Spaces:
Runtime error
Runtime error
losses
Browse files- Time_TravelRephotography/losses/contextual_loss_pytorch/.gitignore +104 -0
- Time_TravelRephotography/losses/contextual_loss_pytorch/LICENSE +21 -0
- Time_TravelRephotography/losses/contextual_loss_pytorch/__init__.py +1 -0
- Time_TravelRephotography/losses/contextual_loss_pytorch/config.py +2 -0
- Time_TravelRephotography/losses/contextual_loss_pytorch/functional.py +198 -0
- Time_TravelRephotography/losses/contextual_loss_pytorch/modules/__init__.py +5 -0
- Time_TravelRephotography/losses/contextual_loss_pytorch/modules/contextual.py +122 -0
- Time_TravelRephotography/losses/contextual_loss_pytorch/modules/contextual_bilateral.py +69 -0
- Time_TravelRephotography/losses/contextual_loss_pytorch/modules/vgg.py +48 -0
Time_TravelRephotography/losses/contextual_loss_pytorch/.gitignore
ADDED
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Byte-compiled / optimized / DLL files
|
2 |
+
__pycache__/
|
3 |
+
*.py[cod]
|
4 |
+
*$py.class
|
5 |
+
|
6 |
+
# C extensions
|
7 |
+
*.so
|
8 |
+
|
9 |
+
# Distribution / packaging
|
10 |
+
.Python
|
11 |
+
build/
|
12 |
+
develop-eggs/
|
13 |
+
dist/
|
14 |
+
downloads/
|
15 |
+
eggs/
|
16 |
+
.eggs/
|
17 |
+
lib/
|
18 |
+
lib64/
|
19 |
+
parts/
|
20 |
+
sdist/
|
21 |
+
var/
|
22 |
+
wheels/
|
23 |
+
*.egg-info/
|
24 |
+
.installed.cfg
|
25 |
+
*.egg
|
26 |
+
MANIFEST
|
27 |
+
|
28 |
+
# PyInstaller
|
29 |
+
# Usually these files are written by a python script from a template
|
30 |
+
# before PyInstaller builds the exe, so as to inject date/other infos into it.
|
31 |
+
*.manifest
|
32 |
+
*.spec
|
33 |
+
|
34 |
+
# Installer logs
|
35 |
+
pip-log.txt
|
36 |
+
pip-delete-this-directory.txt
|
37 |
+
|
38 |
+
# Unit test / coverage reports
|
39 |
+
htmlcov/
|
40 |
+
.tox/
|
41 |
+
.coverage
|
42 |
+
.coverage.*
|
43 |
+
.cache
|
44 |
+
nosetests.xml
|
45 |
+
coverage.xml
|
46 |
+
*.cover
|
47 |
+
.hypothesis/
|
48 |
+
.pytest_cache/
|
49 |
+
|
50 |
+
# Translations
|
51 |
+
*.mo
|
52 |
+
*.pot
|
53 |
+
|
54 |
+
# Django stuff:
|
55 |
+
*.log
|
56 |
+
local_settings.py
|
57 |
+
db.sqlite3
|
58 |
+
|
59 |
+
# Flask stuff:
|
60 |
+
instance/
|
61 |
+
.webassets-cache
|
62 |
+
|
63 |
+
# Scrapy stuff:
|
64 |
+
.scrapy
|
65 |
+
|
66 |
+
# Sphinx documentation
|
67 |
+
docs/_build/
|
68 |
+
|
69 |
+
# PyBuilder
|
70 |
+
target/
|
71 |
+
|
72 |
+
# Jupyter Notebook
|
73 |
+
.ipynb_checkpoints
|
74 |
+
|
75 |
+
# pyenv
|
76 |
+
.python-version
|
77 |
+
|
78 |
+
# celery beat schedule file
|
79 |
+
celerybeat-schedule
|
80 |
+
|
81 |
+
# SageMath parsed files
|
82 |
+
*.sage.py
|
83 |
+
|
84 |
+
# Environments
|
85 |
+
.env
|
86 |
+
.venv
|
87 |
+
env/
|
88 |
+
venv/
|
89 |
+
ENV/
|
90 |
+
env.bak/
|
91 |
+
venv.bak/
|
92 |
+
|
93 |
+
# Spyder project settings
|
94 |
+
.spyderproject
|
95 |
+
.spyproject
|
96 |
+
|
97 |
+
# Rope project settings
|
98 |
+
.ropeproject
|
99 |
+
|
100 |
+
# mkdocs documentation
|
101 |
+
/site
|
102 |
+
|
103 |
+
# mypy
|
104 |
+
.mypy_cache/
|
Time_TravelRephotography/losses/contextual_loss_pytorch/LICENSE
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
MIT License
|
2 |
+
|
3 |
+
Copyright (c) 2019 Sou Uchida
|
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.
