Spaces:
Runtime error
Runtime error
Commit
·
7d6f241
1
Parent(s):
e4fc05d
Init dcunet and dptnet
Browse files- cfg/model/audio_diffusion.yaml +2 -2
- cfg/model/classifier.yaml +1 -1
- cfg/model/dcunet.yaml +22 -0
- cfg/model/demucs.yaml +1 -1
- cfg/model/dptnet.yaml +18 -0
- cfg/model/umx.yaml +2 -2
- remfx/cnn14.py +138 -0
- remfx/datasets.py +1 -0
- remfx/dcunet.py +649 -0
- remfx/dptnet.py +460 -0
- remfx/models.py +39 -199
- remfx/utils.py +92 -0
- scripts/test.py +0 -1
cfg/model/audio_diffusion.yaml
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
# @package _global_
|
2 |
-
model:
|
3 |
-
_target_: remfx.models.
|
4 |
lr: 1e-4
|
5 |
lr_beta1: 0.95
|
6 |
lr_beta2: 0.999
|
|
|
1 |
# @package _global_
|
2 |
+
model:
|
3 |
+
_target_: remfx.models.RemFx
|
4 |
lr: 1e-4
|
5 |
lr_beta1: 0.95
|
6 |
lr_beta2: 0.999
|
cfg/model/classifier.yaml
CHANGED
@@ -5,7 +5,7 @@ model:
|
|
5 |
lr_weight_decay: 1e-3
|
6 |
sample_rate: ${sample_rate}
|
7 |
network:
|
8 |
-
_target_: remfx.
|
9 |
num_classes: ${num_classes}
|
10 |
n_fft: 4096
|
11 |
hop_length: 512
|
|
|
5 |
lr_weight_decay: 1e-3
|
6 |
sample_rate: ${sample_rate}
|
7 |
network:
|
8 |
+
_target_: remfx.cnn14.Cnn14
|
9 |
num_classes: ${num_classes}
|
10 |
n_fft: 4096
|
11 |
hop_length: 512
|
cfg/model/dcunet.yaml
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# @package _global_
|
2 |
+
model:
|
3 |
+
_target_: remfx.models.RemFx
|
4 |
+
lr: 1e-4
|
5 |
+
lr_beta1: 0.95
|
6 |
+
lr_beta2: 0.999
|
7 |
+
lr_eps: 1e-6
|
8 |
+
lr_weight_decay: 1e-3
|
9 |
+
sample_rate: ${sample_rate}
|
10 |
+
network:
|
11 |
+
_target_: remfx.models.DCUNetModel
|
12 |
+
spec_dim: 256 + 1
|
13 |
+
hidden_dim: 768
|
14 |
+
filter_len: 512
|
15 |
+
hop_len: 64
|
16 |
+
block_layers: 4
|
17 |
+
layers: 4
|
18 |
+
kernel_size: 3
|
19 |
+
refine_layers: 1
|
20 |
+
is_mask: True
|
21 |
+
norm: 'ins'
|
22 |
+
act: 'comp'
|
cfg/model/demucs.yaml
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
# @package _global_
|
2 |
model:
|
3 |
-
_target_: remfx.models.
|
4 |
lr: 1e-4
|
5 |
lr_beta1: 0.95
|
6 |
lr_beta2: 0.999
|
|
|
1 |
# @package _global_
|
2 |
model:
|
3 |
+
_target_: remfx.models.RemFx
|
4 |
lr: 1e-4
|
5 |
lr_beta1: 0.95
|
6 |
lr_beta2: 0.999
|
cfg/model/dptnet.yaml
ADDED
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# @package _global_
|
2 |
+
model:
|
3 |
+
_target_: remfx.models.RemFx
|
4 |
+
lr: 1e-4
|
5 |
+
lr_beta1: 0.95
|
6 |
+
lr_beta2: 0.999
|
7 |
+
lr_eps: 1e-6
|
8 |
+
lr_weight_decay: 1e-3
|
9 |
+
sample_rate: ${sample_rate}
|
10 |
+
network:
|
11 |
+
_target_: remfx.models.DPTNetModel
|
12 |
+
enc_dim: 256
|
13 |
+
feature_dim: 64
|
14 |
+
hidden_dim: 128
|
15 |
+
layer: 6
|
16 |
+
segment_size: 250
|
17 |
+
nspk: 1
|
18 |
+
win_len: 2
|
cfg/model/umx.yaml
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
# @package _global_
|
2 |
-
model:
|
3 |
-
_target_: remfx.models.
