-
Notifications
You must be signed in to change notification settings - Fork 72
/
Network.py
237 lines (182 loc) · 8.36 KB
/
Network.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
import torch
import torch.nn as nn
from Modules import *
from GST import GST
from Hyperparameters import Hyperparameters as hp
class Tacotron(nn.Module):
'''
input:
texts: [N, T_x]
mels: [N, T_y/r, n_mels*r]
output:
mels --- [N, T_y/r, n_mels*r]
mags --- [N, T_y, 1+n_fft//2]
attn_weights --- [N, T_y/r, T_x]
'''
def __init__(self):
super().__init__()
self.embedding = nn.Embedding(len(hp.vocab), hp.E)
self.encoder = Encoder()
self.decoder = Decoder()
self.gst = GST()
def forward(self, texts, mels, ref_mels):
embedded = self.embedding(texts) # [N, T_x, E]
memory, encoder_hidden = self.encoder(embedded) # [N, T_x, E]
style_embed = self.gst(ref_mels) # [N, 256]
style_embed = style_embed.expand_as(memory)
memory = memory + style_embed
mels_hat, mags_hat, attn_weights = self.decoder(mels, memory)
return mels_hat, mags_hat, attn_weights
class Encoder(nn.Module):
'''
input:
inputs: [N, T_x, E]
output:
outputs: [N, T_x, E]
hidden: [2, N, E//2]
'''
def __init__(self):
super().__init__()
self.prenet = PreNet(in_features=hp.E) # [N, T, E//2]
self.conv1d_bank = Conv1dBank(K=hp.K, in_channels=hp.E // 2, out_channels=hp.E // 2) # [N, T, E//2 * K]
self.conv1d_1 = Conv1d(in_channels=hp.K * hp.E // 2, out_channels=hp.E // 2, kernel_size=3) # [N, T, E//2]
self.conv1d_2 = Conv1d(in_channels=hp.E // 2, out_channels=hp.E // 2, kernel_size=3) # [N, T, E//2]
self.bn1 = BatchNorm1d(num_features=hp.E // 2)
self.bn2 = BatchNorm1d(num_features=hp.E // 2)
self.highways = nn.ModuleList()
for i in range(hp.num_highways):
self.highways.append(Highway(in_features=hp.E // 2, out_features=hp.E // 2))
self.gru = nn.GRU(input_size=hp.E // 2, hidden_size=hp.E // 2, num_layers=2, bidirectional=True, batch_first=True)
def forward(self, inputs, prev_hidden=None):
# prenet
inputs = self.prenet(inputs) # [N, T, E//2]
# CBHG
# conv1d bank
outputs = self.conv1d_bank(inputs) # [N, T, E//2 * K]
outputs = max_pool1d(outputs, kernel_size=2) # [N, T, E//2 * K]
# conv1d projections
outputs = self.conv1d_1(outputs) # [N, T, E//2]
outputs = self.bn1(outputs)
outputs = nn.functional.relu(outputs) # [N, T, E//2]
outputs = self.conv1d_2(outputs) # [N, T, E//2]
outputs = self.bn2(outputs)
outputs = outputs + inputs # residual connect
# highway
for layer in self.highways:
outputs = layer(outputs)
# outputs = nn.functional.relu(outputs) # [N, T, E//2]
# outputs = torch.transpose(outputs, 0, 1) # [T, N, E//2]
self.gru.flatten_parameters()
outputs, hidden = self.gru(outputs, prev_hidden) # outputs [N, T, E]
return outputs, hidden
class Decoder(nn.Module):
'''
input:
inputs --- [N, T_y/r, n_mels * r]
memory --- [N, T_x, E]
output:
mels --- [N, T_y/r, n_mels*r]
mags --- [N, T_y, 1+n_fft//2]
attn_weights --- [N, T_y/r, T_x]
'''
def __init__(self):
super().__init__()
self.prenet = PreNet(hp.n_mels)
self.attn_rnn = AttentionRNN()
self.attn_projection = nn.Linear(in_features=2 * hp.E, out_features=hp.E)
self.gru1 = nn.GRU(input_size=hp.E, hidden_size=hp.E, batch_first=True, bidirectional=False)
self.gru2 = nn.GRU(input_size=hp.E, hidden_size=hp.E, batch_first=True, bidirectional=False)
self.fc1 = nn.Linear(in_features=hp.E, out_features=hp.n_mels * hp.r)
self.cbhg = DecoderCBHG() # Deng
self.fc2 = nn.Linear(in_features=hp.E, out_features=1 + hp.n_fft // 2) # Deng
def forward(self, inputs, memory):
if self.training:
# prenet
outputs = self.prenet(inputs) # [N, T_y/r, E//2]