|
Time_TravelRephotography/losses/contextual_loss_pytorch/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
from .modules import *
|
Time_TravelRephotography/losses/contextual_loss_pytorch/config.py
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
# TODO: add supports for L1, L2 etc.
|
2 |
+
LOSS_TYPES = ['cosine', 'l1', 'l2']
|
Time_TravelRephotography/losses/contextual_loss_pytorch/functional.py
ADDED
@@ -0,0 +1,198 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as F
|
3 |
+
|
4 |
+
from .config import LOSS_TYPES
|
5 |
+
|
6 |
+
__all__ = ['contextual_loss', 'contextual_bilateral_loss']
|
7 |
+
|
8 |
+
|
9 |
+
def contextual_loss(x: torch.Tensor,
|
10 |
+
y: torch.Tensor,
|
11 |
+
band_width: float = 0.5,
|
12 |
+
loss_type: str = 'cosine',
|
13 |
+
all_dist: bool = False):
|
14 |
+
"""
|
15 |
+
Computes contextual loss between x and y.
|
16 |
+
The most of this code is copied from
|
17 |
+
https://gist.github.com/yunjey/3105146c736f9c1055463c33b4c989da.
|
18 |
+
|
19 |
+
Parameters
|
20 |
+
---
|
21 |
+
x : torch.Tensor
|
22 |
+
features of shape (N, C, H, W).
|
23 |
+
y : torch.Tensor
|
24 |
+
features of shape (N, C, H, W).
|
25 |
+
band_width : float, optional
|
26 |
+
a band-width parameter used to convert distance to similarity.
|
27 |
+
in the paper, this is described as :math:`h`.
|
28 |
+
loss_type : str, optional
|
29 |
+
a loss type to measure the distance between features.
|
30 |
+
Note: `l1` and `l2` frequently raises OOM.
|
31 |
+
|
32 |
+
Returns
|
33 |
+
---
|
34 |
+
cx_loss : torch.Tensor
|
35 |
+
contextual loss between x and y (Eq (1) in the paper)
|
36 |
+
"""
|
37 |
+
|
38 |
+
assert x.size() == y.size(), 'input tensor must have the same size.'
|
39 |
+
assert loss_type in LOSS_TYPES, f'select a loss type from {LOSS_TYPES}.'
|
40 |
+
|
41 |
+
N, C, H, W = x.size()
|
42 |
+
|
43 |
+
if loss_type == 'cosine':
|
44 |
+
dist_raw = compute_cosine_distance(x, y)
|
45 |
+
elif loss_type == 'l1':
|
46 |
+
dist_raw = compute_l1_distance(x, y)
|
47 |
+
elif loss_type == 'l2':
|
48 |
+
dist_raw = compute_l2_distance(x, y)
|
49 |
+
|
50 |
+
dist_tilde = compute_relative_distance(dist_raw)
|
51 |
+
cx = compute_cx(dist_tilde, band_width)
|
52 |
+
if all_dist:
|
53 |
+
return cx
|
54 |
+
|
55 |
+
cx = torch.mean(torch.max(cx, dim=1)[0], dim=1) # Eq(1)
|
56 |
+
cx_loss = torch.mean(-torch.log(cx + 1e-5)) # Eq(5)
|
57 |
+
|
58 |
+
return cx_loss
|
59 |
+
|
60 |
+
|
61 |
+
# TODO: Operation check
|
62 |
+
def contextual_bilateral_loss(x: torch.Tensor,
|
63 |
+
y: torch.Tensor,
|
64 |
+
weight_sp: float = 0.1,
|
65 |
+
band_width: float = 1.,
|
66 |
+
loss_type: str = 'cosine'):
|
67 |
+
"""
|
68 |
+
Computes Contextual Bilateral (CoBi) Loss between x and y,
|
69 |
+
proposed in https://arxiv.org/pdf/1905.05169.pdf.