|
4 |
lr: 1e-4
|
5 |
lr_beta1: 0.95
|
6 |
lr_beta2: 0.999
|
|
|
1 |
# @package _global_
|
2 |
+
model:
|
3 |
+
_target_: remfx.models.RemFx
|
4 |
lr: 1e-4
|
5 |
lr_beta1: 0.95
|
6 |
lr_beta2: 0.999
|
remfx/cnn14.py
ADDED
@@ -0,0 +1,138 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torchaudio
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from utils import init_bn, init_layer
|
6 |
+
|
7 |
+
# adapted from https://github.com/qiuqiangkong/audioset_tagging_cnn/blob/master/pytorch/models.py
|
8 |
+
|
9 |
+
|
10 |
+
class Cnn14(nn.Module):
|
11 |
+
def __init__(
|
12 |
+
self,
|
13 |
+
num_classes: int,
|
14 |
+
sample_rate: float,
|
15 |
+
n_fft: int = 2048,
|
16 |
+
hop_length: int = 512,
|
17 |
+
n_mels: int = 128,
|
18 |
+
):
|
19 |
+
super().__init__()
|
20 |
+
self.num_classes = num_classes
|
21 |
+
self.n_fft = n_fft
|
22 |
+
self.hop_length = hop_length
|
23 |
+
|
24 |
+
window = torch.hann_window(n_fft)
|
25 |
+
self.register_buffer("window", window)
|
26 |
+
|
27 |
+
self.melspec = torchaudio.transforms.MelSpectrogram(
|
28 |
+
sample_rate,
|
29 |
+
n_fft,
|
30 |
+
hop_length=hop_length,
|
31 |
+
n_mels=n_mels,
|
32 |
+
)
|
33 |
+
|
34 |
+
self.bn0 = nn.BatchNorm2d(n_mels)
|
35 |
+
|
36 |
+
self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
|
37 |
+
self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
|
38 |
+
self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
|
39 |
+
self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
|
40 |
+
self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
|
41 |
+
self.conv_block6 = ConvBlock(in_channels=1024, out_channels=2048)
|
42 |
+
|
43 |
+
self.fc1 = nn.Linear(2048, 2048, bias=True)
|
44 |
+
self.fc_audioset = nn.Linear(2048, num_classes, bias=True)
|
45 |
+
|
46 |
+
self.init_weight()
|
47 |
+
|
48 |
+
def init_weight(self):
|
49 |
+
init_bn(self.bn0)
|
50 |
+
init_layer(self.fc1)
|
51 |
+
init_layer(self.fc_audioset)
|
52 |
+
|
53 |
+
def forward(self, x: torch.Tensor):
|
54 |
+
"""
|
55 |
+
Input: (batch_size, data_length)"""
|
56 |
+
|
57 |
+
x = self.melspec(x)
|
58 |
+
x = x.permute(0, 2, 1, 3)
|
59 |
+
x = self.bn0(x)
|
60 |
+
x = x.permute(0, 2, 1, 3)
|
61 |
+
|
62 |
+
if self.training:
|
63 |
+
pass
|
64 |
+
# x = self.spec_augmenter(x)
|
65 |
+
|
66 |
+
x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
|
67 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
68 |
+
x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
|
69 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
70 |
+
x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
|
71 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
72 |
+
x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
|
73 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
74 |
+
x = self.conv_block5(x, pool_size=(2, 2), pool_type="avg")
|
75 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
76 |
+
x = self.conv_block6(x, pool_size=(1, 1), pool_type="avg")
|
77 |
+
x = F.dropout(x, p=0.2, training=self.training)
|
78 |
+
x = torch.mean(x, dim=3)
|
79 |
+
|
80 |
+
(x1, _) = torch.max(x, dim=2)
|
81 |
+
x2 = torch.mean(x, dim=2)
|
82 |
+
x = x1 + x2
|
83 |
+
