", "fc.bias"}
36
+ # remove output layer
37
+ model = nn.Sequential(*list(model.children())[:-1]).cuda()
38
+ return model
39
+
40
+ def extract_feats(self, x):
41
+ x = F.interpolate(x, size=224)
42
+ x_feats = self.model(x)
43
+ x_feats = nn.functional.normalize(x_feats, dim=1)
44
+ x_feats = x_feats.squeeze()
45
+ return x_feats
46
+
47
+ def forward(self, y_hat, y, x):
48
+ n_samples = x.shape[0]
49
+ x_feats = self.extract_feats(x)
50
+ y_feats = self.extract_feats(y)
51
+ y_hat_feats = self.extract_feats(y_hat)
52
+ y_feats = y_feats.detach()
53
+ loss = 0
54
+ sim_improvement = 0
55
+ sim_logs = []
56
+ count = 0
57
+ for i in range(n_samples):
58
+ diff_target = y_hat_feats[i].dot(y_feats[i])
59
+ diff_input = y_hat_feats[i].dot(x_feats[i])
60
+ diff_views = y_feats[i].dot(x_feats[i])
61
+ sim_logs.append({'diff_target': float(diff_target),
62
+ 'diff_input': float(diff_input),
63
+ 'diff_views': float(diff_views)})
64
+ loss += 1 - diff_target
65
+ sim_diff = float(diff_target) - float(diff_views)
66
+ sim_improvement += sim_diff
67
+ count += 1
68
+
69
+ return loss / count, sim_improvement / count, sim_logs
interface/pixel2style2pixel/criteria/w_norm.py ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+
4
+
5
+ class WNormLoss(nn.Module):
6
+
7
+ def __init__(self, start_from_latent_avg=True):
8
+ super(WNormLoss, self).__init__()
9
+ self.start_from_latent_avg = start_from_latent_avg
10
+
11
+ def forward(self, latent, latent_avg=None):
12
+ if self.start_from_latent_avg:
13
+ latent = latent - latent_avg
14
+ return torch.sum(latent.norm(2, dim=(1, 2))) / latent.shape[0]
interface/pixel2style2pixel/datasets/__init__.py ADDED
File without changes
interface/pixel2style2pixel/datasets/augmentations.py ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import torch
3
+ from torch import nn
4
+ from torch.nn import functional as F
5
+ from torchvision import transforms
6
+
7
+
8
+ class ToOneHot(object):
9
+ """ Convert the input PIL image to a one-hot torch tensor """
10
+ def __init__(self, n_classes=None):
11
+ self.n_classes = n_classes
12
+
13
+ def onehot_initialization(self, a):
14
+ if self.n_classes is None:
15
+ self.n_classes = len(np.unique(a))
16
+ out = np.zeros(a.shape + (self.n_classes, ), dtype=int)
17
+ out[self.__all_idx(a, axis=2)] = 1
18
+ return out
19
+
20
+ def __all_idx(self, idx, axis):
21
+ grid = np.ogrid[tuple(map(slice, idx.shape))]
22
+ grid.insert(axis, idx)
23
+ return tuple(grid)
24
+
25
+ def __call__(self, img):
26
+ img = np.array(img)
27
+ one_hot = self.onehot_initialization(img)
28
+ return one_hot
29
+
30
+
31
+ class BilinearResize(object):
32
+ def __init__(self, factors=[1, 2, 4, 8, 16, 32]):
33
+ self.factors = factors
34
+
35
+ def __call__(self, image):
36
+ factor = np.random.choice(self.factors, size=1)[0]
37
+ D = BicubicDownSample(factor=factor, cuda=False)
38
+ img_tensor = transforms.ToTensor()(image).unsqueeze(0)
39
+ img_tensor_lr = D(img_tensor)[0].clamp(0, 1)
40
+ img_low_res = transforms.ToPILImage()(img_tensor_lr)
41
+ return img_low_res
42
+
43
+
44
+ class BicubicDownSample(nn.Module):
45
+ def bicubic_kernel(self, x, a=-0.50):
46
+ """
47
+ This equation is exactly copied from the website below:
48
+ https://clouard.users.greyc.fr/Pantheon/experiments/rescaling/index-en.html#bicubic
49
+ """
50
+ abs_x = torch.abs(x)
51
+ if abs_x <= 1.:
52
+ return (a + 2.) * torch.pow(abs_x, 3.) - (a + 3.) * torch.pow(abs_x, 2.) + 1
53
+ elif 1. < abs_x < 2.:
54
+ return a * torch.pow(abs_x, 3) - 5. * a * torch.pow(abs_x, 2.) + 8. * a * abs_x - 4. * a
55
+ else:
56
+ return 0.0
57
+
58
+ def __init__(self, factor=4, cuda=True, padding='reflect'):
59
+ super().__init__()
60
+ self.factor = factor
61
+ size = factor * 4
62
+ k = torch.tensor([self.bicubic_kernel((i - torch.floor(torch.tensor(size / 2)) + 0.5) / factor)
63
+ for i in range(size)], dtype=torch.float32)
64
+ k = k / torch.sum(k)
65
+ k1 = torch.reshape(k, shape=(1, 1, size, 1))
66
+ self.k1 = torch.cat([k1, k1, k1], dim=0)
67
+ k2 = torch.reshape(k, shape=(1, 1, 1, size))
68
+ self.k2 = torch.cat([k2, k2, k2], dim=0)
69
+ self.cuda = '.cuda' if cuda else ''
70
+ self.padding = padding
71
+ for param in self.parameters():
72
+ param.requires_grad = False
73
+
74
+ def forward(self, x, nhwc=False, clip_round=False, byte_output=False):
75
+ filter_height = self.factor * 4
76
+ filter_width = self.factor * 4
77
+ stride = self.factor
78
+
79
+ pad_along_height = max(filter_height - stride, 0)
80
+ pad_along_width = max(filter_width - stride, 0)
81
+ filters1 = self.k1.type('torch{}.FloatTensor'.format(self.cuda))
82
+ filters2 = self.k2.type('torch{}.FloatTensor'.format(self.cuda))
83
+
84
+ # compute actual padding values for each side
85
+ pad_top = pad_along_height // 2
86
+ pad_bottom = pad_along_height - pad_top
87
+ pad_left = pad_along_width // 2
88
+ pad_right = pad_along_width - pad_left
89
+
90
+ # apply mirror padding
91
+ if nhwc:
92
+ x = torch.transpose(torch.transpose(x, 2, 3), 1, 2) # NHWC to NCHW
93
+
94
+ # downscaling performed by 1-d convolution
95
+ x = F.pad(x, (0, 0, pad_top, pad_bottom), self.padding)
96
+ x = F.conv2d(input=x, weight=filters1, stride=(stride, 1), groups=3)
97
+ if clip_round:
98
+ x = torch.clamp(torch.round(x), 0.0, 255.)