|
70 |
+
|
71 |
+
Parameters
|
72 |
+
---
|
73 |
+
x : torch.Tensor
|
74 |
+
features of shape (N, C, H, W).
|
75 |
+
y : torch.Tensor
|
76 |
+
features of shape (N, C, H, W).
|
77 |
+
band_width : float, optional
|
78 |
+
a band-width parameter used to convert distance to similarity.
|
79 |
+
in the paper, this is described as :math:`h`.
|
80 |
+
loss_type : str, optional
|
81 |
+
a loss type to measure the distance between features.
|
82 |
+
Note: `l1` and `l2` frequently raises OOM.
|
83 |
+
|
84 |
+
Returns
|
85 |
+
---
|
86 |
+
cx_loss : torch.Tensor
|
87 |
+
contextual loss between x and y (Eq (1) in the paper).
|
88 |
+
k_arg_max_NC : torch.Tensor
|
89 |
+
indices to maximize similarity over channels.
|
90 |
+
"""
|
91 |
+
|
92 |
+
assert x.size() == y.size(), 'input tensor must have the same size.'
|
93 |
+
assert loss_type in LOSS_TYPES, f'select a loss type from {LOSS_TYPES}.'
|
94 |
+
|
95 |
+
# spatial loss
|
96 |
+
grid = compute_meshgrid(x.shape).to(x.device)
|
97 |
+
dist_raw = compute_l2_distance(grid, grid)
|
98 |
+
dist_tilde = compute_relative_distance(dist_raw)
|
99 |
+
cx_sp = compute_cx(dist_tilde, band_width)
|
100 |
+
|
101 |
+
# feature loss
|
102 |
+
if loss_type == 'cosine':
|
103 |
+
dist_raw = compute_cosine_distance(x, y)
|
104 |
+
elif loss_type == 'l1':
|
105 |
+
dist_raw = compute_l1_distance(x, y)
|
106 |
+
elif loss_type == 'l2':
|
107 |
+
dist_raw = compute_l2_distance(x, y)
|
108 |
+
dist_tilde = compute_relative_distance(dist_raw)
|
109 |
+
cx_feat = compute_cx(dist_tilde, band_width)
|
110 |
+
|
111 |
+
# combined loss
|
112 |
+
cx_combine = (1. - weight_sp) * cx_feat + weight_sp * cx_sp
|
113 |
+
|
114 |
+
k_max_NC, _ = torch.max(cx_combine, dim=2, keepdim=True)
|
115 |
+
|
116 |
+
cx = k_max_NC.mean(dim=1)
|
117 |
+
cx_loss = torch.mean(-torch.log(cx + 1e-5))
|
118 |
+
|
119 |
+
return cx_loss
|
120 |
+
|
121 |
+
|
122 |
+
def compute_cx(dist_tilde, band_width):
|
123 |
+
w = torch.exp((1 - dist_tilde) / band_width) # Eq(3)
|
124 |
+
cx = w / torch.sum(w, dim=2, keepdim=True) # Eq(4)
|
125 |
+
return cx
|
126 |
+
|
127 |
+
|
128 |
+
def compute_relative_distance(dist_raw):
|
129 |
+
dist_min, _ = torch.min(dist_raw, dim=2, keepdim=True)
|
130 |
+
dist_tilde = dist_raw / (dist_min + 1e-5)
|
131 |
+
return dist_tilde
|
132 |
+
|
133 |
+
|
134 |
+
def compute_cosine_distance(x, y):
