-
Notifications
You must be signed in to change notification settings - Fork 8
/
InvertibleResnet.py
250 lines (225 loc) · 10.6 KB
/
InvertibleResnet.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
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
import torch.nn.functional as F
import torch
import torch.nn as nn
import numpy as np
from SpectralNormGouk import *
def run_module_with_logdet(module, x, num_logdet_iter):
if 'block' in str(module.__class__):
x, dlogdet = module(x, logdet=True, reverse=False, num_logdet_iter=num_logdet_iter)
else:
x, dlogdet = module(x, logdet=True, reverse=False)
return x, dlogdet
logabs = lambda x: torch.log(torch.abs(x))
class ActNorm(nn.Module):
def __init__(self, in_channel):
super().__init__()
self.loc = nn.Parameter(torch.zeros(1, in_channel, 1, 1))
self.scale = nn.Parameter(torch.ones(1, in_channel, 1, 1))
self.initialized = False
def initialize(self, input):
with torch.no_grad():
if len(input.shape) == 2: # linear
flatten = input.permute(1, 0).contiguous().view(input.shape[1], -1)
mean = (
flatten.mean(1)
.unsqueeze(1)
.permute(1, 0,)
)
std = (
flatten.std(1)
.unsqueeze(1)
.permute(1, 0)
)
self.loc.data.copy_(-mean.view_as(self.loc))
self.scale.data.copy_(1 / (std.view_as(self.scale) + 1e-6))
elif len(input.shape) == 4: # conv
flatten = input.permute(1, 0, 2, 3).contiguous().view(input.shape[1], -1)
mean = (
flatten.mean(1)
.unsqueeze(1)
.unsqueeze(2)
.unsqueeze(3)
.permute(1, 0, 2, 3)
)
std = (
flatten.std(1)
.unsqueeze(1)
.unsqueeze(2)
.unsqueeze(3)
.permute(1, 0, 2, 3)
)
self.loc.data.copy_(-mean)
self.scale.data.copy_(1 / (std + 1e-6))
else:
raise 'Input shape not supported {}'.format(input.shape)
def forward(self, input, logdet=False, reverse=False):
scale = self.scale if len(input.shape) == 4 else self.scale.view(1, -1)
loc = self.loc if len(input.shape) == 4 else self.loc.view(1, -1)
if reverse:
return input / scale - loc
if not self.initialized:
self.initialize(input)
self.initialized = True
log_abs = logabs(scale) #return logdet PER sample, not adjusting for number of dimensions!
if len(input.shape) == 4:
_, _, height, width = input.shape
else:
height, width = 1,1
dlogdet = height * width * torch.sum(log_abs)
if logdet:
return scale * (input + loc), dlogdet
else:
return scale * (input + loc)
def vjp(ys, xs, v):
vJ = torch.autograd.grad(ys, xs, grad_outputs=v, create_graph=True, retain_graph=True, allow_unused=True)
return tuple([j for j in vJ])
class block(nn.Module):
def __init__(self, net, reverse_iterations=40):
super(block, self).__init__()
self.reverse_iterations = reverse_iterations
self.net = net # residual neural network
self.normalize(self.net)
def normalize(self, net):
for n in net.modules():
for k,hook in n._forward_pre_hooks.items():
if isinstance(hook,SpectralNorm):
hook(n, None)
def calcG(self, x):
return self.net(x)
def forward(self, x, logdet=False, reverse=False, num_logdet_iter=1,power_seq_len=10, reverse_iterations=None):
if reverse:
y = x
for count in range(reverse_iterations if reverse_iterations else self.reverse_iterations):
x = y - self.calcG(x)
return x
else:
if logdet:
g = self.calcG(x)
y = g + x
temp_training = self.training
self.eval()
logdet = 0
for i in range(0,num_logdet_iter):
v = y.detach().clone().normal_()
w = v
for k in range(1,power_seq_len):
w = vjp(g,x,w)[0]
logdet += (-1)**(k+1) * torch.dot(w.flatten(), v.flatten()) / k
logdet /= num_logdet_iter
if temp_training:
self.train()
return y, logdet/ y.shape[0]
else:
y = self.calcG(x) + x
return y
def apply_module_reverse(module, x, reverse_iterations=None):
if 'block' in str(module.__class__):
return module(x, reverse=True, reverse_iterations=reverse_iterations)
else:
return module(x, reverse=True)
class InvertibleResnetLinear(nn.Module):
def __init__(self, dim, num_blocks,magnitude=0.7, hidden_dim=600, reverse_iterations=40, bias=False, n_power_iterations=5):
