Spaces:
Runtime error
Runtime error
Commit
·
ca6b6f7
1
Parent(s):
79a7f1b
Remove previous DPTNet/DCUNet implementations
Browse files- cfg/model/dptnet.yaml +5 -4
- cfg/model/tcn.yaml +1 -1
- remfx/datasets.py +1 -1
- remfx/dcunet.py +0 -649
- remfx/dptnet.py +0 -459
- remfx/models.py +2 -2
cfg/model/dptnet.yaml
CHANGED
@@ -10,12 +10,13 @@ model:
|
|
10 |
network:
|
11 |
_target_: remfx.models.DPTNetModel
|
12 |
n_src: 1
|
13 |
-
|
14 |
-
|
15 |
chunk_size: 100
|
16 |
-
n_repeats:
|
17 |
fb_name: "free"
|
18 |
kernel_size: 16
|
19 |
-
n_filters:
|
|
|
20 |
sample_rate: ${sample_rate}
|
21 |
num_bins: 1025
|
|
|
10 |
network:
|
11 |
_target_: remfx.models.DPTNetModel
|
12 |
n_src: 1
|
13 |
+
in_chan: 64
|
14 |
+
out_chan: 64
|
15 |
chunk_size: 100
|
16 |
+
n_repeats: 2
|
17 |
fb_name: "free"
|
18 |
kernel_size: 16
|
19 |
+
n_filters: 64
|
20 |
+
stride: 8
|
21 |
sample_rate: ${sample_rate}
|
22 |
num_bins: 1025
|
cfg/model/tcn.yaml
CHANGED
@@ -13,7 +13,7 @@ model:
|
|
13 |
noutputs: 1
|
14 |
nblocks: 20
|
15 |
channel_growth: 0
|
16 |
-
channel_width:
|
17 |
kernel_size: 7
|
18 |
stack_size: 10
|
19 |
dilation_growth: 2
|
|
|
13 |
noutputs: 1
|
14 |
nblocks: 20
|
15 |
channel_growth: 0
|
16 |
+
channel_width: 64
|
17 |
kernel_size: 7
|
18 |
stack_size: 10
|
19 |
dilation_growth: 2
|
remfx/datasets.py
CHANGED
@@ -295,7 +295,7 @@ class EffectDataset(Dataset):
|
|
295 |
|
296 |
# Up to max_kept_effects
|
297 |
if self.max_kept_effects != -1:
|
298 |
-
num_kept_effects = int(torch.rand(1).item() * (self.max_kept_effects))
|
299 |
else:
|
300 |
num_kept_effects = len(self.effects_to_keep)
|
301 |
effect_indices = effect_indices[:num_kept_effects]
|
|
|
295 |
|
296 |
# Up to max_kept_effects
|
297 |
if self.max_kept_effects != -1:
|
298 |
+
num_kept_effects = int(torch.rand(1).item() * (self.max_kept_effects))
|
299 |
else:
|
300 |
num_kept_effects = len(self.effects_to_keep)
|
301 |
effect_indices = effect_indices[:num_kept_effects]
|
remfx/dcunet.py
DELETED
@@ -1,649 +0,0 @@
|
|
1 |
-
# Adapted from https://github.com/AppleHolic/source_separation/tree/master/source_separation
|
2 |
-
|
3 |
-
|
4 |
-
import torch
|
5 |
-
import torch.nn as nn
|
6 |
-
import torch.nn.functional as F
|
7 |
-
import numpy as np
|
8 |
-
from torch.nn.init import calculate_gain
|
9 |
-
from typing import Tuple
|
10 |
-
from scipy.signal import get_window
|
11 |
-
from librosa.util import pad_center
|
12 |
-
from remfx.utils import single, concat_complex
|
13 |
-
|
14 |
-
|
15 |
-
class ComplexConvBlock(nn.Module):
|
16 |
-
"""
|
17 |
-
Convolution block
|
18 |
-
"""
|
19 |
-
|
20 |
-
def __init__(
|
21 |
-
self,
|
22 |
-
in_channels: int,
|
23 |
-
out_channels: int,
|
24 |
-
kernel_size: int,
|
25 |
-
padding: int = 0,
|
26 |
-
layers: int = 4,
|
27 |
-
bn_func=nn.BatchNorm1d,
|
28 |
-
act_func=nn.LeakyReLU,
|
29 |
-
skip_res: bool = False,
|
30 |
-
):
|
31 |
-
super().__init__()
|
32 |
-
# modules
|
33 |
-
self.blocks = nn.ModuleList()
|
34 |
-
self.skip_res = skip_res
|
35 |
-
|
36 |
-
for idx in range(layers):
|
37 |
-
in_ = in_channels if idx == 0 else out_channels
|
38 |
-
self.blocks.append(
|
39 |
-
nn.Sequential(
|
40 |
-
*[
|
41 |
-
bn_func(in_),
|
42 |
-
act_func(),
|
43 |
-
ComplexConv1d(in_, out_channels, kernel_size, padding=padding),
|
44 |
-
]
|
45 |
-
)
|
46 |
-
)
|
47 |
-
|
48 |
-
def forward(self, x: torch.tensor) -> torch.tensor:
|
49 |
-
temp = x
|
50 |
-
for idx, block in enumerate(self.blocks):
|
51 |
-
x = block(x)
|
52 |
-
|
53 |
-
if temp.size() != x.size() or self.skip_res:
|
54 |
-
return x
|
55 |
-
else:
|
56 |
-
return x + temp
|
57 |
-
|
58 |
-
|
59 |
-
class SpectrogramUnet(nn.Module):
|
60 |
-
def __init__(
|
61 |
-
self,
|
62 |
-
spec_dim: int,
|
63 |
-
hidden_dim: int,
|
64 |
-
filter_len: int,
|
65 |
-
hop_len: int,
|
66 |
-
layers: int = 3,
|
67 |
-
block_layers: int = 3,
|
68 |
-
kernel_size: int = 5,
|
69 |
-
is_mask: bool = False,
|
70 |
-
norm: str = "bn",
|
71 |
-
act: str = "tanh",
|
72 |
-
):
|
73 |
-
super().__init__()
|
74 |
-
self.layers = layers
|
75 |
-
self.is_mask = is_mask
|
76 |
-
|
77 |
-
# stft modules
|
78 |
-
self.stft = STFT(filter_len, hop_len)
|
79 |
-
|
80 |
-
if norm == "bn":
|
81 |
-
self.bn_func = nn.BatchNorm1d
|
82 |
-
elif norm == "ins":
|
83 |
-
self.bn_func = lambda x: nn.InstanceNorm1d(x, affine=True)
|
84 |
-
else:
|
85 |
-
raise NotImplementedError("{} is not implemented !".format(norm))
|
86 |
-
|
87 |
-
if act == "tanh":
|
88 |
-
self.act_func = nn.Tanh
|
89 |
-
self.act_out = nn.Tanh
|
90 |
-
elif act == "comp":
|
91 |
-
self.act_func = ComplexActLayer
|
92 |
-
self.act_out = lambda: ComplexActLayer(is_out=True)
|
93 |
-
else:
|
94 |
-
raise NotImplementedError("{} is not implemented !".format(act))
|
95 |
-
|
96 |
-
# prev conv
|
97 |
-
self.prev_conv = ComplexConv1d(spec_dim * 2, hidden_dim, 1)
|
98 |
-
|
99 |
-
# down
|
100 |
-
self.down = nn.ModuleList()
|
101 |
-
self.down_pool = nn.MaxPool1d(3, stride=2, padding=1)
|
102 |
-
for idx in range(self.layers):