|
135 |
+
# mean shifting by channel-wise mean of `y`.
|
136 |
+
y_mu = y.mean(dim=(0, 2, 3), keepdim=True)
|
137 |
+
x_centered = x - y_mu
|
138 |
+
y_centered = y - y_mu
|
139 |
+
|
140 |
+
# L2 normalization
|
141 |
+
x_normalized = F.normalize(x_centered, p=2, dim=1)
|
142 |
+
y_normalized = F.normalize(y_centered, p=2, dim=1)
|
143 |
+
|
144 |
+
# channel-wise vectorization
|
145 |
+
N, C, *_ = x.size()
|
146 |
+
x_normalized = x_normalized.reshape(N, C, -1) # (N, C, H*W)
|
147 |
+
y_normalized = y_normalized.reshape(N, C, -1) # (N, C, H*W)
|
148 |
+
|
149 |
+
# consine similarity
|
150 |
+
cosine_sim = torch.bmm(x_normalized.transpose(1, 2),
|
151 |
+
y_normalized) # (N, H*W, H*W)
|
152 |
+
|
153 |
+
# convert to distance
|
154 |
+
dist = 1 - cosine_sim
|
155 |
+
|
156 |
+
return dist
|
157 |
+
|
158 |
+
|
159 |
+
# TODO: Considering avoiding OOM.
|
160 |
+
def compute_l1_distance(x: torch.Tensor, y: torch.Tensor):
|
161 |
+
N, C, H, W = x.size()
|
162 |
+
x_vec = x.view(N, C, -1)
|
163 |
+
y_vec = y.view(N, C, -1)
|
164 |
+
|
165 |
+
dist = x_vec.unsqueeze(2) - y_vec.unsqueeze(3)
|
166 |
+
dist = dist.abs().sum(dim=1)
|
167 |
+
dist = dist.transpose(1, 2).reshape(N, H*W, H*W)
|
168 |
+
dist = dist.clamp(min=0.)
|
169 |
+
|
170 |
+
return dist
|
171 |
+
|
172 |
+
|
173 |
+
# TODO: Considering avoiding OOM.
|
174 |
+
def compute_l2_distance(x, y):
|
175 |
+
N, C, H, W = x.size()
|
176 |
+
x_vec = x.view(N, C, -1)
|
177 |
+
y_vec = y.view(N, C, -1)
|
178 |
+
x_s = torch.sum(x_vec ** 2, dim=1)
|
179 |
+
y_s = torch.sum(y_vec ** 2, dim=1)
|
180 |
+
|
181 |
+
A = y_vec.transpose(1, 2) @ x_vec
|
182 |
+
dist = y_s - 2 * A + x_s.transpose(0, 1)
|
183 |
+
dist = dist.transpose(1, 2).reshape(N, H*W, H*W)
|
184 |
+
dist = dist.clamp(min=0.)
|
185 |
+
|
186 |
+
return dist
|
187 |
+
|
188 |
+
|
189 |
+
def compute_meshgrid(shape):
|
190 |
+
N, C, H, W = shape
|
191 |
+
rows = torch.arange(0, H, dtype=torch.float32) / (H + 1)
|
192 |
+
cols = torch.arange(0, W, dtype=torch.float32) / (W + 1)
|
193 |
+
|
194 |
+
feature_grid = torch.meshgrid(rows, cols)
|
195 |
+
feature_grid = torch.stack(feature_grid).unsqueeze(0)
|
196 |
+
feature_grid = torch.cat([feature_grid for _ in range(N)], dim=0)
|
197 |
+
|
198 |
+
return feature_grid
|
Time_TravelRephotography/losses/contextual_loss_pytorch/modules/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .contextual import ContextualLoss
|
2 |
+
from .contextual_bilateral import ContextualBilateralLoss
|
3 |
+
from .vgg import VGG19
|
4 |
+
|
5 |
+
__all__ = ['ContextualLoss', 'ContextualBilateralLoss', 'VGG19']
|
Time_TravelRephotography/losses/contextual_loss_pytorch/modules/contextual.py
ADDED
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import random
|
2 |
+
from typing import (
|
3 |
+
Iterable,
|
4 |
+
List,
|
5 |
+
Optional,
|
6 |
+
)
|
7 |
+
|
8 |
+
import numpy as np
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
|
12 |
+
from .vgg import VGG19
|
13 |
+
from .. import functional as F
|
14 |
+
from ..config import LOSS_TYPES
|
15 |
+
|
16 |
+
|
17 |
+
class ContextualLoss(nn.Module):
|
18 |
+
"""
|
19 |
+
Creates a criterion that measures the contextual loss.
