Create models.py
Browse files
Hiformer_Checkpoint_Libri_24khz/models.py
ADDED
@@ -0,0 +1,943 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as F
|
3 |
+
import torch.nn as nn
|
4 |
+
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
|
5 |
+
from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
|
6 |
+
from utils import init_weights, get_padding
|
7 |
+
import numpy as np
|
8 |
+
from stft import TorchSTFT
|
9 |
+
import torchaudio
|
10 |
+
from nnAudio import features
|
11 |
+
from einops import rearrange
|
12 |
+
from norm2d import NormConv2d
|
13 |
+
from utils import get_padding
|
14 |
+
from munch import Munch
|
15 |
+
from conformer import Conformer
|
16 |
+
|
17 |
+
LRELU_SLOPE = 0.1
|
18 |
+
|
19 |
+
|
20 |
+
def get_2d_padding(kernel_size, dilation=(1, 1)):
|
21 |
+
return (
|
22 |
+
((kernel_size[0] - 1) * dilation[0]) // 2,
|
23 |
+
((kernel_size[1] - 1) * dilation[1]) // 2,
|
24 |
+
)
|
25 |
+
|
26 |
+
|
27 |
+
|
28 |
+
class ResBlock1(torch.nn.Module):
|
29 |
+
def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
|
30 |
+
super(ResBlock1, self).__init__()
|
31 |
+
self.h = h
|
32 |
+
self.convs1 = nn.ModuleList([
|
33 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
|
34 |
+
padding=get_padding(kernel_size, dilation[0]))),
|
35 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
|
36 |
+
padding=get_padding(kernel_size, dilation[1]))),
|
37 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
|
38 |
+
padding=get_padding(kernel_size, dilation[2])))
|
39 |
+
])
|
40 |
+
self.convs1.apply(init_weights)
|
41 |
+
|
42 |
+
self.convs2 = nn.ModuleList([
|
43 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
44 |
+
padding=get_padding(kernel_size, 1))),
|
45 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
46 |
+
padding=get_padding(kernel_size, 1))),
|
47 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
48 |
+
padding=get_padding(kernel_size, 1)))
|
49 |
+
])
|
50 |
+
self.convs2.apply(init_weights)
|
51 |
+
|
52 |
+
self.alpha1 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs1))])
|
53 |
+
self.alpha2 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs2))])
|
54 |
+
|
55 |
+
|
56 |
+
def forward(self, x):
|
57 |
+
for c1, c2, a1, a2 in zip(self.convs1, self.convs2, self.alpha1, self.alpha2):
|
58 |
+
xt = x + (1 / a1) * (torch.sin(a1 * x) ** 2) # Snake1D
|
59 |
+
xt = c1(xt)
|
60 |
+
xt = xt + (1 / a2) * (torch.sin(a2 * xt) ** 2) # Snake1D
|
61 |
+
xt = c2(xt)
|
62 |
+
x = xt + x
|
63 |
+
return x
|
64 |
+
|
65 |
+
def remove_weight_norm(self):
|
66 |
+
for l in self.convs1:
|
67 |
+
remove_weight_norm(l)
|
68 |
+
for l in self.convs2:
|
69 |
+
remove_weight_norm(l)
|
70 |
+
|
71 |
+
class ResBlock1_old(torch.nn.Module):
|
72 |
+
def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
|
73 |
+
super(ResBlock1, self).__init__()
|
74 |
+
self.h = h
|
75 |
+
self.convs1 = nn.ModuleList([
|
76 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
|
77 |
+
padding=get_padding(kernel_size, dilation[0]))),
|
78 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
|
79 |
+
padding=get_padding(kernel_size, dilation[1]))),
|
80 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
|
81 |
+
padding=get_padding(kernel_size, dilation[2])))
|
82 |
+
])
|
83 |
+
self.convs1.apply(init_weights)
|
84 |
+
|
85 |
+
self.convs2 = nn.ModuleList([
|
86 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
87 |
+
padding=get_padding(kernel_size, 1))),
|
88 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
89 |
+
padding=get_padding(kernel_size, 1))),
|
90 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
91 |
+
padding=get_padding(kernel_size, 1)))
|
92 |
+
])
|
93 |
+
self.convs2.apply(init_weights)
|
94 |
+
|
95 |
+
def forward(self, x):
|
96 |
+
for c1, c2 in zip(self.convs1, self.convs2):
|
97 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
98 |
+
xt = c1(xt)
|
99 |
+
xt = F.leaky_relu(xt, LRELU_SLOPE)
|
100 |
+
xt = c2(xt)
|
101 |
+
x = xt + x
|
102 |
+
return x
|
103 |
+
|
104 |
+
def remove_weight_norm(self):
|
105 |
+
for l in self.convs1:
|
106 |
+
remove_weight_norm(l)
|
107 |
+
for l in self.convs2:
|
108 |
+
remove_weight_norm(l)
|
109 |
+
|
110 |
+
|
111 |
+
class ResBlock2(torch.nn.Module):
|
112 |
+
def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)):
|
113 |
+
super(ResBlock2, self).__init__()
|
114 |
+
self.h = h
|
115 |
+
self.convs = nn.ModuleList([
|
116 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
|
117 |
+
padding=get_padding(kernel_size, dilation[0]))),
|
118 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
|
119 |
+
padding=get_padding(kernel_size, dilation[1])))
|
120 |
+
])
|
121 |
+
self.convs.apply(init_weights)