99
+
100
+ x = F.pad(x, (pad_left, pad_right, 0, 0), self.padding)
101
+ x = F.conv2d(input=x, weight=filters2, stride=(1, stride), groups=3)
102
+ if clip_round:
103
+ x = torch.clamp(torch.round(x), 0.0, 255.)
104
+
105
+ if nhwc:
106
+ x = torch.transpose(torch.transpose(x, 1, 3), 1, 2)
107
+ if byte_output:
108
+ return x.type('torch.ByteTensor'.format(self.cuda))
109
+ else:
110
+ return x
interface/pixel2style2pixel/datasets/gt_res_dataset.py ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/python
2
+ # encoding: utf-8
3
+ import os
4
+ from torch.utils.data import Dataset
5
+ from PIL import Image
6
+
7
+
8
+ class GTResDataset(Dataset):
9
+
10
+ def __init__(self, root_path, gt_dir=None, transform=None, transform_train=None):
11
+ self.pairs = []
12
+ for f in os.listdir(root_path):
13
+ image_path = os.path.join(root_path, f)
14
+ gt_path = os.path.join(gt_dir, f)
15
+ if f.endswith(".jpg") or f.endswith(".png"):
16
+ self.pairs.append([image_path, gt_path.replace('.png', '.jpg'), None])
17
+ self.transform = transform
18
+ self.transform_train = transform_train
19
+
20
+ def __len__(self):
21
+ return len(self.pairs)
22
+
23
+ def __getitem__(self, index):
24
+ from_path, to_path, _ = self.pairs[index]
25
+ from_im = Image.open(from_path).convert('RGB')
26
+ to_im = Image.open(to_path).convert('RGB')
27
+
28
+ if self.transform:
29
+ to_im = self.transform(to_im)
30
+ from_im = self.transform(from_im)
31
+
32
+ return from_im, to_im
interface/pixel2style2pixel/datasets/images_dataset.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from torch.utils.data import Dataset
2
+ from PIL import Image
3
+ from utils import data_utils
4
+
5
+
6
+ class ImagesDataset(Dataset):
7
+
8
+ def __init__(self, source_root, target_root, opts, target_transform=None, source_transform=None):
9
+ self.source_paths = sorted(data_utils.make_dataset(source_root))
10
+ self.target_paths = sorted(data_utils.make_dataset(target_root))
11
+ self.source_transform = source_transform
12
+ self.target_transform = target_transform
13
+ self.opts = opts
14
+
15
+ def __len__(self):
16
+ return len(self.source_paths)
17
+
18
+ def __getitem__(self, index):
19
+ from_path = self.source_paths[index]
20
+ from_im = Image.open(from_path)
21
+ from_im = from_im.convert('RGB') if self.opts.label_nc == 0 else from_im.convert('L')
22
+
23
+ to_path = self.target_paths[index]
24
+ to_im = Image.open(to_path).convert('RGB')
25
+ if self.target_transform:
26
+ to_im = self.target_transform(to_im)
27
+
28
+ if self.source_transform:
29
+ from_im = self.source_transform(from_im)
30
+ else:
31
+ from_im = to_im
32
+
33
+ return from_im, to_im
interface/pixel2style2pixel/datasets/inference_dataset.py ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from torch.utils.data import Dataset
2
+ from PIL import Image
3
+ from utils import data_utils
4
+
5
+
6
+ class InferenceDataset(Dataset):
7
+
8
+ def __init__(self, root, opts, transform=None):
9
+ self.paths = sorted(data_utils.make_dataset(root))
10
+ self.transform = transform
11
+ self.opts = opts
12
+
13
+ def __len__(self):
14
+ return len(self.paths)
15
+
16
+ def __getitem__(self, index):
17
+ from_path = self.paths[index]
18
+ from_im = Image.open(from_path)
19
+ from_im = from_im.convert('RGB') if self.opts.label_nc == 0 else from_im.convert('L')
20
+ if self.transform:
21
+ from_im = self.transform(from_im)
22
+ return from_im
interface/pixel2style2pixel/docs/encoding_inputs.jpg ADDED
interface/pixel2style2pixel/docs/encoding_outputs.jpg ADDED
interface/pixel2style2pixel/docs/frontalization_inputs.jpg ADDED
interface/pixel2style2pixel/docs/frontalization_outputs.jpg ADDED
interface/pixel2style2pixel/docs/seg2image.png ADDED

Git LFS Details

