forked from graphcore/examples
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdeep_voice_model.py
271 lines (214 loc) · 14.6 KB
/
deep_voice_model.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
# Copyright (c) 2020 Graphcore Ltd. All rights reserved.
import numpy as np
import logging_util
from deep_voice_component import Component
# set up logging
logger = logging_util.get_basic_logger(__name__)
class Encoder(Component):
def __init__(self, conf, builder, graph_initial_weights=None, graph_name_to_tensor_map=None):
self.num_symbols = conf.num_symbols
self.encoder_channels = conf.encoder_channels
self.speaker_embedding_dim = conf.speaker_embedding_dim
self.character_embedding_dim = conf.character_embedding_dim
self.num_speakers = conf.num_speakers
self.num_encoder_conv_blocks = conf.num_encoder_conv_blocks
self.dropout_rate = conf.dropout_rate
super(Encoder, self).__init__(conf, builder, conv_type="same",
graph_initial_weights=graph_initial_weights,
graph_name_to_tensor_map=graph_name_to_tensor_map)
def __call__(self, x_text, speaker_embedding, name_to_tensor):
self.name_to_tensor = name_to_tensor
return self.__build_graph(x_text, speaker_embedding)
def __build_graph(self, x_text, speaker_embedding):
logger.info("Building Encoder Graph")
# embedding layer
with self.namescope("text_embedding"):
h_e, character_embedding_matrix = self.embedding(x_text, self.num_symbols, self.character_embedding_dim,
"text_embedding")
if speaker_embedding:
h_e = self.apply_speaker_embedding(h_e, self.character_embedding_dim,
speaker_embedding, self.speaker_embedding_dim,
"encoder_pre_apply_speaker_embedding")
# encoder prenet
with self.namescope("encoder_prenet"):
x = self.temp_distributed_FC(h_e, self.character_embedding_dim, self.encoder_channels,
"encoder_prenet", activation="relu")
with self.namescope("encoder_conv_blocks"):
for block_ind in range(self.num_encoder_conv_blocks):
block_name = "encoder_conv_block_" + str(block_ind)
x = self.gated_residual_conv_block(x, self.encoder_channels, block_name,
speaker_embedding, self.speaker_embedding_dim,
dropout_rate=self.dropout_rate)
# encoder postnet
with self.namescope("encoder_postnet"):
# attention key vectors
h_k = self.temp_distributed_FC(x, self.encoder_channels, self.character_embedding_dim,
"encoder_postnet", activation="relu")
if speaker_embedding:
h_k = self.apply_speaker_embedding(h_k, self.character_embedding_dim,
speaker_embedding, self.speaker_embedding_dim,
"encoder_post_apply_speaker_embedding")
# attention value vectors
h_v = self.builder.aiOnnx.mul([self.get_constant(np.sqrt(0.5)),
self.builder.aiOnnx.add([h_e, h_k])])
return h_k, h_v
class Decoder(Component):
def __init__(self, conf, builder, graph_initial_weights=None, graph_name_to_tensor_map=None, for_inference=False):
self.character_embedding_dim = conf.character_embedding_dim
self.max_text_sequence_length = conf.max_text_sequence_length
self.speaker_embedding_dim = conf.speaker_embedding_dim
self.decoder_channels = conf.decoder_channels
self.num_decoder_conv_blocks = conf.num_decoder_conv_blocks
self.decoder_attention_flags = conf.decoder_attention_flags
self.dropout_rate = conf.dropout_rate
self.decoder_prenet_sizes = conf.decoder_prenet_sizes
self.attention_hidden_size = conf.attention_hidden_size
self.mel_bands = conf.mel_bands
self.n_frames_per_pred = conf.n_frames_per_pred
self.num_speakers = conf.num_speakers
self.for_inference = for_inference
super(Decoder, self).__init__(conf, builder, conv_type="causal",