|
122 |
+
|
123 |
+
def forward(self, x):
|
124 |
+
for c in self.convs:
|
125 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
126 |
+
xt = c(xt)
|
127 |
+
x = xt + x
|
128 |
+
return x
|
129 |
+
|
130 |
+
def remove_weight_norm(self):
|
131 |
+
for l in self.convs:
|
132 |
+
remove_weight_norm(l)
|
133 |
+
|
134 |
+
|
135 |
+
class SineGen(torch.nn.Module):
|
136 |
+
""" Definition of sine generator
|
137 |
+
SineGen(samp_rate, harmonic_num = 0,
|
138 |
+
sine_amp = 0.1, noise_std = 0.003,
|
139 |
+
voiced_threshold = 0,
|
140 |
+
flag_for_pulse=False)
|
141 |
+
samp_rate: sampling rate in Hz
|
142 |
+
harmonic_num: number of harmonic overtones (default 0)
|
143 |
+
sine_amp: amplitude of sine-wavefrom (default 0.1)
|
144 |
+
noise_std: std of Gaussian noise (default 0.003)
|
145 |
+
voiced_thoreshold: F0 threshold for U/V classification (default 0)
|
146 |
+
flag_for_pulse: this SinGen is used inside PulseGen (default False)
|
147 |
+
Note: when flag_for_pulse is True, the first time step of a voiced
|
148 |
+
segment is always sin(np.pi) or cos(0)
|
149 |
+
"""
|
150 |
+
|
151 |
+
def __init__(self, samp_rate, upsample_scale, harmonic_num=0,
|
152 |
+
sine_amp=0.1, noise_std=0.003,
|
153 |
+
voiced_threshold=0,
|
154 |
+
flag_for_pulse=False):
|
155 |
+
super(SineGen, self).__init__()
|
156 |
+
self.sine_amp = sine_amp
|
157 |
+
self.noise_std = noise_std
|
158 |
+
self.harmonic_num = harmonic_num
|
159 |
+
self.dim = self.harmonic_num + 1
|
160 |
+
self.sampling_rate = samp_rate
|
161 |
+
self.voiced_threshold = voiced_threshold
|
162 |
+
self.flag_for_pulse = flag_for_pulse
|
163 |
+
self.upsample_scale = upsample_scale
|
164 |
+
|
165 |
+
def _f02uv(self, f0):
|
166 |
+
# generate uv signal
|
167 |
+
uv = (f0 > self.voiced_threshold).type(torch.float32)
|
168 |
+
return uv
|
169 |
+
|
170 |
+
def _f02sine(self, f0_values):
|
171 |
+
""" f0_values: (batchsize, length, dim)
|
172 |
+
where dim indicates fundamental tone and overtones
|
173 |
+
"""
|
174 |
+
# convert to F0 in rad. The interger part n can be ignored
|
175 |
+
# because 2 * np.pi * n doesn't affect phase
|
176 |
+
rad_values = (f0_values / self.sampling_rate) % 1
|
177 |
+
|
178 |
+
# initial phase noise (no noise for fundamental component)
|
179 |
+
rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], \
|
180 |
+
device=f0_values.device)
|
181 |
+
rand_ini[:, 0] = 0
|
182 |
+
rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
|
183 |
+
|
184 |
+
# instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad)
|
185 |
+
if not self.flag_for_pulse:
|
186 |
+
# # for normal case
|
187 |
+
|
188 |
+
# # To prevent torch.cumsum numerical overflow,
|
189 |
+
# # it is necessary to add -1 whenever \sum_k=1^n rad_value_k > 1.
|
190 |
+
# # Buffer tmp_over_one_idx indicates the time step to add -1.
|
191 |
+
# # This will not change F0 of sine because (x-1) * 2*pi = x * 2*pi
|
192 |
+
# tmp_over_one = torch.cumsum(rad_values, 1) % 1
|
193 |
+
# tmp_over_one_idx = (padDiff(tmp_over_one)) < 0
|
194 |
+
# cumsum_shift = torch.zeros_like(rad_values)
|
195 |
+
# cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
|
196 |
+
|
197 |
+
# phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
|
198 |
+
rad_values = torch.nn.functional.interpolate(rad_values.transpose(1, 2),
|
199 |
+
scale_factor=1/self.upsample_scale,
|
200 |
+
mode="linear").transpose(1, 2)
|
201 |
+
|
202 |
+
# tmp_over_one = torch.cumsum(rad_values, 1) % 1
|
203 |
+
# tmp_over_one_idx = (padDiff(tmp_over_one)) < 0
|
204 |
+
# cumsum_shift = torch.zeros_like(rad_values)
|
205 |
+
# cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
|
206 |
+
|
207 |
+
phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
|
208 |
+
phase = torch.nn.functional.interpolate(phase.transpose(1, 2) * self.upsample_scale,
|
209 |
+
scale_factor=self.upsample_scale, mode="linear").transpose(1, 2)
|
210 |
+
sines = torch.sin(phase)
|
211 |
+
|
212 |
+
else:
|
213 |
+
# If necessary, make sure that the first time step of every
|
214 |
+
# voiced segments is sin(pi) or cos(0)
|
215 |
+
# This is used for pulse-train generation
|
216 |
+
|
217 |
+
# identify the last time step in unvoiced segments
|
218 |
+
uv = self._f02uv(f0_values)
|
219 |
+
uv_1 = torch.roll(uv, shifts=-1, dims=1)
|
220 |
+
uv_1[:, -1, :] = 1
|
221 |
+
u_loc = (uv < 1) * (uv_1 > 0)
|
222 |
+
|
223 |
+
# get the instantanouse phase
|
224 |
+
tmp_cumsum = torch.cumsum(rad_values, dim=1)
|
225 |
+
# different batch needs to be processed differently
|
226 |
+
for idx in range(f0_values.shape[0]):
|
227 |
+
temp_sum = tmp_cumsum[idx, u_loc[idx, :, 0], :]
|
228 |
+
temp_sum[1:, :] = temp_sum[1:, :] - temp_sum[0:-1, :]
