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
|
from functools import partial
from typing import List
import chainer
from .config import DiscriminatorModelConfig
from .config import ModelConfig
class Convolution1D(chainer.links.ConvolutionND):
def __init__(self, in_channels, out_channels, ksize, stride=1, pad=0,
nobias=False, initialW=None, initial_bias=None,
cover_all=False):
super().__init__(
ndim=1,
in_channels=in_channels,
out_channels=out_channels,
ksize=ksize,
stride=stride,
pad=pad,
nobias=nobias,
initialW=initialW,
initial_bias=initial_bias,
cover_all=cover_all,
)
class LegacyConvolution1D(chainer.links.Convolution2D):
def __init__(self, in_channels, out_channels, ksize=None, stride=1, pad=0,
nobias=False, initialW=None, initial_bias=None, **kwargs):
assert ksize is None or isinstance(ksize, int)
assert isinstance(stride, int)
assert isinstance(pad, int)
super().__init__(
in_channels=in_channels,
out_channels=out_channels,
ksize=(ksize, 1),
stride=(stride, 1),
pad=(pad, 0),
nobias=nobias,
initialW=initialW,
initial_bias=initial_bias,
**kwargs,
)
def __call__(self, x):
assert x.shape[-1] == 1
return super().__call__(x)
class ConvHighway(chainer.link.Chain):
def __init__(self, in_out_size, nobias=False, activate=chainer.functions.relu,
init_Wh=None, init_Wt=None, init_bh=None, init_bt=-1):
super().__init__()
self.activate = activate
with self.init_scope():
self.plain = Convolution1D(
in_out_size, in_out_size, 1, nobias=nobias,
initialW=init_Wh, initial_bias=init_bh)
self.transform = Convolution1D(
in_out_size, in_out_size, 1, nobias=nobias,
initialW=init_Wt, initial_bias=init_bt)
def __call__(self, x):
out_plain = self.activate(self.plain(x))
out_transform = chainer.functions.sigmoid(self.transform(x))
y = out_plain * out_transform + x * (1 - out_transform)
return y
class PreNet(chainer.link.Chain):
def __init__(self, in_channels: int, hidden_channels: int, out_channels: int):
super().__init__()
with self.init_scope():
self.conv1 = Convolution1D(in_channels, hidden_channels, 1)
self.conv2 = Convolution1D(hidden_channels, out_channels, 1)
def __call__(self, x):
h = x
h = chainer.functions.dropout((chainer.functions.relu(self.conv1(h)), 0.5))
h = chainer.functions.dropout((chainer.functions.relu(self.conv2(h)), 0.5))
return h
class Conv1DBank(chainer.link.Chain):
def __init__(self, in_channels: int, out_channels: int, k: int):
super().__init__()
self.stacked_channels = out_channels * k
self.pads = [
partial(chainer.functions.pad, pad_width=((0, 0), (0, 0), (i // 2, (i + 1) // 2)), mode='constant')
for i in range(k)
]
with self.init_scope():
self.convs = chainer.link.ChainList(
*(Convolution1D(in_channels, out_channels, i + 1, nobias=True) for i in range(k))
)
self.bn = chainer.links.BatchNormalization(out_channels * k)
def __call__(self, x):
h = x
h = chainer.functions.concat([conv(pad(h)) for pad, conv in zip(self.pads, self.convs)])
h = chainer.functions.relu(self.bn(h))
return h
class Conv1DProjections(chainer.link.Chain):
def __init__(self, in_channels: int, hidden_channels: int, out_channels: int):
super().__init__()
with self.init_scope():
self.conv1 = Convolution1D(in_channels, hidden_channels, 3, pad=1, nobias=True)
self.bn1 = chainer.links.BatchNormalization(hidden_channels)
self.conv2 = Convolution1D(hidden_channels, out_channels, 3, pad=1, nobias=True)
self.bn2 = chainer.links.BatchNormalization(out_channels)
def __call__(self, x):
h = x
h = chainer.functions.relu(self.bn1(self.conv1(h)))
h = chainer.functions.relu(self.bn2(self.conv2(h)))
return h
class CBHG(chainer.link.Chain):
def __init__(
self,
in_channels: int,
conv_bank_out_channels: int,
conv_bank_k: int,
max_pooling_k: int,
conv_projections_hidden_channels: int,
highway_layers: int,
out_channels: int,
disable_last_rnn: bool,
):
super().__init__()
self.max_pooling_padding = partial(
chainer.functions.pad,
pad_width=((0, 0), (0, 0), ((max_pooling_k - 1) // 2, max_pooling_k // 2)),
mode='constant',
)
