-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
189 lines (156 loc) · 6.7 KB
/
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
# Copyright 2020 LMNT, Inc. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from math import sqrt
Linear = nn.Linear
ConvTranspose2d = nn.ConvTranspose2d
def Conv1d(*args, **kwargs):
layer = nn.Conv1d(*args, **kwargs)
nn.init.kaiming_normal_(layer.weight)
return layer
@torch.jit.script
def silu(x):
return x * torch.sigmoid(x)
class DiffusionEmbedding(nn.Module):
def __init__(self, max_steps):
super().__init__()
self.register_buffer('embedding', self._build_embedding(max_steps), persistent=False)
self.projection1 = Linear(128, 512)
self.projection2 = Linear(512, 512)
def forward(self, diffusion_step):
if diffusion_step.dtype in [torch.int32, torch.int64]:
x = self.embedding[diffusion_step]
else:
x = self._lerp_embedding(diffusion_step)
x = self.projection1(x)
x = silu(x)
x = self.projection2(x)
x = silu(x)
return x
def _lerp_embedding(self, t):
low_idx = torch.floor(t).long()
high_idx = torch.ceil(t).long()
low = self.embedding[low_idx]
high = self.embedding[high_idx]
return low + (high - low) * (t - low_idx)
def _build_embedding(self, max_steps):
steps = torch.arange(max_steps).unsqueeze(1) # [T,1]
dims = torch.arange(64).unsqueeze(0) # [1,64]
table = steps * 10.0**(dims * 4.0 / 63.0) # [T,64]
table = torch.cat([torch.sin(table), torch.cos(table)], dim=1)
return table
class SpectrogramUpsampler(nn.Module):
def __init__(self, n_mels):
super().__init__()
#self.conv1 = ConvTranspose2d(1, 1, [3, 32], stride=[1, 16], padding=[1, 8])
#self.conv2 = ConvTranspose2d(1, 1, [3, 32], stride=[1, 16], padding=[1, 8])
self.conv1 = nn.ConvTranspose2d(1, 1, (3, 12), stride=(1, 6), padding=(1, 3))
self.conv2 = nn.ConvTranspose2d(1, 1, (3, 4), stride=(1, 2), padding=(1, 1))
def forward(self, x):
# [batch size, frequency, time] -> [batch size, 1, frequency, time] add a channel dimension
# [16, 180, 200] -> [16, 1, 180, 200]
x = torch.unsqueeze(x, 1)
# [16, 1, 180, 200] -> [16, 1, 180, 1200]
x = self.conv1(x)
x = F.leaky_relu(x, 0.4)
x = self.conv2(x)
x = F.leaky_relu(x, 0.4)
x = torch.squeeze(x, 1)
return x
class ResidualBlock(nn.Module):
def __init__(self, n_mels, residual_channels, dilation, uncond=False):
'''
:param n_mels: inplanes of conv1x1 for spectrogram conditional
:param residual_channels: audio conv
:param dilation: audio conv dilation
:param uncond: disable spectrogram conditional
'''
super().__init__()
self.dilated_conv = Conv1d(residual_channels, 2 * residual_channels, 3, padding=dilation, dilation=dilation)
self.diffusion_projection = Linear(512, residual_channels)
if not uncond: # conditional model
self.conditioner_projection = Conv1d(n_mels, 2 * residual_channels, 1)
else: # unconditional model
self.conditioner_projection = None
self.output_projection = Conv1d(residual_channels, 2 * residual_channels, 1)
def forward(self, x, diffusion_step, conditioner=None):
assert (conditioner is None and self.conditioner_projection is None) or \
(conditioner is not None and self.conditioner_projection is not None)
# [16, 512] -> [16, 64] -> [16, 64, 1]
diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1)
# x: [16, 64, 2400]
# diffusion_step: [16, 64, 1]
# 因为有一个是1，所以可以应用广播机制
y = x + diffusion_step
if self.conditioner_projection is None: # using a unconditional model
y = self.dilated_conv(y)
else:
#print(conditioner.shape)
conditioner = self.conditioner_projection(conditioner)
#print(self.dilated_conv(y).shape, conditioner.shape)
y = self.dilated_conv(y) + conditioner
gate, filter = torch.chunk(y, 2, dim=1)
y = torch.sigmoid(gate) * torch.tanh(filter)
y = self.output_projection(y)
residual, skip = torch.chunk(y, 2, dim=1)
return (x + residual) / sqrt(2.0), skip
class DiffWave(nn.Module):
def __init__(self, params):
super().__init__()
self.params = params
# define input_projection, class DiffusionEmbedding, spectrogram_upsampler
self.input_projection = Conv1d(1, params.residual_channels, 1)
self.diffusion_embedding = DiffusionEmbedding(len(params.noise_schedule))
if self.params.unconditional: # use unconditional model
self.spectrogram_upsampler = None
else:
self.spectrogram_upsampler = SpectrogramUpsampler(params.n_mels)
# define class residual_layers
self.residual_layers = nn.ModuleList([
ResidualBlock(params.n_mels, params.residual_channels, 2**(i % params.dilation_cycle_length), uncond=params.unconditional)
for i in range(params.residual_layers)
])
# residual_channels -> residual channels
self.skip_projection = Conv1d(params.residual_channels, params.residual_channels, 1)
# residual_channels -> 1
self.output_projection = Conv1d(params.residual_channels, 1, 1)
nn.init.zeros_(self.output_projection.weight)
def forward(self, audio, diffusion_step, spectrogram=None):
assert (spectrogram is None and self.spectrogram_upsampler is None) or \
(spectrogram is not None and self.spectrogram_upsampler is not None)
# [batch size, length] -> [batch size, dimension, length]
# [16, 2400] -> [16, 1, 2400]
x = audio.unsqueeze(1)
# [16, 1, 2400] -> [16, residual channel, 2400] -> [16, 64, 2400]
x = self.input_projection(x)
x = F.relu(x)
# t: [batch size, 1] -> [batch size, dimension]
# [16, 1] -> [16, 512]
diffusion_step = self.diffusion_embedding(diffusion_step)
# use conditional model
if self.spectrogram_upsampler:
spectrogram = self.spectrogram_upsampler(spectrogram)
skip = None
for layer in self.residual_layers:
x, skip_connection = layer(x, diffusion_step, spectrogram)
skip = skip_connection if skip is None else skip_connection + skip
x = skip / sqrt(len(self.residual_layers))
x = self.skip_projection(x)
x = F.relu(x)
x = self.output_projection(x)
return x