/
generator.py
272 lines (256 loc) · 12 KB
/
generator.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
from __future__ import print_function
import tensorflow as tf
from tensorflow.contrib.layers import batch_norm, fully_connected, flatten
from tensorflow.contrib.layers import xavier_initializer
from ops import *
import numpy as np
class Generator(object):
def __init__(self, segan):
self.segan = segan
def __call__(self, noisy_w, is_ref, spk=None):
""" Build the graph propagating (noisy_w) --> x
On first pass will make variables.
"""
segan = self.segan
def make_z(shape, mean=0., std=1., name='z'):
if is_ref:
with tf.variable_scope(name) as scope:
z_init = tf.random_normal_initializer(mean=mean, stddev=std)
z = tf.get_variable("z", shape,
initializer=z_init,
trainable=False
)
if z.device != "/device:GPU:0":
# this has to be created into gpu0
print('z.device is {}'.format(z.device))
assert False
else:
z = tf.random_normal(shape, mean=mean, stddev=std,
name=name, dtype=tf.float32)
return z
if hasattr(segan, 'generator_built'):
tf.get_variable_scope().reuse_variables()
make_vars = False
else:
make_vars = True
print('*** Building Generator ***')
in_dims = noisy_w.get_shape().as_list()
h_i = noisy_w
if len(in_dims) == 2:
h_i = tf.expand_dims(noisy_w, -1)
elif len(in_dims) < 2 or len(in_dims) > 3:
raise ValueError('Generator input must be 2-D or 3-D')
kwidth = 3
z = make_z([segan.batch_size, h_i.get_shape().as_list()[1],
segan.g_enc_depths[-1]])
h_i = tf.concat(2, [h_i, z])
skip_out = True
skips = []
for block_idx, dilation in enumerate(segan.g_dilated_blocks):
name = 'g_residual_block_{}'.format(block_idx)
if block_idx >= len(segan.g_dilated_blocks) - 1:
skip_out = False
if skip_out:
res_i, skip_i = residual_block(h_i,
dilation, kwidth, num_kernels=32,
bias_init=None, stddev=0.02,
do_skip = True,
name=name)
else:
res_i = residual_block(h_i,
dilation, kwidth, num_kernels=32,
bias_init=None, stddev=0.02,
do_skip = False,
name=name)
# feed the residual output to the next block
h_i = res_i
if segan.keep_prob < 1:
print('Adding dropout w/ keep prob {} '
'to G'.format(segan.keep_prob))
h_i = tf.nn.dropout(h_i, segan.keep_prob_var)
if skip_out:
# accumulate the skip connections
skips.append(skip_i)
else:
# for last block, the residual output is appended
skips.append(res_i)
print('Amount of skip connections: ', len(skips))
# TODO: last pooling for actual wave
with tf.variable_scope('g_wave_pooling'):
skip_T = tf.stack(skips, axis=0)
skips_sum = tf.reduce_sum(skip_T, axis=0)
skips_sum = leakyrelu(skips_sum)
wave_a = conv1d(skips_sum, kwidth=1, num_kernels=1,
init=tf.truncated_normal_initializer(stddev=0.02))
wave = tf.tanh(wave_a)
segan.gen_wave_summ = histogram_summary('gen_wave', wave)
print('Last residual wave shape: ', res_i.get_shape())
print('*************************')
segan.generator_built = True
return wave, z
class AEGenerator(object):
def __init__(self, segan):
self.segan = segan
def __call__(self, noisy_w, is_ref, spk=None, z_on=True, do_prelu=False):
# TODO: remove c_vec
""" Build the graph propagating (noisy_w) --> x
On first pass will make variables.
"""
segan = self.segan
def make_z(shape, mean=0., std=1., name='z'):
if is_ref:
with tf.variable_scope(name) as scope:
z_init = tf.random_normal_initializer(mean=mean, stddev=std)
z = tf.get_variable("z", shape,
initializer=z_init,
trainable=False
)
if z.device != "/device:GPU:0":
# this has to be created into gpu0
print('z.device is {}'.format(z.device))
assert False
else:
z = tf.random_normal(shape, mean=mean, stddev=std,
name=name, dtype=tf.float32)
return z
if hasattr(segan, 'generator_built'):
tf.get_variable_scope().reuse_variables()
make_vars = False
else:
make_vars = True
if is_ref:
print('*** Building Generator ***')
in_dims = noisy_w.get_shape().as_list()
h_i = noisy_w
if len(in_dims) == 2:
h_i = tf.expand_dims(noisy_w, -1)
elif len(in_dims) < 2 or len(in_dims) > 3:
raise ValueError('Generator input must be 2-D or 3-D')
kwidth = 31
enc_layers = 7
skips = []
if is_ref and do_prelu:
#keep track of prelu activations
alphas = []
with tf.variable_scope('g_ae'):
#AE to be built is shaped:
# enc ~ [16384x1, 8192x16, 4096x32, 2048x32, 1024x64, 512x64, 256x128, 128x128, 64x256, 32x256, 16x512, 8x1024]
# dec ~ [8x2048, 16x1024, 32x512, 64x512, 8x256, 256x256, 512x128, 1024x128, 2048x64, 4096x64, 8192x32, 16384x1]
#FIRST ENCODER
for layer_idx, layer_depth in enumerate(segan.g_enc_depths):
bias_init = None
if segan.bias_downconv:
if is_ref:
print('Biasing downconv in G')
bias_init = tf.constant_initializer(0.)