|
103 |
-
block = ComplexConvBlock(
|
104 |
-
hidden_dim,
|
105 |
-
hidden_dim,
|
106 |
-
kernel_size=kernel_size,
|
107 |
-
padding=kernel_size // 2,
|
108 |
-
bn_func=self.bn_func,
|
109 |
-
act_func=self.act_func,
|
110 |
-
layers=block_layers,
|
111 |
-
)
|
112 |
-
self.down.append(block)
|
113 |
-
|
114 |
-
# up
|
115 |
-
self.up = nn.ModuleList()
|
116 |
-
for idx in range(self.layers):
|
117 |
-
in_c = hidden_dim if idx == 0 else hidden_dim * 2
|
118 |
-
self.up.append(
|
119 |
-
nn.Sequential(
|
120 |
-
ComplexConvBlock(
|
121 |
-
in_c,
|
122 |
-
hidden_dim,
|
123 |
-
kernel_size=kernel_size,
|
124 |
-
padding=kernel_size // 2,
|
125 |
-
bn_func=self.bn_func,
|
126 |
-
act_func=self.act_func,
|
127 |
-
layers=block_layers,
|
128 |
-
),
|
129 |
-
self.bn_func(hidden_dim),
|
130 |
-
self.act_func(),
|
131 |
-
ComplexTransposedConv1d(
|
132 |
-
hidden_dim, hidden_dim, kernel_size=2, stride=2
|
133 |
-
),
|
134 |
-
)
|
135 |
-
)
|
136 |
-
|
137 |
-
# out_conv
|
138 |
-
self.out_conv = nn.Sequential(
|
139 |
-
ComplexConvBlock(
|
140 |
-
hidden_dim * 2,
|
141 |
-
spec_dim * 2,
|
142 |
-
kernel_size=kernel_size,
|
143 |
-
padding=kernel_size // 2,
|
144 |
-
bn_func=self.bn_func,
|
145 |
-
act_func=self.act_func,
|
146 |
-
),
|
147 |
-
self.bn_func(spec_dim * 2),
|
148 |
-
self.act_func(),
|
149 |
-
)
|
150 |
-
|
151 |
-
# refine conv
|
152 |
-
self.refine_conv = nn.Sequential(
|
153 |
-
ComplexConvBlock(
|
154 |
-
spec_dim * 4,
|
155 |
-
spec_dim * 2,
|
156 |
-
kernel_size=kernel_size,
|
157 |
-
padding=kernel_size // 2,
|
158 |
-
bn_func=self.bn_func,
|
159 |
-
act_func=self.act_func,
|
160 |
-
),
|
161 |
-
self.bn_func(spec_dim * 2),
|
162 |
-
self.act_func(),
|
163 |
-
)
|
164 |
-
|
165 |
-
def log_stft(self, wav):
|
166 |
-
# stft
|
167 |
-
mag, phase = self.stft.transform(wav)
|
168 |
-
return torch.log(mag + 1), phase
|
169 |
-
|
170 |
-
def exp_istft(self, log_mag, phase):
|
171 |
-
# exp
|
172 |
-
mag = np.e**log_mag - 1
|
173 |
-
# istft
|
174 |
-
wav = self.stft.inverse(mag, phase)
|
175 |
-
return wav
|
176 |
-
|
177 |
-
def adjust_diff(self, x, target):
|
178 |
-
size_diff = target.size()[-1] - x.size()[-1]
|
179 |
-
assert size_diff >= 0
|
180 |
-
if size_diff > 0:
|
181 |
-
x = F.pad(
|
182 |
-
x.unsqueeze(1), (size_diff // 2, size_diff // 2), "reflect"
|
183 |
-
).squeeze(1)
|
184 |
-
return x
|
185 |
-
|
186 |
-
def masking(self, mag, phase, origin_mag, origin_phase):
|
187 |
-
abs_mag = torch.abs(mag)
|
188 |
-
mag_mask = torch.tanh(abs_mag)
|
189 |
-
phase_mask = mag / abs_mag
|
190 |
-
|
191 |
-
# masking
|
192 |
-
mag = mag_mask * origin_mag
|
193 |
-
phase = phase_mask * (origin_phase + phase)
|
194 |
-
return mag, phase
|
195 |
-
|
196 |
-
def forward(self, wav):
|
197 |
-
# stft
|
198 |
-
origin_mag, origin_phase = self.log_stft(wav)
|
199 |
-
origin_x = torch.cat([origin_mag, origin_phase], dim=1)
|
200 |
-
|
201 |
-
# prev
|
202 |
-
x = self.prev_conv(origin_x)
|
203 |
-
|
204 |
-
# body
|
205 |
-
# down
|
206 |
-
down_cache = []
|
207 |
-
for idx, block in enumerate(self.down):
|
208 |
-
x = block(x)
|
209 |
-
down_cache.append(x)
|
210 |
-
x = self.down_pool(x)
|
211 |
-
|
212 |
-
# up
|
213 |
-
for idx, block in enumerate(self.up):
|
214 |
-
x = block(x)
|
215 |
-
res = F.interpolate(
|
216 |
-
down_cache[self.layers - (idx + 1)],
|
217 |
-
size=[x.size()[2]],
|
218 |
-
mode="linear",
|
219 |
-
align_corners=False,
|
220 |
-
)
|
221 |
-
x = concat_complex(x, res, dim=1)
|
222 |
-
|
223 |
-
# match spec dimension
|
224 |
-
x = self.out_conv(x)
|
225 |
-
if origin_mag.size(2) != x.size(2):
|
226 |
-
x = F.interpolate(
|
227 |
-
x, size=[origin_mag.size(2)], mode="linear", align_corners=False
|
228 |
-
)
|
229 |
-
|
230 |
-
# refine
|
231 |
-
x = self.refine_conv(concat_complex(x, origin_x))
|
232 |
-
|
233 |
-
def to_wav(stft):
|
234 |
-
mag, phase = stft.chunk(2, 1)
|
235 |
-
if self.is_mask:
|
236 |
-
mag, phase = self.masking(mag, phase, origin_mag, origin_phase)
|
237 |
-
out = self.exp_istft(mag, phase)
|
238 |
-
out = self.adjust_diff(out, wav)
|
239 |
-
return out
|
240 |
-
|
241 |
-
refine_wav = to_wav(x)
|
242 |
-
|
243 |
-
return refine_wav
|
244 |
-
|
245 |
-
|
246 |
-
class RefineSpectrogramUnet(SpectrogramUnet):
|
247 |
-
def __init__(
|
248 |
-
self,
|
249 |
-
spec_dim: int,
|
250 |
-
hidden_dim: int,
|
251 |
-
filter_len: int,
|
252 |
-
hop_len: int,
|
253 |
-
layers: int = 4,
|
254 |
-
block_layers: int = 4,
|
255 |
-
kernel_size: int = 3,
|
256 |
-
is_mask: bool = True,
|
257 |
-
norm: str = "ins",
|
258 |
-
act: str = "comp",
|
259 |
-
refine_layers: int = 1,
|
260 |
-
add_spec_results: bool = False,
|
261 |
-
):
|
262 |
-
super().__init__(
|
263 |
-
spec_dim,
|
264 |
-
hidden_dim,
|
265 |
-
filter_len,
|
266 |
-
hop_len,
|
267 |
-
layers,
|
268 |
-
block_layers,
|
269 |
-
kernel_size,
|
270 |
-
is_mask,
|
271 |
-
norm,
|
272 |
-
act,
|
273 |
-
)
|
274 |
-
self.add_spec_results = add_spec_results
|
275 |
-
# refine conv
|
276 |
-
