-
Notifications
You must be signed in to change notification settings - Fork 1
/
RealNVP.py
213 lines (185 loc) · 8.12 KB
/
RealNVP.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import numpy as np
# weighted norm convolutional 2D normalization
class WNConv2d(nn.Module):
def __init__(self, in_dim, out_dim, kernel_size, stride=1, padding=0,
bias=True):
super(WNConv2d, self).__init__()
self.conv = nn.utils.weight_norm(
nn.Conv2d(in_dim, out_dim, kernel_size, stride=stride, padding=padding,
bias=bias))
def forward(self, x):
return self.conv(x)
# Resnet block using the weighted norm
class ResnetBlock(nn.Module):
def __init__(self, filters):
super(ResnetBlock, self).__init__()
self.block = nn.Sequential(
WNConv2d(filters, filters, (1, 1), stride=1, padding=0),
nn.ReLU(),
WNConv2d(filters, filters, (3, 3), stride=1, padding=1),
nn.ReLU(),
WNConv2d(filters, filters, (1, 1), stride=1, padding=0))
def forward(self, x):
return x + self.block(x)
# Activation normalization that adds scale and translation to each channel and using data dependent normalization
class ActNorm(nn.Module):
def __init__(self, n_channels):
super(ActNorm, self).__init__()
self.log_scale = nn.Parameter(torch.zeros(1, n_channels, 1, 1), requires_grad = True) # scale factor (s) in paper
self.bias = nn.Parameter(torch.zeros(1, n_channels, 1, 1), requires_grad = True) # translation factor
self.channels = n_channels
self.initialized = False
def forward(self, x, reverse = False):
if reverse:
return (x - self.bias) * torch.exp(-self.log_scale), self.log_scale
if not self.initialized:
self.log_scale.data = -torch.log(torch.std(x.permute(1, 0, 2, 3).reshape(self.channels, -1), dim = 1)).view(1, self.channels, 1, 1)
self.bias.data = -torch.mean(x.permute(1, 0, 2, 3).reshape(self.channels, -1), dim = 1).view(1, self.channels, 1, 1)
self.initialized = True
return x * torch.exp(self.log_scale) + self.bias, self.log_scale
class Resnet(nn.Module):
def __init__(self, in_channels = 3, out_channels = 6, filters = 128, blocks = 3):
super(Resnet, self).__init__()
layers = []
layers.extend([WNConv2d(in_channels, filters, (3, 3), stride = 1, padding = 1),
nn.ReLU()])
for _ in range(blocks):
layers.append(ResnetBlock(filters))
layers.extend([nn.ReLU(),
WNConv2d(filters, out_channels, (3, 3), stride = 1, padding = 1)])
self.resnet = nn.Sequential(*layers)
def forward(self, x):
return self.resnet(x)
class AffineCheckerboardTransform(nn.Module):
def __init__(self, type=1.0):
super(AffineCheckerboardTransform, self).__init__()
self.mask = self.build_mask(type=type)
self.scale = nn.Parameter(torch.zeros(1), requires_grad=True)
self.scale_shift = nn.Parameter(torch.zeros(1), requires_grad=True)
self.resnet = Resnet()
def build_mask(self, type=1.0):
# if type == 1.0, the top left corner will be 1.0 else on type == 0.0 it will be 0.0
mask = np.arange(32).reshape(-1, 1) + np.arange(32)
mask = np.mod(type + mask, 2)
mask = mask.reshape(-1, 1, 32, 32)
return torch.tensor(mask.astype('float32')).to(device)
def forward(self, x, reverse=False):
# returns transform(x), log_det
batch_size, n_channels, _, _ = x.shape
mask = self.mask.repeat(batch_size, 1, 1, 1)
x_ = x * mask
# from pseudo-code provided
log_s, t = self.resnet(x_).split(n_channels, dim=1)
log_s = self.scale * torch.tanh(log_s) + self.scale_shift # both scale and scale_shift learnable params
t = t * (1.0 - mask) # for the other half of the x
log_s = log_s * (1.0 - mask) # for the other half of the x