attn_weights, outputs, attn_hidden = self.attn_rnn(outputs, memory)
attn_apply = torch.bmm(attn_weights, memory) # [N, T_y/r, E]
attn_project = self.attn_projection(torch.cat([attn_apply, outputs], dim=2)) # [N, T_y/r, E]
# GRU1
self.gru1.flatten_parameters()
outputs1, gru1_hidden = self.gru1(attn_project) # outputs1--[N, T_y/r, E] gru1_hidden--[1, N, E]
gru_outputs1 = outputs1 + attn_project # [N, T_y/r, E]
# GRU2
self.gru2.flatten_parameters()
outputs2, gru2_hidden = self.gru2(gru_outputs1) # outputs2--[N, T_y/r, E] gru2_hidden--[1, N, E]
gru_outputs2 = outputs2 + gru_outputs1
# generate log melspectrogram
mels = self.fc1(gru_outputs2) # [N, T_y/r, n_mels*r]
# CBHG
out, cbhg_hidden = self.cbhg(mels) # out -- [N, T_y, E]
# generate linear spectrogram
mags = self.fc2(out) # out -- [N, T_y, 1+n_fft//2]
return mels, mags, attn_weights
else:
# inputs = Go_frame [1, 1, n_mels*r]
attn_hidden = None
gru1_hidden = None
gru2_hidden = None
mels = []
mags = []
attn_weights = []
for i in range(hp.max_Ty):
inputs = self.prenet(inputs)
attn_weight, outputs, attn_hidden = self.attn_rnn(inputs, memory, attn_hidden)
attn_weights.append(attn_weight) # attn_weight: [1, 1, T_x]
attn_apply = torch.bmm(attn_weight, memory) # [1, 1, E]
attn_project = self.attn_projection(torch.cat([attn_apply, outputs], dim=-1)) # [1, 1, E]
# GRU1
self.gru1.flatten_parameters()
outputs1, gru1_hidden = self.gru1(attn_project, gru1_hidden) # outputs1--[1, 1, E] gru1_hidden--[1, 1, E]
outputs1 = outputs1 + attn_project # [1, T_y/r, E]
# GRU2
self.gru2.flatten_parameters()
outputs2, gru2_hidden = self.gru2(outputs1, gru2_hidden) # outputs2--[1, T_y/r, E] gru2_hidden--[1, 1, E]
outputs2 = outputs2 + outputs1
# generate log melspectrogram
mel = self.fc1(outputs2) # [1, 1, n_mels*r]
inputs = mel[:, :, -hp.n_mels:] # get last frame
mels.append(mel)
mels = torch.cat(mels, dim=1) # [1, max_iter, n_mels*r]
attn_weights = torch.cat(attn_weights, dim=1) # [1, T, T_x]
out, cbhg_hidden = self.cbhg(mels)
mags = self.fc2(out)
return mels, mags, attn_weights
class DecoderCBHG(nn.Module):
'''
input:
inputs: [N, T/r, n_mels * r]
output:
outputs: [N, T, E]
hidden: [2, N, E//2]
'''
def __init__(self):
super().__init__()
self.conv1d_bank = Conv1dBank(K=hp.decoder_K, in_channels=hp.n_mels, out_channels=hp.E // 2)
self.conv1d_1 = Conv1d(in_channels=hp.decoder_K * hp.E // 2, out_channels=hp.E, kernel_size=3)
self.bn1 = BatchNorm1d(hp.E)
self.conv1d_2 = Conv1d(in_channels=hp.E, out_channels=hp.n_mels, kernel_size=3)
self.bn2 = BatchNorm1d(hp.n_mels)
self.highways = nn.ModuleList()
for i in range(hp.num_highways):
self.highways.append(Highway(in_features=hp.n_mels, out_features=hp.n_mels))
self.gru = nn.GRU(input_size=hp.n_mels, hidden_size=hp.E // 2, num_layers=2, bidirectional=True, batch_first=True)
def forward(self, inputs, prev_hidden=None):
inputs = inputs.view(inputs.size(0), -1, hp.n_mels) # [N, T, n_mels]
# conv1d bank
outputs = self.conv1d_bank(inputs) # [N, T, E//2 * K]
outputs = max_pool1d(outputs, kernel_size=2)
# conv1d projections
outputs = self.conv1d_1(outputs) # [N, T, E]
outputs = self.bn1(outputs)
outputs = nn.functional.relu(outputs)
outputs = self.conv1d_2(outputs) # [N, T, n_mels]
outputs = self.bn2(outputs)
outputs = outputs + inputs # residual connect [N, T, n_mels]
# highway net
for layer in self.highways:
outputs = layer(outputs) # [N, T, n_mels]
# bidirection gru
self.gru.flatten_parameters()
outputs, hidden = self.gru(outputs, prev_hidden) # outputs: [N, T, E]
return outputs, hidden