|
229 |
+
# stores the accumulation of i.phase within
|
230 |
+
# each voiced segments
|
231 |
+
tmp_cumsum[idx, :, :] = 0
|
232 |
+
tmp_cumsum[idx, u_loc[idx, :, 0], :] = temp_sum
|
233 |
+
|
234 |
+
# rad_values - tmp_cumsum: remove the accumulation of i.phase
|
235 |
+
# within the previous voiced segment.
|
236 |
+
i_phase = torch.cumsum(rad_values - tmp_cumsum, dim=1)
|
237 |
+
|
238 |
+
# get the sines
|
239 |
+
sines = torch.cos(i_phase * 2 * np.pi)
|
240 |
+
return sines
|
241 |
+
|
242 |
+
def forward(self, f0):
|
243 |
+
""" sine_tensor, uv = forward(f0)
|
244 |
+
input F0: tensor(batchsize=1, length, dim=1)
|
245 |
+
f0 for unvoiced steps should be 0
|
246 |
+
output sine_tensor: tensor(batchsize=1, length, dim)
|
247 |
+
output uv: tensor(batchsize=1, length, 1)
|
248 |
+
"""
|
249 |
+
f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim,
|
250 |
+
device=f0.device)
|
251 |
+
# fundamental component
|
252 |
+
fn = torch.multiply(f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device))
|
253 |
+
|
254 |
+
# generate sine waveforms
|
255 |
+
sine_waves = self._f02sine(fn) * self.sine_amp
|
256 |
+
|
257 |
+
# generate uv signal
|
258 |
+
# uv = torch.ones(f0.shape)
|
259 |
+
# uv = uv * (f0 > self.voiced_threshold)
|
260 |
+
uv = self._f02uv(f0)
|
261 |
+
|
262 |
+
# noise: for unvoiced should be similar to sine_amp
|
263 |
+
# std = self.sine_amp/3 -> max value ~ self.sine_amp
|
264 |
+
# . for voiced regions is self.noise_std
|
265 |
+
noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
|
266 |
+
noise = noise_amp * torch.randn_like(sine_waves)
|
267 |
+
|
268 |
+
# first: set the unvoiced part to 0 by uv
|
269 |
+
# then: additive noise
|
270 |
+
sine_waves = sine_waves * uv + noise
|
271 |
+
return sine_waves, uv, noise
|
272 |
+
|
273 |
+
|
274 |
+
class SourceModuleHnNSF(torch.nn.Module):
|
275 |
+
""" SourceModule for hn-nsf
|
276 |
+
SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
|
277 |
+
add_noise_std=0.003, voiced_threshod=0)
|
278 |
+
sampling_rate: sampling_rate in Hz
|
279 |
+
harmonic_num: number of harmonic above F0 (default: 0)
|
280 |
+
sine_amp: amplitude of sine source signal (default: 0.1)
|
281 |
+
add_noise_std: std of additive Gaussian noise (default: 0.003)
|
282 |
+
note that amplitude of noise in unvoiced is decided
|
283 |
+
by sine_amp
|
284 |
+
voiced_threshold: threhold to set U/V given F0 (default: 0)
|
285 |
+
Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
|
286 |
+
F0_sampled (batchsize, length, 1)
|
287 |
+
Sine_source (batchsize, length, 1)
|
288 |
+
noise_source (batchsize, length 1)
|
289 |
+
uv (batchsize, length, 1)
|
290 |
+
"""
|
291 |
+
|
292 |
+
def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
|
293 |
+
add_noise_std=0.003, voiced_threshod=0):
|
294 |
+
super(SourceModuleHnNSF, self).__init__()
|
295 |
+
|
296 |
+
self.sine_amp = sine_amp
|
297 |
+
self.noise_std = add_noise_std
|
298 |
+
|
299 |
+
# to produce sine waveforms
|
300 |
+
self.l_sin_gen = SineGen(sampling_rate, upsample_scale, harmonic_num,
|
301 |
+
sine_amp, add_noise_std, voiced_threshod)
|
302 |
+
|
303 |
+
# to merge source harmonics into a single excitation
|
304 |
+
self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
|
305 |
+
self.l_tanh = torch.nn.Tanh()
|
306 |
+
|
307 |
+
def forward(self, x):
|
308 |
+
"""
|
309 |
+
Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
|
310 |
+
F0_sampled (batchsize, length, 1)
|
311 |
+
Sine_source (batchsize, length, 1)
|
312 |
+
noise_source (batchsize, length 1)
|
313 |
+
"""
|
314 |
+
# source for harmonic branch
|
315 |
+
with torch.no_grad():
|
316 |
+
sine_wavs, uv, _ = self.l_sin_gen(x)
|
317 |
+
sine_merge = self.l_tanh(self.l_linear(sine_wavs))
|
318 |
+
|
319 |
+
# source for noise branch, in the same shape as uv
|
320 |
+
noise = torch.randn_like(uv) * self.sine_amp / 3
|
321 |
+
return sine_merge, noise, uv
|
322 |
+
def padDiff(x):
|
323 |
+
return F.pad(F.pad(x, (0,0,-1,1), 'constant', 0) - x, (0,0,0,-1), 'constant', 0)
|
324 |
+
|
325 |
+
|
326 |
+
|
327 |
+
class Generator(torch.nn.Module):
|
328 |
+
def __init__(self, h, F0_model):
|
329 |
+
super(Generator, self).__init__()
|
330 |
+
self.h = h
|
331 |
+
self.num_kernels = len(h.resblock_kernel_sizes)
|
332 |
+
self.num_upsamples = len(h.upsample_rates)
|
333 |
+
self.conv_pre = weight_norm(Conv1d(80, h.upsample_initial_channel, 7, 1, padding=3))
|
334 |
+
|
335 |
+
|
336 |
+
|
337 |
+
resblock = ResBlock1 if h.resblock == '1' else ResBlock2
|
338 |
+
|
339 |
+
self.m_source = SourceModuleHnNSF(
|
340 |
+
sampling_rate=h.sampling_rate,
|
341 |
+
