-
Notifications
You must be signed in to change notification settings - Fork 0
/
modules_realnvp.py
370 lines (348 loc) · 16 KB
/
modules_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
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
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
# Utilizing Normalizing Flows for Anime Face Generation
#
# Deep Learning Summer 2022 - Final Project
# Hasso-Plattner Institute
#
# Code adapted by Alisher Turubayev, M.Sc. in Digital Health Student
#
# References to algorithms:
# https://arxiv.org/pdf/1605.08803.pdf - RealNVP
# https://arxiv.org/pdf/1511.06434.pdf - DCGAN
#
# Code references:
# https://github.com/ikostrikov/pytorch-flows/,
# https://github.com/pytorch/tutorials/blob/master/beginner_source/dcgan_faces_tutorial.py,
# https://github.com/fmu2/realNVP
#
# All code utilitzed in this project is a property of the respective authors. Code was used in good faith
# for learning purposes and for the completion of the final project. The author of this notice does not
# claim any rights of ownership and/or originality.
#
# Code by Ilya Kostrikov (ikostrikov) and Fangzhou Mu (fmu2) is licensed under MIT License.
# Code by Nathan Inkawhich (inkawich) is licensed under BSD 3-Clause License.
import numpy as np
import torch
import torch.nn as nn
# RealNVP modules by @fmu2 - https://github.com/fmu2/realNVP
#
# While the author (of the final project) generally follows the idea of the RealNVP, some of the aspects were not fully understood.
# Coupled with the lack of time, the decision was made to not comment this file, instead focusing on code clean-up and final report.
# To that end, most of the code here is a direct copy from the @fmu2 implementation, with two major differences:
# 1. The author (of project) removed additive coupling (from the Glow paper, it seems that the model with affine coupling converges
# faster and has lower negative log-likelihood)
# 2. Due to removal of additive coupling, supporting classes ChannlewiseCoupling and CheckerboardCoupling were removed for cleaner code.
class WeightNormConv2d(nn.Module):
def __init__(self, in_dim, out_dim, kernel_size, stride=1, padding=0,
bias=True, weight_norm=True, scale=False):
"""Intializes a Conv2d augmented with weight normalization.
(See torch.nn.utils.weight_norm for detail.)
Args:
in_dim: number of input channels.
out_dim: number of output channels.
kernel_size: size of convolving kernel.
stride: stride of convolution.
padding: zero-padding added to both sides of input.
bias: True if include learnable bias parameters, False otherwise.
weight_norm: True if apply weight normalization, False otherwise.
scale: True if include magnitude parameters, False otherwise.
"""
super(WeightNormConv2d, self).__init__()
if weight_norm:
self.conv = nn.utils.weight_norm(
nn.Conv2d(in_dim, out_dim, kernel_size,
stride=stride, padding=padding, bias=bias))
if not scale:
self.conv.weight_g.data = torch.ones_like(self.conv.weight_g.data)
self.conv.weight_g.requires_grad = False # freeze scaling
else:
self.conv = nn.Conv2d(in_dim, out_dim, kernel_size,
stride=stride, padding=padding, bias=bias)
def forward(self, x):
"""Forward pass.
Args:
x: input tensor.
Returns:
transformed tensor.
"""
return self.conv(x)
class ResidualBlock(nn.Module):
def __init__(self, dim, bottleneck, weight_norm):
"""Initializes a ResidualBlock.
Args:
dim: number of input and output features.
bottleneck: True if use bottleneck, False otherwise.
weight_norm: True if apply weight normalization, False otherwise.
"""
super(ResidualBlock, self).__init__()
self.in_block = nn.Sequential(
nn.BatchNorm2d(dim),
nn.ReLU())
if bottleneck:
self.res_block = nn.Sequential(
WeightNormConv2d(dim, dim, (1, 1), stride=1, padding=0,
bias=False, weight_norm=weight_norm, scale=False),
nn.BatchNorm2d(dim),
nn.ReLU(),
WeightNormConv2d(dim, dim, (3, 3), stride=1, padding=1,
bias=False, weight_norm=weight_norm, scale=False),
nn.BatchNorm2d(dim),
nn.ReLU(),
WeightNormConv2d(dim, dim, (1, 1), stride=1, padding=0,
bias=True, weight_norm=weight_norm, scale=True))
else:
self.res_block = nn.Sequential(
WeightNormConv2d(dim, dim, (3, 3), stride=1, padding=1,
bias=False, weight_norm=weight_norm, scale=False),
nn.BatchNorm2d(dim),
nn.ReLU(),
WeightNormConv2d(dim, dim, (3, 3), stride=1, padding=1,
bias=True, weight_norm=weight_norm, scale=True))
def forward(self, x):
"""Forward pass.
Args:
x: input tensor.
Returns:
transformed tensor.
"""
return x + self.res_block(self.in_block(x))
class ResidualModule(nn.Module):
def __init__(self, in_dim, dim, out_dim,
res_blocks, bottleneck, skip, weight_norm):
"""Initializes a ResidualModule.
