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
| import numpy as np
import torch
import torch.nn as nn
class ConvBNReLU(nn.Module):
def __init__(self, in_nc, out_nc, k=3, s=2, nlayer=1):
super().__init__()
p = (k - 1) // 2
module_list = list()
for i in range(nlayer):
mid_nc = out_nc if i == nlayer - 1 else in_nc
cur_s = s if i == nlayer - 1 else 1
module_list.extend([
nn.Conv2d(in_nc, mid_nc, kernel_size=k, stride=cur_s, padding=p),
nn.BatchNorm2d(mid_nc),
nn.ReLU(inplace=True)
])
self.block = nn.Sequential(*module_list)
def forward(self, x):
out = self.block(x)
return out
class DeconvBNReLU(nn.Module):
def __init__(self, in_nc, out_nc, k=2, s=2, nlayer=1):
super().__init__()
p = (k - 1) // 2
module_list = list()
for i in range(nlayer):
if i == 0:
module_list.extend([
nn.ConvTranspose2d(in_nc, out_nc, kernel_size=k, stride=s, padding=p),
nn.BatchNorm2d(out_nc),
nn.ReLU(inplace=True)
])
else:
module_list.extend([
nn.Conv2d(out_nc, out_nc, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(out_nc),
nn.ReLU(inplace=True)
])
self.block = nn.Sequential(*module_list)
def forward(self, x):
out = self.block(x)
return out
class MLP(nn.Module):
def __init__(self, in_nc, mid_nc, out_nc, nlayer=2, is_flatten=True):
super().__init__()
mlp_blocks = list()
for i in range(nlayer):
cur_in = in_nc if i == 0 else mid_nc
cur_out = out_nc if i == nlayer - 1 else mid_nc
mlp_blocks.append(nn.Linear(cur_in, cur_out))
if i < nlayer - 1:
mlp_blocks.append(nn.BatchNorm1d(cur_out))
mlp_blocks.append(nn.ReLU(cur_out))
self.mlp = nn.Sequential(*mlp_blocks)
self.is_flatten = is_flatten
def forward(self, x):
if self.is_flatten:
vec = torch.flatten(x, start_dim=1)
else:
vec = x
out = self.mlp(vec)
return out
class Encoder(nn.Module):
def __init__(self, in_nc=1, in_size=(28, 28),
base_ch=8, num_layer=2,
mlp_nc=128, latent_dim=64):
super().__init__()
in_h, in_w = in_size
self.out_size = (base_ch * (2 ** (num_layer - 1)), in_h // 2 ** num_layer, in_w // 2 ** num_layer)
mlp_in_nc = self.out_size[0] * self.out_size[1] * self.out_size[2]
basic_blocks = list()
for i in range(num_layer):
cur_in = in_nc if i == 0 else base_ch * (2 ** (i - 1))
cur_out = base_ch * (2 ** i)
basic_blocks.append(
ConvBNReLU(cur_in, cur_out, k=3, s=2, nlayer=3)
)
self.encode_block = nn.Sequential(*basic_blocks)
self.calc_mean = MLP(mlp_in_nc, mlp_nc, latent_dim, is_flatten=True)
self.calc_logvar = MLP(mlp_in_nc, mlp_nc, latent_dim, is_flatten=True)
def forward(self, x):
feat = self.encode_block(x)
mean = self.calc_mean(feat)
logvar = self.calc_logvar(feat)
return mean, logvar
class Decoder(nn.Module):
def __init__(self, in_size=(64, 7, 7), out_nc=1,
base_ch=8, num_layer=2,
mlp_nc=128, latent_dim=64):
super().__init__()
mlp_out_nc = in_size[0] * in_size[1] * in_size[2]
self.in_size = in_size
self.mlp = MLP(latent_dim, mlp_nc, mlp_out_nc, nlayer=2, is_flatten=False)
basic_blocks = list()
for i in range(num_layer):
cur_in = base_ch * (2 ** (num_layer - 1 - i))
cur_out = out_nc if i == num_layer - 1 else base_ch * (2 ** (num_layer - 2 - i))
basic_blocks.append(
DeconvBNReLU(cur_in, cur_out, k=2, s=2, nlayer=3)
)
self.decode_block = nn.Sequential(*basic_blocks)
def forward(self, x):
feat = self.mlp(x)
feat = feat.view(-1, *self.in_size)
out = self.decode_block(feat)
return out
class VariationalAutoEncoder(nn.Module):
def __init__(self, in_nc=1, in_size=(28, 28),
base_ch=8, num_layer_enc=2, num_layer_dec=2,
mlp_nc=128, latent_dim=64, decoder_in_size=(16, 7, 7)):
super().__init__()
self.latent_dim = latent_dim
self.encoder = Encoder(in_nc, in_size, base_ch, num_layer_enc, mlp_nc, latent_dim)
self.decoder = Decoder(decoder_in_size, in_nc, base_ch, num_layer_dec, mlp_nc, latent_dim)
def reparam(self, mu, logvar):
std = torch.exp(logvar / 2)
eps = torch.randn_like(std)
return mu + std * eps
def forward(self, x):
mu, logvar = self.encoder(x)
z = self.reparam(mu, logvar)
output = self.decoder(z)
return output, mu, logvar
def sample(self, n_samples):
z = torch.randn(n_samples, self.latent_dim)
samples = self.decoder(z)
return samples
if __name__ == "__main__":
vae = VariationalAutoEncoder()
dummy_input = torch.randn(4, 1, 28, 28)
out = vae(dummy_input)
print(out.size())
|