  • SHA256: 4ab1c1e95833c45509444fc7a81f7bc995c8d38dd8ede2fe2a80d6dc6f954a7c
  • Pointer size: 132 Bytes
  • Size of remote file: 1.89 MB
interface/pixel2style2pixel/docs/sketch2image.png ADDED

Git LFS Details

  • SHA256: 03bd36027bb05fd1c7a132ad6cebb5e2527d911fa517e4f8561667de5eaa18de
  • Pointer size: 132 Bytes
  • Size of remote file: 2.09 MB
interface/pixel2style2pixel/docs/super_res_32.jpg ADDED
interface/pixel2style2pixel/docs/super_res_style_mixing.jpg ADDED
interface/pixel2style2pixel/docs/teaser.png ADDED

Git LFS Details

  • SHA256: 98445b52153dffaadf92a491424e8cbd805bd0ca25385afcd8d40f4403254a76
  • Pointer size: 132 Bytes
  • Size of remote file: 5.16 MB
interface/pixel2style2pixel/docs/toonify_input.jpg ADDED
interface/pixel2style2pixel/docs/toonify_output.jpg ADDED
interface/pixel2style2pixel/download-weights.sh ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/sh
2
+ mkdir pretrained_models
3
+ cd pretrained_models
4
+
5
+ wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1lB7wk7MwtdxL-LL4Z_T76DuCfk00aSXA' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1lB7wk7MwtdxL-LL4Z_T76DuCfk00aSXA" -O psp_celebs_sketch_to_face.pt && rm -rf /tmp/cookies.txt
6
+ wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1_S4THAzXb-97DbpXmanjHtXRyKxqjARv' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1_S4THAzXb-97DbpXmanjHtXRyKxqjARv" -O psp_ffhq_frontalization.pt && rm -rf /tmp/cookies.txt
7
+ wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1ZpmSXBpJ9pFEov6-jjQstAlfYbkebECu' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1ZpmSXBpJ9pFEov6-jjQstAlfYbkebECu" -O psp_celebs_super_resolution.pt && rm -rf /tmp/cookies.txt
8
+ wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1YKoiVuFaqdvzDP5CZaqa3k5phL-VDmyz' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1YKoiVuFaqdvzDP5CZaqa3k5phL-VDmyz" -O psp_ffhq_toonify.pt && rm -rf /tmp/cookies.txt
9
+
10
+ cd ..
11
+ wget http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2
12
+ bunzip2 shape_predictor_68_face_landmarks.dat.bz2
interface/pixel2style2pixel/environment/psp_env.yaml ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: psp_env
2
+ channels:
3
+ - conda-forge
4
+ - defaults
5
+ dependencies:
6
+ - _libgcc_mutex=0.1=main
7
+ - ca-certificates=2020.4.5.1=hecc5488_0
8
+ - certifi=2020.4.5.1=py36h9f0ad1d_0
9
+ - libedit=3.1.20181209=hc058e9b_0
10
+ - libffi=3.2.1=hd88cf55_4
11
+ - libgcc-ng=9.1.0=hdf63c60_0
12
+ - libstdcxx-ng=9.1.0=hdf63c60_0
13
+ - ncurses=6.2=he6710b0_1
14
+ - ninja=1.10.0=hc9558a2_0
15
+ - openssl=1.1.1g=h516909a_0
16
+ - pip=20.0.2=py36_3
17
+ - python=3.6.7=h0371630_0
18
+ - python_abi=3.6=1_cp36m
19
+ - readline=7.0=h7b6447c_5
20
+ - setuptools=46.4.0=py36_0
21
+ - sqlite=3.31.1=h62c20be_1
22
+ - tk=8.6.8=hbc83047_0
23
+ - wheel=0.34.2=py36_0
24
+ - xz=5.2.5=h7b6447c_0
25
+ - zlib=1.2.11=h7b6447c_3
26
+ - pip:
27
+ - scipy==1.4.1
28
+ - matplotlib==3.2.1
29
+ - tqdm==4.46.0
30
+ - numpy==1.18.4
31
+ - opencv-python==4.2.0.34
32
+ - pillow==7.1.2
33
+ - tensorboard==2.2.1
34
+ - torch==1.6.0
35
+ - torchvision==0.4.2
36
+ prefix: ~/anaconda3/envs/psp_env
37
+
interface/pixel2style2pixel/licenses/LICENSE_HuangYG123 ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2020 HuangYG123
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
interface/pixel2style2pixel/licenses/LICENSE_S-aiueo32 ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ BSD 2-Clause License
2
+
3
+ Copyright (c) 2020, Sou Uchida
4
+ All rights reserved.
5
+
6
+ Redistribution and use in source and binary forms, with or without
7
+ modification, are permitted provided that the following conditions are met:
8
+
9
+ 1. Redistributions of source code must retain the above copyright notice, this
10
+ list of conditions and the following disclaimer.
