-
Notifications
You must be signed in to change notification settings - Fork 0
/
ImageTransformationNetwork.py
242 lines (200 loc) · 9.63 KB
/
ImageTransformationNetwork.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from config import options
import random
#############
# StyPath ###
#############
class SEStyEncNet(torch.nn.Module):
def __init__(self, num_styles):
super(SEStyEncNet, self).__init__()
self.conv_1 = CondConvolution(3, 32, 3, 1, (1, 1), num_styles)
self.se1 = SE_Block(32)
self.conv_2 = CondConvolution(32, 64, 3, 1, (2, 2), num_styles)
self.se2 = SE_Block(64)
self.conv_3 = CondConvolution(64, 128, 3, 1, (2, 2), num_styles)
self.se3 = SE_Block(128)
self.res_block1 = ResBlock(num_styles)
self.se4 = SE_Block(128)
self.res_block2 = ResBlock(num_styles)
self.se5 = SE_Block(128)
# self.upsample1 = Upsampling(128, 64, 3, 1, (1, 1), num_styles)
# self.se6 = SE_Block(64)
# self.upsample2 = Upsampling(64, 32, 3, 1, (1, 1), num_styles)
# self.se7 = SE_Block(32)
self.upsample1 = UpsampleConvLayer(128, 64, kernel_size=3, stride=1, upsample=2, num_styles=num_styles)
self.se6 = SE_Block(64)
self.upsample2 = UpsampleConvLayer(64, 32, kernel_size=3, stride=1, upsample=2, num_styles=num_styles)
self.se7 = SE_Block(32)
self.conv_4 = CondConvolution(32, 3, 3, 1, (1, 1), num_styles, False)
def forward(self, x, style_no, style_no2=None, style_no3=None, alphas=None):
x = self.conv_1(x, style_no, style_no2, style_no3, alphas)
x = self.se1(x)
x = self.conv_2(x, style_no, style_no2, style_no3, alphas)
x = self.se2(x)
x = self.conv_3(x, style_no, style_no2, style_no3, alphas)
x = self.se3(x)
x = self.res_block1(x, style_no, style_no2, style_no3, alphas)
x = self.se4(x)
x = self.res_block2(x, style_no, style_no2, style_no3, alphas)
x = self.se5(x)
x = self.upsample1(x, style_no, style_no2, style_no3, alphas)
x = self.se6(x)
x = self.upsample2(x, style_no, style_no2, style_no3, alphas)
x = self.se7(x)
x = self.conv_4(x, style_no, style_no2, style_no3, alphas)
x = F.sigmoid(x)
return x
class SE_Block(nn.Module):
"credits: https://github.com/moskomule/senet.pytorch/blob/master/senet/se_module.py#L4"
def __init__(self, c, r=16):
super().__init__()
self.squeeze = nn.AdaptiveAvgPool2d(1)
self.excitation = nn.Sequential(
nn.Linear(c, c // r, bias=False),
nn.ReLU(inplace=True),
nn.Linear(c // r, c, bias=False),
nn.Sigmoid()
)
def forward(self, x):
bs, c, _, _ = x.shape
y = self.squeeze(x).view(bs, c)
y = self.excitation(y).view(bs, c, 1, 1)
return x * y.expand_as(x)
class UpsampleConvLayer(torch.nn.Module):
"""UpsampleConvLayer
Upsamples the input and then does a convolution. This method gives better results
compared to ConvTranspose2d.
ref: http://distill.pub/2016/deconv-checkerboard/
"""
def __init__(self, in_channels, out_channels, kernel_size, stride, upsample=None, num_styles=None):
super(UpsampleConvLayer, self).__init__()
self.upsample = upsample
reflection_padding = kernel_size // 2
self.reflection_pad = torch.nn.ReflectionPad2d(reflection_padding)
self.conv2d = CondConvolution(in_channels, out_channels, kernel_size, 0, stride, num_styles)
def forward(self, x, style_no, style_no2=None, style_no3=None, alphas=None):
x_in = x
if self.upsample:
x_in = torch.nn.functional.interpolate(x_in, mode='nearest', scale_factor=self.upsample)
out = self.reflection_pad(x_in)
out = self.conv2d(out, style_no, style_no2, style_no3, alphas)
return out
class CondConvolution(nn.Module):
def __init__(self, input_filters, output_filters, kernel_size, padding, stride, num_styles, act=True):
super(CondConvolution, self).__init__()
self.reflection2d = nn.ReflectionPad2d(padding)
self.conv = nn.Conv2d(input_filters, output_filters, kernel_size=kernel_size, stride=stride)
self.instnorm = nn.InstanceNorm2d(output_filters, affine=True)
self.act = act
# Conditional Instance Parameters - Allows for transferring of multiple styles
self.gamma = torch.nn.Parameter(data=torch.Tensor(num_styles, output_filters), requires_grad=True)
nn.init.normal_(self.gamma, mean=0.0, std=0.1) # 0.01
self.beta = torch.nn.Parameter(data=torch.Tensor(num_styles, output_filters), requires_grad=True)
nn.init.normal_(self.beta, mean=0.0, std=0.1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.normal_(m.weight, mean=0, std=0.01)