Args:
in_dim: number of input features.
dim: number of features in residual blocks.
out_dim: number of output features.
res_blocks: number of residual blocks to use.
bottleneck: True if use bottleneck, False otherwise.
skip: True if use skip architecture, False otherwise.
weight_norm: True if apply weight normalization, False otherwise.
"""
super(ResidualModule, self).__init__()
self.res_blocks = res_blocks
self.skip = skip
if res_blocks > 0:
self.in_block = WeightNormConv2d(in_dim, dim, (3, 3), stride=1,
padding=1, bias=True, weight_norm=weight_norm, scale=False)
self.core_block = nn.ModuleList(
[ResidualBlock(dim, bottleneck, weight_norm)
for _ in range(res_blocks)])
self.out_block = nn.Sequential(
nn.BatchNorm2d(dim),
nn.ReLU(),
WeightNormConv2d(dim, out_dim, (1, 1), stride=1, padding=0,
bias=True, weight_norm=weight_norm, scale=True))
if skip:
self.in_skip = WeightNormConv2d(dim, dim, (1, 1), stride=1,
padding=0, bias=True, weight_norm=weight_norm, scale=True)
self.core_skips = nn.ModuleList(
[WeightNormConv2d(
dim, dim, (1, 1), stride=1, padding=0, bias=True,
weight_norm=weight_norm, scale=True)
for _ in range(res_blocks)])
else:
if bottleneck:
self.block = nn.Sequential(
WeightNormConv2d(in_dim, dim, (1, 1), stride=1, padding=0,
bias=False, weight_norm=weight_norm, scale=False),
nn.BatchNorm2d(dim),
nn.ReLU(),
WeightNormConv2d(dim, dim, (3, 3), stride=1, padding=1,
bias=False, weight_norm=weight_norm, scale=False),
nn.BatchNorm2d(dim),
nn.ReLU(),
WeightNormConv2d(dim, out_dim, (1, 1), stride=1, padding=0,
bias=True, weight_norm=weight_norm, scale=True))
else:
self.block = nn.Sequential(
WeightNormConv2d(in_dim, dim, (3, 3), stride=1, padding=1,
bias=False, weight_norm=weight_norm, scale=False),
nn.BatchNorm2d(dim),
nn.ReLU(),
WeightNormConv2d(dim, out_dim, (3, 3), stride=1, padding=1,
bias=True, weight_norm=weight_norm, scale=True))
def forward(self, x):
"""Forward pass.
Args:
x: input tensor.
Returns:
transformed tensor.
"""
if self.res_blocks > 0:
x = self.in_block(x)
if self.skip:
out = self.in_skip(x)
for i in range(len(self.core_block)):
x = self.core_block[i](x)
if self.skip:
out = out + self.core_skips[i](x)
if self.skip:
x = out
return self.out_block(x)
else:
return self.block(x)
class AbstractCoupling(nn.Module):
def __init__(self, mask_config, hps):
"""Initializes an AbstractCoupling.
Args:
mask_config: mask configuration (see build_mask() for more detail).
hps: the set of hyperparameters.
"""
super(AbstractCoupling, self).__init__()
self.mask_config = mask_config
self.res_blocks = hps.res_blocks
self.bottleneck = hps.bottleneck
self.skip = hps.skip
self.weight_norm = hps.weight_norm
self.coupling_bn = hps.coupling_bn
def build_mask(self, size, config=1.):
"""Builds a binary checkerboard mask.
(Only for constructing masks for checkerboard coupling layers.)
Args:
size: height/width of features.
config: mask configuration that determines which pixels to mask up.
if 1: if 0:
1 0 0 1
0 1 1 0
Returns:
a binary mask (1: pixel on, 0: pixel off).
"""
mask = np.arange(size).reshape(-1, 1) + np.arange(size)
mask = np.mod(config + mask, 2)
mask = mask.reshape(-1, 1, size, size)
return torch.tensor(mask.astype('float32'))
def batch_stat(self, x):
"""Compute (spatial) batch statistics.
Args:
x: input minibatch.
Returns:
batch mean and variance.
"""
mean = torch.mean(x, dim=(0, 2, 3), keepdim=True)
var = torch.mean((x - mean)**2, dim=(0, 2, 3), keepdim=True)
return mean, var
class CheckerboardAffineCoupling(AbstractCoupling):
def __init__(self, in_out_dim, mid_dim, size, mask_config, hps):
"""Initializes a CheckerboardAffineCoupling.
Args:
in_out_dim: number of input and output features.
mid_dim: number of features in residual blocks.
size: height/width of features.
mask_config: mask configuration (see build_mask() for more detail).
hps: the set of hyperparameters.
"""
super(CheckerboardAffineCoupling, self).__init__(mask_config, hps)
try:
self.mask = self.build_mask(size, config = mask_config).cuda()
except AssertionError:
self.mask = self.build_mask(size, config = mask_config)
self.scale = nn.Parameter(torch.zeros(1), requires_grad=True)
self.scale_shift = nn.Parameter(torch.zeros(1), requires_grad=True)
self.in_bn = nn.BatchNorm2d(in_out_dim)
self.block = nn.Sequential( # 1st half of resnet: shift
nn.ReLU(), # 2nd half of resnet: log_rescale
ResidualModule(2*in_out_dim+1, mid_dim, 2*in_out_dim,
self.res_blocks, self.bottleneck, self.skip, self.weight_norm))
self.out_bn = nn.BatchNorm2d(in_out_dim, affine=False)
def forward(self, x, reverse=False):
"""Forward pass.