11
+
12
+ 2. Redistributions in binary form must reproduce the above copyright notice,
13
+ this list of conditions and the following disclaimer in the documentation
14
+ and/or other materials provided with the distribution.
15
+
16
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
17
+ AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18
+ IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
20
+ FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21
+ DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22
+ SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23
+ CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24
+ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
25
+ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
interface/pixel2style2pixel/licenses/LICENSE_TreB1eN ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2018 TreB1eN
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
interface/pixel2style2pixel/licenses/LICENSE_lessw2020 ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Apache License
2
+ Version 2.0, January 2004
3
+ http://www.apache.org/licenses/
4
+
5
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6
+
7
+ 1. Definitions.
8
+
9
+ "License" shall mean the terms and conditions for use, reproduction,
10
+ and distribution as defined by Sections 1 through 9 of this document.
11
+
12
+ "Licensor" shall mean the copyright owner or entity authorized by
13
+ the copyright owner that is granting the License.
14
+
15
+ "Legal Entity" shall mean the union of the acting entity and all
16
+ other entities that control, are controlled by, or are under common
17
+ control with that entity. For the purposes of this definition,
18
+ "control" means (i) the power, direct or indirect, to cause the
19
+ direction or management of such entity, whether by contract or
20
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
21
+ outstanding shares, or (iii) beneficial ownership of such entity.
22
+
23
+ "You" (or "Your") shall mean an individual or Legal Entity
24
+ exercising permissions granted by this License.
25
+
26
+ "Source" form shall mean the preferred form for making modifications,
27
+ including but not limited to software source code, documentation
28
+ source, and configuration files.
29
+
30
+ "Object" form shall mean any form resulting from mechanical
31
+ transformation or translation of a Source form, including but
32
+ not limited to compiled object code, generated documentation,
33
+ and conversions to other media types.
34
+
35
+ "Work" shall mean the work of authorship, whether in Source or
36
+ Object form, made available under the License, as indicated by a
37
+ copyright notice that is included in or attached to the work
38
+ (an example is provided in the Appendix below).
39
+
40
+ "Derivative Works" shall mean any work, whether in Source or Object
41
+ form, that is based on (or derived from) the Work and for which the
42
+ editorial revisions, annotations, elaborations, or other modifications
43
+ represent, as a whole, an original work of authorship. For the purposes
44
+ of this License, Derivative Works shall not include works that remain
45
+ separable from, or merely link (or bind by name) to the interfaces of,
46
+ the Work and Derivative Works thereof.
47
+
48
+ "Contribution" shall mean any work of authorship, including
49
+ the original version of the Work and any modifications or additions
50
+ to that Work or Derivative Works thereof, that is intentionally
51
+ submitted to Licensor for inclusion in the Work by the copyright owner
52
+ or by an individual or Legal Entity authorized to submit on behalf of
53
+ the copyright owner. For the purposes of this definition, "submitted"
54
+ means any form of electronic, verbal, or written communication sent
55
+ to the Licensor or its representatives, including but not limited to
56
+ communication on electronic mailing lists, source code control systems,
57
+ and issue tracking systems that are managed by, or on behalf of, the
58
+ Licensor for the purpose of discussing and improving the Work, but
59
+ excluding communication that is conspicuously marked or otherwise
60
+ designated in writing by the copyright owner as "Not a Contribution."
61
+
62
+ "Contributor" shall mean Licensor and any individual or Legal Entity
63
+ on behalf of whom a Contribution has been received by Licensor and
64
+ subsequently incorporated within the Work.
65
+
66
+ 2. Grant of Copyright License. Subject to the terms and conditions of
67
+ this License, each Contributor hereby grants to You a perpetual,
68
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69
+ copyright license to reproduce, prepare Derivative Works of,
70
+ publicly display, publicly perform, sublicense, and distribute the
71
+ Work and such Derivative Works in Source or Object form.
72
+
73
+ 3. Grant of Patent License. Subject to the terms and conditions of
74
+ this License, each Contributor hereby grants to You a perpetual,
75
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76
+ (except as stated in this section) patent license to make, have made,
77
+ use, offer to sell, sell, import, and otherwise transfer the Work,
78
+ where such license applies only to those patent claims licensable
79
+ by such Contributor that are necessarily infringed by their
80
+ Contribution(s) alone or by combination of their Contribution(s)
81
+ with the Work to which such Contribution(s) was submitted. If You
82
+ institute patent litigation against any entity (including a
83
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
84
+ or a Contribution incorporated within the Work constitutes direct
85
+ or contributory patent infringement, then any patent licenses
86
+ granted to You under this License for that Work shall terminate
87
+ as of the date such litigation is filed.
88
+
89
+ 4. Redistribution. You may reproduce and distribute copies of the
90
+ Work or Derivative Works thereof in any medium, with or without
91
+ modifications, and in Source or Object form, provided that You
92
+ meet the following conditions:
93
+
94
+ (a) You must give any other recipients of the Work or
95
+ Derivative Works a copy of this License; and
96
+
97
+ (b) You must cause any modified files to carry prominent notices
98
+ stating that You changed the files; and
99
+
100
+ (c) You must retain, in the Source form of any Derivative Works
101
+ that You distribute, all copyright, patent, trademark, and
102
+ attribution notices from the Source form of the Work,
103
+ excluding those notices that do not pertain to any part of
104
+ the Derivative Works; and
105
+
106
+ (d) If the Work includes a "NOTICE" text file as part of its
107
+ distribution, then any Derivative Works that You distribute must
108
+ include a readable copy of the attribution notices contained
109
+ within such NOTICE file, excluding those notices that do not
110
+ pertain to any part of the Derivative Works, in at least one
111
+ of the following places: within a NOTICE text file distributed
112
+ as part of the Derivative Works; within the Source form or
113
+ documentation, if provided along with the Derivative Works; or,
114
+ within a display generated by the Derivative Works, if and
115
+ wherever such third-party notices normally appear. The contents
116
+ of the NOTICE file are for informational purposes only and
117
+ do not modify the License. You may add Your own attribution
118
+ notices within Derivative Works that You distribute, alongside
119
+ or as an addendum to the NOTICE text from the Work, provided
120
+ that such additional attribution notices cannot be construed
121
+ as modifying the License.
