-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathEffCRN23.py
246 lines (200 loc) · 11.1 KB
/
EffCRN23.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
238
239
240
241
242
243
244
245
246
# -*- coding: utf-8 -*-
import math
import torch
import torch.nn as nn
import numpy as np
from torchinfo import summary
from ConvLSTM import ConvLSTM
class ConvBlock(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, stride, padding="same", activate=True):
super(ConvBlock, self).__init__()
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
if activate:
self.convblock = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding=0),
nn.LeakyReLU(0.2))
else:
self.convblock = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding=0),
nn.Identity())
def forward(self, x):
global pad_x
if self.padding == "same":
pad = self.kernel_size[0] - 1
pad_x = torch.nn.functional.pad(x, (0, 0, 0, pad), "constant", 0)
elif self.padding == "valid":
pad_x = x
output = self.convblock(pad_x)
return output
class DeconvBlock(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, stride):
super(DeconvBlock, self).__init__()
self.kernel_size = kernel_size
self.stride = stride
self.deconvblock = nn.Sequential(
nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride, padding=((kernel_size[0] - 2) // 2, 0)),
nn.LeakyReLU(0.2))
def forward(self, x):
output = self.deconvblock(x)
return output
class GRUBlock(nn.Module):
def __init__(self, in_channels, hidden_channels, out_channels, bidirectional=False):
super(GRUBlock, self).__init__()
self.gru = nn.GRU(in_channels, hidden_channels, bidirectional=bidirectional, bias=True)
self.linear = nn.Linear(hidden_channels, out_channels)
def forward(self, x, rnn_state):
gru_out, rnn_state = self.gru(x, rnn_state)
output = self.linear(gru_out)
return output, rnn_state
class EffCRN23(nn.Module):
def __init__(self, in_channels, out_channels, F, N):
super(EffCRN23, self).__init__()
self.N = N
self.en_convblock_1 = ConvBlock(in_channels, F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.en_downblock_1 = ConvBlock(F, F, kernel_size=(N, 1), stride=(2, 1), padding="valid")
self.en_convblock_2 = ConvBlock(F, 2 * F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.en_downblock_2 = ConvBlock(2 * F, 2 * F, kernel_size=(N, 1), stride=(2, 1), padding="valid")
self.en_convblock_3 = ConvBlock(2 * F, 3 * F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.en_downblock_3 = ConvBlock(3 * F, 3 * F, kernel_size=(N, 1), stride=(2, 1), padding="valid")
self.en_convblock_4 = ConvBlock(3 * F, 4 * F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.en_downblock_4 = ConvBlock(4 * F, 4 * F, kernel_size=(N, 1), stride=(2, 1), padding="valid")
self.en_convblock_5 = ConvBlock(4 * F, 5 * F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.en_downblock_5 = ConvBlock(5 * F, 5 * F, kernel_size=(N, 1), stride=(2, 1), padding="valid")
self.convlstm = ConvLSTM(9, 5 * F, F, kernel_size=N, cnn_dropout=0.2, rnn_dropout=0.2, batch_first=True, bias=False)
self.gru = GRUBlock(F, 9 * F, F, bidirectional=False)
self.de_upblock_1 = DeconvBlock(F, 5 * F, kernel_size=(N, 1), stride=(2, 1))
self.de_skip_conv_1 = ConvBlock(5 * F, 5 * F, kernel_size=(1, 1), stride=(1, 1), padding="valid")