|
20 |
+
|
21 |
+
Parameters
|
22 |
+
---
|
23 |
+
band_width : int, optional
|
24 |
+
a band_width parameter described as :math:`h` in the paper.
|
25 |
+
use_vgg : bool, optional
|
26 |
+
if you want to use VGG feature, set this `True`.
|
27 |
+
vgg_layer : str, optional
|
28 |
+
intermidiate layer name for VGG feature.
|
29 |
+
Now we support layer names:
|
30 |
+
`['relu1_2', 'relu2_2', 'relu3_4', 'relu4_4', 'relu5_4']`
|
31 |
+
"""
|
32 |
+
|
33 |
+
def __init__(
|
34 |
+
self,
|
35 |
+
band_width: float = 0.5,
|
36 |
+
loss_type: str = 'cosine',
|
37 |
+
use_vgg: bool = False,
|
38 |
+
vgg_model: nn.Module = None,
|
39 |
+
vgg_layers: List[str] = ['relu3_4'],
|
40 |
+
feature_1d_size: int = 64,
|
41 |
+
):
|
42 |
+
|
43 |
+
super().__init__()
|
44 |
+
|
45 |
+
assert band_width > 0, 'band_width parameter must be positive.'
|
46 |
+
assert loss_type in LOSS_TYPES,\
|
47 |
+
f'select a loss type from {LOSS_TYPES}.'
|
48 |
+
|
49 |
+
self.loss_type = loss_type
|
50 |
+
self.band_width = band_width
|
51 |
+
self.feature_1d_size = feature_1d_size
|
52 |
+
|
53 |
+
if use_vgg:
|
54 |
+
self.vgg_model = VGG19() if vgg_model is None else vgg_model
|
55 |
+
self.vgg_layers = vgg_layers
|
56 |
+
self.register_buffer(
|
57 |
+
name='vgg_mean',
|
58 |
+
tensor=torch.tensor(
|
59 |
+
[[[0.485]], [[0.456]], [[0.406]]], requires_grad=False)
|
60 |
+
)
|
61 |
+
self.register_buffer(
|
62 |
+
name='vgg_std',
|
63 |
+
tensor=torch.tensor(
|
64 |
+
[[[0.229]], [[0.224]], [[0.225]]], requires_grad=False)
|
65 |
+
)
|
66 |
+
|
67 |
+
def forward(self, x: torch.Tensor, y: torch.Tensor, all_dist: bool = False):
|
68 |
+
if not hasattr(self, 'vgg_model'):
|
69 |
+
return self.contextual_loss(x, y, self.feature_1d_size, self.band_width, all_dist=all_dist)
|
70 |
+
|
71 |
+
|
72 |
+
x = self.forward_vgg(x)
|
73 |
+
y = self.forward_vgg(y)
|
74 |
+
|
75 |
+
loss = 0
|
76 |
+
for layer in self.vgg_layers:
|
77 |
+
# picking up vgg feature maps
|
78 |
+
fx = getattr(x, layer)
|
79 |
+
fy = getattr(y, layer)
|
80 |
+
loss = loss + self.contextual_loss(
|
81 |
+
fx, fy, self.feature_1d_size, self.band_width, all_dist=all_dist, loss_type=self.loss_type
|
82 |
+
)
|
83 |
+
return loss
|
84 |
+
|
85 |
+
def forward_vgg(self, x: torch.Tensor):
|
86 |
+
assert x.shape[1] == 3, 'VGG model takes 3 chennel images.'