h_i_dwn = downconv(h_i, layer_depth, kwidth=kwidth,
init=tf.truncated_normal_initializer(stddev=0.02),
bias_init=bias_init,
name='enc_{}'.format(layer_idx))
if is_ref:
print('Downconv {} -> {}'.format(h_i.get_shape(),
h_i_dwn.get_shape()))
h_i = h_i_dwn
if layer_idx < len(segan.g_enc_depths) - 1:
if is_ref:
print('Adding skip connection downconv '
'{}'.format(layer_idx))
# store skip connection
# last one is not stored cause it's the code
skips.append(h_i)
if do_prelu:
if is_ref:
print('-- Enc: prelu activation --')
h_i = prelu(h_i, ref=is_ref, name='enc_prelu_{}'.format(layer_idx))
if is_ref:
# split h_i into its components
alpha_i = h_i[1]
h_i = h_i[0]
alphas.append(alpha_i)
else:
if is_ref:
print('-- Enc: leakyrelu activation --')
h_i = leakyrelu(h_i)
if z_on:
# random code is fused with intermediate representation
z = make_z([segan.batch_size, h_i.get_shape().as_list()[1],
segan.g_enc_depths[-1]])
h_i = tf.concat(2, [z, h_i])
#SECOND DECODER (reverse order)
g_dec_depths = segan.g_enc_depths[:-1][::-1] + [1]
if is_ref:
print('g_dec_depths: ', g_dec_depths)
for layer_idx, layer_depth in enumerate(g_dec_depths):
h_i_dim = h_i.get_shape().as_list()
out_shape = [h_i_dim[0], h_i_dim[1] * 2, layer_depth]
bias_init = None
# deconv
if segan.deconv_type == 'deconv':
if is_ref:
print('-- Transposed deconvolution type --')
if segan.bias_deconv:
print('Biasing deconv in G')
if segan.bias_deconv:
bias_init = tf.constant_initializer(0.)
h_i_dcv = deconv(h_i, out_shape, kwidth=kwidth, dilation=2,
init=tf.truncated_normal_initializer(stddev=0.02),
bias_init=bias_init,
name='dec_{}'.format(layer_idx))
elif segan.deconv_type == 'nn_deconv':
if is_ref:
print('-- NN interpolated deconvolution type --')
if segan.bias_deconv:
print('Biasing deconv in G')
if segan.bias_deconv:
bias_init = 0.
h_i_dcv = nn_deconv(h_i, kwidth=kwidth, dilation=2,
init=tf.truncated_normal_initializer(stddev=0.02),
bias_init=bias_init,
name='dec_{}'.format(layer_idx))
else:
raise ValueError('Unknown deconv type {}'.format(segan.deconv_type))
if is_ref:
print('Deconv {} -> {}'.format(h_i.get_shape(),
h_i_dcv.get_shape()))
h_i = h_i_dcv
if layer_idx < len(g_dec_depths) - 1:
if do_prelu:
if is_ref:
print('-- Dec: prelu activation --')
h_i = prelu(h_i, ref=is_ref,
name='dec_prelu_{}'.format(layer_idx))
if is_ref:
# split h_i into its components
alpha_i = h_i[1]
h_i = h_i[0]
alphas.append(alpha_i)
else:
if is_ref:
print('-- Dec: leakyrelu activation --')
h_i = leakyrelu(h_i)
# fuse skip connection
skip_ = skips[-(layer_idx + 1)]
if is_ref:
print('Fusing skip connection of '
'shape {}'.format(skip_.get_shape()))
h_i = tf.concat(2, [h_i, skip_])
else:
if is_ref:
print('-- Dec: tanh activation --')
h_i = tf.tanh(h_i)
wave = h_i
if is_ref and do_prelu:
print('Amount of alpha vectors: ', len(alphas))
segan.gen_wave_summ = histogram_summary('gen_wave', wave)
if is_ref:
print('Amount of skip connections: ', len(skips))
print('Last wave shape: ', wave.get_shape())
print('*************************')
segan.generator_built = True
# ret feats contains the features refs to be returned
ret_feats = [wave]
if z_on:
ret_feats.append(z)
if is_ref and do_prelu:
ret_feats += alphas
return ret_feats