-
Notifications
You must be signed in to change notification settings - Fork 4
/
models.py
200 lines (149 loc) · 7.39 KB
/
models.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
import torch
import torch.nn as nn
import torch.nn.functional as F
class MaskedConv2d(nn.Conv2d):
def __init__(self, mask_type, *args, **kwargs):
super(MaskedConv2d, self).__init__(*args, **kwargs)
assert mask_type in {'A', 'B'}
self.register_buffer('mask', self.weight.data.clone())
_,_,kH,kW = self.weight.size()
self.mask.fill_(1)
self.mask[:,:,kH//2,kW//2 + (mask_type == 'B'):] = 0
self.mask[:,:,kH//2 + 1:] = 0
def forward(self, x):
self.weight.data *= self.mask
return super(MaskedConv2d, self).forward(x)
class ConvnetBlock(nn.Module):
def __init__(self, filters, *args, **kwargs):
super(ConvnetBlock, self).__init__(*args, **kwargs)
def forward(self, x):
#
# Problem 5a: Implement a residual convnet block as described in Lecture 7.
# Use a kernel size of 3. Do not implement 1x1 convolutions.
#
raise NotImplementedError
class MaskedConvnetBlock(nn.Module):
def __init__(self, filters, *args, **kwargs):
super(MaskedConvnetBlock, self).__init__(*args, **kwargs)
def forward(self, x):
#
# Problem 6a: Implement a masked residual convnet block as described in Lecture 7.
# Use a kernel size of 3. Do not implement 1x1 convolutions.
# Use the MaskedConv2d to implement a masked convolution.
# You'll want to use mask-type B.
#
raise NotImplementedError
class PixelCNN(nn.Module):
def __init__(self, capacity=32, depth=9, *args, **kwargs):
super(PixelCNN, self).__init__(*args, **kwargs)
self.capacity = capacity
self.embed = MaskedConv2d('A', 1, capacity, 9, padding=4, bias=False)
self.resnet = nn.ModuleList()
for i in range(depth): self.resnet.append(MaskedConvnetBlock(capacity))
self.image = MaskedConv2d('B', capacity, 1, 3, padding=1, bias=True)
self.bias = nn.Parameter(torch.Tensor(28,28))
for name, parm in self.named_parameters():
if name.endswith('weight'): nn.init.normal_(parm, 0, .01)
if name.endswith('bias'): nn.init.constant_(parm, 0.0)
def sample(self, n):
x = torch.zeros(n,1,28,28).cuda()
for i in range(28):
for j in range(28):
p = torch.sigmoid(self.forward(x).detach()[:,:,i,j])
x[:,:,i,j] = torch.bernoulli(p)
return x
def forward(self, x):
zx = F.relu(self.embed(x))
for layer in self.resnet: zx = layer(zx)
return self.image(zx) + self.bias[None,None,:,:]
class GaussianVAEDecoder(nn.Module):
def __init__(self, capacity=32, depth=51, autoregress=False, *args, **kwargs):
super(GaussianVAEDecoder, self).__init__(*args, **kwargs)
self.capacity = capacity
self.embed = nn.Linear(49, capacity*7*7, bias=False)
self.resnet = nn.ModuleList()
for i in range(depth): self.resnet.append(ConvnetBlock(capacity))
self.image = nn.ConvTranspose2d(capacity, 1, 4, stride=4, bias=True)
self.bias = nn.Parameter(torch.Tensor(28,28))
for name, parm in self.named_parameters():
if name.endswith('weight'): nn.init.normal_(parm, 0, .01)
if name.endswith('bias'): nn.init.constant_(parm, 0.0)
def sample(self, z, sigma):
return torch.normal(self(z), sigma).clamp(0,1)
def forward(self, s):
zx = F.relu(self.embed(s.view(-1,49)).view(-1,self.capacity,7,7))
