-
Notifications
You must be signed in to change notification settings - Fork 0
/
losses.py
222 lines (172 loc) · 7.48 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
import torch
import torch.nn.functional as F
from torch import device, nn
import torchvision.models as models
from pytorch_lightning import LightningModule
import copy
from pathlib import Path
class MSELoss(nn.Module):
def __init__(self):
super(MSELoss, self).__init__()
self.loss = nn.MSELoss(reduction='mean')
def forward(self, inputs, targets):
loss = self.loss(inputs['rgb_coarse'], targets)
if 'rgb_fine' in inputs:
loss += self.loss(inputs['rgb_fine'], targets)
return loss
class FeatureLoss(LightningModule):
'''Given a content style reference images will find the style and content loss'''
def __init__(self,
style_img,
style_weight,
content_weight
) -> None:
super(FeatureLoss, self).__init__()
self.style_weight = style_weight
self.content_weight = content_weight
content_layers_default = ['conv_4']
style_layers_default = ['conv_1', 'conv_2', 'conv_3', 'conv_4', 'conv_5']
normalization_mean_default = torch.tensor([0.485, 0.456, 0.406], device=self.device)
normalization_std_default = torch.tensor([0.229, 0.224, 0.225], device=self.device)
# Load the VGG
# model_file = 'vgg19.pt'
# if Path(model_file).is_file():
# print(f'Loading model {model_file}')
# cnn = torch.load(model_file)
# else:
print('Downloading vgg19')
cnn = models.vgg19(pretrained=True).to(self.device)
# torch.save(cnn, model_file)
self.cnn = cnn.features.eval()
self.style_model, self.style_losses, self.content_features = self.get_style_model_and_losses(
self.cnn,
style_img,
normalization_mean=normalization_mean_default,
normalization_std=normalization_std_default,
content_layers=content_layers_default,
style_layers=style_layers_default
)
def forward(self, input_img, content_img):
# collect feature loss tensors for the target/content image
self.style_model(content_img)
target_content_features = [cf.content_feature for cf in self.content_features]
# another forward pass for the input image
self.style_model(input_img)
style_score = 0
content_score = 0
# style loss
for sl in self.style_losses:
style_score += sl.loss
# content loss
input_content_features = [cf.content_feature for cf in self.content_features]
for in_cl_feat, target_cl_feat in zip(input_content_features, target_content_features):
content_score += F.mse_loss(in_cl_feat, target_cl_feat)
# weight the loss and combine
style_score *= self.style_weight
content_score *= self.content_weight
loss = style_score + content_score
return loss, style_score, content_score
def get_style_model_and_losses(self,
cnn,
style_img,
normalization_mean,
normalization_std,
content_layers,
style_layers
):
'''Build model to be used for style/content loss'''
cnn = copy.deepcopy(cnn)
# normalization module
normalization = Normalization(
normalization_mean,
normalization_std
)
# to have iterable access to a list of content features and style losses
content_features = []
style_losses = []
# assuming that cnn is a nn.Sequential, we make a new nn.Sequential
# to put in modules that are supposed to be activated sequentially
model = nn.Sequential(normalization)
i = 0 # increment every time we see a conv
for layer in cnn.children():
if isinstance(layer, nn.Conv2d):
i += 1
name = 'conv_{}'.format(i)
elif isinstance(layer, nn.ReLU):
name = 'relu_{}'.format(i)
# The in-place version doesn't play very nicely with the ContentLoss
# and StyleLoss we insert below. So we replace with out-of-place
# ones here.
layer = nn.ReLU(inplace=False)
elif isinstance(layer, nn.MaxPool2d):
name = 'pool_{}'.format(i)
elif isinstance(layer, nn.BatchNorm2d):
name = 'bn_{}'.format(i)
else:
raise RuntimeError('Unrecognized layer: {}'.format(layer.__class__.__name__))
model.add_module(name, layer)
# add content loss to the model:
if name in content_layers:
# Get the feature map of the content image using the half built model
content_feature = ContentFeature()
model.add_module("content_feat_{}".format(i), content_feature)
content_features.append(content_feature)
# add style loss to the model:
if name in style_layers:
# add style loss:
target_feature = model(style_img).detach()
style_loss = StyleLoss(target_feature)
model.add_module("style_loss_{}".format(i), style_loss)
style_losses.append(style_loss)
# now we trim off the layers after the last content and style losses
for i in range(len(model) - 1, -1, -1):
if isinstance(model[i], ContentFeature) or isinstance(model[i], StyleLoss):
break
model = model[:(i + 1)]
return model, style_losses, content_features
class ContentFeature(LightningModule):
'''Extract the feature map'''
def __init__(self):
super(ContentFeature, self).__init__()
def forward(self, input):
# Save the content feature
self.content_feature = input
return input
class StyleLoss(LightningModule):
'''Compute the style loss using Gram matrices'''
def __init__(self, target_feature):
super(StyleLoss, self).__init__()
self.target = self._gram_matrix(target_feature).detach()
def forward(self, input):
if self.target.device.type == 'cpu':
# TODO avoid doing this
self.target = self.target.to(self.device)
G = self._gram_matrix(input)
self.loss = F.mse_loss(G, self.target)
return input
def _gram_matrix(self, input):
# a=batch size(=1)
# b=number of feature maps
# (c,d)=dimensions of a f. map (N=c*d)
a, b, c, d = input.size()
features = input.view(a * b, c * d) # resise F_XL into \hat F_XL
G = torch.mm(features, features.t()) # compute the gram product
# 'normalize' the gram matrix by dividing by size of the feature map
return G.div(a * b * c * d)
class Normalization(LightningModule):
'''VGG Normalisation (pretrained models expect this)'''
def __init__(self, mean, std):
super(Normalization, self).__init__()
# .view the mean and std to make them [C x 1 x 1] so that they can
# directly work with image Tensor of shape [B x C x H x W].
# B is batch size. C is number of channels. H is height and W is width.
self.mean = mean.view(-1, 1, 1).to(self.device)
self.std = std.view(-1, 1, 1).to(self.device)
def forward(self, img):
# normalize img
if self.mean.device.type == 'cpu':
# TODO avoid doing this
self.mean = self.mean.to(self.device)
self.std = self.std.to(self.device)
return (img - self.mean) / self.std
loss_dict = {'mse': MSELoss}