x = F.dropout(x, p=0.5, training=self.training)
|
84 |
+
x = F.relu_(self.fc1(x))
|
85 |
+
clipwise_output = self.fc_audioset(x)
|
86 |
+
|
87 |
+
return clipwise_output
|
88 |
+
|
89 |
+
|
90 |
+
class ConvBlock(nn.Module):
|
91 |
+
def __init__(self, in_channels, out_channels):
|
92 |
+
super(ConvBlock, self).__init__()
|
93 |
+
|
94 |
+
self.conv1 = nn.Conv2d(
|
95 |
+
in_channels=in_channels,
|
96 |
+
out_channels=out_channels,
|
97 |
+
kernel_size=(3, 3),
|
98 |
+
stride=(1, 1),
|
99 |
+
padding=(1, 1),
|
100 |
+
bias=False,
|
101 |
+
)
|
102 |
+
|
103 |
+
self.conv2 = nn.Conv2d(
|
104 |
+
in_channels=out_channels,
|
105 |
+
out_channels=out_channels,
|
106 |
+
kernel_size=(3, 3),
|
107 |
+
stride=(1, 1),
|
108 |
+
padding=(1, 1),
|
109 |
+
bias=False,
|
110 |
+
)
|
111 |
+
|
112 |
+
self.bn1 = nn.BatchNorm2d(out_channels)
|
113 |
+
self.bn2 = nn.BatchNorm2d(out_channels)
|
114 |
+
|
115 |
+
self.init_weight()
|
116 |
+
|
117 |
+
def init_weight(self):
|
118 |
+
init_layer(self.conv1)
|
119 |
+
init_layer(self.conv2)
|
120 |
+
init_bn(self.bn1)
|
121 |
+
init_bn(self.bn2)
|
122 |
+
|
123 |
+
def forward(self, input, pool_size=(2, 2), pool_type="avg"):
|
124 |
+
x = input
|
125 |
+
x = F.relu_(self.bn1(self.conv1(x)))
|
126 |
+
x = F.relu_(self.bn2(self.conv2(x)))
|
127 |
+
if pool_type == "max":
|
128 |
+
x = F.max_pool2d(x, kernel_size=pool_size)
|
129 |
+
elif pool_type == "avg":
|
130 |
+
x = F.avg_pool2d(x, kernel_size=pool_size)
|
131 |
+
elif pool_type == "avg+max":
|
132 |
+
x1 = F.avg_pool2d(x, kernel_size=pool_size)
|
133 |
+
x2 = F.max_pool2d(x, kernel_size=pool_size)
|
134 |
+
x = x1 + x2
|
135 |
+
else:
|
136 |
+
raise Exception("Incorrect argument!")
|
137 |
+
|
138 |
+
return x
|
remfx/datasets.py
CHANGED
@@ -250,6 +250,7 @@ class VocalSet(Dataset):
|
|
250 |
# Normalize
|
251 |
normalized_dry = self.normalize(dry)
|
252 |
normalized_wet = self.normalize(wet)
|
|
|
253 |
return normalized_dry, normalized_wet, dry_labels_tensor, wet_labels_tensor
|
254 |
|
255 |
|
|
|
250 |
# Normalize
|
251 |
normalized_dry = self.normalize(dry)
|
252 |
normalized_wet = self.normalize(wet)
|
253 |
+
|
254 |
return normalized_dry, normalized_wet, dry_labels_tensor, wet_labels_tensor
|
255 |
|
256 |
|
remfx/dcunet.py
ADDED
@@ -0,0 +1,649 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 utils import single, concat_complex
|
9 |
+
from torch.nn.init import calculate_gain
|
10 |
+
from typing import Tuple
|
11 |
+
from scipy.signal import get_window
|
12 |
+
from librosa.util import pad_center
|
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
ADDED
@@ -0,0 +1,460 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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, batch):
|
61 |
+
"""
|
62 |
+
mix: shape (batch, T)
|
63 |
+
"""
|
64 |
+
mix, target = batch
|
65 |
+
batch_size = mix.shape[0]
|
66 |
+
mix = self.dpt_encoder(mix) # (B, E, L)
|
67 |
+
|
68 |
+
score_ = self.enc_LN(mix) # B, E, L
|
69 |
+
score_ = self.dpt_separation(score_) # B, nspk, T, N
|
70 |
+
score_ = (
|
71 |
+
score_.view(batch_size * self.num_spk, -1, self.feature_dim)
|
72 |
+
.transpose(1, 2)
|
73 |
+
.contiguous()
|
74 |
+
