-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodels.py
269 lines (223 loc) · 9.97 KB
/
models.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
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
import torch
from torch import nn
from einops import rearrange
from torch.nn.utils import spectral_norm
import torch.nn.functional as F
from torchvision.models import vgg16
import torchvision.utils as vutils
import os
#####################################################################
# functions #
#####################################################################
def Attention(content_features, style_features):
b, c, h, w = content_features.shape
softmax = nn.Softmax(dim=-1)
q = content_features.view(b, c, -1) # (b, c, h * w)
k = style_features.view(b, c, -1)
v = content_features.view(b, c, -1)
att = torch.bmm(q, k.transpose(1, 2)) # (b, c, c)
att = softmax(att)
out = torch.bmm(att, v) # (b, c, h * w)
out = rearrange(out, 'b c (h w) -> b c h w', h=h, w=w) # (b, c, h, w)
return out
def AdaIN(content_features, style_features, eps=1e-6):
content_mean, content_std = torch.mean(content_features, dim=[2, 3], keepdim=True), \
torch.std(content_features, dim=[2, 3], keepdim=True)
style_mean = torch.mean(style_features, dim=[2, 3], keepdim=True)
style_std = torch.std(style_features, dim=[2, 3], keepdim=True)
normalized_content_features = (content_features - content_mean) / (content_std + eps)
normalized_features = normalized_content_features * style_std + style_mean
return normalized_features
def AdaIN_N(content_features, style_mean, style_std, eps=1e-6):
content_mean, content_std = torch.mean(content_features, dim=[2, 3], keepdim=True), \
torch.std(content_features, dim=[2, 3], keepdim=True)
style_mean = style_mean.repeat(content_features.shape[0], 1, 1, 1)
style_std = style_std.repeat(content_features.shape[0], 1, 1, 1)
normalized_content_features = (content_features - content_mean) / (content_std + eps)
normalized_features = normalized_content_features * style_std + style_mean
return normalized_features
class Adaptive_pool(nn.Module):
'''
take a input tensor of size: B x C' X C'
output a maxpooled tensor of size: B x C x H x W
'''
def __init__(self, channel_out, hw_out):
super().__init__()
self.channel_out = channel_out
self.hw_out = hw_out
self.pool = nn.AdaptiveAvgPool2d((channel_out, hw_out ** 2))
def forward(self, input):
if len(input.shape) == 3:
input.unsqueeze_(1)
return self.pool(input).view(-1, self.channel_out, self.hw_out, self.hw_out)
#####################################################################
# models #
#####################################################################
class VGGSimple(nn.Module):
def __init__(self):
super(VGGSimple, self).__init__()
self.backbone = vgg16(pretrained=True).features
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.norm_mean = torch.Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1)
self.norm_std = torch.Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1)
def forward(self, x):
# re-normalize from [-1, 1] to [0, 1] then to the range used for vgg
feat = (((x + 1) * 0.5) - self.norm_mean.to(x.device)) / self.norm_std.to(x.device)
# the layer numbers used to extract features
cut_points = [3, 8, 15, 22]
for idx, module in enumerate(self.backbone):
feat = module(feat)
if idx == cut_points[0]:
layer_1 = feat # (B, 64, 128, 128)
if idx == cut_points[1]:
layer_2 = feat # (B, 128, 64, 64)
if idx == cut_points[2]:
layer_3 = feat # (B, 256, 32, 32)
if idx == cut_points[3]:
layer_4 = feat # (B, 512, 16, 16)
feat = torch.flatten(self.avgpool(feat), 1)
return layer_1, layer_2, layer_3, layer_4, feat
# class ConvBlock(nn.Module):
# def __init__(self, in_channel, out_channel, is_up=True):
# super().__init__()
# self.is_up = is_up
# self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
# self.block1 = nn.Sequential(
# nn.ReflectionPad2d(1),
# spectral_norm(nn.Conv2d(in_channel, in_channel // 2, 3, 1, 0, bias=True)),
# nn.BatchNorm2d(in_channel // 2),
# nn.LeakyReLU(0.01, inplace=True)
# )
# self.block2 = nn.Sequential(
# nn.ReflectionPad2d(1),
# spectral_norm(nn.Conv2d(in_channel // 2, out_channel, 3, 1, 0, bias=True)),
# nn.BatchNorm2d(out_channel),
# nn.LeakyReLU(0.01, inplace=True)
# )
#
# def forward(self, x):
# if self.is_up:
# return self.block2(self.block1(self.up(x)))
# else:
# return self.block2(self.block1(x))
#
#
# class Generator(nn.Module):
# def __init__(self, nfc=256, ch_out=3):
# super(Generator, self).