122
+
123
+ You may add Your own copyright statement to Your modifications and
124
+ may provide additional or different license terms and conditions
125
+ for use, reproduction, or distribution of Your modifications, or
126
+ for any such Derivative Works as a whole, provided Your use,
127
+ reproduction, and distribution of the Work otherwise complies with
128
+ the conditions stated in this License.
129
+
130
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
131
+ any Contribution intentionally submitted for inclusion in the Work
132
+ by You to the Licensor shall be under the terms and conditions of
133
+ this License, without any additional terms or conditions.
134
+ Notwithstanding the above, nothing herein shall supersede or modify
135
+ the terms of any separate license agreement you may have executed
136
+ with Licensor regarding such Contributions.
137
+
138
+ 6. Trademarks. This License does not grant permission to use the trade
139
+ names, trademarks, service marks, or product names of the Licensor,
140
+ except as required for reasonable and customary use in describing the
141
+ origin of the Work and reproducing the content of the NOTICE file.
142
+
143
+ 7. Disclaimer of Warranty. Unless required by applicable law or
144
+ agreed to in writing, Licensor provides the Work (and each
145
+ Contributor provides its Contributions) on an "AS IS" BASIS,
146
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147
+ implied, including, without limitation, any warranties or conditions
148
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149
+ PARTICULAR PURPOSE. You are solely responsible for determining the
150
+ appropriateness of using or redistributing the Work and assume any
151
+ risks associated with Your exercise of permissions under this License.
152
+
153
+ 8. Limitation of Liability. In no event and under no legal theory,
154
+ whether in tort (including negligence), contract, or otherwise,
155
+ unless required by applicable law (such as deliberate and grossly
156
+ negligent acts) or agreed to in writing, shall any Contributor be
157
+ liable to You for damages, including any direct, indirect, special,
158
+ incidental, or consequential damages of any character arising as a
159
+ result of this License or out of the use or inability to use the
160
+ Work (including but not limited to damages for loss of goodwill,
161
+ work stoppage, computer failure or malfunction, or any and all
162
+ other commercial damages or losses), even if such Contributor
163
+ has been advised of the possibility of such damages.
164
+
165
+ 9. Accepting Warranty or Additional Liability. While redistributing
166
+ the Work or Derivative Works thereof, You may choose to offer,
167
+ and charge a fee for, acceptance of support, warranty, indemnity,
168
+ or other liability obligations and/or rights consistent with this
169
+ License. However, in accepting such obligations, You may act only
170
+ on Your own behalf and on Your sole responsibility, not on behalf
171
+ of any other Contributor, and only if You agree to indemnify,
172
+ defend, and hold each Contributor harmless for any liability
173
+ incurred by, or claims asserted against, such Contributor by reason
174
+ of your accepting any such warranty or additional liability.
175
+
176
+ END OF TERMS AND CONDITIONS
177
+
178
+ APPENDIX: How to apply the Apache License to your work.
179
+
180
+ To apply the Apache License to your work, attach the following
181
+ boilerplate notice, with the fields enclosed by brackets "[]"
182
+ replaced with your own identifying information. (Don't include
183
+ the brackets!) The text should be enclosed in the appropriate
184
+ comment syntax for the file format. We also recommend that a
185
+ file or class name and description of purpose be included on the
186
+ same "printed page" as the copyright notice for easier
187
+ identification within third-party archives.
188
+
189
+ Copyright [yyyy] [name of copyright owner]
190
+
191
+ Licensed under the Apache License, Version 2.0 (the "License");
192
+ you may not use this file except in compliance with the License.
193
+ You may obtain a copy of the License at
194
+
195
+ http://www.apache.org/licenses/LICENSE-2.0
196
+
197
+ Unless required by applicable law or agreed to in writing, software
198
+ distributed under the License is distributed on an "AS IS" BASIS,
199
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200
+ See the License for the specific language governing permissions and
201
+ limitations under the License.