self.refine_conv = nn.ModuleList(
|
277 |
-
[
|
278 |
-
nn.Sequential(
|
279 |
-
ComplexConvBlock(
|
280 |
-
spec_dim * 2,
|
281 |
-
spec_dim * 2,
|
282 |
-
kernel_size=kernel_size,
|
283 |
-
padding=kernel_size // 2,
|
284 |
-
bn_func=self.bn_func,
|
285 |
-
act_func=self.act_func,
|
286 |
-
),
|
287 |
-
self.bn_func(spec_dim * 2),
|
288 |
-
self.act_func(),
|
289 |
-
)
|
290 |
-
]
|
291 |
-
* refine_layers
|
292 |
-
)
|
293 |
-
|
294 |
-
def forward(self, wav):
|
295 |
-
# stft
|
296 |
-
origin_mag, origin_phase = self.log_stft(wav)
|
297 |
-
origin_x = torch.cat([origin_mag, origin_phase], dim=1)
|
298 |
-
|
299 |
-
# prev
|
300 |
-
x = self.prev_conv(origin_x)
|
301 |
-
|
302 |
-
# body
|
303 |
-
# down
|
304 |
-
down_cache = []
|
305 |
-
for idx, block in enumerate(self.down):
|
306 |
-
x = block(x)
|
307 |
-
down_cache.append(x)
|
308 |
-
x = self.down_pool(x)
|
309 |
-
|
310 |
-
# up
|
311 |
-
for idx, block in enumerate(self.up):
|
312 |
-
x = block(x)
|
313 |
-
res = F.interpolate(
|
314 |
-
down_cache[self.layers - (idx + 1)],
|
315 |
-
size=[x.size()[2]],
|
316 |
-
mode="linear",
|
317 |
-
align_corners=False,
|
318 |
-
)
|
319 |
-
x = concat_complex(x, res, dim=1)
|
320 |
-
|
321 |
-
# match spec dimension
|
322 |
-
x = self.out_conv(x)
|
323 |
-
if origin_mag.size(2) != x.size(2):
|
324 |
-
x = F.interpolate(
|
325 |
-
x, size=[origin_mag.size(2)], mode="linear", align_corners=False
|
326 |
-
)
|
327 |
-
|
328 |
-
# refine
|
329 |
-
for idx, refine_module in enumerate(self.refine_conv):
|
330 |
-
x = refine_module(x)
|
331 |
-
mag, phase = x.chunk(2, 1)
|
332 |
-
mag, phase = self.masking(mag, phase, origin_mag, origin_phase)
|
333 |
-
if idx < len(self.refine_conv) - 1:
|
334 |
-
x = torch.cat([mag, phase], dim=1)
|
335 |
-
|
336 |
-
# clamp phase
|
337 |
-
phase = phase.clamp(-np.pi, np.pi)
|
338 |
-
|
339 |
-
out = self.exp_istft(mag, phase)
|
340 |
-
out = self.adjust_diff(out, wav)
|
341 |
-
|
342 |
-
if self.add_spec_results:
|
343 |
-
out = (out, mag, phase)
|
344 |
-
|
345 |
-
return out
|
346 |
-
|
347 |
-
|
348 |
-
class _ComplexConvNd(nn.Module):
|
349 |
-
"""
|
350 |
-
Implement Complex Convolution
|
351 |
-
A: real weight
|
352 |
-
B: img weight
|
353 |
-
"""
|
354 |
-
|
355 |
-
def __init__(
|
356 |
-
self,
|
357 |
-
in_channels,
|
358 |
-
out_channels,
|
359 |
-
kernel_size,
|
360 |
-
stride,
|
361 |
-
padding,
|
362 |
-
dilation,
|
363 |
-
transposed,
|
364 |
-
output_padding,
|
365 |
-
):
|
366 |
-
super().__init__()
|
367 |
-
self.in_channels = in_channels
|
368 |
-
self.out_channels = out_channels
|
369 |
-
self.kernel_size = kernel_size
|
370 |
-
self.stride = stride
|
371 |
-
self.padding = padding
|
372 |
-
self.dilation = dilation
|
373 |
-
self.output_padding = output_padding
|
374 |
-
self.transposed = transposed
|
375 |
-
|
376 |
-
self.A = self.make_weight(in_channels, out_channels, kernel_size)
|
377 |
-
self.B = self.make_weight(in_channels, out_channels, kernel_size)
|
378 |
-
|
379 |
-
self.reset_parameters()
|
380 |
-
|
381 |
-
def make_weight(self, in_ch, out_ch, kernel_size):
|
382 |
-
if self.transposed:
|
383 |
-
tensor = nn.Parameter(torch.Tensor(in_ch, out_ch // 2, *kernel_size))
|
384 |
-
else:
|
385 |
-
tensor = nn.Parameter(torch.Tensor(out_ch, in_ch // 2, *kernel_size))
|
386 |
-
return tensor
|
387 |
-
|
388 |
-
def reset_parameters(self):
|
389 |
-
# init real weight
|
390 |
-
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.A)
|
391 |
-
|
392 |
-
# init A
|
393 |
-
gain = calculate_gain("leaky_relu", 0)
|
394 |
-
std = gain / np.sqrt(fan_in)
|
395 |
-
bound = np.sqrt(3.0) * std
|
396 |
-
|
397 |
-
with torch.no_grad():
|
398 |
-
# TODO: find more stable initial values
|
399 |
-
self.A.uniform_(-bound * (1 / (np.pi**2)), bound * (1 / (np.pi**2)))
|
400 |
-
#
|
401 |
-
# B is initialized by pi
|
402 |
-
# -pi and pi is too big, so it is powed by -1
|
403 |
-
self.B.uniform_(-1 / np.pi, 1 / np.pi)
|
404 |
-
|
405 |
-
|
406 |
-
class ComplexConv1d(_ComplexConvNd):
|
407 |
-
"""
|
408 |
-
Complex Convolution 1d
|
409 |
-
"""
|
410 |
-
|
411 |
-
def __init__(
|
412 |
-
self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1
|
413 |
-
):
|
414 |
-
kernel_size = single(kernel_size)
|
415 |
-
stride = single(stride)
|
416 |
-
# edit padding
|
417 |
-
padding = padding
|
418 |
-
dilation = single(dilation)
|
419 |
-
super(ComplexConv1d, self).__init__(
|
420 |
-
in_channels,
|
421 |
-
out_channels,
|
422 |
-
kernel_size,
|
423 |
-
stride,
|
424 |
-
padding,
|
425 |
-
dilation,
|
426 |
-
False,
|
427 |
-
single(0),
|
428 |
-
)
|
429 |
-
|
430 |
-
def forward(self, x):
|
431 |
-
"""
|
432 |
-
Implemented complex convolution using combining 'grouped convolution' and
|
433 |
-
'real / img weight'
|
434 |
-
:param x: data (N, C, T) C is concatenated with C/2 real channels and C/2 idea channels
|
435 |
-
:return: complex conved result
|
436 |
-
"""
|
437 |
-
# adopt reflect padding
|
438 |
-
if self.padding:
|
439 |
-
x = F.pad(x, (self.padding, self.padding), "reflect")
