-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy pathTasNET_model.py
397 lines (354 loc) · 13.9 KB
/
TasNET_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
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
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
import torch
import torch.nn as nn
import torch.nn.functional as F
from utils import overlap_and_add
EPS = 1e-8
class TasNET(nn.Module):
def __init__(self, N=256, L=20, B=256, H=512, P=3,
X=8, R=4, C=2, norm_type="gLN", causal=False,
mask_nonlinear='relu'):
"""
Args:
N: Number of filters in autoencoder
L: Length of the filters (in samples)
B: Number of channels in bottleneck 1 × 1-conv block
H: Number of channels in convolutional blocks
P: Kernel size in convolutional blocks
X: Number of convolutional blocks in each repeat
R: Number of repeats
C: Number of speakers
norm_type: BN, gLN, cLN
causal: causal or non-causal
mask_nonlinear: use which non-linear function to generate mask
"""
super(TasNET, self).__init__()
# Hyper-parameter
self.N, self.L, self.B, self.H, self.P, self.X, self.R, self.C = N, L, B, H, P, X, R, C
self.norm_type = norm_type
self.causal = causal
self.mask_nonlinear = mask_nonlinear
# Components
self.encoder = Encoder(L, N)
self.separator = TemporalConvNet(N, B, H, P, X, R, C, norm_type, causal, mask_nonlinear)
self.decoder = Decoder(N, L)
# init
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_normal_(p)
def forward(self, mixture):
"""
Args:
mixture: [M, N, L], M is batch size, #T is #samples
Returns:
x1: [M, T]
x2: [M, T]
#est_source: [M, C, T]
"""
mixture_w = self.encoder(mixture)
est_mask = self.separator(mixture_w)
est_source = self.decoder(mixture_w, est_mask)
# T changed after conv1d in encoder, fix it here
T_origin = mixture.size(-1)*mixture.size(-2)
T_conv = est_source.size(-1)
est_source = F.pad(est_source, (0, T_origin - T_conv))
M = est_source.size()[0]
x1 = est_source[:,0,:]
x1 = x1.view(M, -1)
x2 = est_source[:,1,:]
x2 = x2.view(M, -1)
# print(x1.size())
return x1, x2
@classmethod
def load_model(cls, path):
# Load to CPU
package = torch.load(path, map_location=lambda storage, loc: storage)
model = cls.load_model_from_package(package)
return model
@classmethod
def load_model_from_package(cls, package):
model = cls(package['N'], package['L'], package['B'], package['H'],
package['P'], package['X'], package['R'], package['C'],
norm_type=package['norm_type'], causal=package['causal'],
mask_nonlinear=package['mask_nonlinear'])
model.load_state_dict(package['state_dict'])
return model
@staticmethod
def serialize(model, optimizer, epoch, tr_loss=None, cv_loss=None):
package = {
# hyper-parameter
'N': model.N, 'L': model.L, 'B': model.B, 'H': model.H,
'P': model.P, 'X': model.X, 'R': model.R, 'C': model.C,
'norm_type': model.norm_type, 'causal': model.causal,
'mask_nonlinear': model.mask_nonlinear,
# state
'state_dict': model.state_dict(),
'optim_dict': optimizer.state_dict(),
'epoch': epoch
}
if tr_loss is not None:
package['tr_loss'] = tr_loss
package['cv_loss'] = cv_loss
return package
class Encoder(nn.Module):
"""Estimation of the nonnegative mixture weight by a 1-D conv layer.