interface/pixel2style2pixel/licenses/LICENSE_rosinality ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2019 Kim Seonghyeon
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
interface/pixel2style2pixel/models/__init__.py ADDED
File without changes
interface/pixel2style2pixel/models/encoders/__init__.py ADDED
File without changes
interface/pixel2style2pixel/models/encoders/helpers.py ADDED
@@ -0,0 +1,119 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from collections import namedtuple
2
+ import torch
3
+ from torch.nn import Conv2d, BatchNorm2d, PReLU, ReLU, Sigmoid, MaxPool2d, AdaptiveAvgPool2d, Sequential, Module
4
+
5
+ """
6
+ ArcFace implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
7
+ """
8
+
9
+
10
+ class Flatten(Module):
11
+ def forward(self, input):
12
+ return input.view(input.size(0), -1)
13
+
14
+
15
+ def l2_norm(input, axis=1):
16
+ norm = torch.norm(input, 2, axis, True)
17
+ output = torch.div(input, norm)
18
+ return output
19
+
20
+
21
+ class Bottleneck(namedtuple('Block', ['in_channel', 'depth', 'stride'])):
22
+ """ A named tuple describing a ResNet block. """
23
+
24
+
25
+ def get_block(in_channel, depth, num_units, stride=2):
26
+ return [Bottleneck(in_channel, depth, stride)] + [Bottleneck(depth, depth, 1) for i in range(num_units - 1)]
27
+
28
+
29
+ def get_blocks(num_layers):
30
+ if num_layers == 50:
31
+ blocks = [
32
+ get_block(in_channel=64, depth=64, num_units=3),
33
+ get_block(in_channel=64, depth=128, num_units=4),
34
+ get_block(in_channel=128, depth=256, num_units=14),
35
+ get_block(in_channel=256, depth=512, num_units=3)
36
+ ]
37
+ elif num_layers == 100:
38
+ blocks = [
39
+ get_block(in_channel=64, depth=64, num_units=3),
40
+ get_block(in_channel=64, depth=128, num_units=13),
41
+ get_block(in_channel=128, depth=256, num_units=30),
42
+ get_block(in_channel=256, depth=512, num_units=3)
43
+ ]
44
+ elif num_layers == 152:
45
+ blocks = [
46
+ get_block(in_channel=64, depth=64, num_units=3),
47
+ get_block(in_channel=64, depth=128, num_units=8),
48
+ get_block(in_channel=128, depth=256, num_units=36),
49
+ get_block(in_channel=256, depth=512, num_units=3)
50
+ ]
51
+ else:
52
+ raise ValueError("Invalid number of layers: {}. Must be one of [50, 100, 152]".format(num_layers))
53
+ return blocks
54
+
55
+
56
+ class SEModule(Module):
57
+ def __init__(self, channels, reduction):
58
+ super(SEModule, self).__init__()
59
+ self.avg_pool = AdaptiveAvgPool2d(1)
60
+ self.fc1 = Conv2d(channels, channels // reduction, kernel_size=1, padding=0, bias=False)
61
+ self.relu = ReLU(inplace=True)
62
+ self.fc2 = Conv2d(channels // reduction, channels, kernel_size=1, padding=0, bias=False)
63
+ self.sigmoid = Sigmoid()
64
+
65
+ def forward(self, x):
66
+ module_input = x
67
+ x = self.avg_pool(x)
68
+ x = self.fc1(x)
69
+ x = self.relu(x)
70
+ x = self.fc2(x)
71
+ x = self.sigmoid(x)
72
+ return module_input * x
73
+
74
+
75
+ class bottleneck_IR(Module):
76
+ def __init__(self, in_channel, depth, stride):
77
+ super(bottleneck_IR, self).__init__()
78
+ if in_channel == depth:
79
+ self.shortcut_layer = MaxPool2d(1, stride)
80
+ else:
81
+ self.shortcut_layer = Sequential(
82
+ Conv2d(in_channel, depth, (1, 1), stride, bias=False),
83
+ BatchNorm2d(depth)
84
+ )
85
+ self.res_layer = Sequential(
86
+ BatchNorm2d(in_channel),
87
+ Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), PReLU(depth),
88
+ Conv2d(depth, depth, (3, 3), stride, 1, bias=False), BatchNorm2d(depth)
89
+ )
90
+
91
+ def forward(self, x):
92
+ shortcut = self.shortcut_layer(x)
93
+ res = self.res_layer(x)
94
+ return res + shortcut
95
+
96
+
97
+ class bottleneck_IR_SE(Module):
98
+ def __init__(self, in_channel, depth, stride):
99
+ super(bottleneck_IR_SE, self).__init__()
100
+ if in_channel == depth:
101
+ self.shortcut_layer = MaxPool2d(1, stride)
102
+ else:
103
+ self.shortcut_layer = Sequential(
104
+ Conv2d(in_channel, depth, (1, 1), stride, bias=False),
105
+ BatchNorm2d(depth)
106
+ )
107
+ self.res_layer = Sequential(
108
+ BatchNorm2d(in_channel),
109
+ Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False),
110
+ PReLU(depth),
111
+ Conv2d(depth, depth, (3, 3), stride, 1, bias=False),
112
+ BatchNorm2d(depth),
113
+ SEModule(depth, 16)
114
+ )
115
+
116
+ def forward(self, x):
117
+ shortcut = self.shortcut_layer(x)
118
+ res = self.res_layer(x)
119
+ return res + shortcut
interface/pixel2style2pixel/models/encoders/model_irse.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, Dropout, Sequential, Module
2
+ from models.encoders.helpers import get_blocks, Flatten, bottleneck_IR, bottleneck_IR_SE, l2_norm
3
+
4
+ """
5
+ Modified Backbone implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
6
+ """
7
+
8
+
9
+ class Backbone(Module):
10
+ def __init__(self, input_size, num_layers, mode='ir', drop_ratio=0.4, affine=True):
11
+ super(Backbone, self).__init__()
12
+ assert input_size in [112, 224], "input_size should be 112 or 224"
13
+ assert num_layers in [50, 100, 152], "num_layers should be 50, 100 or 152"
14
+ assert mode in ['ir', 'ir_se'], "mode should be ir or ir_se"
15
+ blocks = get_blocks(num_layers)
16
+ if mode == 'ir':
17
+ unit_module = bottleneck_IR
18
+ elif mode == 'ir_se':
19
+ unit_module = bottleneck_IR_SE
20
+ self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False),
21
+ BatchNorm2d(64),
22
+ PReLU(64))
23
+ if input_size == 112:
24
+ self.output_layer = Sequential(BatchNorm2d(512),
25
+ Dropout(drop_ratio),
26
+ Flatten(),
27
+ Linear(512 * 7 * 7, 512),
28
+ BatchNorm1d(512, affine=affine))
29
+ else:
30
+ self.output_layer = Sequential(BatchNorm2d(512),
31
+ Dropout(drop_ratio),
32
+ Flatten(),
33
+ Linear(512 * 14 * 14, 512),
34
+ BatchNorm1d(512, affine=affine))
35
+
36
+ modules = []
37
+ for block in blocks:
38
+ for bottleneck in block:
39
+ modules.append(unit_module(bottleneck.in_channel,
40
+ bottleneck.depth,
41
+ bottleneck.stride))
42
+ self.body = Sequential(*modules)
43
+
44
+ def forward(self, x):
45
+ x = self.input_layer(x)
46
+ x = self.body(x)
47
+ x = self.output_layer(x)
48
+ return l2_norm(x)
49
+
50
+
51
+ def IR_50(input_size):
52
+ """Constructs a ir-50 model."""
