-
Notifications
You must be signed in to change notification settings - Fork 2
/
VAE.py
159 lines (131 loc) · 5.13 KB
/
VAE.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
from collections import OrderedDict
import numpy as np
import torch
from torch import nn
## implementations from cs294
## VAE for 2D
class MLP(nn.Module):
def __init__(self, input_shape, output_shape, hiddens=[]):
super().__init__()
if isinstance(input_shape, int):
input_shape = (input_shape,)
if isinstance(output_shape, int):
output_shape = (output_shape,)
self.input_shape = input_shape
self.output_shape = output_shape
self.hiddens = hiddens
model = []
prev_h = np.prod(input_shape)
for h in hiddens + [np.prod(output_shape)]:
model.append(nn.Linear(prev_h, h))
model.append(nn.ReLU())
prev_h = h
model.pop()
self.net = nn.Sequential(*model)
def forward(self, x):
b = x.shape[0]
x = x.view(b, -1)
return self.net(x).view(b, *self.output_shape)
class FullyConnectedVAE(nn.Module):
def __init__(self, input_dim, latent_dim, enc_hidden_sizes=[], dec_hidden_sizes=[]):
super().__init__()
self.latent_dim = latent_dim
self.encoder = MLP(input_dim, 2 * latent_dim, enc_hidden_sizes)
self.decoder = MLP(latent_dim, 2 * input_dim, dec_hidden_sizes)
def loss(self, x):
mu_z, log_std_z = self.encoder(x).chunk(2, dim=1)
z = torch.randn_like(mu_z) * log_std_z.exp() + mu_z
mu_x, log_std_x = self.decoder(z).chunk(2, dim=1)
# Compute reconstruction loss - Note that it may be easier for you
# to use torch.distributions.normal to compute the log_prob
recon_loss = (
0.5 * np.log(2 * np.pi)
+ log_std_x
+ (x - mu_x) ** 2 * torch.exp(-2 * log_std_x) * 0.5
)
recon_loss = recon_loss.sum(1).mean()
# Compute KL
kl_loss = -log_std_z - 0.5 + (torch.exp(2 * log_std_z) + mu_z ** 2) * 0.5
kl_loss = kl_loss.sum(1).mean()
return OrderedDict(
loss=recon_loss + kl_loss, recon_loss=recon_loss, kl_loss=kl_loss
)
def sample(self, n, noise=True):
with torch.no_grad():
z = torch.randn(n, self.latent_dim).cuda()
mu, log_std = self.decoder(z).chunk(2, dim=1)
if noise:
z = torch.randn_like(mu) * log_std.exp() + mu
else:
z = mu
return z.cpu().numpy()
## VAE for images
class ConvDecoder(nn.Module):
def __init__(self, latent_dim, output_shape):
super().__init__()
self.latent_dim = latent_dim
self.output_shape = output_shape
self.base_size = (128, output_shape[1] // 8, output_shape[2] // 8)
self.fc = nn.Linear(latent_dim, np.prod(self.base_size))
self.deconvs = nn.Sequential(
nn.ReLU(),
nn.ConvTranspose2d(128, 128, 4, stride=2, padding=1),
nn.ReLU(),
nn.ConvTranspose2d(128, 64, 4, stride=2, padding=1),
nn.ReLU(),
nn.ConvTranspose2d(64, 32, 4, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(32, output_shape[0], 3, padding=1),
)
def forward(self, z):
out = self.fc(z)
out = out.view(out.shape[0], *self.base_size)
out = self.deconvs(out)
return out
class ConvEncoder(nn.Module):
def __init__(self, input_shape, latent_dim):
super().__init__()
self.input_shape = input_shape
self.latent_dim = latent_dim
self.convs = nn.Sequential(
nn.Conv2d(input_shape[0], 32, 3, padding=1),
nn.ReLU(),
nn.Conv2d(32, 64, 3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(64, 128, 3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(128, 256, 3, stride=2, padding=1),
)
conv_out_dim = input_shape[1] // 8 * input_shape[2] // 8 * 256
self.fc = nn.Linear(conv_out_dim, 2 * latent_dim)
def forward(self, x):
out = self.convs(x)
out = out.view(out.shape[0], -1)
mu, log_std = self.fc(out).chunk(2, dim=1)
return mu, log_std
class ConvVAE(nn.Module):
def __init__(self, input_shape, latent_size):
super().__init__()
assert len(input_shape) == 3
self.input_shape = input_shape
self.latent_size = latent_size
self.encoder = ConvEncoder(input_shape, latent_size)
self.decoder = ConvDecoder(latent_size, input_shape)
def loss(self, x):
x = 2 * x - 1
mu, log_std = self.encoder(x)
z = torch.randn_like(mu) * log_std.exp() + mu
x_recon = self.decoder(z)
recon_loss = (
F.mse_loss(x, x_recon, reduction="none").view(x.shape[0], -1).sum(1).mean()
)
kl_loss = -log_std - 0.5 + (torch.exp(2 * log_std) + mu ** 2) * 0.5
kl_loss = kl_loss.sum(1).mean()
return OrderedDict(
loss=recon_loss + kl_loss, recon_loss=recon_loss, kl_loss=kl_loss
)
def sample(self, n):
with torch.no_grad():
z = torch.randn(n, self.latent_size).cuda()
samples = torch.clamp(self.decoder(z), -1, 1)
return samples.cpu().permute(0, 2, 3, 1).numpy() * 0.5 + 0.5