) # B*nspk, N, T
|
75 |
+
score = self.mask_conv1x1(score_) # [B*nspk, N, L] -> [B*nspk, E, L]
|
76 |
+
score = score.view(
|
77 |
+
batch_size, self.num_spk, self.enc_dim, -1
|
78 |
+
) # [B*nspk, E, L] -> [B, nspk, E, L]
|
79 |
+
est_mask = F.relu(score)
|
80 |
+
|
81 |
+
est_source = self.decoder(
|
82 |
+
mix, est_mask
|
83 |
+
) # [B, E, L] + [B, nspk, E, L]--> [B, nspk, T]
|
84 |
+
|
85 |
+
return est_source
|
86 |
+
|
87 |
+
|
88 |
+
class DPTEncoder(nn.Module):
|
89 |
+
def __init__(self, n_filters: int = 64, window_size: int = 2):
|
90 |
+
super().__init__()
|
91 |
+
self.conv = nn.Conv1d(
|
92 |
+
1, n_filters, kernel_size=window_size, stride=window_size // 2, bias=False
|
93 |
+
)
|
94 |
+
|
95 |
+
def forward(self, x):
|
96 |
+
x = x.unsqueeze(1)
|
97 |
+
x = F.relu(self.conv(x))
|
98 |
+
return x
|
99 |
+
|
100 |
+
|
101 |
+
class TransformerEncoderLayer(torch.nn.Module):
|
102 |
+
def __init__(
|
103 |
+
self, d_model, nhead, hidden_size, dim_feedforward, dropout, activation="relu"
|
104 |
+
):
|
105 |
+
super(TransformerEncoderLayer, self).__init__()
|
106 |
+
self.self_attn = MultiheadAttention(d_model, nhead, dropout=dropout)
|
107 |
+
|
108 |
+
# Implementation of improved part
|
109 |
+
self.lstm = LSTM(d_model, hidden_size, 1, bidirectional=True)
|
110 |
+
self.dropout = Dropout(dropout)
|
111 |
+
self.linear = Linear(hidden_size * 2, d_model)
|
112 |
+
|
113 |
+
self.norm1 = LayerNorm(d_model)
|
114 |
+
self.norm2 = LayerNorm(d_model)
|
115 |
+
self.dropout1 = Dropout(dropout)
|
116 |
+
self.dropout2 = Dropout(dropout)
|
117 |
+
|
118 |
+
self.activation = _get_activation_fn(activation)
|
119 |
+
|
120 |
+
def __setstate__(self, state):
|
121 |
+
if "activation" not in state:
|
122 |
+
state["activation"] = F.relu
|
123 |
+
super(TransformerEncoderLayer, self).__setstate__(state)
|
124 |
+
|
125 |
+
def forward(self, src, src_mask=None, src_key_padding_mask=None):
|
126 |
+
r"""Pass the input through the encoder layer.
|
127 |
+
Args:
|
128 |
+
src: the sequnce to the encoder layer (required).
|
129 |
+
src_mask: the mask for the src sequence (optional).
|
130 |
+
src_key_padding_mask: the mask for the src keys per batch (optional).
|
131 |
+
Shape:
|
132 |
+
see the docs in Transformer class.
|
133 |
+
"""
|
134 |
+
src2 = self.self_attn(
|
135 |
+
src, src, src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask
|
136 |
+
)[0]
|
137 |
+
src = src + self.dropout1(src2)
|
138 |
+
src = self.norm1(src)
|
139 |
+
src2 = self.linear(self.dropout(self.activation(self.lstm(src)[0])))
|
140 |
+
src = src + self.dropout2(src2)
|
141 |
+
src = self.norm2(src)
|
142 |
+
return src
|
143 |
+
|
144 |
+
|
145 |
+
def _get_clones(module, N):
|
146 |
+
return ModuleList([copy.deepcopy(module) for i in range(N)])
|
147 |
+
|
148 |
+
|
149 |
+
def _get_activation_fn(activation):
|
150 |
+
if activation == "relu":
|
151 |
+
return F.relu
|
152 |
+
elif activation == "gelu":
|
153 |
+
return F.gelu
|
154 |
+
|
155 |
+
raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
|
156 |
+
|
157 |
+
|
158 |
+
class SingleTransformer(nn.Module):
|
159 |
+
"""
|
160 |
+
Container module for a single Transformer layer.
