-
Notifications
You must be signed in to change notification settings - Fork 0
/
VAECNN.py
222 lines (179 loc) · 7.89 KB
/
VAECNN.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
'''ResNet in PyTorch.
For Pre-activation ResNet, see 'preact_resnet.py'.
Reference:
[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
Deep Residual Learning for Image Recognition. arXiv:1512.03385
'''
import torch
import torch.nn as nn
import torch.nn.functional as F
def kl_loss(mu, log_var):
return torch.mean(-0.5 * torch.sum(1 + log_var - mu ** 2 - log_var.exp(), dim = 1))
def sample_image(mu, log_var):
# log_var = self.conv1_log_var(x)
# mu = self.conv1_mu(x)
std = torch.exp(torch.mul(log_var, 0.5))
eps = torch.randn_like(std)
return eps * std + mu
class VAEBasicBlock(nn.Module):
expansion = 1
def __init__(self, in_planes, planes, stride=1):
super(VAEBasicBlock, self).__init__()
self.stride = stride
self.in_planes = in_planes
self.planes = planes
self.conv1_mu = nn.Conv2d(
in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.conv1_log_var = nn.Conv2d(
in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2_mu = nn.Conv2d(planes, planes, kernel_size=3,
stride=1, padding=1, bias=False)
self.conv2_log_var = nn.Conv2d(planes, planes, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion*planes:
self.conv_mu_shortcut = nn.Conv2d(in_planes, self.expansion*planes,
kernel_size=1, stride=stride, bias=False)
self.conv_logvar_shortcut = nn.Conv2d(in_planes, self.expansion*planes,
kernel_size=1, stride=stride, bias=False)
self.bn_shortcut = nn.BatchNorm2d(self.expansion*planes)
def forward(self, x):
mu1 = self.conv1_mu(x)
log_var1 = self.conv1_log_var(x)
std1 = torch.exp(torch.mul(log_var1, 0.5))
eps1 = torch.randn_like(std1)
conv1 = eps1 * std1 + mu1
out = F.relu(self.bn1(conv1))
log_var2 = self.conv2_log_var(out)
mu2 = self.conv2_mu(out)
std2 = torch.exp(torch.mul(log_var2, 0.5))
eps2 = torch.randn_like(std2)
conv2 = eps2 * std2 + mu2
out = self.bn2(conv2)
# out += self.shortcut(x)
if self.stride != 1 or self.in_planes != self.expansion*self.planes:
mu_shortcut = self.conv_mu_shortcut(x)
logvar_shortcut = self.conv_logvar_shortcut(x)
std_shortcut = torch.exp(torch.mul(logvar_shortcut, 0.5))
eps_shortcut = torch.randn_like(std_shortcut)
conv_shortcut = eps_shortcut * std_shortcut + mu_shortcut
out += self.bn_shortcut(conv_shortcut)
else:
out += self.shortcut(x)
out = F.relu(out)
return out, kl_loss(mu=mu1, log_var=log_var1) + kl_loss(mu=mu2, log_var=log_var2)
class VAEResNet(nn.Module):
def __init__(self, block, num_blocks, num_classes=10):
super(VAEResNet, self).__init__()
self.in_planes = 64
self.conv1_mu = nn.Conv2d(3, 64, kernel_size=3,
stride=1, padding=1, bias=False)
self.conv1_log_var = nn.Conv2d(3, 64, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
self.linear = nn.Linear(512*block.expansion, num_classes)
def _make_layer(self, block, planes, num_blocks, stride):
strides = [stride] + [1]*(num_blocks-1)
layers = []
# for stride in strides:
layers.append(block(self.in_planes, planes, stride))
self.in_planes = planes * block.expansion
return nn.Sequential(*layers)
def kl_loss(self, mu, log_var):
return torch.mean(-0.5 * torch.sum(1 + log_var - mu ** 2 - log_var.exp(), dim = 1))
def forward(self, x, _eval=False):
if _eval:
# switch to eval mode
self.eval()
else:
self.train()
log_var = self.conv1_log_var(x)
mu = self.conv1_mu(x)
std = torch.exp(torch.mul(log_var, 0.5))
eps = torch.randn_like(std)
conv1 = eps * std + mu
out = F.relu(self.bn1(conv1))
out, kl_loss1 = self.layer1(out)
out, kl_loss2 = self.layer2(out)
out, kl_loss3 = self.layer3(out)
out, kl_loss4 = self.layer4(out)
out = F.avg_pool2d(out, 4)
out = out.view(out.size(0), -1)
out = self.linear(out)
self.train()
return [out, self.kl_loss(mu, log_var) + kl_loss1 + kl_loss2 + kl_loss3 + kl_loss4]
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, in_planes, planes, stride=1):
super(BasicBlock, self).__init__()
self.conv1 = nn.Conv2d(
in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion*planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion*planes,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion*planes)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out += self.shortcut(x)
out = F.relu(out)
return out
class VAECNNFirstLayerChanged(nn.Module):
def __init__(self, block, num_blocks, num_classes=10):
super(VAECNNFirstLayerChanged, self).__init__()
self.in_planes = 64
self.conv1_mu = nn.Conv2d(3, 64, kernel_size=3,
stride=1, padding=1, bias=False)
self.conv1_log_var = nn.Conv2d(3, 64, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
self.linear = nn.Linear(512*block.expansion, num_classes)
def _make_layer(self, block, planes, num_blocks, stride):
strides = [stride] + [1]*(num_blocks-1)
layers = []
for stride in strides:
layers.append(block(self.in_planes, planes, stride))
self.in_planes = planes * block.expansion
return nn.Sequential(*layers)
def forward(self, x, _eval=False):
if _eval:
# switch to eval mode
self.eval()
else:
self.train()
log_var = self.conv1_log_var(x)
mu = self.conv1_mu(x)
std = torch.exp(torch.mul(log_var, 0.5))
eps = torch.randn_like(std)
conv1 = eps * std + mu
out = F.relu(self.bn1(conv1))
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = self.layer4(out)
out = F.avg_pool2d(out, 4)
out = out.view(out.size(0), -1)
out = self.linear(out)
self.train()
return out , kl_loss(mu, log_var)
def VAEResNet18FirstLayerChanged():
return VAECNNFirstLayerChanged(BasicBlock, [2,2,2,2])
def VAEResNet18():
return VAEResNet(VAEBasicBlock, [2,2,2,2])