|
440 |
-
|
441 |
-
# forward real
|
442 |
-
real_part = F.conv1d(
|
443 |
-
x,
|
444 |
-
self.A,
|
445 |
-
None,
|
446 |
-
stride=self.stride,
|
447 |
-
padding=0,
|
448 |
-
dilation=self.dilation,
|
449 |
-
groups=2,
|
450 |
-
)
|
451 |
-
|
452 |
-
# forward idea
|
453 |
-
spl = self.in_channels // 2
|
454 |
-
weight_B = torch.cat([self.B[:spl].data * (-1), self.B[spl:].data])
|
455 |
-
idea_part = F.conv1d(
|
456 |
-
x,
|
457 |
-
weight_B,
|
458 |
-
None,
|
459 |
-
stride=self.stride,
|
460 |
-
padding=0,
|
461 |
-
dilation=self.dilation,
|
462 |
-
groups=2,
|
463 |
-
)
|
464 |
-
|
465 |
-
return real_part + idea_part
|
466 |
-
|
467 |
-
|
468 |
-
class ComplexTransposedConv1d(_ComplexConvNd):
|
469 |
-
"""
|
470 |
-
Complex Transposed Convolution 1d
|
471 |
-
"""
|
472 |
-
|
473 |
-
def __init__(
|
474 |
-
self,
|
475 |
-
in_channels,
|
476 |
-
out_channels,
|
477 |
-
kernel_size,
|
478 |
-
stride=1,
|
479 |
-
padding=0,
|
480 |
-
output_padding=0,
|
481 |
-
dilation=1,
|
482 |
-
):
|
483 |
-
kernel_size = single(kernel_size)
|
484 |
-
stride = single(stride)
|
485 |
-
padding = padding
|
486 |
-
dilation = single(dilation)
|
487 |
-
super().__init__(
|
488 |
-
in_channels,
|
489 |
-
out_channels,
|
490 |
-
kernel_size,
|
491 |
-
stride,
|
492 |
-
padding,
|
493 |
-
dilation,
|
494 |
-
True,
|
495 |
-
output_padding,
|
496 |
-
)
|
497 |
-
|
498 |
-
def forward(self, x, output_size=None):
|
499 |
-
"""
|
500 |
-
Implemented complex transposed convolution using combining 'grouped convolution'
|
501 |
-
and 'real / img weight'
|
502 |
-
:param x: data (N, C, T) C is concatenated with C/2 real channels and C/2 idea channels
|
503 |
-
:return: complex transposed convolution result
|
504 |
-
"""
|
505 |
-
# forward real
|
506 |
-
if self.padding:
|
507 |
-
x = F.pad(x, (self.padding, self.padding), "reflect")
|
508 |
-
|
509 |
-
real_part = F.conv_transpose1d(
|
510 |
-
x,
|
511 |
-
self.A,
|
512 |
-
None,
|
513 |
-
stride=self.stride,
|
514 |
-
padding=0,
|
515 |
-
dilation=self.dilation,
|
516 |
-
groups=2,
|
517 |
-
)
|
518 |
-
|
519 |
-
# forward idea
|
520 |
-
spl = self.out_channels // 2
|
521 |
-
weight_B = torch.cat([self.B[:spl] * (-1), self.B[spl:]])
|
522 |
-
idea_part = F.conv_transpose1d(
|
523 |
-
x,
|
524 |
-
weight_B,
|
525 |
-
None,
|
526 |
-
stride=self.stride,
|
527 |
-
padding=0,
|
528 |
-
dilation=self.dilation,
|
529 |
-
groups=2,
|
530 |
-
)
|
531 |
-
|
532 |
-
if self.output_padding:
|
533 |
-
real_part = F.pad(
|
534 |
-
real_part, (self.output_padding, self.output_padding), "reflect"
|
535 |
-
)
|
536 |
-
idea_part = F.pad(
|
537 |
-
idea_part, (self.output_padding, self.output_padding), "reflect"
|
538 |
-
)
|
539 |
-
|
540 |
-
return real_part + idea_part
|
541 |
-
|
542 |
-
|
543 |
-
class ComplexActLayer(nn.Module):
|
544 |
-
"""
|
545 |
-
Activation differently 'real' part and 'img' part
|
546 |
-
In implemented DCUnet on this repository, Real part is activated to log space.
|
547 |
-
And Phase(img) part, it is distributed in [-pi, pi]...
|
548 |
-
"""
|
549 |
-
|
550 |
-
def forward(self, x):
|
551 |
-
real, img = x.chunk(2, 1)
|
552 |
-
return torch.cat([F.leaky_relu(real), torch.tanh(img) * np.pi], dim=1)
|
553 |
-
|
554 |
-
|
555 |
-
class STFT(nn.Module):
|
556 |
-
"""
|
557 |
-
Re-construct stft for calculating backward operation
|
558 |
-
refer on : https://github.com/pseeth/torch-stft/blob/master/torch_stft/stft.py
|
559 |
-
"""
|
560 |
-
|
561 |
-
def __init__(
|
562 |
-
self,
|
563 |
-
filter_length: int = 1024,
|
564 |
-
hop_length: int = 512,
|
565 |
-
win_length: int = None,
|
566 |
-
window: str = "hann",
|
567 |
-
):
|
568 |
-
super().__init__()
|
569 |
-
self.filter_length = filter_length
|
570 |
-
self.hop_length = hop_length
|
571 |
-
self.win_length = win_length if win_length else filter_length
|
572 |
-
self.window = window
|
573 |
-
self.pad_amount = self.filter_length // 2
|
574 |
-
|
575 |
-
# make fft window
|
576 |
-
assert filter_length >= self.win_length
|
577 |
-
# get window and zero center pad it to filter_length
|
578 |
-
fft_window = get_window(window, self.win_length, fftbins=True)
|
579 |
-
fft_window = pad_center(fft_window, filter_length)
|
580 |
-
fft_window = torch.from_numpy(fft_window).float()
|
581 |
-
|
582 |
-
# calculate fourer_basis
|
583 |
-
cut_off = int((self.filter_length / 2 + 1))
|
584 |
-
fourier_basis = np.fft.fft(np.eye(self.filter_length))
|
585 |
-
fourier_basis = np.vstack(
|
586 |
-
[np.real(fourier_basis[:cut_off, :]), np.imag(fourier_basis[:cut_off, :])]
|
587 |
-
)
|
588 |
-
|
589 |
-
# make forward & inverse basis
|
590 |
-
self.register_buffer("square_window", fft_window**2)
|
591 |
-
|
592 |
-
forward_basis = torch.FloatTensor(fourier_basis[:, np.newaxis, :]) * fft_window
|
593 |
-
inverse_basis = (
|
594 |
-
torch.FloatTensor(
|
595 |
-
np.linalg.pinv(self.filter_length / self.hop_length * fourier_basis).T[
|
596 |
-
:, np.newaxis, :
|
597 |
-
]
|
598 |
-
)
|
599 |
-
* fft_window
|
600 |
-
)