self.de_convblock_1 = ConvBlock(5 * F, 5 * F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.de_upblock_2 = DeconvBlock(5 * F, 4 * F, kernel_size=(N, 1), stride=(2, 1))
self.de_skip_conv_2 = ConvBlock(4 * F, 4 * F, kernel_size=(1, 1), stride=(1, 1), padding="valid")
self.de_convblock_2 = ConvBlock(4 * F, 4 * F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.de_upblock_3 = DeconvBlock(4 * F, 3 * F, kernel_size=(N, 1), stride=(2, 1))
self.de_skip_conv_3 = ConvBlock(3 * F, 3 * F, kernel_size=(1, 1), stride=(1, 1), padding="valid")
self.de_convblock_3 = ConvBlock(3 * F, 3 * F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.de_upblock_4 = DeconvBlock(3 * F, 2 * F, kernel_size=(N, 1), stride=(2, 1))
self.de_skip_conv_4 = ConvBlock(2 * F, 2 * F, kernel_size=(1, 1), stride=(1, 1), padding="valid")
self.de_convblock_4 = ConvBlock(2 * F, 2 * F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.de_upblock_5 = DeconvBlock(2 * F, F, kernel_size=(N, 1), stride=(2, 1))
self.de_skip_conv_5 = ConvBlock(F, F, kernel_size=(1, 1), stride=(1, 1), padding="valid")
self.de_convblock_5 = ConvBlock(F, F, kernel_size=(N, 1), stride=(1, 1), padding="same")
self.out_convblock = ConvBlock(F, out_channels, kernel_size=(N, 1), stride=(1, 1), padding="same", activate=False)
def mask_func(self, noisy, estimate):
r, i = noisy[:, 0, :, :], noisy[:, 1, :, :]
r_hat, i_hat = estimate[:, 0, :, :], noisy[:, 1, :, :]
mag_mask = torch.sqrt(r_hat ** 2 + i_hat ** 2)
phase_rotate = torch.atan2(i_hat, r_hat)
mag_mask = torch.tanh(mag_mask)
mag = mag_mask * torch.sqrt(r ** 2 + i ** 2)
phase = phase_rotate + torch.atan2(i, r)
# return real, imag
return mag * torch.cos(phase), mag * torch.sin(phase)
def forward(self, x, rnn_state_1=None, rnn_state_2=None):
input = x
pad = torch.nn.functional.pad(input, (0, 0, 0, 3), "constant", 0)
en_conv_1 = self.en_convblock_1(pad)
pad_en1 = torch.nn.functional.pad(en_conv_1, (0, 0, 0, self.N - 2), "constant", 0)
en_down_1 = self.en_downblock_1(pad_en1)
en_conv_2 = self.en_convblock_2(en_down_1)
pad_en2_1 = torch.nn.functional.pad(en_conv_2, (0, 0, 0, self.N - 2), "constant", 0)
en_down_2 = self.en_downblock_2(pad_en2_1)
pad_en2_2 = torch.nn.functional.pad(en_down_2, (0, 0, 0, 1), "constant", 0)
en_conv_3 = self.en_convblock_3(pad_en2_2)
pad_en3_1 = torch.nn.functional.pad(en_conv_3, (0, 0, 0, self.N - 2), "constant", 0)
en_down_3 = self.en_downblock_3(pad_en3_1)
pad_en3_2 = torch.nn.functional.pad(en_down_3, (0, 0, 0, 1), "constant", 0)
en_conv_4 = self.en_convblock_4(pad_en3_2)
pad_en4_1 = torch.nn.functional.pad(en_conv_4, (0, 0, 0, self.N - 2), "constant", 0)
en_down_4 = self.en_downblock_4(pad_en4_1)
pad_en4_2 = torch.nn.functional.pad(en_down_4, (0, 0, 0, 1), "constant", 0)
en_conv_5 = self.en_convblock_5(pad_en4_2)
pad_en5_1 = torch.nn.functional.pad(en_conv_5, (0, 0, 0, self.N - 2), "constant", 0)
en_down_5 = self.en_downblock_5(pad_en5_1)
conv_lstm, state_out_1 = self.convlstm(en_down_5.squeeze(-1).unsqueeze(1), rnn_state_1)
conv_lstm = conv_lstm.squeeze(1).transpose(1, 2)
gru_out, state_out_2 = self.gru(conv_lstm, rnn_state_2)
gru_out = gru_out.transpose(1, 2).unsqueeze(-1)
de_up_1 = self.de_upblock_1(gru_out)
de_skip_1 = self.de_skip_conv_1(en_conv_5)