"""
def __init__(self, L, N):
super(Encoder, self).__init__()
# Hyper-parameter
self.L, self.N = L, N
# Components
# 50% overlap
self.conv1d_U = nn.Conv1d(1, N, kernel_size=L, stride=L // 2, bias=False)
def forward(self, mixture):
"""
Args:
mixture: [M, T], M is batch size, T is #samples
Returns:
mixture_w: [M, N, K], where K = (T-L)/(L/2)+1 = 2T/L-1
"""
M, _, _ = mixture.size()
mixture = mixture.view(M, -1)
mixture = torch.unsqueeze(mixture, 1) # [M, 1, T]
mixture_w = F.relu(self.conv1d_U(mixture)) # [M, N, K]
return mixture_w
class Decoder(nn.Module):
def __init__(self, N, L):
super(Decoder, self).__init__()
# Hyper-parameter
self.N, self.L = N, L
# Components
self.basis_signals = nn.Linear(N, L, bias=False)
def forward(self, mixture_w, est_mask):
"""
Args:
mixture_w: [M, N, K]
est_mask: [M, C, N, K]
Returns:
est_source: [M, C, T]
"""
# D = W * M
source_w = torch.unsqueeze(mixture_w, 1) * est_mask # [M, C, N, K]
source_w = torch.transpose(source_w, 2, 3) # [M, C, K, N]
# S = DV
est_source = self.basis_signals(source_w) # [M, C, K, L]
est_source = overlap_and_add(est_source, self.L//2) # M x C x T
return est_source
class TemporalConvNet(nn.Module):
def __init__(self, N, B, H, P, X, R, C, norm_type="gLN", causal=False,
mask_nonlinear='relu'):
"""
Args:
N: Number of filters in autoencoder
B: Number of channels in bottleneck 1 × 1-conv block
H: Number of channels in convolutional blocks
P: Kernel size in convolutional blocks
X: Number of convolutional blocks in each repeat
R: Number of repeats
C: Number of speakers
norm_type: BN, gLN, cLN
causal: causal or non-causal
mask_nonlinear: use which non-linear function to generate mask
"""
super(TemporalConvNet, self).__init__()
# Hyper-parameter
self.C = C
self.mask_nonlinear = mask_nonlinear
# Components
# [M, N, K] -> [M, N, K]
layer_norm = ChannelwiseLayerNorm(N)
# [M, N, K] -> [M, B, K]
bottleneck_conv1x1 = nn.Conv1d(N, B, 1, bias=False)
# [M, B, K] -> [M, B, K]
repeats = []
for r in range(R):
blocks = []
for x in range(X):
dilation = 2**x
padding = (P - 1) * dilation if causal else (P - 1) * dilation // 2
blocks += [TemporalBlock(B, H, P, stride=1,
padding=padding,
dilation=dilation,
norm_type=norm_type,
causal=causal)]
repeats += [nn.Sequential(*blocks)]
temporal_conv_net = nn.Sequential(*repeats)
# [M, B, K] -> [M, C*N, K]
mask_conv1x1 = nn.Conv1d(B, C*N, 1, bias=False)
# Put together
self.network = nn.Sequential(layer_norm,
bottleneck_conv1x1,
temporal_conv_net,
mask_conv1x1)
def forward(self, mixture_w):
"""
Keep this API same with TasNet
Args:
mixture_w: [M, N, K], M is batch size
returns:
est_mask: [M, C, N, K]
"""
M, N, K = mixture_w.size()
score = self.network(mixture_w) # [M, N, K] -> [M, C*N, K]
score = score.view(M, self.C, N, K) # [M, C*N, K] -> [M, C, N, K]
if self.mask_nonlinear == 'softmax':
est_mask = F.softmax(score, dim=1)
elif self.mask_nonlinear == 'relu':
est_mask = F.relu(score)
else:
raise ValueError("Unsupported mask non-linear function")
return est_mask
class TemporalBlock(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size,
stride, padding, dilation, norm_type="gLN", causal=False):
super(TemporalBlock, self).__init__()
# [M, B, K] -> [M, H, K]
conv1x1 = nn.Conv1d(in_channels, out_channels, 1, bias=False)
prelu = nn.PReLU()
norm = chose_norm(norm_type, out_channels)
# [M, H, K] -> [M, B, K]
dsconv = DepthwiseSeparableConv(out_channels, in_channels, kernel_size,
stride, padding, dilation, norm_type,
causal)
# Put together
self.net = nn.Sequential(conv1x1, prelu, norm, dsconv)
def forward(self, x):
"""
Args:
x: [M, B, K]
Returns:
[M, B, K]
"""
residual = x
out = self.net(x)
# TODO: when P = 3 here works fine, but when P = 2 maybe need to pad?
return out + residual # look like w/o F.relu is better than w/ F.relu
# return F.relu(out + residual)
class DepthwiseSeparableConv(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size,
stride, padding, dilation, norm_type="gLN", causal=False):
super(DepthwiseSeparableConv, self).__init__()
# Use `groups` option to implement depthwise convolution
# [M, H, K] -> [M, H, K]
depthwise_conv = nn.Conv1d(in_channels, in_channels, kernel_size,
stride=stride, padding=padding,
dilation=dilation, groups=in_channels,
bias=False)
if causal:
chomp = Chomp1d(padding)
prelu = nn.PReLU()
norm = chose_norm(norm_type, in_channels)
# [M, H, K] -> [M, B, K]
pointwise_conv = nn.Conv1d(in_channels, out_channels, 1, bias=False)
# Put together
if causal:
self.net = nn.Sequential(depthwise_conv, chomp, prelu, norm,
pointwise_conv)
else:
self.net = nn.Sequential(depthwise_conv, prelu, norm,
pointwise_conv)
def forward(self, x):
"""
Args:
x: [M, H, K]
Returns:
result: [M, B, K]
"""
return self.net(x)
class Chomp1d(nn.Module):
"""To ensure the output length is the same as the input.