upsample_scale=np.prod(h.upsample_rates) * h.gen_istft_hop_size,
|
342 |
+
harmonic_num=8, voiced_threshod=10)
|
343 |
+
|
344 |
+
self.f0_upsamp = torch.nn.Upsample(
|
345 |
+
scale_factor=np.prod(h.upsample_rates) * h.gen_istft_hop_size)
|
346 |
+
self.noise_convs = nn.ModuleList()
|
347 |
+
self.noise_res = nn.ModuleList()
|
348 |
+
|
349 |
+
self.F0_model = F0_model
|
350 |
+
|
351 |
+
self.ups = nn.ModuleList()
|
352 |
+
for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
|
353 |
+
self.ups.append(weight_norm(
|
354 |
+
ConvTranspose1d(h.upsample_initial_channel//(2**i),
|
355 |
+
h.upsample_initial_channel//(2**(i+1)),
|
356 |
+
k,
|
357 |
+
u,
|
358 |
+
padding=(k-u)//2)))
|
359 |
+
|
360 |
+
c_cur = h.upsample_initial_channel // (2 ** (i + 1))
|
361 |
+
|
362 |
+
if i + 1 < len(h.upsample_rates): #
|
363 |
+
stride_f0 = np.prod(h.upsample_rates[i + 1:])
|
364 |
+
self.noise_convs.append(Conv1d(
|
365 |
+
h.gen_istft_n_fft + 2, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=(stride_f0+1) // 2))
|
366 |
+
self.noise_res.append(resblock(h, c_cur, 7, [1,3,5]))
|
367 |
+
else:
|
368 |
+
self.noise_convs.append(Conv1d(h.gen_istft_n_fft + 2, c_cur, kernel_size=1))
|
369 |
+
self.noise_res.append(resblock(h, c_cur, 11, [1,3,5]))
|
370 |
+
|
371 |
+
self.alphas = nn.ParameterList()
|
372 |
+
self.alphas.append(nn.Parameter(torch.ones(1, h.upsample_initial_channel, 1)))
|
373 |
+
self.resblocks = nn.ModuleList()
|
374 |
+
for i in range(len(self.ups)):
|
375 |
+
ch = h.upsample_initial_channel//(2**(i+1))
|
376 |
+
self.alphas.append(nn.Parameter(torch.ones(1, ch, 1)))
|
377 |
+
for j, (k, d) in enumerate(
|
378 |
+
zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
|
379 |
+
self.resblocks.append(resblock(h, ch, k, d))
|
380 |
+
|
381 |
+
|
382 |
+
self.conformers = nn.ModuleList()
|
383 |
+
self.post_n_fft = h.gen_istft_n_fft
|
384 |
+
self.conv_post = weight_norm(Conv1d(128, self.post_n_fft + 2, 7, 1, padding=3))
|
385 |
+
|
386 |
+
for i in range(len(self.ups)):
|
387 |
+
ch = h.upsample_initial_channel // (2**i)
|
388 |
+
self.conformers.append(
|
389 |
+
Conformer(
|
390 |
+
dim=ch,
|
391 |
+
depth=2,
|
392 |
+
dim_head=64,
|
393 |
+
heads=8,
|
394 |
+
ff_mult=4,
|
395 |
+
conv_expansion_factor=2,
|
396 |
+
conv_kernel_size=31,
|
397 |
+
attn_dropout=0.1,
|
398 |
+
ff_dropout=0.1,
|
399 |
+
conv_dropout=0.1,
|
400 |
+
# device=self.device
|
401 |
+
)
|
402 |
+
)
|
403 |
+
|
404 |
+
self.ups.apply(init_weights)
|
405 |
+
self.conv_post.apply(init_weights)
|
406 |
+
self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
|
407 |
+
self.stft = TorchSTFT(filter_length=h.gen_istft_n_fft,
|
408 |
+
hop_length=h.gen_istft_hop_size,
|
409 |
+
win_length=h.gen_istft_n_fft)
|
410 |
+
|
411 |
+
|
412 |
+
|
413 |
+
def forward(self, x):
|
414 |
+
|
415 |
+
|
416 |
+
|
417 |
+
f0, _, _ = self.F0_model(x.unsqueeze(1))
|
418 |
+
if len(f0.shape) == 1:
|
419 |
+
f0 = f0.unsqueeze(0)
|
420 |
+
|
421 |
+
f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2) # bs,n,t
|
422 |
+
|
423 |
+
har_source, _, _ = self.m_source(f0)
|
424 |
+
har_source = har_source.transpose(1, 2).squeeze(1)
|
425 |
+
har_spec, har_phase = self.stft.transform(har_source)
|
426 |
+
har = torch.cat([har_spec, har_phase], dim=1)
|
427 |
+
|
428 |
+
|
429 |
+
x = self.conv_pre(x)
|
430 |
+
|
431 |
+
for i in range(self.num_upsamples):
|
432 |
+
|
433 |
+
x = x + (1 / self.alphas[i]) * (torch.sin(self.alphas[i] * x) ** 2)
|
434 |
+
x = rearrange(x, "b f t -> b t f")
|
435 |
+
|
436 |
+
x = self.conformers[i](x)
|
437 |
+
|
438 |
+
x = rearrange(x, "b t f -> b f t")
|
439 |
+
|
440 |
+
# x = F.leaky_relu(x, LRELU_SLOPE)
|
441 |
+
x_source = self.noise_convs[i](har)
|
442 |
+
x_source = self.noise_res[i](x_source)
|
443 |
+
|
444 |
+
x = self.ups[i](x)
|
445 |
+
if i == self.num_upsamples - 1:
|
446 |
+
x = self.reflection_pad(x)
|
447 |
+
|
448 |
+
x = x + x_source
|
449 |
+
|
450 |
+
|
451 |
+
xs = None
|
452 |
+
for j in range(self.num_kernels):
|
453 |
+
if xs is None:
|
454 |
+
xs = self.resblocks[i*self.num_kernels+j](x)
|
455 |
+
else:
|
456 |
+
xs += self.resblocks[i*self.num_kernels+j](x)
|
457 |
+
x = xs / self.num_kernels
|
458 |
+
# x = F.leaky_relu(x)
|
459 |
+
|
460 |
+
|
461 |
+
x = x + (1 / self.alphas[i + 1]) * (torch.sin(self.alphas[i + 1] * x) ** 2)
|
462 |
+
|
463 |
+
x = self.conv_post(x)
|
464 |
+
spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :]).to(x.device)
|
465 |
+
phase = torch.sin(x[:, self.post_n_fft // 2 + 1:, :]).to(x.device)
|
466 |
+
|
467 |
+
return spec, phase
|
468 |
+
|
469 |
+
def remove_weight_norm(self):
|
470 |
+
print("Removing weight norm...")