de_conv_1 = self.de_convblock_1(de_up_1 + de_skip_1)
de_up_2 = self.de_upblock_2(de_conv_1[:, :, :-1, :])
de_skip_2 = self.de_skip_conv_2(en_conv_4)
de_conv_2 = self.de_convblock_2(de_up_2 + de_skip_2)
de_up_3 = self.de_upblock_3(de_conv_2[:, :, :-1, :])
de_skip_3 = self.de_skip_conv_3(en_conv_3)
de_conv_3 = self.de_convblock_3(de_up_3 + de_skip_3)
de_up_4 = self.de_upblock_4(de_conv_3[:, :, :-1, :])
de_skip_4 = self.de_skip_conv_4(en_conv_2)
de_conv_4 = self.de_convblock_4(de_up_4 + de_skip_4)
de_up_5 = self.de_upblock_5(de_conv_4)
de_skip_5 = self.de_skip_conv_5(en_conv_1)
de_conv_5 = self.de_convblock_5(de_up_5 + de_skip_5)
out = self.out_convblock(de_conv_5)[:, :, :-3, :]
real, imag = self.mask_func(x, out)
est = torch.cat([real.unsqueeze(1), imag.unsqueeze(1)], dim=1)
return est, state_out_1, state_out_2
class TorchSignalToFrames(object):
def __init__(self, frame_size=128, frame_shift=64):
super(TorchSignalToFrames, self).__init__()
self.frame_size = frame_size
self.frame_shift = frame_shift
def __call__(self, in_sig):
sig_len = in_sig.shape[-1]
nframes = math.ceil((sig_len - self.frame_size) / self.frame_shift + 1)
a = torch.zeros(tuple(in_sig.shape[:-1]) + (nframes, self.frame_size), device=in_sig.device)
start = 0
end = start + self.frame_size
k = 0
for i in range(nframes):
if end < sig_len:
a[..., i, :] = in_sig[..., start:end]
k += 1
else:
tail_size = sig_len - start
a[..., i, :tail_size] = in_sig[..., start:]
start = start + self.frame_shift
end = start + self.frame_size
return a
class TorchOLA(object):
def __init__(self, frame_shift=256):
super(TorchOLA, self).__init__()
self.frame_shift = frame_shift
def __call__(self, inputs):
nframes = inputs.shape[-2]
frame_size = inputs.shape[-1]
frame_step = self.frame_shift
sig_length = (nframes - 1) * frame_step + frame_size
sig = torch.zeros(list(inputs.shape[:-2]) + [sig_length], dtype=inputs.dtype, device=inputs.device,
requires_grad=False)
ones = torch.zeros_like(sig)
start = 0
end = start + frame_size
for i in range(nframes):
sig[..., start:end] += inputs[..., i, :]
ones[..., start:end] += 1.
start = start + frame_step
end = start + frame_size
return sig / ones
if __name__ == '__main__':
n_fft = 512
hop_length = 256
x = torch.randn(1, 16000)
model = EffCRN23(in_channels=2, out_channels=2, F=27, N=4)
TO_SAMPLES = TorchOLA(frame_shift=hop_length)
TO_FRAME = TorchSignalToFrames(frame_size=n_fft, frame_shift=hop_length)
h_zeros_1 = torch.zeros(1, 27, 9, dtype=x.dtype, device=x.device)
c_zeros_1 = torch.zeros(1, 27, 9, dtype=x.dtype, device=x.device)
h_zeros_2 = torch.zeros(1, 9, 27 * 9, dtype=x.dtype, device=x.device)
state_list_1 = [h_zeros_1, c_zeros_1]
state_list_2 = h_zeros_2
input = TO_FRAME(x)
window = torch.sqrt(torch.from_numpy(np.hanning(n_fft).astype(np.float32)))
spec = torch.view_as_real(torch.fft.rfft(input)).permute(0, 3, 2, 1).contiguous()
_, _, _, t = spec.size()
est_out = torch.zeros(spec.size(), dtype=spec.dtype, device=spec.device)
for i in range(t):
input_frame = spec[:, :, :, i].unsqueeze(-1)
est_out_t, state_list_1, state_list_2 = model(input_frame, state_list_1, state_list_2)
est_out[:, :, :, i] = est_out_t.squeeze(-1)
summary(model, input_size=(1, 2, 257, 1))