super(InvertibleResnetLinear, self).__init__()
l = []
self.num_blocks = num_blocks
self.dim = dim
self.reverse_iterations = reverse_iterations
for i in range(0,num_blocks):
net = nn.Sequential(nn.ELU(),
spectral_norm(nn.Linear(dim, hidden_dim, bias=bias), n_power_iterations=n_power_iterations, magnitude=magnitude),
nn.ELU(),
spectral_norm(nn.Linear(hidden_dim,hidden_dim, bias=bias), n_power_iterations=n_power_iterations, magnitude=magnitude),
nn.ELU(),
spectral_norm(nn.Linear(hidden_dim,hidden_dim, bias=bias), n_power_iterations=n_power_iterations, magnitude=magnitude),
nn.ELU(),
spectral_norm(nn.Linear(hidden_dim,hidden_dim, bias=bias), n_power_iterations=n_power_iterations, magnitude=magnitude),
nn.ELU(),
spectral_norm(nn.Linear(hidden_dim, dim, bias=bias), n_power_iterations=n_power_iterations, magnitude=magnitude))
b = block(net, reverse_iterations=reverse_iterations)
l.append(b)
l.append(ActNorm(dim))
self.net = nn.Sequential(*l)
def forward(self, x, return_logdet=False, reverse=False, num_logdet_iter=1, reverse_iterations=None):
if reverse:
for module in self.net[::-1]:
x = apply_module_reverse(module, x, reverse_iterations)
return x
else:
if return_logdet:
logdet = 0
x.requires_grad =True
for module in self.net:
if 'block' in str(module.__class__):
x, dlogdet = module(x, logdet=True, reverse=False, num_logdet_iter=num_logdet_iter)
else:
x, dlogdet = module(x, logdet=True, reverse=False)
logdet += dlogdet
return x, logdet
else:
return self.net(x)
class SqueezeLayer(nn.Module):
def __init__(self):
super(SqueezeLayer, self).__init__()
def forward(self, x, reverse=False, logdet=None):
assert len(x.shape) == 4, 'Input must be 4dim, currently {}'.format(x.shape)
if reverse:
ret = x.view(x.shape[0], x.shape[1]//2//2, 2, 2, x.shape[2], x.shape[3]).permute(0,1,4,2,5,3).contiguous().view(x.shape[0], x.shape[1]//2//2, x.shape[2] * 2, x.shape[3] * 2)
else:
ret = x.view(x.shape[0], x.shape[1], x.shape[2] // 2, 2, x.shape[3] // 2, 2).permute(0,1,3,5,2,4).contiguous().view(x.shape[0], x.shape[1]*2*2, x.shape[2] // 2, x.shape[3] // 2)
if logdet:
return ret, 0
else:
return ret
class InvertibleResnetConv(nn.Module):
def __init__(self, dim, hidden_dim = 32, list_num_blocks=(32,32,32),magnitude=0.7, reverse_iterations=10, bias=False, n_power_iterations=5):
super(InvertibleResnetConv, self).__init__()
self.dim = dim
self.reverse_iterations = reverse_iterations
self.nets = nn.ModuleList()
for num_blocks in list_num_blocks:
l = nn.ModuleList()
for i in range(0,num_blocks):
net = nn.Sequential(nn.ELU(),
spectral_norm(nn.Conv2d(dim, hidden_dim, 3, padding=1, bias=bias), n_power_iterations=n_power_iterations, magnitude=magnitude),
nn.ELU(),
spectral_norm(nn.Conv2d(hidden_dim, hidden_dim, 1, bias=bias), n_power_iterations=n_power_iterations, magnitude=magnitude),
nn.ELU(),
spectral_norm(nn.Conv2d(hidden_dim, dim, 3, padding=1, bias=bias), n_power_iterations=n_power_iterations, magnitude=magnitude),
)
b = block(net, reverse_iterations=reverse_iterations)
l.append(ActNorm(dim))
l.append(b)
l.append(SqueezeLayer())
dim *= 2
self.nets.append(nn.Sequential(*l))
def forward(self, x_list, return_logdet=False, reverse=False, num_logdet_iter=1,reverse_iterations=None):
if reverse:
for i, net in enumerate(self.nets[::-1]):
if i == 0:
x = x_list[len(self.nets)-1-i]
else:
x = torch.cat([x, x_list[len(self.nets)-1-i]], dim=1)
for module in net[::-1]:
x = apply_module_reverse(module, x, reverse_iterations)
return x
else:
logdet = 0
y_list = []
x = x_list
if return_logdet:
x.requires_grad =True
for i, net in enumerate(self.nets):
if return_logdet:
for module in net:
x, dlogdet = run_module_with_logdet(module, x, num_logdet_iter)
logdet += dlogdet
else:
x = net(x)
if i < len(self.nets)-1:
y_list.append(x[:,x.shape[1]//2:])
x = x[:,:x.shape[1]//2]
y_list.append(x)
if return_logdet:
return y_list, logdet
else:
return y_list