if reverse: # inverting the transformation
x = (x - t) * torch.exp(-log_s) # for inverse
else:
x = x * torch.exp(log_s) + t # for forward
return x, log_s
class AffineChannelTransform(nn.Module):
def __init__(self, modify_top):
'''
modify_top : Signifies which half of x is activated
'''
super(AffineChannelTransform, self).__init__()
self.modify_top = modify_top
self.scale = nn.Parameter(torch.zeros(1), requires_grad=True)
self.scale_shift = nn.Parameter(torch.zeros(1), requires_grad=True)
self.resnet = Resnet(in_channels=6, out_channels=12)
def forward(self, x, reverse=False):
batch_size, n_channels, _, _ = x.shape
if self.modify_top:
on, off = x.split(n_channels // 2, dim=1)
else:
off, on = x.split(n_channels // 2, dim=1)
log_s, t = self.resnet(off).split(n_channels // 2, dim=1)
log_s = self.scale * torch.tanh(log_s) + self.scale_shift
if reverse: # inverting the transformation
on = (on - t) * torch.exp(-log_s)
else:
on = on * torch.exp(log_s) + t
if self.modify_top:
return torch.cat([on, off], dim=1), torch.cat([log_s, torch.zeros_like(log_s)], dim=1)
else:
return torch.cat([off, on], dim=1), torch.cat([torch.zeros_like(log_s), log_s], dim=1)
class RealNVP(nn.Module):
def __init__(self):
super(RealNVP, self).__init__()
self.prior = torch.distributions.Normal(torch.tensor(0.).to(device), torch.tensor(1.).to(device)) # standard normal distribution
self.checker_transforms1 = nn.ModuleList([
AffineCheckerboardTransform(1.0),
ActNorm(3),
AffineCheckerboardTransform(0.0),
ActNorm(3),
AffineCheckerboardTransform(1.0),
ActNorm(3),
AffineCheckerboardTransform(0.0)
])
self.channel_transforms = nn.ModuleList([
AffineChannelTransform(True),
ActNorm(12),
AffineChannelTransform(False),
ActNorm(12),
AffineChannelTransform(True),
])
self.checker_transforms2 = nn.ModuleList([
AffineCheckerboardTransform(1.0),
ActNorm(3),
AffineCheckerboardTransform(0.0),
ActNorm(3),
AffineCheckerboardTransform(1.0)
])
def squeeze(self, x):
# C x H x W -> 4C x H/2 x W/2
B, C, H, W = x.size()
x = x.reshape(B, C, H // 2, 2, W // 2, 2)
x = x.permute(0, 1, 3, 5, 2, 4)
x = x.reshape(B, C * 4, H // 2, W // 2)
return x
def unsqueeze(self, x):
# 4C x H/2 x W/2 -> C x H x W
B, C, H, W = x.size()
x = x.reshape(B, C // 4, 2, 2, H, W)
x = x.permute(0, 1, 4, 2, 5, 3)
x = x.reshape(B, C // 4, H * 2, W * 2)
return x
def g(self, z):
# z -> x (inverse of f)
x = z
for op in reversed(self.checker_transforms2):
x, _ = op.forward(x, reverse=True)
x = self.squeeze(x)
for op in reversed(self.channel_transforms):
x, _ = op.forward(x, reverse=True)
x = self.unsqueeze(x)
for op in reversed(self.checker_transforms1):
x, _ = op.forward(x, reverse=True)
return x
def f(self, x):
# maps x -> z, and returns the log determinant (not reduced)
z, log_det = x, torch.zeros_like(x)
for op in self.checker_transforms1:
z, delta_log_det = op.forward(z)
log_det += delta_log_det
z, log_det = self.squeeze(z), self.squeeze(log_det)
for op in self.channel_transforms:
z, delta_log_det = op.forward(z)
log_det += delta_log_det
z, log_det = self.unsqueeze(z), self.unsqueeze(log_det)
for op in self.checker_transforms2:
z, delta_log_det = op.forward(z)
log_det += delta_log_det
return z, log_det
def log_prob(self, x):
z, log_det = self.f(x)
# equation 3 RealNVP paper
return torch.sum(log_det, dim = [1, 2, 3]) + torch.sum(self.prior.log_prob(z), dim = [1, 2, 3])
def sample(self, num_samples):
z = self.prior.sample([num_samples, 3, 32, 32])
return self.g(z)