|
601 |
-
# torch.pinverse has a bug, so at this time, it is separated into two parts..
|
602 |
-
self.register_buffer("forward_basis", forward_basis)
|
603 |
-
self.register_buffer("inverse_basis", inverse_basis)
|
604 |
-
|
605 |
-
def transform(self, wav: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
606 |
-
# reflect padding
|
607 |
-
wav = wav.unsqueeze(1).unsqueeze(1)
|
608 |
-
wav = F.pad(
|
609 |
-
wav, (self.pad_amount, self.pad_amount, 0, 0), mode="reflect"
|
610 |
-
).squeeze(1)
|
611 |
-
|
612 |
-
# conv
|
613 |
-
forward_trans = F.conv1d(
|
614 |
-
wav, self.forward_basis, stride=self.hop_length, padding=0
|
615 |
-
)
|
616 |
-
real_part, imag_part = forward_trans.chunk(2, 1)
|
617 |
-
|
618 |
-
return torch.sqrt(real_part**2 + imag_part**2), torch.atan2(
|
619 |
-
imag_part.data, real_part.data
|
620 |
-
)
|
621 |
-
|
622 |
-
def inverse(
|
623 |
-
self, magnitude: torch.Tensor, phase: torch.Tensor, eps: float = 1e-9
|
624 |
-
) -> torch.Tensor:
|
625 |
-
comp = torch.cat(
|
626 |
-
[magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
|
627 |
-
)
|
628 |
-
inverse_transform = F.conv_transpose1d(
|
629 |
-
comp, self.inverse_basis, stride=self.hop_length, padding=0
|
630 |
-
)
|
631 |
-
|
632 |
-
# remove window effect
|
633 |
-
n_frames = comp.size(-1)
|
634 |
-
inverse_size = inverse_transform.size(-1)
|
635 |
-
|
636 |
-
window_filter = torch.ones(1, 1, n_frames).type_as(inverse_transform)
|
637 |
-
|
638 |
-
weight = self.square_window[: self.filter_length].unsqueeze(0).unsqueeze(0)
|
639 |
-
window_filter = F.conv_transpose1d(
|
640 |
-
window_filter, weight, stride=self.hop_length, padding=0
|
641 |
-
)
|
642 |
-
window_filter = window_filter.squeeze()[:inverse_size] + eps
|
643 |
-
|
644 |
-
inverse_transform /= window_filter
|
645 |
-
|
646 |
-
# scale by hop ratio
|
647 |
-
inverse_transform *= self.filter_length / self.hop_length
|
648 |
-
|
649 |
-
return inverse_transform[..., self.pad_amount : -self.pad_amount].squeeze(1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
remfx/dptnet.py
DELETED
@@ -1,459 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
import torch.nn as nn
|
3 |
-
import torch.nn.functional as F
|
4 |
-
from torch.nn.modules.container import ModuleList
|
5 |
-
from torch.nn.modules.activation import MultiheadAttention
|
6 |
-
from torch.nn.modules.dropout import Dropout
|
7 |
-
from torch.nn.modules.linear import Linear
|
8 |
-
from torch.nn.modules.rnn import LSTM
|
9 |
-
from torch.nn.modules.normalization import LayerNorm
|
10 |
-
from torch.autograd import Variable
|
11 |
-
import copy
|
12 |
-
import math
|
13 |
-
|
14 |
-
|
15 |
-
# adapted from https://github.com/ujscjj/DPTNet
|
16 |
-
|
17 |
-
|
18 |
-
class DPTNet_base(nn.Module):
|
19 |
-
def __init__(
|
20 |
-
self,
|
21 |
-
enc_dim,
|
22 |
-
feature_dim,
|
23 |
-
hidden_dim,
|
24 |
-
layer,
|
25 |
-
segment_size=250,
|
26 |
-
nspk=2,
|
27 |
-
win_len=2,
|
28 |
-
):
|
29 |
-
super().__init__()
|
30 |
-
# parameters
|
31 |
-
self.window = win_len
|
32 |
-
self.stride = self.window // 2
|
33 |
-
|
34 |
-
self.enc_dim = enc_dim
|
35 |
-
self.feature_dim = feature_dim
|
36 |
-
self.hidden_dim = hidden_dim
|
37 |
-
self.segment_size = segment_size
|
38 |
-
|
39 |
-
self.layer = layer
|
40 |
-
self.num_spk = nspk
|
41 |
-
self.eps = 1e-8
|
42 |
-
|
43 |
-
self.dpt_encoder = DPTEncoder(
|
44 |
-
n_filters=enc_dim,
|
45 |
-
window_size=win_len,
|
46 |
-
)
|
47 |
-
self.enc_LN = nn.GroupNorm(1, self.enc_dim, eps=1e-8)
|
48 |
-
self.dpt_separation = DPTSeparation(
|
49 |
-
self.enc_dim,
|
50 |
-
self.feature_dim,
|
51 |
-
self.hidden_dim,
|
52 |
-
self.num_spk,
|
53 |
-
self.layer,
|
54 |
-
self.segment_size,
|
55 |
-
)
|
56 |
-
|
57 |
-
self.mask_conv1x1 = nn.Conv1d(self.feature_dim, self.enc_dim, 1, bias=False)
|
58 |
-
self.decoder = DPTDecoder(n_filters=enc_dim, window_size=win_len)
|
59 |
-
|
60 |
-
def forward(self, mix):
|
61 |
-
"""
|
62 |
-
mix: shape (batch, T)
|
63 |
-
"""
|
64 |
-
batch_size = mix.shape[0]
|
65 |
-
mix = self.dpt_encoder(mix) # (B, E, L)
|
66 |
-
|
67 |
-
score_ = self.enc_LN(mix) # B, E, L
|
68 |
-
score_ = self.dpt_separation(score_) # B, nspk, T, N
|
69 |
-
score_ = (
|
70 |
-
score_.view(batch_size * self.num_spk, -1, self.feature_dim)
|
71 |
-
.transpose(1, 2)
|
72 |
-
.contiguous()
|
73 |
-
) # B*nspk, N, T
|
74 |
-
score = self.mask_conv1x1(score_) # [B*nspk, N, L] -> [B*nspk, E, L]