graph_initial_weights=graph_initial_weights,
graph_name_to_tensor_map=graph_name_to_tensor_map)
def __call__(self, h_k, h_v, x_spectrogram, speaker_embedding, name_to_tensor):
self.name_to_tensor = name_to_tensor
if not self.for_inference:
return self.__build_graph(h_k, h_v, x_spectrogram, speaker_embedding)
else:
return self.__build_graph_for_inference(h_k, h_v, speaker_embedding)
def __build_graph(self, h_k, h_v, x_spectrogram, speaker_embedding):
logger.info("Building Decoder Graph")
# get positional encodings
self.keys_positional_encodings = \
self.get_constant(sinusoidal_position_encoding(self.conf.max_text_sequence_length,
self.conf.character_embedding_dim,
position_rate=self.conf.key_position_rate,
position_weight=1.0))
self.queries_positional_encodings = \
self.get_constant(sinusoidal_position_encoding(self.conf.max_spectrogram_length,
self.conf.decoder_channels,
position_rate=self.conf.query_position_rate,
position_weight=1.0))
with self.namescope("decoder_prenet"):
x = x_spectrogram
for dec_pre_ind, dps in enumerate(self.decoder_prenet_sizes):
if self.dropout_rate > 0.0:
x = self.builder.aiOnnx.dropout([x], 1, self.dropout_rate)[0]
if dec_pre_ind == 0:
x = self.apply_speaker_embedding(x, self.mel_bands,
speaker_embedding, self.speaker_embedding_dim,
"decoder_pre_{}_apply_speaker_embedding".format(dec_pre_ind))
x = self.temp_distributed_FC(x, self.mel_bands, dps,
"decoder_pre_" + str(dec_pre_ind), activation="relu")
else:
x = self.apply_speaker_embedding(x, self.decoder_prenet_sizes[dec_pre_ind-1],
speaker_embedding, self.speaker_embedding_dim,
"decoder_pre_{}_apply_speaker_embedding".format(dec_pre_ind))
x = self.temp_distributed_FC(x, self.decoder_prenet_sizes[dec_pre_ind-1], dps,
"decoder_pre_" + str(dec_pre_ind),
activation="relu")
# list to store attention scores from each block
attention_scores_arrays = []
with self.namescope("decoder_conv_blocks"):
for block_ind in range(self.num_decoder_conv_blocks):
block_name = "deconv_conv_block_" + str(block_ind)
x = self.gated_residual_conv_block(x, self.decoder_channels, block_name,
speaker_embedding, self.speaker_embedding_dim,
dropout_rate=self.dropout_rate)
if self.decoder_attention_flags[block_ind]:
attention_block_name = "decoder_conv_block_attention_" + str(block_ind)
context_vecs, attention_scores = \
self.attention_block(h_k, h_v, x,
self.character_embedding_dim, self.character_embedding_dim,
self.decoder_channels, self.max_text_sequence_length,
self.attention_hidden_size, attention_block_name,
attention_dropout_rate=self.dropout_rate,
keys_positional_encodings=self.keys_positional_encodings,
queries_positional_encodings=self.queries_positional_encodings)
x = self.builder.aiOnnx.mul([self.get_constant(np.sqrt(0.5)),
self.builder.aiOnnx.add([x, context_vecs])])
attention_scores_arrays.append(attention_scores)
hid = x
with self.namescope("decoder_postnet"):
out_size = self.mel_bands * self.n_frames_per_pred
# gated linear unit
x = self.temp_distributed_FC(x, self.decoder_channels, 2 * out_size, "decoder_postnet")
xs1, xs2 = self.builder.aiOnnx.split([x], num_outputs=2, axis=1)
xs2_gated = self.builder.aiOnnx.sigmoid([xs2])
x = self.builder.aiOnnx.mul([xs1, xs2_gated])
x = self.builder.aiOnnx.sigmoid([x])
done_flags = self.temp_distributed_FC(hid, self.decoder_channels, 1,
"decoder_done_block", activation='sigmoid')
return x, attention_scores_arrays, hid, done_flags
def __build_graph_for_inference(self, h_k, h_v, speaker_embedding):
raise NotImplementedError("Autoregressive Inference not implemented yet!")