for layer in self.resnet: zx = layer(zx)
return torch.sigmoid(self.image(zx) + self.bias[None,None,:,:])
class DiscreteVAEDecoder(nn.Module):
def __init__(self, capacity=32, depth=51, autoregress=False, *args, **kwargs):
super(DiscreteVAEDecoder, self).__init__(*args, **kwargs)
self.capacity = capacity
self.embed = nn.Linear(49, capacity*7*7, bias=False)
self.resnet = nn.ModuleList()
for i in range(depth): self.resnet.append(ConvnetBlock(capacity))
self.image = nn.ConvTranspose2d(capacity, 1, 4, stride=4, bias=True)
self.bias = nn.Parameter(torch.Tensor(28,28))
# regardless whether we autoregress, warm up without it
self.autoregress = False
if autoregress:
self.pixelcnn = PixelCNN()
for name, parm in self.named_parameters():
if name.endswith('weight'): nn.init.normal_(parm, 0, .01)
if name.endswith('bias'): nn.init.constant_(parm, 0.0)
def sample(self, z):
if self.autoregress:
x = torch.zeros(z.shape[0],1,28,28).cuda()
for i in range(28):
for j in range(28):
p = torch.sigmoid(self.forward(z,x).detach()[:,:,i,j])
x[:,:,i,j] = torch.bernoulli(p)
else:
p = torch.sigmoid(self(z).detach())
x = torch.bernoulli(p)
return x
def forward(self, s, x=None):
zx = F.relu(self.embed(s.view(-1,49)).view(-1,self.capacity,7,7))
for layer in self.resnet: zx = layer(zx)
xr = self.image(zx) + self.bias[None,None,:,:]
return xr if not self.autoregress else self.pixelcnn(x) + xr
class IAF(nn.Module):
def __init__(self, filters, depth=9, *args, **kwargs):
super(IAF, self).__init__(*args, **kwargs)
self.embedz = MaskedConv2d('A',1,filters,3,1,1,bias=False)
self.embedh = nn.Conv2d(filters,filters,3,1,1,bias=False)
resnet = []
for i in range(9): resnet.append(MaskedConvnetBlock(filters))
self.resnet = nn.Sequential(*resnet)
self.m = MaskedConv2d('B',filters,1,3,padding=1,bias=True)
self.s = MaskedConv2d('B',filters,1,3,padding=1,bias=True)
def forward(self, z, h):
u = F.relu(self.embedz(z) + self.embedh(h))
u = self.resnet(u)
return self.m(u), 1 + self.s(u)
class VAEEncoder(nn.Module):
def __init__(self, capacity=32, depth=9, flows=0, *args, **kwargs):
super(VAEEncoder, self).__init__(*args, **kwargs)
self.capacity = capacity
self.embed = nn.Conv2d(1, capacity, 7, padding=3, stride=4, bias=False)
self.resnet = nn.ModuleList()
for i in range(depth): self.resnet.append(ConvnetBlock(capacity))
self.mu = nn.Conv2d(capacity, 1, 3, padding=1, bias=True)
self.var = nn.Conv2d(capacity, 1, 3, padding=1, bias=True)
self.flows = nn.ModuleList()
for f in range(flows): self.flows.append(IAF(capacity))
for name, parm in self.named_parameters():
if name.endswith('weight'): nn.init.normal_(parm, 0, .01)
if name.endswith('bias'): nn.init.constant_(parm, 0.0)
def forward(self, x, epsilon):
h = F.relu(self.embed(x))
for layer in self.resnet: h = layer(h)
mu, logvar = self.mu(h), self.var(h)
z = mu + torch.exp(.5*logvar)*epsilon
logqzx = 0.5*(logvar + torch.pow((z-mu),2)/logvar.exp()).view(-1,49).sum(1)
for t, flow in enumerate(self.flows):
#
# Problem 6d: Calculate an inverse autoregressive flow z.
# While calculating z, accumulate a calculation of its probability.
#
raise NotImplementedError
return z.view(-1,49), logqzx, mu.view(-1,49), logvar.view(-1,49)