|
75 |
-
score = score.view(
|
76 |
-
batch_size, self.num_spk, self.enc_dim, -1
|
77 |
-
) # [B*nspk, E, L] -> [B, nspk, E, L]
|
78 |
-
est_mask = F.relu(score)
|
79 |
-
|
80 |
-
est_source = self.decoder(
|
81 |
-
mix, est_mask
|
82 |
-
) # [B, E, L] + [B, nspk, E, L]--> [B, nspk, T]
|
83 |
-
|
84 |
-
return est_source
|
85 |
-
|
86 |
-
|
87 |
-
class DPTEncoder(nn.Module):
|
88 |
-
def __init__(self, n_filters: int = 64, window_size: int = 2):
|
89 |
-
super().__init__()
|
90 |
-
self.conv = nn.Conv1d(
|
91 |
-
1, n_filters, kernel_size=window_size, stride=window_size // 2, bias=False
|
92 |
-
)
|
93 |
-
|
94 |
-
def forward(self, x):
|
95 |
-
x = x.unsqueeze(1)
|
96 |
-
x = F.relu(self.conv(x))
|
97 |
-
return x
|
98 |
-
|
99 |
-
|
100 |
-
class TransformerEncoderLayer(torch.nn.Module):
|
101 |
-
def __init__(
|
102 |
-
self, d_model, nhead, hidden_size, dim_feedforward, dropout, activation="relu"
|
103 |
-
):
|
104 |
-
super(TransformerEncoderLayer, self).__init__()
|
105 |
-
self.self_attn = MultiheadAttention(d_model, nhead, dropout=dropout)
|
106 |
-
|
107 |
-
# Implementation of improved part
|
108 |
-
self.lstm = LSTM(d_model, hidden_size, 1, bidirectional=True)
|
109 |
-
self.dropout = Dropout(dropout)
|
110 |
-
self.linear = Linear(hidden_size * 2, d_model)
|
111 |
-
|
112 |
-
self.norm1 = LayerNorm(d_model)
|
113 |
-
self.norm2 = LayerNorm(d_model)
|
114 |
-
self.dropout1 = Dropout(dropout)
|
115 |
-
self.dropout2 = Dropout(dropout)
|
116 |
-
|
117 |
-
self.activation = _get_activation_fn(activation)
|
118 |
-
|
119 |
-
def __setstate__(self, state):
|
120 |
-
if "activation" not in state:
|
121 |
-
state["activation"] = F.relu
|
122 |
-
super(TransformerEncoderLayer, self).__setstate__(state)
|
123 |
-
|
124 |
-
def forward(self, src, src_mask=None, src_key_padding_mask=None):
|
125 |
-
r"""Pass the input through the encoder layer.
|
126 |
-
Args:
|
127 |
-
src: the sequnce to the encoder layer (required).
|
128 |
-
src_mask: the mask for the src sequence (optional).
|
129 |
-
src_key_padding_mask: the mask for the src keys per batch (optional).
|
130 |
-
Shape:
|
131 |
-
see the docs in Transformer class.
|
132 |
-
"""
|
133 |
-
src2 = self.self_attn(
|
134 |
-
src, src, src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask
|
135 |
-
)[0]
|
136 |
-
src = src + self.dropout1(src2)
|
137 |
-
src = self.norm1(src)
|
138 |
-
src2 = self.linear(self.dropout(self.activation(self.lstm(src)[0])))
|
139 |
-
src = src + self.dropout2(src2)
|
140 |
-
src = self.norm2(src)
|
141 |
-
return src
|
142 |
-
|
143 |
-
|
144 |
-
def _get_clones(module, N):
|
145 |
-
return ModuleList([copy.deepcopy(module) for i in range(N)])
|
146 |
-
|
147 |
-
|
148 |
-
def _get_activation_fn(activation):
|
149 |
-
if activation == "relu":
|
150 |
-
return F.relu
|
151 |
-
elif activation == "gelu":
|
152 |
-
return F.gelu
|
153 |
-
|
154 |
-
raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
|
155 |
-
|
156 |
-
|
157 |
-
class SingleTransformer(nn.Module):
|
158 |
-
"""
|
159 |
-
Container module for a single Transformer layer.
|
160 |
-
args: input_size: int, dimension of the input feature.
|
161 |
-
The input should have shape (batch, seq_len, input_size).
|
162 |
-
"""
|
163 |
-
|
164 |
-
def __init__(self, input_size, hidden_size, dropout):
|
165 |
-
super(SingleTransformer, self).__init__()
|
166 |
-
self.transformer = TransformerEncoderLayer(
|
167 |
-
d_model=input_size,
|
168 |
-
nhead=4,
|
169 |
-
hidden_size=hidden_size,
|
170 |
-
dim_feedforward=hidden_size * 2,
|
171 |
-
dropout=dropout,
|
172 |
-
)
|
173 |
-
|
174 |
-
def forward(self, input):
|
175 |
-
# input shape: batch, seq, dim
|
176 |
-
output = input
|
177 |
-
transformer_output = (
|
178 |
-
self.transformer(output.permute(1, 0, 2).contiguous())
|
179 |
-
.permute(1, 0, 2)
|
180 |
-
.contiguous()
|
181 |
-
)
|
182 |
-
return transformer_output
|
183 |
-
|
184 |
-
|
185 |
-
# dual-path transformer
|
186 |
-
class DPT(nn.Module):
|
187 |
-
"""
|
188 |
-
Deep dual-path transformer.
|
189 |
-
args:
|
190 |
-
input_size: int, dimension of the input feature. The input should have shape
|
191 |
-
(batch, seq_len, input_size).
|
192 |
-
hidden_size: int, dimension of the hidden state.
|
193 |
-
output_size: int, dimension of the output size.
|
194 |
-
num_layers: int, number of stacked Transformer layers. Default is 1.
|
195 |
-
dropout: float, dropout ratio. Default is 0.
