-
Notifications
You must be signed in to change notification settings - Fork 1
/
layers.py
162 lines (138 loc) · 5.87 KB
/
layers.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
import torch
from librosa.filters import mel as librosa_mel_fn
from audio_processing import dynamic_range_compression
from audio_processing import dynamic_range_decompression
from stft import STFT
from ge2e_hparams import hparams
from torch import nn
from torch.nn import init
import torch.functional as F
class LinearNorm(torch.nn.Module):
def __init__(self, in_dim, out_dim, bias=True, w_init_gain='linear'):
super(LinearNorm, self).__init__()
self.linear_layer = torch.nn.Linear(in_dim, out_dim, bias=bias)
torch.nn.init.xavier_uniform_(
self.linear_layer.weight,
gain=torch.nn.init.calculate_gain(w_init_gain))
def forward(self, x):
return self.linear_layer(x)
class ConvNorm(torch.nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=1, stride=1,
padding=None, dilation=1, bias=True, w_init_gain='linear'):
super(ConvNorm, self).__init__()
if padding is None:
assert (kernel_size % 2 == 1)
padding = int(dilation * (kernel_size - 1) / 2)
self.conv = torch.nn.Conv1d(in_channels, out_channels,
kernel_size=kernel_size, stride=stride,
padding=padding, dilation=dilation,
bias=bias)
torch.nn.init.xavier_uniform_(
self.conv.weight, gain=torch.nn.init.calculate_gain(w_init_gain))
def forward(self, signal):
conv_signal = self.conv(signal)
return conv_signal
class TacotronSTFT(torch.nn.Module):
def __init__(self, filter_length=1024, hop_length=256, win_length=1024,
n_mel_channels=80, sampling_rate=22050, mel_fmin=0.0,
mel_fmax=None):
super(TacotronSTFT, self).__init__()
self.n_mel_channels = n_mel_channels
self.sampling_rate = sampling_rate
self.stft_fn = STFT(filter_length, hop_length, win_length)
mel_basis = librosa_mel_fn(
sampling_rate, filter_length, n_mel_channels, mel_fmin, mel_fmax)
mel_basis = torch.from_numpy(mel_basis).float()
self.register_buffer('mel_basis', mel_basis)
def spectral_normalize(self, magnitudes):
output = dynamic_range_compression(magnitudes)
return output
def spectral_de_normalize(self, magnitudes):
output = dynamic_range_decompression(magnitudes)
return output
def mel_spectrogram(self, y):
"""Computes mel-spectrograms from a batch of waves
PARAMS
------
y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
RETURNS
-------
mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
"""
assert (torch.min(y.data) >= -1)
assert (torch.max(y.data) <= 1)
magnitudes, phases = self.stft_fn.transform(y)
magnitudes = magnitudes.data
mel_output = torch.matmul(self.mel_basis, magnitudes)
mel_output = self.spectral_normalize(mel_output)
return mel_output
class SpeakerEncoder(nn.Module):
def __init__(self, input_size, n=hparams.N, m=hparams.M, hidden_size=768, project_size=256):
super(SpeakerEncoder, self).__init__()
self.w = nn.Parameter(torch.tensor(10.0))
self.b = nn.Parameter(torch.tensor(-5.0))
self.N = n
self.M = m
if hparams.mode == 'TD-SV':
hidden_size = hparams.hidden_size_tdsv
project_size = hparams.project_size_tdsv
else:
hidden_size = hparams.hidden_size_tisv
project_size = hparams.project_size_tisv
self.lstm1 = nn.LSTM(input_size=input_size, hidden_size=hidden_size, dropout=0.5,
batch_first=False)
self.project1 = nn.Linear(hidden_size, project_size)
self.lstm2 = nn.LSTM(input_size=project_size, hidden_size=hidden_size, dropout=0.5,
batch_first=False)
self.project2 = nn.Linear(hidden_size, project_size)
self.lstm3 = nn.LSTM(input_size=project_size, hidden_size=hidden_size, dropout=0.5,
batch_first=False)
self.project3 = nn.Linear(hidden_size, project_size)
# self.bn1 = nn.BatchNorm1d(hidden_size)
self.init()
def init_lstm(self, lstm):
for layer in lstm.all_weights:
for p in layer:
if len(p.size()) >= 2:
init.orthogonal_(p)
def init(self):
self.init_lstm(self.lstm1)
self.init_lstm(self.lstm2)
self.init_lstm(self.lstm3)
init.normal_(self.project1.weight.data, 0, 0.02)
init.normal_(self.project2.weight.data, 0, 0.02)
init.normal_(self.project3.weight.data, 0, 0.02)
def similarity_matrix(self, x):
N, M = self.N, self.M
# x [N*M,d] B=N*M,d is a vector
yy = x.unsqueeze(0).repeat(N, 1, 1)
c = torch.stack(x.split([M] * N), 0).mean(1, keepdim=True)
cc = c.repeat(1, M * N, 1)
cc = cc.permute(1, 0, 2)
yy = yy.permute(1, 0, 2)
sim = F.cosine_similarity(cc, yy, dim=-1)
similarity = self.w * sim + self.b
return similarity
def forward(self, x, return_sim=True):
x, (h1, c1) = self.lstm1(x)
x = x.permute(1, 0, 2)
x = self.project1(x)
x = x.permute(1, 0, 2)
x, (h2, c2) = self.lstm2(x)
x = x.permute(1, 0, 2)
x = self.project2(x)
x = x.permute(1, 0, 2)
x, (h3, c3) = self.lstm3(x)
x = x.permute(1, 0, 2)
x = self.project3(x)
x = x.permute(1, 0, 2)
x = x[-1, :, :]
# l2 norm
x = x / torch.norm(x)
if not return_sim:
return x, None
sim = self.similarity_matrix(x)
return x, sim
def load_model(self, checkpoint_path):
checkpoint = torch.load(checkpoint_path)
self.load_state_dict(checkpoint['state_dict'])