class Converter(Component):
def __init__(self, conf, builder, graph_initial_weights=None, graph_name_to_tensor_map=None):
super(Converter, self).__init__(conf, builder, conv_type="same",
graph_initial_weights=graph_initial_weights,
graph_name_to_tensor_map=graph_name_to_tensor_map)
self.num_converter_conv_blocks = conf.num_converter_conv_blocks
self.converter_channels = conf.converter_channels
self.n_fft = conf.n_fft
self.n_frames_per_pred = conf.n_frames_per_pred
self.dropout_rate = conf.dropout_rate
def __call__(self, x, name_to_tensor):
self.name_to_tensor = name_to_tensor
return self.__build_graph(x)
def __build_graph(self, x):
logger.info("Building Converter Graph")
with self.namescope("converter"):
for block_ind in range(self.num_converter_conv_blocks):
block_name = "converter_conv_block_" + str(block_ind)
x = self.gated_residual_conv_block(x, self.converter_channels,
block_name, dropout_rate=self.dropout_rate)
x = self.temp_distributed_FC(x, self.converter_channels, self.converter_channels,
"converter_post_1", activation="relu")
x = self.temp_distributed_FC(x, self.converter_channels,
self.n_frames_per_pred * (self.n_fft//2 + 1),
"converter_post_2", activation="sigmoid")
return x
class PopartDeepVoice(Component):
def __init__(self, conf, builder, graph_initial_weights=None, graph_name_to_tensor_map=dict(), for_inference=False):
super(PopartDeepVoice, self).__init__(conf, builder,
graph_initial_weights=graph_initial_weights,
graph_name_to_tensor_map=graph_name_to_tensor_map)
self.for_inference = for_inference
self.encoder = Encoder(conf, builder,
graph_initial_weights=graph_initial_weights,
graph_name_to_tensor_map=graph_name_to_tensor_map)
self.decoder = Decoder(conf, builder,
graph_initial_weights=graph_initial_weights,
graph_name_to_tensor_map=graph_name_to_tensor_map,
for_inference=for_inference)
self.converter = Converter(conf, builder,
graph_initial_weights=graph_initial_weights,
graph_name_to_tensor_map=graph_name_to_tensor_map)
def __call__(self, x_text, x_spectrogram, speaker_id):
self.name_to_tensor = dict()
self.speaker_embedding, speaker_embedding_matrix = self.embedding(speaker_id,
self.conf.num_speakers,
self.conf.speaker_embedding_dim,
"speaker_embedding")
h_k, h_v = self.encoder(x_text, self.speaker_embedding, self.name_to_tensor)
# to build inference graph, x_spectrogram must be set to None
# note that the attention scores arrays format is different for training and inference
decoder_output, attention_scores_arrays, hid, done_flags = self.decoder(h_k, h_v,
x_spectrogram,
self.speaker_embedding,
self.name_to_tensor)
mag_spec_out = self.converter(hid, self.name_to_tensor)
main_outputs = {"mel_spec_output": decoder_output,
"mag_spec_output": mag_spec_out,
"done_flag_output": done_flags}
aux_outputs = {"attention_scores_arrays": attention_scores_arrays,
"speaker_embedding_matrix": speaker_embedding_matrix}
return main_outputs, aux_outputs, self.name_to_tensor
def sinusoidal_position_encoding(num_positions, num_channels, position_rate=1.0, position_weight=1.0):
""" Returns a sinusoidal position encoding table """
position_encoding = np.array([
[position_rate * pos / np.power(10000, 2 * (i // 2) / num_channels) for i in range(num_channels)]
if pos != 0 else np.zeros(num_channels) for pos in range(num_positions)])
position_encoding[:, 0::2] = np.sin(position_encoding[:, 0::2]) # even i
position_encoding[:, 1::2] = np.cos(position_encoding[:, 1::2]) # odd i
return position_weight * position_encoding.T