53
+ model = Backbone(input_size, num_layers=50, mode='ir', drop_ratio=0.4, affine=False)
54
+ return model
55
+
56
+
57
+ def IR_101(input_size):
58
+ """Constructs a ir-101 model."""
59
+ model = Backbone(input_size, num_layers=100, mode='ir', drop_ratio=0.4, affine=False)
60
+ return model
61
+
62
+
63
+ def IR_152(input_size):
64
+ """Constructs a ir-152 model."""
65
+ model = Backbone(input_size, num_layers=152, mode='ir', drop_ratio=0.4, affine=False)
66
+ return model
67
+
68
+
69
+ def IR_SE_50(input_size):
70
+ """Constructs a ir_se-50 model."""
71
+ model = Backbone(input_size, num_layers=50, mode='ir_se', drop_ratio=0.4, affine=False)
72
+ return model
73
+
74
+
75
+ def IR_SE_101(input_size):
76
+ """Constructs a ir_se-101 model."""
77
+ model = Backbone(input_size, num_layers=100, mode='ir_se', drop_ratio=0.4, affine=False)
78
+ return model
79
+
80
+
81
+ def IR_SE_152(input_size):
82
+ """Constructs a ir_se-152 model."""
83
+ model = Backbone(input_size, num_layers=152, mode='ir_se', drop_ratio=0.4, affine=False)
84
+ return model
interface/pixel2style2pixel/models/encoders/psp_encoders.py ADDED
@@ -0,0 +1,186 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import torch
3
+ import torch.nn.functional as F
4
+ from torch import nn
5
+ from torch.nn import Linear, Conv2d, BatchNorm2d, PReLU, Sequential, Module
6
+
7
+ from pixel2style2pixel.models.encoders.helpers import get_blocks, Flatten, bottleneck_IR, bottleneck_IR_SE
8
+ from pixel2style2pixel.models.stylegan2.model import EqualLinear
9
+
10
+
11
+ class GradualStyleBlock(Module):
12
+ def __init__(self, in_c, out_c, spatial):
13
+ super(GradualStyleBlock, self).__init__()
14
+ self.out_c = out_c
15
+ self.spatial = spatial
16
+ num_pools = int(np.log2(spatial))
17
+ modules = []
18
+ modules += [Conv2d(in_c, out_c, kernel_size=3, stride=2, padding=1),
19
+ nn.LeakyReLU()]
20
+ for i in range(num_pools - 1):
21
+ modules += [
22
+ Conv2d(out_c, out_c, kernel_size=3, stride=2, padding=1),
23
+ nn.LeakyReLU()
24
+ ]
25
+ self.convs = nn.Sequential(*modules)
26
+ self.linear = EqualLinear(out_c, out_c, lr_mul=1)
27
+
28
+ def forward(self, x):
29
+ x = self.convs(x)
30
+ x = x.view(-1, self.out_c)
31
+ x = self.linear(x)
32
+ return x
33
+
34
+
35
+ class GradualStyleEncoder(Module):
36
+ def __init__(self, num_layers, mode='ir', opts=None):
37
+ super(GradualStyleEncoder, self).__init__()
38
+ assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152'
39
+ assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
40
+ blocks = get_blocks(num_layers)
41
+ if mode == 'ir':
42
+ unit_module = bottleneck_IR
43
+ elif mode == 'ir_se':
44
+ unit_module = bottleneck_IR_SE
45
+ self.input_layer = Sequential(Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False),
46
+ BatchNorm2d(64),
47
+ PReLU(64))
48
+ modules = []
49
+ for block in blocks:
50
+ for bottleneck in block:
51
+ modules.append(unit_module(bottleneck.in_channel,
52
+ bottleneck.depth,
53
+ bottleneck.stride))
54
+ self.body = Sequential(*modules)
55
+
56
+ self.styles = nn.ModuleList()
57
+ self.style_count = opts.n_styles
58
+ self.coarse_ind = 3
59
+ self.middle_ind = 7
60
+ for i in range(self.style_count):
61
+ if i < self.coarse_ind:
62
+ style = GradualStyleBlock(512, 512, 16)
63
+ elif i < self.middle_ind:
64
+ style = GradualStyleBlock(512, 512, 32)
65
+ else:
66
+ style = GradualStyleBlock(512, 512, 64)
67
+ self.styles.append(style)
68
+ self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0)
69
+ self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0)
70
+
71
+ def _upsample_add(self, x, y):
72
+ '''Upsample and add two feature maps.