"""
def __init__(self, chomp_size):
super(Chomp1d, self).__init__()
self.chomp_size = chomp_size
def forward(self, x):
"""
Args:
x: [M, H, Kpad]
Returns:
[M, H, K]
"""
return x[:, :, :-self.chomp_size].contiguous()
def chose_norm(norm_type, channel_size):
"""The input of normlization will be (M, C, K), where M is batch size,
C is channel size and K is sequence length.
"""
if norm_type == "gLN":
return GlobalLayerNorm(channel_size)
elif norm_type == "cLN":
return ChannelwiseLayerNorm(channel_size)
else: # norm_type == "BN":
# Given input (M, C, K), nn.BatchNorm1d(C) will accumulate statics
# along M and K, so this BN usage is right.
return nn.BatchNorm1d(channel_size)
# TODO: Use nn.LayerNorm to impl cLN to speed up
class ChannelwiseLayerNorm(nn.Module):
"""Channel-wise Layer Normalization (cLN)"""
def __init__(self, channel_size):
super(ChannelwiseLayerNorm, self).__init__()
self.gamma = nn.Parameter(torch.Tensor(1, channel_size, 1)) # [1, N, 1]
self.beta = nn.Parameter(torch.Tensor(1, channel_size,1 )) # [1, N, 1]
self.reset_parameters()
def reset_parameters(self):
self.gamma.data.fill_(1)
self.beta.data.zero_()
def forward(self, y):
"""
Args:
y: [M, N, K], M is batch size, N is channel size, K is length
Returns:
cLN_y: [M, N, K]
"""
mean = torch.mean(y, dim=1, keepdim=True) # [M, 1, K]
var = torch.var(y, dim=1, keepdim=True, unbiased=False) # [M, 1, K]
cLN_y = self.gamma * (y - mean) / torch.pow(var + EPS, 0.5) + self.beta
return cLN_y
class GlobalLayerNorm(nn.Module):
"""Global Layer Normalization (gLN)"""
def __init__(self, channel_size):
super(GlobalLayerNorm, self).__init__()
self.gamma = nn.Parameter(torch.Tensor(1, channel_size, 1)) # [1, N, 1]
self.beta = nn.Parameter(torch.Tensor(1, channel_size,1 )) # [1, N, 1]
self.reset_parameters()
def reset_parameters(self):
self.gamma.data.fill_(1)
self.beta.data.zero_()
def forward(self, y):
"""
Args:
y: [M, N, K], M is batch size, N is channel size, K is length
Returns:
gLN_y: [M, N, K]
"""
# TODO: in torch 1.0, torch.mean() support dim list
mean = y.mean(dim=1, keepdim=True).mean(dim=2, keepdim=True) #[M, 1, 1]
var = (torch.pow(y-mean, 2)).mean(dim=1, keepdim=True).mean(dim=2, keepdim=True)
gLN_y = self.gamma * (y - mean) / torch.pow(var + EPS, 0.5) + self.beta
return gLN_y
if __name__ == "__main__":
torch.manual_seed(123)
M, N, L, T = 2, 3, 4, 12
K = 2*T//L-1
B, H, P, X, R, C, norm_type, causal = 2, 3, 3, 3, 2, 2, "gLN", False
mixture = torch.randint(3, (M, T))
# test Encoder
encoder = Encoder(L, N)
encoder.conv1d_U.weight.data = torch.randint(2, encoder.conv1d_U.weight.size())
mixture_w = encoder(mixture)
print('mixture', mixture)
print('U', encoder.conv1d_U.weight)
print('mixture_w', mixture_w)
print('mixture_w size', mixture_w.size())
# test TemporalConvNet
separator = TemporalConvNet(N, B, H, P, X, R, C, norm_type=norm_type, causal=causal)
est_mask = separator(mixture_w)
print('est_mask', est_mask)
# test Decoder
decoder = Decoder(N, L)
est_mask = torch.randint(2, (B, K, C, N))
est_source = decoder(mixture_w, est_mask)
print('est_source', est_source)
# test Conv-TasNet
conv_tasnet = TasNET(N, L, B, H, P, X, R, C, norm_type=norm_type)
est_source = conv_tasnet(mixture)
print('est_source', est_source)
print('est_source size', est_source.size())