|
471 |
+
for l in self.ups:
|
472 |
+
remove_weight_norm(l)
|
473 |
+
for l in self.resblocks:
|
474 |
+
l.remove_weight_norm()
|
475 |
+
remove_weight_norm(self.conv_pre)
|
476 |
+
remove_weight_norm(self.conv_post)
|
477 |
+
|
478 |
+
|
479 |
+
|
480 |
+
def stft(x, fft_size, hop_size, win_length, window):
|
481 |
+
"""Perform STFT and convert to magnitude spectrogram.
|
482 |
+
Args:
|
483 |
+
x (Tensor): Input signal tensor (B, T).
|
484 |
+
fft_size (int): FFT size.
|
485 |
+
hop_size (int): Hop size.
|
486 |
+
win_length (int): Window length.
|
487 |
+
window (str): Window function type.
|
488 |
+
Returns:
|
489 |
+
Tensor: Magnitude spectrogram (B, #frames, fft_size // 2 + 1).
|
490 |
+
"""
|
491 |
+
x_stft = torch.stft(x, fft_size, hop_size, win_length, window,
|
492 |
+
return_complex=True)
|
493 |
+
real = x_stft[..., 0]
|
494 |
+
imag = x_stft[..., 1]
|
495 |
+
|
496 |
+
# NOTE(kan-bayashi): clamp is needed to avoid nan or inf
|
497 |
+
return torch.abs(x_stft).transpose(2, 1)
|
498 |
+
|
499 |
+
class SpecDiscriminator(nn.Module):
|
500 |
+
"""docstring for Discriminator."""
|
501 |
+
|
502 |
+
def __init__(self, fft_size=1024, shift_size=120, win_length=600, window="hann_window", use_spectral_norm=False):
|
503 |
+
super(SpecDiscriminator, self).__init__()
|
504 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
505 |
+
self.fft_size = fft_size
|
506 |
+
self.shift_size = shift_size
|
507 |
+
self.win_length = win_length
|
508 |
+
self.window = getattr(torch, window)(win_length)
|
509 |
+
self.discriminators = nn.ModuleList([
|
510 |
+
norm_f(nn.Conv2d(1, 32, kernel_size=(3, 9), padding=(1, 4))),
|
511 |
+
norm_f(nn.Conv2d(32, 32, kernel_size=(3, 9), stride=(1,2), padding=(1, 4))),
|
512 |
+
norm_f(nn.Conv2d(32, 32, kernel_size=(3, 9), stride=(1,2), padding=(1, 4))),
|
513 |
+
norm_f(nn.Conv2d(32, 32, kernel_size=(3, 9), stride=(1,2), padding=(1, 4))),
|
514 |
+
norm_f(nn.Conv2d(32, 32, kernel_size=(3, 3), stride=(1,1), padding=(1, 1))),
|
515 |
+
])
|
516 |
+
|
517 |
+
self.out = norm_f(nn.Conv2d(32, 1, 3, 1, 1))
|
518 |
+
|
519 |
+
def forward(self, y):
|
520 |
+
|
521 |
+
fmap = []
|
522 |
+
y = y.squeeze(1)
|
523 |
+
y = stft(y, self.fft_size, self.shift_size, self.win_length, self.window.to(y.get_device()))
|
524 |
+
y = y.unsqueeze(1)
|
525 |
+
for i, d in enumerate(self.discriminators):
|
526 |
+
y = d(y)
|
527 |
+
y = F.leaky_relu(y, LRELU_SLOPE)
|
528 |
+
fmap.append(y)
|
529 |
+
|
530 |
+
y = self.out(y)
|
531 |
+
fmap.append(y)
|
532 |
+
|
533 |
+
return torch.flatten(y, 1, -1), fmap
|
534 |
+
|
535 |
+
# class MultiResSpecDiscriminator(torch.nn.Module):
|
536 |
+
|
537 |
+
# def __init__(self,
|
538 |
+
# fft_sizes=[1024, 2048, 512],
|
539 |
+
# hop_sizes=[120, 240, 50],
|
540 |
+
# win_lengths=[600, 1200, 240],
|
541 |
+
# window="hann_window"):
|
542 |
+
|
543 |
+
# super(MultiResSpecDiscriminator, self).__init__()
|
544 |
+
# self.discriminators = nn.ModuleList([
|
545 |
+
# SpecDiscriminator(fft_sizes[0], hop_sizes[0], win_lengths[0], window),
|
546 |
+
# SpecDiscriminator(fft_sizes[1], hop_sizes[1], win_lengths[1], window),
|
547 |
+
# SpecDiscriminator(fft_sizes[2], hop_sizes[2], win_lengths[2], window)
|
548 |
+
# ])
|
549 |
+
|
550 |
+
# def forward(self, y, y_hat):
|
551 |
+
# y_d_rs = []
|
552 |
+
# y_d_gs = []
|
553 |
+
# fmap_rs = []
|
554 |
+
# fmap_gs = []
|
555 |
+
# for i, d in enumerate(self.discriminators):
|
556 |
+
# y_d_r, fmap_r = d(y)
|
557 |
+
# y_d_g, fmap_g = d(y_hat)
|
558 |
+
# y_d_rs.append(y_d_r)
|
559 |
+
# fmap_rs.append(fmap_r)
|
560 |
+
# y_d_gs.append(y_d_g)
|
561 |
+
# fmap_gs.append(fmap_g)
|
562 |
+
|
563 |
+
# return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
564 |
+
|
565 |
+
|
566 |
+
class DiscriminatorP(torch.nn.Module):
|
567 |
+
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
|
568 |
+
super(DiscriminatorP, self).__init__()
|
569 |
+
self.period = period
|
570 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
571 |
+
self.convs = nn.ModuleList([
|
572 |
+
norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
|
573 |
+
norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
|
574 |
+
norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
|
575 |
+
norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
|
576 |
+
norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
|
577 |
+
])
|
578 |
+
self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
|
579 |
+
|
580 |
+
def forward(self, x):
|
581 |
+
fmap = []
|
582 |
+
|
583 |
+
# 1d to 2d
|
584 |
+
b, c, t = x.shape
|
585 |
+
if t % self.period != 0: # pad first
|
586 |
+
n_pad = self.period - (t % self.period)
|
587 |
+
x = F.pad(x, (0, n_pad), "reflect")
|
588 |
+
t = t + n_pad
|
589 |
+
x = x.view(b, c, t // self.period, self.period)
|
590 |
+
|
591 |
+
for l in self.convs:
|
592 |
+
x = l(x)
|
593 |
+
x = F.leaky_relu(x, LRELU_SLOPE)
|
594 |
+
fmap.append(x)
|
595 |
+
x = self.conv_post(x)
|
596 |
+
fmap.append(x)
|
597 |
+
x = torch.flatten(x, 1, -1)
|
598 |
+
|
599 |
+
return x, fmap
|
600 |
+
|
601 |
+
|
602 |
+
class MultiPeriodDiscriminator(torch.nn.Module):
|
603 |
+
def __init__(self):
|
604 |
+
super(MultiPeriodDiscriminator, self).__init__()
|
605 |
+
self.discriminators = nn.ModuleList([
|
606 |
+
DiscriminatorP(2),
|
607 |
+
DiscriminatorP(3),
|
608 |
+
DiscriminatorP(5),
|
609 |
+
DiscriminatorP(7),
|
610 |
+
DiscriminatorP(11),
|
611 |
+
])
|
612 |
+
|
613 |
+
def forward(self, y, y_hat):
|
614 |
+
y_d_rs = []
|
615 |
+
y_d_gs = []
|
616 |
+
fmap_rs = []
|
617 |
+
fmap_gs = []
|
618 |
+
for i, d in enumerate(self.discriminators):
|
619 |
+
y_d_r, fmap_r = d(y)
|
620 |
+
y_d_g, fmap_g = d(y_hat)
|
621 |
+
y_d_rs.append(y_d_r)
|
622 |
+
fmap_rs.append(fmap_r)
|
623 |
+
y_d_gs.append(y_d_g)
|
624 |
+
fmap_gs.append(fmap_g)
|
625 |
+
|
626 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
627 |
+
|
628 |
+
|
629 |
+
class DiscriminatorS(torch.nn.Module):
|
630 |
+
def __init__(self, use_spectral_norm=False):
|
631 |
+
super(DiscriminatorS, self).__init__()
|
632 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
633 |
+
self.convs = nn.ModuleList([
|
634 |
+
norm_f(Conv1d(1, 128, 15, 1, padding=7)),
|
635 |
+
norm_f(Conv1d(128, 128, 41, 2, groups=4, padding=20)),
|
636 |
+
norm_f(Conv1d(128, 256, 41, 2, groups=16, padding=20)),
|
637 |
+
norm_f(Conv1d(256, 512, 41, 4, groups=16, padding=20)),
|
638 |
+
norm_f(Conv1d(512, 1024, 41, 4, groups=16, padding=20)),
|
639 |
+
norm_f(Conv1d(1024, 1024, 41, 1, groups=16, padding=20)),
|
640 |
+
norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
|
641 |
+
])
|
642 |
+
self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
|
643 |
+
|
644 |
+
def forward(self, x):
|
645 |
+
fmap = []
|
646 |
+
for l in self.convs:
|
647 |
+
x = l(x)
|
648 |
+
x = F.leaky_relu(x, LRELU_SLOPE)
|
649 |
+
fmap.append(x)
|
650 |
+
x = self.conv_post(x)
|
651 |
+
fmap.append(x)
|
652 |
+
x = torch.flatten(x, 1, -1)
|
653 |
+
|
654 |
+
return x, fmap
|
655 |
+
|
656 |
+
|
657 |
+
class MultiScaleDiscriminator(torch.nn.Module):
|
658 |
+
def __init__(self):
|
659 |
+
super(MultiScaleDiscriminator, self).__init__()
|
660 |
+
self.discriminators = nn.ModuleList([
|
661 |
+
DiscriminatorS(use_spectral_norm=True),
|
662 |
+
DiscriminatorS(),
|
663 |
+
DiscriminatorS(),
|
664 |
+
])
|
665 |
+
self.meanpools = nn.ModuleList([
|
666 |
+
AvgPool1d(4, 2, padding=2),
|
667 |
+
AvgPool1d(4, 2, padding=2)
|
668 |
+
])
|
669 |
+
|
670 |
+
def forward(self, y, y_hat):
|
671 |
+
y_d_rs = []
|
672 |
+
y_d_gs = []
|
673 |
+
fmap_rs = []
|
674 |
+
fmap_gs = []
|
675 |
+
for i, d in enumerate(self.discriminators):
|
676 |
+
if i != 0:
|
677 |
+
y = self.meanpools[i-1](y)
|
678 |
+
y_hat = self.meanpools[i-1](y_hat)
|
679 |
+
y_d_r, fmap_r = d(y)
|
680 |
+
y_d_g, fmap_g = d(y_hat)
|
681 |
+
y_d_rs.append(y_d_r)
|
682 |
+
fmap_rs.append(fmap_r)
|
683 |
+
y_d_gs.append(y_d_g)
|
684 |
+
fmap_gs.append(fmap_g)
|
685 |
+
|
686 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
687 |
+
|
688 |
+
|
689 |
+
|
690 |
+
|
691 |
+
|
692 |
+
########################### from ringformer
|
693 |
+
|
694 |
+
# multiscale_subband_cfg = {
|
695 |
+