Args:
x: input tensor.
reverse: True in inference mode, False in sampling mode.
Returns:
transformed tensor and log of diagonal elements of Jacobian.
"""
[B, C, _, _] = list(x.size())
mask = self.mask.repeat(B, 1, 1, 1)
x_ = self.in_bn(x * mask)
x_ = torch.cat((x_, -x_), dim=1)
x_ = torch.cat((x_, mask), dim=1) # 2C+1 channels
(shift, log_rescale) = self.block(x_).split(C, dim=1)
log_rescale = self.scale * torch.tanh(log_rescale) + self.scale_shift
shift = shift * (1. - mask)
log_rescale = log_rescale * (1. - mask)
log_diag_J = log_rescale # See Eq(6) in real NVP
# See Eq(7) and Eq(8) and Section 3.7 in real NVP
if reverse:
if self.coupling_bn:
mean, var = self.out_bn.running_mean, self.out_bn.running_var
mean = mean.reshape(-1, 1, 1, 1).transpose(0, 1)
var = var.reshape(-1, 1, 1, 1).transpose(0, 1)
x = x * torch.exp(0.5 * torch.log(var + 1e-5) * (1. - mask)) \
+ mean * (1. - mask)
x = (x - shift) * torch.exp(-log_rescale)
else:
x = x * torch.exp(log_rescale) + shift
if self.coupling_bn:
if self.training:
_, var = self.batch_stat(x)
else:
var = self.out_bn.running_var
var = var.reshape(-1, 1, 1, 1).transpose(0, 1)
x = self.out_bn(x) * (1. - mask) + x * mask
log_diag_J = log_diag_J - 0.5 * torch.log(var + 1e-5) * (1. - mask)
return x, log_diag_J
class ChannelwiseAffineCoupling(AbstractCoupling):
def __init__(self, in_out_dim, mid_dim, mask_config, hps):
"""Initializes a ChannelwiseAffineCoupling.
Args:
in_out_dim: number of input and output features.
mid_dim: number of features in residual blocks.
mask_config: 1 if change the top half, 0 if change the bottom half.
hps: the set of hyperparameters.
"""
super(ChannelwiseAffineCoupling, self).__init__(mask_config, hps)
self.scale = nn.Parameter(torch.zeros(1), requires_grad=True)
self.scale_shift = nn.Parameter(torch.zeros(1), requires_grad=True)
self.in_bn = nn.BatchNorm2d(in_out_dim//2)
self.block = nn.Sequential( # 1st half of resnet: shift
nn.ReLU(), # 2nd half of resnet: log_rescale
ResidualModule(in_out_dim, mid_dim, in_out_dim,
self.res_blocks, self.bottleneck, self.skip, self.weight_norm))
self.out_bn = nn.BatchNorm2d(in_out_dim//2, affine=False)
def forward(self, x, reverse=False):
"""Forward pass.
Args:
x: input tensor.
reverse: True in inference mode, False in sampling mode.
Returns:
transformed tensor and log of diagonal elements of Jacobian.
"""
[_, C, _, _] = list(x.size())
if self.mask_config:
(on, off) = x.split(C//2, dim=1)
else:
(off, on) = x.split(C//2, dim=1)
off_ = self.in_bn(off)
off_ = torch.cat((off_, -off_), dim=1) # C channels
out = self.block(off_)
(shift, log_rescale) = out.split(C//2, dim=1)
log_rescale = self.scale * torch.tanh(log_rescale) + self.scale_shift
log_diag_J = log_rescale # See Eq(6) in real NVP
# See Eq(7) and Eq(8) and Section 3.7 in real NVP
if reverse:
if self.coupling_bn:
mean, var = self.out_bn.running_mean, self.out_bn.running_var
mean = mean.reshape(-1, 1, 1, 1).transpose(0, 1)
var = var.reshape(-1, 1, 1, 1).transpose(0, 1)
on = on * torch.exp(0.5 * torch.log(var + 1e-5)) + mean
on = (on - shift) * torch.exp(-log_rescale)
else:
on = on * torch.exp(log_rescale) + shift
if self.coupling_bn:
if self.training:
_, var = self.batch_stat(on)
else:
var = self.out_bn.running_var
var = var.reshape(-1, 1, 1, 1).transpose(0, 1)
on = self.out_bn(on)
log_diag_J = log_diag_J - 0.5 * torch.log(var + 1e-5)
if self.mask_config:
x = torch.cat((on, off), dim=1)
log_diag_J = torch.cat((log_diag_J, torch.zeros_like(log_diag_J)),
dim=1)
else:
x = torch.cat((off, on), dim=1)
log_diag_J = torch.cat((torch.zeros_like(log_diag_J), log_diag_J),
dim=1)
return x, log_diag_J