-
Notifications
You must be signed in to change notification settings - Fork 39
/
losses.py
248 lines (205 loc) · 8.27 KB
/
losses.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
import torch
import torch.nn.functional as F
import torch.nn as nn
from models.vgg import Vgg19
from utils.image_processing import gram
def to_gray_scale(image):
# https://github.com/pytorch/vision/blob/main/torchvision/transforms/v2/functional/_color.py#L33
# Image are assum in range 1, -1
image = (image + 1.0) / 2.0 # To [0, 1]
r, g, b = image.unbind(dim=-3)
l_img = r.mul(0.2989).add_(g, alpha=0.587).add_(b, alpha=0.114)
l_img = l_img.unsqueeze(dim=-3)
l_img = l_img.to(image.dtype)
l_img = l_img.expand(image.shape)
l_img = l_img / 0.5 - 1.0 # To [-1, 1]
return l_img
class ColorLoss(nn.Module):
def __init__(self):
super(ColorLoss, self).__init__()
self.l1 = nn.L1Loss()
self.huber = nn.SmoothL1Loss()
# self._rgb_to_yuv_kernel = torch.tensor([
# [0.299, -0.14714119, 0.61497538],
# [0.587, -0.28886916, -0.51496512],
# [0.114, 0.43601035, -0.10001026]
# ]).float()
self._rgb_to_yuv_kernel = torch.tensor([
[0.299, 0.587, 0.114],
[-0.14714119, -0.28886916, 0.43601035],
[0.61497538, -0.51496512, -0.10001026],
]).float()
def to(self, device):
new_self = super(ColorLoss, self).to(device)
new_self._rgb_to_yuv_kernel = new_self._rgb_to_yuv_kernel.to(device)
return new_self
def rgb_to_yuv(self, image):
'''
https://en.wikipedia.org/wiki/YUV
output: Image of shape (H, W, C) (channel last)
'''
# -1 1 -> 0 1
image = (image + 1.0) / 2.0
image = image.permute(0, 2, 3, 1) # To channel last
yuv_img = image @ self._rgb_to_yuv_kernel.T
return yuv_img
def forward(self, image, image_g):
image = self.rgb_to_yuv(image)
image_g = self.rgb_to_yuv(image_g)
# After convert to yuv, both images have channel last
return (
self.l1(image[:, :, :, 0], image_g[:, :, :, 0])
+ self.huber(image[:, :, :, 1], image_g[:, :, :, 1])
+ self.huber(image[:, :, :, 2], image_g[:, :, :, 2])
)
class AnimeGanLoss:
def __init__(self, args, device, gray_adv=False):
if isinstance(device, str):
device = torch.device(device)
self.content_loss = nn.L1Loss().to(device)
self.gram_loss = nn.L1Loss().to(device)
self.color_loss = ColorLoss().to(device)
self.wadvg = args.wadvg
self.wadvd = args.wadvd
self.wcon = args.wcon
self.wgra = args.wgra
self.wcol = args.wcol
self.wtvar = args.wtvar
# If true, use gray scale image to calculate adversarial loss
self.gray_adv = gray_adv
self.vgg19 = Vgg19().to(device).eval()
self.adv_type = args.gan_loss
self.bce_loss = nn.BCEWithLogitsLoss()
def compute_loss_G(self, fake_img, img, fake_logit, anime_gray):
'''
Compute loss for Generator
@Args:
- fake_img: generated image
- img: real image
- fake_logit: output of Discriminator given fake image
- anime_gray: grayscale of anime image
@Returns:
- Adversarial Loss of fake logits
- Content loss between real and fake features (vgg19)
- Gram loss between anime and fake features (Vgg19)
- Color loss between image and fake image
- Total variation loss of fake image
'''
fake_feat = self.vgg19(fake_img)
gray_feat = self.vgg19(anime_gray)
img_feat = self.vgg19(img)