|
196 |
-
"""
|
197 |
-
|
198 |
-
def __init__(self, input_size, hidden_size, output_size, num_layers=1, dropout=0):
|
199 |
-
super(DPT, self).__init__()
|
200 |
-
|
201 |
-
self.input_size = input_size
|
202 |
-
self.output_size = output_size
|
203 |
-
self.hidden_size = hidden_size
|
204 |
-
|
205 |
-
# dual-path transformer
|
206 |
-
self.row_transformer = nn.ModuleList([])
|
207 |
-
self.col_transformer = nn.ModuleList([])
|
208 |
-
for i in range(num_layers):
|
209 |
-
self.row_transformer.append(
|
210 |
-
SingleTransformer(input_size, hidden_size, dropout)
|
211 |
-
)
|
212 |
-
self.col_transformer.append(
|
213 |
-
SingleTransformer(input_size, hidden_size, dropout)
|
214 |
-
)
|
215 |
-
|
216 |
-
# output layer
|
217 |
-
self.output = nn.Sequential(nn.PReLU(), nn.Conv2d(input_size, output_size, 1))
|
218 |
-
|
219 |
-
def forward(self, input):
|
220 |
-
# input shape: batch, N, dim1, dim2
|
221 |
-
# apply transformer on dim1 first and then dim2
|
222 |
-
# output shape: B, output_size, dim1, dim2
|
223 |
-
# input = input.to(device)
|
224 |
-
batch_size, _, dim1, dim2 = input.shape
|
225 |
-
output = input
|
226 |
-
for i in range(len(self.row_transformer)):
|
227 |
-
row_input = (
|
228 |
-
output.permute(0, 3, 2, 1)
|
229 |
-
.contiguous()
|
230 |
-
.view(batch_size * dim2, dim1, -1)
|
231 |
-
) # B*dim2, dim1, N
|
232 |
-
row_output = self.row_transformer[i](row_input) # B*dim2, dim1, H
|
233 |
-
row_output = (
|
234 |
-
row_output.view(batch_size, dim2, dim1, -1)
|
235 |
-
.permute(0, 3, 2, 1)
|
236 |
-
.contiguous()
|
237 |
-
) # B, N, dim1, dim2
|
238 |
-
output = row_output
|
239 |
-
|
240 |
-
col_input = (
|
241 |
-
output.permute(0, 2, 3, 1)
|
242 |
-
.contiguous()
|
243 |
-
.view(batch_size * dim1, dim2, -1)
|
244 |
-
) # B*dim1, dim2, N
|
245 |
-
col_output = self.col_transformer[i](col_input) # B*dim1, dim2, H
|
246 |
-
col_output = (
|
247 |
-
col_output.view(batch_size, dim1, dim2, -1)
|
248 |
-
.permute(0, 3, 1, 2)
|
249 |
-
.contiguous()
|
250 |
-
) # B, N, dim1, dim2
|
251 |
-
output = col_output
|
252 |
-
|
253 |
-
output = self.output(output) # B, output_size, dim1, dim2
|
254 |
-
|
255 |
-
return output
|
256 |
-
|
257 |
-
|
258 |
-
# base module for deep DPT
|
259 |
-
class DPT_base(nn.Module):
|
260 |
-
def __init__(
|
261 |
-
self, input_dim, feature_dim, hidden_dim, num_spk=2, layer=6, segment_size=250
|
262 |
-
):
|
263 |
-
super(DPT_base, self).__init__()
|
264 |
-
|
265 |
-
self.input_dim = input_dim
|
266 |
-
self.feature_dim = feature_dim
|
267 |
-
self.hidden_dim = hidden_dim
|
268 |
-
|
269 |
-
self.layer = layer
|
270 |
-
self.segment_size = segment_size
|
271 |
-
self.num_spk = num_spk
|
272 |
-
|
273 |
-
self.eps = 1e-8
|
274 |
-
|
275 |
-
# bottleneck
|
276 |
-
self.BN = nn.Conv1d(self.input_dim, self.feature_dim, 1, bias=False)
|
277 |
-
|
278 |
-
# DPT model
|
279 |
-
self.DPT = DPT(
|
280 |
-
self.feature_dim,
|
281 |
-
self.hidden_dim,
|
282 |
-
self.feature_dim * self.num_spk,
|
283 |
-
num_layers=layer,
|
284 |
-
)
|
285 |
-
|
286 |
-
def pad_segment(self, input, segment_size):
|
287 |
-
# input is the features: (B, N, T)
|
288 |
-
batch_size, dim, seq_len = input.shape
|
289 |
-
segment_stride = segment_size // 2
|
290 |
-
|
291 |
-
rest = segment_size - (segment_stride + seq_len % segment_size) % segment_size
|
292 |
-
if rest > 0:
|
293 |
-
pad = Variable(torch.zeros(batch_size, dim, rest)).type(input.type())
|
294 |
-
input = torch.cat([input, pad], 2)
|
295 |
-
|
296 |
-
pad_aux = Variable(torch.zeros(batch_size, dim, segment_stride)).type(
|
297 |
-
input.type()
|
298 |
-
)
|
299 |
-
input = torch.cat([pad_aux, input, pad_aux], 2)
|
300 |
-
|
301 |
-
return input, rest
|
302 |
-
|
303 |
-
def split_feature(self, input, segment_size):
|
304 |
-
# split the feature into chunks of segment size
|
305 |
-
# input is the features: (B, N, T)
|
306 |
-
|
307 |
-
input, rest = self.pad_segment(input, segment_size)
|
308 |
-
batch_size, dim, seq_len = input.shape
|
309 |
-
segment_stride = segment_size // 2
|
310 |
-
|
311 |
-
segments1 = (
|
312 |
-
input[:, :, :-segment_stride]
|
313 |
-
.contiguous()
|
314 |
-
.view(batch_size, dim, -1, segment_size)
|
315 |
-
)
|
316 |
-
segments2 = (
|
317 |
-
input[:, :, segment_stride:]
|
318 |
-
.contiguous()
|
319 |
-
.view(batch_size, dim, -1, segment_size)
|
320 |
-
)
|
321 |
-
segments = (
|
322 |
-
torch.cat([segments1, segments2], 3)
|
323 |
-
.view(batch_size, dim, -1, segment_size)
|
324 |
-
.transpose(2, 3)
|
325 |
-
)
|
326 |
-
|
327 |
-
return segments.contiguous(), rest
|
328 |
-
|
329 |
-
def merge_feature(self, input, rest):
|
330 |
-
# merge the splitted features into full utterance
|
331 |
-
# input is the features: (B, N, L, K)
|
332 |
-
|
333 |
-
batch_size, dim, segment_size, _ = input.shape
|
334 |
-
segment_stride = segment_size // 2
|
335 |
-
input = (
|
336 |
-
input.transpose(2, 3)
|
337 |
-
.contiguous()
|
338 |
-
.view(batch_size, dim, -1, segment_size * 2)
|
339 |
-
) # B, N, K, L
|
340 |
-
|
341 |
-
input1 = (
|
342 |
-
input[:, :, :, :segment_size]
|
343 |
-
.contiguous()
|
344 |
-
.view(batch_size, dim, -1)[:, :, segment_stride:]
|
345 |
-
)
|
346 |
-
input2 = (
|
347 |
-
input[:, :, :, segment_size:]
|
348 |
-
.contiguous()
|
349 |
-
.view(batch_size, dim, -1)[:, :, :-segment_stride]
|
350 |
-
)
|
351 |
-
|
352 |
-
output = input1 + input2
|
353 |
-
if rest > 0:
|
354 |
-
output = output[:, :, :-rest]
|
355 |
-
|
356 |
-
return output.contiguous() # B, N, T
|
357 |
-
|
358 |
-
def forward(self, input):
|
359 |
-