# "hop_lengths": [1024, 512, 512], # Doubled to maintain similar time resolution
|
696 |
+
# "sampling_rate": 44100, # New sampling rate
|
697 |
+
# "filters": 32, # Kept same as it controls initial feature dimension
|
698 |
+
# "max_filters": 1024, # Kept same as it's a maximum limit
|
699 |
+
# "filters_scale": 1, # Kept same as it's a scaling factor
|
700 |
+
# "dilations": [1, 2, 4], # Kept same as they control receptive field growth
|
701 |
+
# "in_channels": 1, # Kept same (mono audio)
|
702 |
+
# "out_channels": 1, # Kept same (mono audio)
|
703 |
+
# "n_octaves": [10, 10, 10], # Increased by 1 to handle higher frequency range
|
704 |
+
# "bins_per_octaves": [24, 36, 48], # Kept same as they control frequency resolution
|
705 |
+
# }
|
706 |
+
|
707 |
+
|
708 |
+
|
709 |
+
multiscale_subband_cfg = {
|
710 |
+
"hop_lengths": [512, 256, 256],
|
711 |
+
"sampling_rate": 24000,
|
712 |
+
"filters": 32,
|
713 |
+
"max_filters": 1024,
|
714 |
+
"filters_scale": 1,
|
715 |
+
"dilations": [1, 2, 4],
|
716 |
+
"in_channels": 1,
|
717 |
+
"out_channels": 1,
|
718 |
+
"n_octaves": [9, 9, 9],
|
719 |
+
"bins_per_octaves": [24, 36, 48],
|
720 |
+
}
|
721 |
+
|
722 |
+
class DiscriminatorCQT(nn.Module):
|
723 |
+
def __init__(self, cfg, hop_length, n_octaves, bins_per_octave):
|
724 |
+
super(DiscriminatorCQT, self).__init__()
|
725 |
+
self.cfg = cfg
|
726 |
+
|
727 |
+
self.filters = cfg.filters
|
728 |
+
self.max_filters = cfg.max_filters
|
729 |
+
self.filters_scale = cfg.filters_scale
|
730 |
+
self.kernel_size = (3, 9)
|
731 |
+
self.dilations = cfg.dilations
|
732 |
+
self.stride = (1, 2)
|
733 |
+
|
734 |
+
self.in_channels = cfg.in_channels
|
735 |
+
self.out_channels = cfg.out_channels
|
736 |
+
self.fs = cfg.sampling_rate
|
737 |
+
self.hop_length = hop_length
|
738 |
+
self.n_octaves = n_octaves
|
739 |
+
self.bins_per_octave = bins_per_octave
|
740 |
+
|
741 |
+
self.cqt_transform = features.cqt.CQT2010v2(
|
742 |
+
sr=self.fs * 2,
|
743 |
+
hop_length=self.hop_length,
|
744 |
+
n_bins=self.bins_per_octave * self.n_octaves,
|
745 |
+
bins_per_octave=self.bins_per_octave,
|
746 |
+
output_format="Complex",
|
747 |
+
pad_mode="constant",
|
748 |
+
)
|
749 |
+
|
750 |
+
self.conv_pres = nn.ModuleList()
|
751 |
+
for i in range(self.n_octaves):
|
752 |
+
self.conv_pres.append(
|
753 |
+
NormConv2d(
|
754 |
+
self.in_channels * 2,
|
755 |
+
self.in_channels * 2,
|
756 |
+
kernel_size=self.kernel_size,
|
757 |
+
padding=get_2d_padding(self.kernel_size),
|
758 |
+
)
|
759 |
+
)
|
760 |
+
|
761 |
+
self.convs = nn.ModuleList()
|
762 |
+
|
763 |
+
self.convs.append(
|
764 |
+
NormConv2d(
|
765 |
+
self.in_channels * 2,
|
766 |
+
self.filters,
|
767 |
+
kernel_size=self.kernel_size,
|
768 |
+
padding=get_2d_padding(self.kernel_size),
|
769 |
+
)
|
770 |
+
)
|
771 |
+
|
772 |
+
in_chs = min(self.filters_scale * self.filters, self.max_filters)
|
773 |
+
for i, dilation in enumerate(self.dilations):
|
774 |
+
out_chs = min(
|
775 |
+
(self.filters_scale ** (i + 1)) * self.filters, self.max_filters
|
776 |
+
)
|
777 |
+
self.convs.append(
|
778 |
+
NormConv2d(
|
779 |
+
in_chs,
|
780 |
+
out_chs,
|
781 |
+
kernel_size=self.kernel_size,
|
782 |
+
stride=self.stride,
|
783 |
+
dilation=(dilation, 1),
|
784 |
+
padding=get_2d_padding(self.kernel_size, (dilation, 1)),
|
785 |
+
norm="weight_norm",
|
786 |
+
)
|
787 |
+
)
|
788 |
+
in_chs = out_chs
|
789 |
+
out_chs = min(
|
790 |
+
(self.filters_scale ** (len(self.dilations) + 1)) * self.filters,
|
791 |
+
self.max_filters,
|
792 |
+
)
|
793 |
+
self.convs.append(
|
794 |
+
NormConv2d(
|
795 |
+
in_chs,
|
796 |
+
out_chs,
|
797 |
+
kernel_size=(self.kernel_size[0], self.kernel_size[0]),
|
798 |
+
padding=get_2d_padding((self.kernel_size[0], self.kernel_size[0])),
|
799 |
+
norm="weight_norm",
|
800 |
+
)
|
801 |
+
)
|
802 |
+
|
803 |
+
self.conv_post = NormConv2d(
|
804 |
+
out_chs,
|
805 |
+
self.out_channels,
|
806 |
+
kernel_size=(self.kernel_size[0], self.kernel_size[0]),
|
807 |
+
padding=get_2d_padding((self.kernel_size[0], self.kernel_size[0])),
|
808 |
+
norm="weight_norm",
|
809 |
+
)
|
810 |