|
87 |
+
# [-1, 1] -> [0, 1]
|
88 |
+
x = (x + 1) * 0.5
|
89 |
+
|
90 |
+
# normalization
|
91 |
+
x = x.sub(self.vgg_mean.detach()).div(self.vgg_std)
|
92 |
+
return self.vgg_model(x)
|
93 |
+
|
94 |
+
@classmethod
|
95 |
+
def contextual_loss(
|
96 |
+
cls,
|
97 |
+
x: torch.Tensor, y: torch.Tensor,
|
98 |
+
feature_1d_size: int,
|
99 |
+
band_width: int,
|
100 |
+
all_dist: bool = False,
|
101 |
+
loss_type: str = 'cosine',
|
102 |
+
) -> torch.Tensor:
|
103 |
+
feature_size = feature_1d_size ** 2
|
104 |
+
if np.prod(x.shape[2:]) > feature_size or np.prod(y.shape[2:]) > feature_size:
|
105 |
+
x, indices = cls.random_sampling(x, feature_1d_size=feature_1d_size)
|
106 |
+
y, _ = cls.random_sampling(y, feature_1d_size=feature_1d_size, indices=indices)
|
107 |
+
|
108 |
+
return F.contextual_loss(x, y, band_width, all_dist=all_dist, loss_type=loss_type)
|
109 |
+
|
110 |
+
@staticmethod
|
111 |
+
def random_sampling(
|
112 |
+
tensor_NCHW: torch.Tensor, feature_1d_size: int, indices: Optional[List] = None
|
113 |
+
):
|
114 |
+
N, C, H, W = tensor_NCHW.shape
|
115 |
+
S = H * W
|
116 |
+
tensor_NCS = tensor_NCHW.reshape([N, C, S])
|
117 |
+
if indices is None:
|
118 |
+
all_indices = list(range(S))
|
119 |
+
random.shuffle(all_indices)
|
120 |
+
indices = all_indices[:feature_1d_size**2]
|
121 |
+
res = tensor_NCS[:, :, indices].reshape(N, -1, feature_1d_size, feature_1d_size)
|
122 |
+
return res, indices
|
Time_TravelRephotography/losses/contextual_loss_pytorch/modules/contextual_bilateral.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
|
4 |
+
from .vgg import VGG19
|
5 |
+
from .. import functional as F
|
6 |
+
from ..config import LOSS_TYPES
|
7 |
+
|
8 |
+
|
9 |
+
class ContextualBilateralLoss(nn.Module):
|
10 |
+
"""
|
11 |
+
Creates a criterion that measures the contextual bilateral loss.
|
12 |
+
|
13 |
+
Parameters
|
14 |
+
---
|
15 |
+
weight_sp : float, optional
|
16 |
+
a balancing weight between spatial and feature loss.
|
17 |
+
band_width : int, optional
|
18 |
+
a band_width parameter described as :math:`h` in the paper.
|
19 |
+
use_vgg : bool, optional
|
20 |
+
if you want to use VGG feature, set this `True`.
|
21 |
+
vgg_layer : str, optional
|
22 |
+
intermidiate layer name for VGG feature.
|
23 |
+
Now we support layer names:
|
24 |
+
`['relu1_2', 'relu2_2', 'relu3_4', 'relu4_4', 'relu5_4']`
|
25 |
+
"""
|
26 |
+
|
27 |
+
def __init__(self,
|
28 |
+
weight_sp: float = 0.1,
|
29 |
+
band_width: float = 0.5,
|
30 |
+
loss_type: str = 'cosine',
|
31 |
+
use_vgg: bool = False,
|
32 |
+
vgg_layer: str = 'relu3_4'):
|
33 |
+
|
34 |
+
super(ContextualBilateralLoss, self).__init__()
|
35 |
+
|
36 |
+
assert band_width > 0, 'band_width parameter must be positive.'
|
37 |
+
assert loss_type in LOSS_TYPES,\
|
38 |
+
f'select a loss type from {LOSS_TYPES}.'