elif isinstance(m, nn.InstanceNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def forward(self, x, style_no1, style_no2=None, style_no3=None, alphas=None):
x = self.reflection2d(x)
x = self.conv(x)
x = self.instnorm(x)
b, d, w, h = x.size()
x = x.view(b, d, w * h)
# Index into gamma and beta parameters to select specific style
if options.mix_styles:
gamma = alphas[0] * self.gamma[style_no1] + alphas[1] * self.gamma[style_no2] + \
alphas[2] * self.gamma[style_no3] # gamma is N by C
beta = alphas[0] * self.beta[style_no1] + alphas[1] * self.beta[style_no2] + \
alphas[2] * self.beta[style_no3]
else:
gamma = self.gamma[style_no1]
beta = self.beta[style_no1]
x = (x * gamma.unsqueeze(-1).expand_as(x) + beta.unsqueeze(-1).expand_as(x)).view(b, d, w, h)
if self.act:
x = F.relu(x)
return x
#############
# Vanilla ###
#############
class TransformerNet(torch.nn.Module):
def __init__(self, num_styles):
super(TransformerNet, self).__init__()
self.conv_1 = CondConvolution(3, 32, 9, 4, (1, 1), num_styles)
self.conv_2 = CondConvolution(32, 64, 3, 1, (2, 2), num_styles)
self.conv_3 = CondConvolution(64, 128, 3, 1, (2, 2), num_styles)
self.res_block1 = ResBlock(num_styles)
self.res_block2 = ResBlock(num_styles)
self.res_block3 = ResBlock(num_styles)
self.res_block4 = ResBlock(num_styles)
self.upsample1 = Upsampling(128, 64, 3, 1, (1, 1), num_styles)
self.upsample2 = Upsampling(64, 32, 3, 1, (1, 1), num_styles)
self.conv_4 = CondConvolution(32, 3, 9, 4, (1, 1), num_styles, False)
def forward(self, x, style_no):
x = self.conv_1(x, style_no)
x = self.conv_2(x, style_no)
x = self.conv_3(x, style_no)
x = self.res_block1(x, style_no)
x = self.res_block2(x, style_no)
x = self.res_block3(x, style_no)
x = self.res_block4(x, style_no)
x = self.upsample1(x, style_no)
x = self.upsample2(x, style_no)
x = self.conv_4(x, style_no)
x = F.sigmoid(x)
return x
"""class CondConvolution(nn.Module):
def __init__(self, input_filters, output_filters, kernel_size, padding, stride, num_styles, act=True):
super(CondConvolution, self).__init__()
self.reflection2d = nn.ReflectionPad2d(padding)
self.conv = nn.Conv2d(input_filters, output_filters, kernel_size=kernel_size, stride=stride)
self.instnorm = nn.InstanceNorm2d(output_filters, affine=True)
self.act = act
# Conditional Instance Parameters - Allows for transferring of multiple styles
self.gamma = torch.nn.Parameter(data=torch.Tensor(num_styles, output_filters), requires_grad=True)
nn.init.normal_(self.gamma, mean=0.0, std=0.1) # 0.01
self.beta = torch.nn.Parameter(data=torch.Tensor(num_styles, output_filters), requires_grad=True)
nn.init.normal_(self.beta, mean=0.0, std=0.1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.normal_(m.weight, mean=0, std=0.01)
elif isinstance(m, nn.InstanceNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def forward(self, x, style_no):
x = self.reflection2d(x)
x = self.conv(x)
x = self.instnorm(x)
b, d, w, h = x.size()
x = x.view(b, d, w * h)
# Index into gamma and beta parameters to select specific style
gamma = self.gamma[style_no] # gamma is N by C
beta = self.beta[style_no]
x = (x * gamma.unsqueeze(-1).expand_as(x) + beta.unsqueeze(-1).expand_as(x)).view(b, d, w, h)
if self.act:
x = F.relu(x)
return x"""
class ResBlock(nn.Module):
def __init__(self, num_styles):
super(ResBlock, self).__init__()
self.res_conv1 = CondConvolution(128, 128, 3, 1, (1, 1), num_styles)
self.res_conv2 = CondConvolution(128, 128, 3, 1, (1, 1), num_styles, False)
def forward(self, x, style_no, style_no2=None, style_no3=None, alphas=None):
residual = x
out = self.res_conv1(x, style_no, style_no2, style_no3, alphas)
out = self.res_conv2(out, style_no, style_no2, style_no3, alphas)
out += residual
return out
class Upsampling(nn.Module):
def __init__(self, input_filters, output_filters, kernel_size, padding, stride, num_styles):
super(Upsampling, self).__init__()
self.upsample = nn.Upsample(scale_factor=2, mode='nearest')
self.conv = CondConvolution(input_filters, output_filters, kernel_size, padding, stride, num_styles)
def forward(self, x, style_no, style_no2=None, style_no3=None, alphas=None):
x = self.upsample(x)
x = self.conv(x, style_no, style_no2, style_no3, alphas)
return x