+
|
811 |
+
self.activation = torch.nn.LeakyReLU(negative_slope=LRELU_SLOPE)
|
812 |
+
self.resample = torchaudio.transforms.Resample(
|
813 |
+
orig_freq=self.fs, new_freq=self.fs * 2
|
814 |
+
)
|
815 |
+
|
816 |
+
def forward(self, x):
|
817 |
+
fmap = []
|
818 |
+
|
819 |
+
x = self.resample(x)
|
820 |
+
|
821 |
+
z = self.cqt_transform(x)
|
822 |
+
|
823 |
+
z_amplitude = z[:, :, :, 0].unsqueeze(1)
|
824 |
+
z_phase = z[:, :, :, 1].unsqueeze(1)
|
825 |
+
|
826 |
+
z = torch.cat([z_amplitude, z_phase], dim=1)
|
827 |
+
z = rearrange(z, "b c w t -> b c t w")
|
828 |
+
|
829 |
+
latent_z = []
|
830 |
+
for i in range(self.n_octaves):
|
831 |
+
latent_z.append(
|
832 |
+
self.conv_pres[i](
|
833 |
+
z[
|
834 |
+
:,
|
835 |
+
:,
|
836 |
+
:,
|
837 |
+
i * self.bins_per_octave : (i + 1) * self.bins_per_octave,
|
838 |
+
]
|
839 |
+
)
|
840 |
+
)
|
841 |
+
latent_z = torch.cat(latent_z, dim=-1)
|
842 |
+
|
843 |
+
for i, l in enumerate(self.convs):
|
844 |
+
latent_z = l(latent_z)
|
845 |
+
|
846 |
+
latent_z = self.activation(latent_z)
|
847 |
+
fmap.append(latent_z)
|
848 |
+
|
849 |
+
latent_z = self.conv_post(latent_z)
|
850 |
+
|
851 |
+
return latent_z, fmap
|
852 |
+
|
853 |
+
|
854 |
+
|
855 |
+
class MultiScaleSubbandCQTDiscriminator(nn.Module): # replacing "MultiResSpecDiscriminator"
|
856 |
+
def __init__(self):
|
857 |
+
super(MultiScaleSubbandCQTDiscriminator, self).__init__()
|
858 |
+
cfg = Munch(multiscale_subband_cfg)
|
859 |
+
self.cfg = cfg
|
860 |
+
self.discriminators = nn.ModuleList(
|
861 |
+
[
|
862 |
+
DiscriminatorCQT(
|
863 |
+
cfg,
|
864 |
+
hop_length=cfg.hop_lengths[i],
|
865 |
+
n_octaves=cfg.n_octaves[i],
|
866 |
+
bins_per_octave=cfg.bins_per_octaves[i],
|
867 |
+
)
|
868 |
+
for i in range(len(cfg.hop_lengths))
|
869 |
+
]
|
870 |
+
)
|
871 |
+
|
872 |
+
def forward(self, y, y_hat):
|
873 |
+
y_d_rs = []
|
874 |
+
y_d_gs = []
|
875 |
+
fmap_rs = []
|
876 |
+
fmap_gs = []
|
877 |
+
|
878 |
+
for disc in self.discriminators:
|
879 |
+
y_d_r, fmap_r = disc(y)
|
880 |
+
y_d_g, fmap_g = disc(y_hat)
|
881 |
+
y_d_rs.append(y_d_r)
|
882 |
+
fmap_rs.append(fmap_r)
|
883 |
+
y_d_gs.append(y_d_g)
|
884 |
+
fmap_gs.append(fmap_g)
|
885 |
+
|
886 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
887 |
+
|
888 |
+
|
889 |
+
|
890 |
+
#############################
|
891 |
+
|
892 |
+
|
893 |
+
|
894 |
+
def feature_loss(fmap_r, fmap_g):
|
895 |
+
loss = 0
|
896 |
+
for dr, dg in zip(fmap_r, fmap_g):
|
897 |
+
for rl, gl in zip(dr, dg):
|
898 |
+
loss += torch.mean(torch.abs(rl - gl))
|
899 |
+
|
900 |
+
return loss*2
|
901 |
+
|
902 |
+
|
903 |
+
def discriminator_loss(disc_real_outputs, disc_generated_outputs):
|
904 |
+
loss = 0
|
905 |
+
r_losses = []
|
906 |
+
g_losses = []
|
907 |
+
for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
|
908 |
+
r_loss = torch.mean((1-dr)**2)
|
909 |
+
g_loss = torch.mean(dg**2)
|
910 |
+
loss += (r_loss + g_loss)
|
911 |
+
r_losses.append(r_loss.item())
|
912 |
+
g_losses.append(g_loss.item())
|
913 |
+
|
914 |
+
return loss, r_losses, g_losses
|
915 |
+
|
916 |
+
|
917 |
+
def generator_loss(disc_outputs):
|
918 |
+
loss = 0
|
919 |
+
gen_losses = []
|
920 |
+
for dg in disc_outputs:
|
921 |
+
l = torch.mean((1-dg)**2)
|
922 |
+
gen_losses.append(l)
|
923 |
+
loss += l
|
924 |
+
|
925 |
+
return loss, gen_losses
|
926 |
+
|
927 |
+
def discriminator_TPRLS_loss(disc_real_outputs, disc_generated_outputs):
|
928 |
+
loss = 0
|
929 |
+
for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
|
930 |
+
tau = 0.04
|
931 |
+
m_DG = torch.median((dr-dg))
|
932 |
+
L_rel = torch.mean((((dr - dg) - m_DG)**2)[dr < dg + m_DG])
|
933 |
+
loss += tau - F.relu(tau - L_rel)
|
934 |
+
return loss
|
935 |
+
|
936 |
+
def generator_TPRLS_loss(disc_real_outputs, disc_generated_outputs):
|
937 |
+
loss = 0
|
938 |
+
for dg, dr in zip(disc_real_outputs, disc_generated_outputs):
|
939 |
+
tau = 0.04
|
940 |
+
m_DG = torch.median((dr-dg))
|
941 |
+
L_rel = torch.mean((((dr - dg) - m_DG)**2)[dr < dg + m_DG])
|
942 |
+
loss += tau - F.relu(tau - L_rel)
|
943 |
+
return loss
|