|
39 |
+
|
40 |
+
self.band_width = band_width
|
41 |
+
|
42 |
+
if use_vgg:
|
43 |
+
self.vgg_model = VGG19()
|
44 |
+
self.vgg_layer = vgg_layer
|
45 |
+
self.register_buffer(
|
46 |
+
name='vgg_mean',
|
47 |
+
tensor=torch.tensor(
|
48 |
+
[[[0.485]], [[0.456]], [[0.406]]], requires_grad=False)
|
49 |
+
)
|
50 |
+
self.register_buffer(
|
51 |
+
name='vgg_std',
|
52 |
+
tensor=torch.tensor(
|
53 |
+
[[[0.229]], [[0.224]], [[0.225]]], requires_grad=False)
|
54 |
+
)
|
55 |
+
|
56 |
+
def forward(self, x, y):
|
57 |
+
if hasattr(self, 'vgg_model'):
|
58 |
+
assert x.shape[1] == 3 and y.shape[1] == 3,\
|
59 |
+
'VGG model takes 3 chennel images.'
|
60 |
+
|
61 |
+
# normalization
|
62 |
+
x = x.sub(self.vgg_mean.detach()).div(self.vgg_std.detach())
|
63 |
+
y = y.sub(self.vgg_mean.detach()).div(self.vgg_std.detach())
|
64 |
+
|
65 |
+
# picking up vgg feature maps
|
66 |
+
x = getattr(self.vgg_model(x), self.vgg_layer)
|
67 |
+
y = getattr(self.vgg_model(y), self.vgg_layer)
|
68 |
+
|
69 |
+
return F.contextual_bilateral_loss(x, y, self.band_width)
|
Time_TravelRephotography/losses/contextual_loss_pytorch/modules/vgg.py
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from collections import namedtuple
|
2 |
+
|
3 |
+
import torch.nn as nn
|
4 |
+
import torchvision.models.vgg as vgg
|
5 |
+
|
6 |
+
|
7 |
+
class VGG19(nn.Module):
|
8 |
+
def __init__(self, requires_grad=False):
|
9 |
+
super(VGG19, self).__init__()
|
10 |
+
vgg_pretrained_features = vgg.vgg19(pretrained=True).features
|
11 |
+
self.slice1 = nn.Sequential()
|
12 |
+
self.slice2 = nn.Sequential()
|
13 |
+
self.slice3 = nn.Sequential()
|
14 |
+
self.slice4 = nn.Sequential()
|
15 |
+
self.slice5 = nn.Sequential()
|
16 |
+
for x in range(4):
|
17 |
+
self.slice1.add_module(str(x), vgg_pretrained_features[x])
|
18 |
+
for x in range(4, 9):
|
19 |
+
self.slice2.add_module(str(x), vgg_pretrained_features[x])
|
20 |
+
for x in range(9, 18):
|
21 |
+
self.slice3.add_module(str(x), vgg_pretrained_features[x])
|
22 |
+
for x in range(18, 27):
|
23 |
+
self.slice4.add_module(str(x), vgg_pretrained_features[x])
|
24 |
+
for x in range(27, 36):
|
25 |
+
self.slice5.add_module(str(x), vgg_pretrained_features[x])
|
26 |
+
if not requires_grad:
|
27 |
+
for param in self.parameters():
|
28 |
+
param.requires_grad = False
|
29 |
+
|
30 |
+
def forward(self, X):
|
31 |
+
h = self.slice1(X)
|
32 |
+
h_relu1_2 = h
|
33 |
+
h = self.slice2(h)
|
34 |
+
h_relu2_2 = h
|
35 |
+
h = self.slice3(h)
|
36 |
+
h_relu3_4 = h
|
37 |
+
h = self.slice4(h)
|
38 |
+
h_relu4_4 = h
|
39 |
+
h = self.slice5(h)
|
40 |
+
h_relu5_4 = h
|
41 |
+
|
42 |
+
vgg_outputs = namedtuple(
|
43 |
+
"VggOutputs", ['relu1_2', 'relu2_2',
|
44 |
+
'relu3_4', 'relu4_4', 'relu5_4'])
|
45 |
+
out = vgg_outputs(h_relu1_2, h_relu2_2,
|
46 |
+
h_relu3_4, h_relu4_4, h_relu5_4)
|
47 |
+
|
48 |
+
return out
|