self.max_pooling = chainer.functions.MaxPoolingND(1, max_pooling_k, 1, cover_all=False)
self.out_size = out_channels * (1 if disable_last_rnn else 2)
with self.init_scope():
self.conv_bank = Conv1DBank(
in_channels=in_channels,
out_channels=conv_bank_out_channels,
k=conv_bank_k,
)
self.conv_projectoins = Conv1DProjections(
in_channels=self.conv_bank.stacked_channels,
hidden_channels=conv_projections_hidden_channels,
out_channels=out_channels,
)
self.highways = chainer.link.ChainList(
*([ConvHighway(out_channels) for _ in range(highway_layers)])
)
if not disable_last_rnn:
self.gru = chainer.links.NStepBiGRU(
n_layers=1,
in_size=out_channels,
out_size=out_channels,
dropout=0.0,
)
def __call__(self, x):
h = x
h = self.conv_bank(h)
h = self.max_pooling(self.max_pooling_padding(h))
h = self.conv_projectoins(h)
h = h + x
for highway in self.highways:
h = highway(h)
if hasattr(self, 'gru'):
h = chainer.functions.separate(chainer.functions.transpose(h, axes=(0, 2, 1)))
_, h = self.gru(None, h)
h = chainer.functions.transpose(chainer.functions.stack(h), axes=(0, 2, 1))
return h
class Predictor(chainer.link.Chain):
def __init__(self, network, out_size: int):
super().__init__()
with self.init_scope():
self.network = network
self.last = Convolution1D(network.out_size, out_size, 1)
def __call__(self, x):
h = x
h = self.network(h)
h = self.last(h)
return h
class Aligner(chainer.link.Chain):
def __init__(self, in_size: int, out_time_length: int):
super().__init__()
with self.init_scope():
self.gru = chainer.links.NStepBiGRU(
n_layers=1,
in_size=in_size,
out_size=in_size // 2,
dropout=0.0,
)
self.last = Convolution1D(in_size // 2 * 2, out_time_length, 1)
def __call__(self, x):
"""
:param x: (batch, channel, timeA)
"""
h = x
h = chainer.functions.separate(chainer.functions.transpose(h, axes=(0, 2, 1))) # h: batch * (timeA, channel)
_, h = self.gru(None, h) # h: batch * (timeA, ?)
h = chainer.functions.transpose(chainer.functions.stack(h), axes=(0, 2, 1)) # h: (batch, ?, timeA)
h = chainer.functions.softmax(self.last(h), axis=1) # h: (batch, timeB, timeA)
h = chainer.functions.matmul(x, h) # h: (batch, channel, time)
return h
class Discriminator(chainer.link.Chain):
def __init__(self, in_channels: int, hidden_channels_list: List[int], last_channels: int):
super().__init__()
with self.init_scope():
self.convs = chainer.link.ChainList(*(
LegacyConvolution1D(i_c, o_c, ksize=2, stride=2)
for i_c, o_c in zip([in_channels] + hidden_channels_list[:-1], hidden_channels_list)
))
self.last_linear = chainer.links.Linear(None, 1)
def __call__(self, x):
"""
:param x: (batch, channel, time)
"""
h = x
h = chainer.functions.reshape(h, h.shape + (1,))
for conv in self.convs.children():
h = chainer.functions.relu(conv(h))
h = chainer.functions.reshape(h, h.shape[:-1])
h = self.last_linear(h)
return h
def create_predictor(config: ModelConfig):
network = CBHG(
in_channels=config.in_channels,
conv_bank_out_channels=config.conv_bank_out_channels,
conv_bank_k=config.conv_bank_k,
max_pooling_k=config.max_pooling_k,
conv_projections_hidden_channels=config.conv_projections_hidden_channels,
highway_layers=config.highway_layers,
out_channels=config.out_channels,
disable_last_rnn=config.disable_last_rnn,
)
predictor = Predictor(
network=network,
out_size=config.out_size,
)
return predictor
def create_aligner(config: ModelConfig):
assert config.enable_aligner
aligner = Aligner(
in_size=config.in_channels,
out_time_length=config.aligner_out_time_length,
)
return aligner
def create_discriminator(config: DiscriminatorModelConfig):
discriminator = Discriminator(
in_channels=config.in_channels,
hidden_channels_list=config.hidden_channels_list,
last_channels=config.last_channels,
)
return discriminator
def create(config: ModelConfig):
predictor = create_predictor(config)
if config.enable_aligner:
aligner = create_aligner(config)
else:
aligner = None
discriminator = create_discriminator(config.discriminator)
return predictor, aligner, discriminator
|