__init__()
# self.decode_28 = ConvBlock(nfc * 2, nfc, is_up=True) # 28 channel 512 -> 256
# self.decode_56 = ConvBlock(nfc * 2, nfc // 2, is_up=True) # 56 channel 256 -> 128
# self.decode_112 = ConvBlock(nfc, nfc // 4, is_up=True) # 112 channel 128 -> 64
# self.decode_224 = ConvBlock(nfc // 2, nfc // 4, is_up=False) # 224 channel 64 -> 32
#
# self.final = nn.Sequential(
# spectral_norm(nn.Conv2d(nfc // 4, ch_out, 3, 1, 1, bias=True)),
# nn.Tanh()
# )
#
# def forward(self, f28, f56, f112, f224):
# decode_28 = self.decode_28(f28)
# decode_56 = self.decode_56(torch.cat([decode_28, f56], dim=1))
# decode_112 = self.decode_112(torch.cat([decode_56, f112], dim=1))
# decode_224 = self.decode_224(torch.cat([decode_112, f224], dim=1))
#
# output = self.final(decode_224)
# return output
class ConvBlock(nn.Module):
def __init__(self, in_channel, out_channel):
super().__init__()
self.pool = nn.MaxPool2d(2, 2)
self.block1 = nn.Sequential(
nn.ReflectionPad2d(1),
spectral_norm(nn.Conv2d(in_channel, in_channel, 3, 1, 0, bias=True)),
nn.BatchNorm2d(in_channel),
nn.LeakyReLU(0.01, inplace=True)
)
self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
self.block2 = nn.Sequential(
nn.ReflectionPad2d(1),
spectral_norm(nn.Conv2d(in_channel * 2, in_channel, 3, 1, 0, bias=True)),
nn.BatchNorm2d(in_channel),
nn.LeakyReLU(0.01, inplace=True)
)
self.block3 = nn.Sequential(
spectral_norm(nn.Conv2d(in_channel, out_channel, 1, 1, 0, bias=True)),
nn.BatchNorm2d(out_channel),
nn.LeakyReLU(0.01, inplace=True)
)
def forward(self, x1, x2, weight):
f1 = self.pool(x1)
f1 = self.block1(f1)
out = self.block2(self.up(torch.cat([f1, x2], dim=1)))
out = self.block3(out + weight * x1)
return out
class Generator(nn.Module):
def __init__(self, nfc=256, ch_out=3):
super(Generator, self).__init__()
self.decode_28 = nn.Sequential(
nn.ReflectionPad2d(1),
spectral_norm(nn.Conv2d(nfc * 2, nfc, 3, 1, 0, bias=True)),
nn.BatchNorm2d(nfc),
nn.LeakyReLU(0.01, inplace=True)
) # 32 channel 512 -> 256
self.decode_56 = ConvBlock(nfc, nfc // 2) # 64 channel 256 -> 128
self.decode_112 = ConvBlock(nfc // 2, nfc // 4) # 128 channel 128 -> 64
self.decode_224 = ConvBlock(nfc // 4, nfc // 8) # 256 channel 64 -> 32
self.final = nn.Sequential(
spectral_norm(nn.Conv2d(nfc // 8, ch_out, 3, 1, 1, bias=True)),
nn.Tanh()
)
self.weight_28 = nn.Parameter(torch.tensor(0.))
self.weight_56 = nn.Parameter(torch.tensor(0.))
self.weight_112 = nn.Parameter(torch.tensor(0.))
def forward(self, f28, f56, f112, f224):
decode_28 = self.decode_28(f28)
decode_56 = self.decode_56(f56, decode_28, self.weight_28)
decode_112 = self.decode_112(f112, decode_56, self.weight_56)
decode_224 = self.decode_224(f224, decode_112, self.weight_112)
output = self.final(decode_224)
return output
class DownConvBlock(nn.Module):
def __init__(self, in_channel, out_channel, down=True):
super().__init__()
if down:
self.block = nn.Sequential(
spectral_norm(nn.Conv2d(in_channel, out_channel, 3, 1, 1, bias=True)),
nn.BatchNorm2d(out_channel),
nn.LeakyReLU(0.1),
nn.AvgPool2d(2, 2)
)
else:
self.block = nn.Sequential(
spectral_norm(nn.Conv2d(in_channel, out_channel, 3, 1, 1, bias=True)),
nn.BatchNorm2d(out_channel),
nn.LeakyReLU(0.1)
)
def forward(self, x):
return self.block(x)
class Discriminator(nn.Module):
def __init__(self, nfc=256):
super(Discriminator, self).__init__()
self.block1 = DownConvBlock(nfc, nfc // 2, down=False)
self.block2 = DownConvBlock(nfc // 2, nfc // 4, down=True)
self.out = spectral_norm(nn.Conv2d(nfc // 4, 1, 4, 2, 0))
def forward(self, input):
out = self.out(self.block2(self.block1(input)))
return out.view(-1)
def train_dis(net, data, label="real"):
pred = net(data)
if label == "real":
loss = F.relu(1 - pred).mean()
else:
loss = F.relu(1 + pred).mean()
loss.backward()
def save_images(vgg, net_g, datas, saved_image_folder, n_iter):
net_g.eval()
with torch.no_grad():
feats = vgg(datas)
imgs = net_g(feats[3], feats[2], feats[1], feats[0])
vutils.save_image(torch.cat([datas, imgs], dim=0),
os.path.join(saved_image_folder, 'iter_%d.png' % (n_iter)), nrow=8)
# range=(-1, 1), normalize=True)
del datas, feats, imgs
net_g.train()