# fake_gray_feat = self.vgg19(to_gray_scale(fake_img))
return [
# Want to be real image.
self.wadvg * self.adv_loss_g(fake_logit),
self.wcon * self.content_loss(img_feat, fake_feat),
self.wgra * self.gram_loss(gram(gray_feat), gram(fake_feat)),
self.wcol * self.color_loss(img, fake_img),
self.wtvar * self.total_variation_loss(fake_img)
]
def compute_loss_D(
self,
fake_img_d,
real_anime_d,
real_anime_gray_d,
real_anime_smooth_gray_d=None
):
if self.gray_adv:
# Treat gray scale image as real
return (
self.adv_loss_d_real(real_anime_gray_d)
+ self.adv_loss_d_fake(fake_img_d)
+ 0.3 * self.adv_loss_d_fake(real_anime_smooth_gray_d)
)
else:
return (
# Classify real anime as real
self.adv_loss_d_real(real_anime_d)
# Classify generated as fake
+ self.adv_loss_d_fake(fake_img_d)
# Classify real anime gray as fake
# + self.adv_loss_d_fake(real_anime_gray_d)
# Classify real anime as fake
# + 0.1 * self.adv_loss_d_fake(real_anime_smooth_gray_d)
)
def total_variation_loss(self, fake_img):
"""
A smooth loss in fact. Like the smooth prior in MRF.
V(y) = || y_{n+1} - y_n ||_2
"""
# Channel first -> channel last
fake_img = fake_img.permute(0, 2, 3, 1)
def _l2(x):
# sum(t ** 2) / 2
return torch.sum(x ** 2) / 2
dh = fake_img[:, :-1, ...] - fake_img[:, 1:, ...]
dw = fake_img[:, :, :-1, ...] - fake_img[:, :, 1:, ...]
return _l2(dh) / dh.numel() + _l2(dw) / dw.numel()
def content_loss_vgg(self, image, recontruction):
feat = self.vgg19(image)
re_feat = self.vgg19(recontruction)
feature_loss = self.content_loss(feat, re_feat)
content_loss = self.content_loss(image, recontruction)
return feature_loss# + 0.5 * content_loss
def adv_loss_d_real(self, pred):
"""Push pred to class 1 (real)"""
if self.adv_type == 'hinge':
return torch.mean(F.relu(1.0 - pred))
elif self.adv_type == 'lsgan':
# pred = torch.sigmoid(pred)
return torch.mean(torch.square(pred - 1.0))
elif self.adv_type == 'bce':
return self.bce_loss(pred, torch.ones_like(pred))
raise ValueError(f'Do not support loss type {self.adv_type}')
def adv_loss_d_fake(self, pred):
"""Push pred to class 0 (fake)"""
if self.adv_type == 'hinge':
return torch.mean(F.relu(1.0 + pred))
elif self.adv_type == 'lsgan':
# pred = torch.sigmoid(pred)
return torch.mean(torch.square(pred))
elif self.adv_type == 'bce':
return self.bce_loss(pred, torch.zeros_like(pred))
raise ValueError(f'Do not support loss type {self.adv_type}')
def adv_loss_g(self, pred):
"""Push pred to class 1 (real)"""
if self.adv_type == 'hinge':
return -torch.mean(pred)
elif self.adv_type == 'lsgan':
# pred = torch.sigmoid(pred)
return torch.mean(torch.square(pred - 1.0))
elif self.adv_type == 'bce':
return self.bce_loss(pred, torch.ones_like(pred))
raise ValueError(f'Do not support loss type {self.adv_type}')
class LossSummary:
def __init__(self):
self.reset()
def reset(self):
self.loss_g_adv = []
self.loss_content = []
self.loss_gram = []
self.loss_color = []
self.loss_d_adv = []
def update_loss_G(self, adv, gram, color, content):
self.loss_g_adv.append(adv.cpu().detach().numpy())
self.loss_gram.append(gram.cpu().detach().numpy())
self.loss_color.append(color.cpu().detach().numpy())
self.loss_content.append(content.cpu().detach().numpy())
def update_loss_D(self, loss):
self.loss_d_adv.append(loss.cpu().detach().numpy())
def avg_loss_G(self):
return (
self._avg(self.loss_g_adv),
self._avg(self.loss_gram),
self._avg(self.loss_color),
self._avg(self.loss_content),
)
def avg_loss_D(self):
return self._avg(self.loss_d_adv)
def get_loss_description(self):
avg_adv, avg_gram, avg_color, avg_content = self.avg_loss_G()
avg_adv_d = self.avg_loss_D()
return f'loss G: adv {avg_adv:2f} con {avg_content:2f} gram {avg_gram:2f} color {avg_color:2f} / loss D: {avg_adv_d:2f}'
@staticmethod
def _avg(losses):
return sum(losses) / len(losses)