pass
|
360 |
-
|
361 |
-
|
362 |
-
class DPTSeparation(DPT_base):
|
363 |
-
def __init__(self, *args, **kwargs):
|
364 |
-
super(DPTSeparation, self).__init__(*args, **kwargs)
|
365 |
-
|
366 |
-
# gated output layer
|
367 |
-
self.output = nn.Sequential(
|
368 |
-
nn.Conv1d(self.feature_dim, self.feature_dim, 1), nn.Tanh()
|
369 |
-
)
|
370 |
-
self.output_gate = nn.Sequential(
|
371 |
-
nn.Conv1d(self.feature_dim, self.feature_dim, 1), nn.Sigmoid()
|
372 |
-
)
|
373 |
-
|
374 |
-
def forward(self, input):
|
375 |
-
# input = input.to(device)
|
376 |
-
# input: (B, E, T)
|
377 |
-
batch_size, E, seq_length = input.shape
|
378 |
-
|
379 |
-
enc_feature = self.BN(input) # (B, E, L)-->(B, N, L)
|
380 |
-
# split the encoder output into overlapped, longer segments
|
381 |
-
enc_segments, enc_rest = self.split_feature(
|
382 |
-
enc_feature, self.segment_size
|
383 |
-
) # B, N, L, K: L is the segment_size
|
384 |
-
# print('enc_segments.shape {}'.format(enc_segments.shape))
|
385 |
-
# pass to DPT
|
386 |
-
output = self.DPT(enc_segments).view(
|
387 |
-
batch_size * self.num_spk, self.feature_dim, self.segment_size, -1
|
388 |
-
) # B*nspk, N, L, K
|
389 |
-
|
390 |
-
# overlap-and-add of the outputs
|
391 |
-
output = self.merge_feature(output, enc_rest) # B*nspk, N, T
|
392 |
-
|
393 |
-
# gated output layer for filter generation
|
394 |
-
bf_filter = self.output(output) * self.output_gate(output) # B*nspk, K, T
|
395 |
-
bf_filter = (
|
396 |
-
bf_filter.transpose(1, 2)
|
397 |
-
.contiguous()
|
398 |
-
.view(batch_size, self.num_spk, -1, self.feature_dim)
|
399 |
-
) # B, nspk, T, N
|
400 |
-
|
401 |
-
return bf_filter
|
402 |
-
|
403 |
-
|
404 |
-
class DPTDecoder(nn.Module):
|
405 |
-
def __init__(self, n_filters: int = 64, window_size: int = 2):
|
406 |
-
super().__init__()
|
407 |
-
self.W = window_size
|
408 |
-
self.basis_signals = nn.Linear(n_filters, window_size, bias=False)
|
409 |
-
|
410 |
-
def forward(self, mixture, mask):
|
411 |
-
"""
|
412 |
-
mixture: (batch, n_filters, L)
|
413 |
-
mask: (batch, sources, n_filters, L)
|
414 |
-
"""
|
415 |
-
source_w = torch.unsqueeze(mixture, 1) * mask # [B, C, E, L]
|
416 |
-
source_w = torch.transpose(source_w, 2, 3) # [B, C, L, E]
|
417 |
-
# S = DV
|
418 |
-
est_source = self.basis_signals(source_w) # [B, C, L, W]
|
419 |
-
est_source = overlap_and_add(est_source, self.W // 2) # B x C x T
|
420 |
-
return est_source
|
421 |
-
|
422 |
-
|
423 |
-
def overlap_and_add(signal, frame_step):
|
424 |
-
"""Reconstructs a signal from a framed representation.
|
425 |
-
Adds potentially overlapping frames of a signal with shape
|
426 |
-
`[..., frames, frame_length]`, offsetting subsequent frames by `frame_step`.
|
427 |
-
The resulting tensor has shape `[..., output_size]` where
|
428 |
-
output_size = (frames - 1) * frame_step + frame_length
|
429 |
-
Args:
|
430 |
-
signal: A [..., frames, frame_length] Tensor.
|
431 |
-
All dimensions may be unknown, and rank must be at least 2.
|
432 |
-
frame_step: An integer denoting overlap offsets. Must be less than or equal to frame_length.
|
433 |
-
Returns:
|
434 |
-
A Tensor with shape [..., output_size] containing the overlap-added frames of signal's
|
435 |
-
inner-most two dimensions.
|
436 |
-
output_size = (frames - 1) * frame_step + frame_length
|
437 |
-
Based on https://github.com/tensorflow/tensorflow/blob/r1.12/tensorflow/contrib/signal/python/ops/reconstruction_ops.py
|
438 |
-
"""
|
439 |
-
outer_dimensions = signal.size()[:-2]
|
440 |
-
frames, frame_length = signal.size()[-2:]
|
441 |
-
|
442 |
-
subframe_length = math.gcd(frame_length, frame_step) # gcd=Greatest Common Divisor
|
443 |
-
subframe_step = frame_step // subframe_length
|
444 |
-
subframes_per_frame = frame_length // subframe_length
|
445 |
-
output_size = frame_step * (frames - 1) + frame_length
|
446 |
-
output_subframes = output_size // subframe_length
|
447 |
-
|
448 |
-
subframe_signal = signal.reshape(*outer_dimensions, -1, subframe_length)
|
449 |
-
|
450 |
-
frame = torch.arange(0, output_subframes).unfold(
|
451 |
-
0, subframes_per_frame, subframe_step
|
452 |
-
)
|
453 |
-
frame = signal.new_tensor(frame).long() # signal may in GPU or CPU
|
454 |
-
frame = frame.contiguous().view(-1)
|
455 |
-
|
456 |
-
result = signal.new_zeros(*outer_dimensions, output_subframes, subframe_length)
|
457 |
-
result.index_add_(-2, frame, subframe_signal)
|
458 |
-
result = result.view(*outer_dimensions, -1)
|
459 |
-
return result
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
remfx/models.py
CHANGED
@@ -226,7 +226,7 @@ class DCUNetModel(nn.Module):
|
|
226 |
|
227 |
def forward(self, batch):
|
228 |
x, target = batch
|
229 |
-
output = self.model(x.squeeze(1)) # B x
|
230 |
# Crop target to match output
|
231 |
if output.shape[-1] < target.shape[-1]:
|
232 |
target = causal_crop(target, output.shape[-1])
|
@@ -234,7 +234,7 @@ class DCUNetModel(nn.Module):
|
|
234 |
return loss, output
|
235 |
|
236 |
def sample(self, x: Tensor) -> Tensor:
|
237 |
-
output = self.model(x.squeeze(1)) # B x
|
238 |
return output
|
239 |
|
240 |
|
|
|
226 |
|
227 |
def forward(self, batch):
|
228 |
x, target = batch
|
229 |
+
output = self.model(x.squeeze(1)) # B x T
|
230 |
# Crop target to match output
|
231 |
if output.shape[-1] < target.shape[-1]:
|
232 |
target = causal_crop(target, output.shape[-1])
|
|
|
234 |
return loss, output
|
235 |
|
236 |
def sample(self, x: Tensor) -> Tensor:
|
237 |
+
output = self.model(x.squeeze(1)) # B x T
|
238 |
return output
|
239 |
|
240 |
|