73
+ Args:
74
+ x: (Variable) top feature map to be upsampled.
75
+ y: (Variable) lateral feature map.
76
+ Returns:
77
+ (Variable) added feature map.
78
+ Note in PyTorch, when input size is odd, the upsampled feature map
79
+ with `F.upsample(..., scale_factor=2, mode='nearest')`
80
+ maybe not equal to the lateral feature map size.
81
+ e.g.
82
+ original input size: [N,_,15,15] ->
83
+ conv2d feature map size: [N,_,8,8] ->
84
+ upsampled feature map size: [N,_,16,16]
85
+ So we choose bilinear upsample which supports arbitrary output sizes.
86
+ '''
87
+ _, _, H, W = y.size()
88
+ return F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True) + y
89
+
90
+ def forward(self, x):
91
+ x = self.input_layer(x)
92
+
93
+ latents = []
94
+ modulelist = list(self.body._modules.values())
95
+ for i, l in enumerate(modulelist):
96
+ x = l(x)
97
+ if i == 6:
98
+ c1 = x
99
+ elif i == 20:
100
+ c2 = x
101
+ elif i == 23:
102
+ c3 = x
103
+
104
+ for j in range(self.coarse_ind):
105
+ latents.append(self.styles[j](c3))
106
+
107
+ p2 = self._upsample_add(c3, self.latlayer1(c2))
108
+ for j in range(self.coarse_ind, self.middle_ind):
109
+ latents.append(self.styles[j](p2))
110
+
111
+ p1 = self._upsample_add(p2, self.latlayer2(c1))
112
+ for j in range(self.middle_ind, self.style_count):
113
+ latents.append(self.styles[j](p1))
114
+
115
+ out = torch.stack(latents, dim=1)
116
+ return out
117
+
118
+
119
+ class BackboneEncoderUsingLastLayerIntoW(Module):
120
+ def __init__(self, num_layers, mode='ir', opts=None):
121
+ super(BackboneEncoderUsingLastLayerIntoW, self).__init__()
122
+ print('Using BackboneEncoderUsingLastLayerIntoW')
123
+ assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152'
124
+ assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
125
+ blocks = get_blocks(num_layers)
126
+ if mode == 'ir':
127
+ unit_module = bottleneck_IR
128
+ elif mode == 'ir_se':
129
+ unit_module = bottleneck_IR_SE
130
+ self.input_layer = Sequential(Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False),
131
+ BatchNorm2d(64),
132
+ PReLU(64))
133
+ self.output_pool = torch.nn.AdaptiveAvgPool2d((1, 1))
134
+ self.linear = EqualLinear(512, 512, lr_mul=1)
135
+ modules = []
136
+ for block in blocks:
137
+ for bottleneck in block:
138
+ modules.append(unit_module(bottleneck.in_channel,
139
+ bottleneck.depth,
140
+ bottleneck.stride))
141
+ self.body = Sequential(*modules)
142
+
143
+ def forward(self, x):
144
+ x = self.input_layer(x)
145
+ x = self.body(x)
146
+ x = self.output_pool(x)
147
+ x = x.view(-1, 512)
148
+ x = self.linear(x)
149
+ return x
150
+
151
+
152
+ class BackboneEncoderUsingLastLayerIntoWPlus(Module):
153
+ def __init__(self, num_layers, mode='ir', opts=None):
154
+ super(BackboneEncoderUsingLastLayerIntoWPlus, self).__init__()
155
+ print('Using BackboneEncoderUsingLastLayerIntoWPlus')
156
+ assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152'
157
+ assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
158
+ blocks = get_blocks(num_layers)
159
+ if mode == 'ir':
160
+ unit_module = bottleneck_IR
161
+ elif mode == 'ir_se':
162
+ unit_module = bottleneck_IR_SE
163
+ self.n_styles = opts.n_styles
164
+ self.input_layer = Sequential(Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False),
165
+ BatchNorm2d(64),
166
+ PReLU(64))
167
+ self.output_layer_2 = Sequential(BatchNorm2d(512),
168
+ torch.nn.AdaptiveAvgPool2d((7, 7)),
169
+ Flatten(),
170
+ Linear(512 * 7 * 7, 512))
171
+ self.linear = EqualLinear(512, 512 * self.n_styles, lr_mul=1)
172
+ modules = []
173
+ for block in blocks:
174
+ for bottleneck in block:
175
+ modules.append(unit_module(bottleneck.in_channel,
176
+ bottleneck.depth,
177
+ bottleneck.stride))
178
+ self.body = Sequential(*modules)
179
+
180
+ def forward(self, x):
181
+ x = self.input_layer(x)
182
+ x = self.body(x)
183
+ x = self.output_layer_2(x)
184
+ x = self.linear(x)
185
+ x = x.view(-1, self.n_styles, 512)
186
+ return x