forked from KevinMIN95/StyleSpeech
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathVarianceAdaptor.py
159 lines (129 loc) · 6.28 KB
/
VarianceAdaptor.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
import torch
import torch.nn as nn
from models.Modules import LinearNorm, ConvNorm, get_sinusoid_encoding_table
import utils
class VarianceAdaptor(nn.Module):
""" Variance Adaptor """
def __init__(self, config):
super(VarianceAdaptor, self).__init__()
self.hidden_dim = config.variance_predictor_filter_size
self.predictor_kernel_size = config.variance_predictor_kernel_size
self.embedding_kernel_size = config.variance_embedding_kernel_size
self.dropout = config.variance_dropout
# Duration
self.duration_predictor = VariancePredictor(self.hidden_dim, self.hidden_dim,
self.predictor_kernel_size, dropout=self.dropout)
# Pitch
self.pitch_predictor = VariancePredictor(self.hidden_dim, self.hidden_dim, self.predictor_kernel_size,
dropout=self.dropout)
self.pitch_embedding = VarianceEmbedding(1, self.hidden_dim, self.embedding_kernel_size, self.dropout)
# Energy
self.energy_predictor = VariancePredictor(self.hidden_dim, self.hidden_dim, self.predictor_kernel_size,
dropout=self.dropout)
self.energy_embedding = VarianceEmbedding(1, self.hidden_dim, self.embedding_kernel_size, self.dropout)
# Phoneme
self.ln = nn.LayerNorm(self.hidden_dim)
# Length regulator
self.length_regulator = LengthRegulator(self.hidden_dim, config.max_seq_len)
def forward(self, x, src_mask, mel_len=None, mel_mask=None,
duration_target=None, pitch_target=None, energy_target=None, max_len=None):
# Duration
log_duration_prediction = self.duration_predictor(x, src_mask)
# Pitch & Energy
pitch_prediction = self.pitch_predictor(x, src_mask)
if pitch_target is not None:
pitch_embedding = self.pitch_embedding(pitch_target.unsqueeze(-1))
else:
pitch_embedding = self.pitch_embedding(pitch_prediction.unsqueeze(-1))
energy_prediction = self.energy_predictor(x, src_mask)
if energy_target is not None:
energy_embedding = self.energy_embedding(energy_target.unsqueeze(-1))
else:
energy_embedding = self.energy_embedding(energy_prediction.unsqueeze(-1))
x = self.ln(x) + pitch_embedding + energy_embedding
# Length regulate
if duration_target is not None:
output, pe, mel_len = self.length_regulator(x, duration_target, max_len)
mel_mask = utils.get_mask_from_lengths(mel_len)
else:
duration_rounded = torch.clamp(torch.round(torch.exp(log_duration_prediction)-1.0), min=0)
duration_rounded = duration_rounded.masked_fill(src_mask, 0).long()
output, pe, mel_len = self.length_regulator(x, duration_rounded)
mel_mask = utils.get_mask_from_lengths(mel_len)
# Phoneme-wise positional encoding
output = output + pe
return output, log_duration_prediction, pitch_prediction, energy_prediction, mel_len, mel_mask
class LengthRegulator(nn.Module):
""" Length Regulator """
def __init__(self, hidden_size, max_pos):
super(LengthRegulator, self).__init__()
self.position_enc = nn.Parameter(
get_sinusoid_encoding_table(max_pos+1, hidden_size), requires_grad=False)
def LR(self, x, duration, max_len):
output = list()
position = list()
mel_len = list()
for batch, expand_target in zip(x, duration):
expanded, pos = self.expand(batch, expand_target)
output.append(expanded)
position.append(pos)
mel_len.append(expanded.shape[0])
if max_len is not None:
output = utils.pad(output, max_len)
position = utils.pad(position, max_len)
else:
output = utils.pad(output)
position = utils.pad(position)
return output, position, torch.LongTensor(mel_len).cuda()
def expand(self, batch, predicted):
out = list()
pos = list()
for i, vec in enumerate(batch):
expand_size = predicted[i].item()
out.append(vec.expand(int(expand_size), -1))
pos.append(self.position_enc[:expand_size, :])
out = torch.cat(out, 0)
pos = torch.cat(pos, 0)
return out, pos
def forward(self, x, duration, max_len=None):
output, position, mel_len = self.LR(x, duration, max_len)
return output, position, mel_len
class VariancePredictor(nn.Module):
""" Variance Predictor """
def __init__(self, input_size, filter_size, kernel_size, output_size=1, n_layers=2, dropout=0.5):
super(VariancePredictor, self).__init__()
convs = [ConvNorm(input_size, filter_size, kernel_size)]
for _ in range(n_layers-1):
convs.append(ConvNorm(filter_size, filter_size, kernel_size))
self.convs = nn.ModuleList(convs)
self.lns = nn.ModuleList([nn.LayerNorm(filter_size) for _ in range(n_layers)])
self.linear_layer = nn.Linear(filter_size, output_size)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(dropout)
def forward(self, x, mask):
for conv, ln in zip(self.convs, self.lns):
x = x.transpose(1,2)
x = self.relu(conv(x))
x = x.transpose(1,2)
x = ln(x)
x = self.dropout(x)
out = self.linear_layer(x)
if mask is not None:
out = out.masked_fill(mask.unsqueeze(-1), 0)
return out.squeeze(-1)
class VarianceEmbedding(nn.Module):
""" Variance Embedding """
def __init__(self, input_size, embed_size, kernel_size, dropout):
super(VarianceEmbedding, self).__init__()
self.conv1 = ConvNorm(input_size, embed_size, kernel_size)
self.conv2 = ConvNorm(embed_size, embed_size, kernel_size)
self.fc = LinearNorm(embed_size, embed_size)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(dropout)
def forward(self, x):
x = x.transpose(1,2)
x = self.dropout(self.relu(self.conv1(x)))
x = self.dropout(self.relu(self.conv2(x)))
x = x.transpose(1,2)
out = self.dropout(self.fc(x))
return out