|
161 |
+
args: input_size: int, dimension of the input feature.
|
162 |
+
The input should have shape (batch, seq_len, input_size).
|
163 |
+
"""
|
164 |
+
|
165 |
+
def __init__(self, input_size, hidden_size, dropout):
|
166 |
+
super(SingleTransformer, self).__init__()
|
167 |
+
self.transformer = TransformerEncoderLayer(
|
168 |
+
d_model=input_size,
|
169 |
+
nhead=4,
|
170 |
+
hidden_size=hidden_size,
|
171 |
+
dim_feedforward=hidden_size * 2,
|
172 |
+
dropout=dropout,
|
173 |
+
)
|
174 |
+
|
175 |
+
def forward(self, input):
|
176 |
+
# input shape: batch, seq, dim
|
177 |
+
output = input
|
178 |
+
transformer_output = (
|
179 |
+
self.transformer(output.permute(1, 0, 2).contiguous())
|
180 |
+
.permute(1, 0, 2)
|
181 |
+
.contiguous()
|
182 |
+
)
|
183 |
+
return transformer_output
|
184 |
+
|
185 |
+
|
186 |
+
# dual-path transformer
|
187 |
+
class DPT(nn.Module):
|
188 |
+
"""
|
189 |
+
Deep dual-path transformer.
|
190 |
+
args:
|
191 |
+
input_size: int, dimension of the input feature. The input should have shape
|
192 |
+
(batch, seq_len, input_size).
|
193 |
+
hidden_size: int, dimension of the hidden state.
|
194 |
+
output_size: int, dimension of the output size.
|
195 |
+
num_layers: int, number of stacked Transformer layers. Default is 1.
|
196 |
+
dropout: float, dropout ratio. Default is 0.
|
197 |
+
"""
|
198 |
+
|
199 |
+
def __init__(self, input_size, hidden_size, output_size, num_layers=1, dropout=0):
|
200 |
+
super(DPT, self).__init__()
|
201 |
+
|
202 |
+
self.input_size = input_size
|
203 |
+
self.output_size = output_size
|
204 |
+
self.hidden_size = hidden_size
|
205 |
+
|
206 |
+
# dual-path transformer
|
207 |
+
self.row_transformer = nn.ModuleList([])
|
208 |
+
self.col_transformer = nn.ModuleList([])
|
209 |
+
for i in range(num_layers):
|
210 |
+
self.row_transformer.append(
|
211 |
+
SingleTransformer(input_size, hidden_size, dropout)
|
212 |
+
)
|
213 |
+
self.col_transformer.append(
|
214 |
+
SingleTransformer(input_size, hidden_size, dropout)
|
215 |
+
)
|
216 |
+
|
217 |
+
# output layer
|
218 |
+
self.output = nn.Sequential(nn.PReLU(), nn.Conv2d(input_size, output_size, 1))
|
219 |
+
|
220 |
+
def forward(self, input):
|
221 |
+
# input shape: batch, N, dim1, dim2
|
222 |
+
# apply transformer on dim1 first and then dim2
|
223 |
+
# output shape: B, output_size, dim1, dim2
|
224 |
+
# input = input.to(device)
|
225 |
+
batch_size, _, dim1, dim2 = input.shape
|
226 |
+
output = input
|
227 |
+
for i in range(len(self.row_transformer)):
|
228 |
+
row_input = (
|
229 |
+
output.permute(0, 3, 2, 1)
|
230 |
+
.contiguous()
|
231 |
+
.view(batch_size * dim2, dim1, -1)
|
232 |
+
) # B*dim2, dim1, N
|
233 |
+
row_output = self.row_transformer[i](row_input) # B*dim2, dim1, H
|
234 |
+
row_output = (
|
235 |
+
row_output.view(batch_size, dim2, dim1, -1)
|
236 |
+
.permute(0, 3, 2, 1)
|
237 |
+
.contiguous()
|
238 |
+
) # B, N, dim1, dim2
|
239 |
+
output = row_output
|
240 |
+
|
241 |
+
col_input = (
|
242 |
+
output.permute(0, 2, 3, 1)
|
243 |
+
.contiguous()
|
244 |
+
.view(batch_size * dim1, dim2, -1)
|
245 |
+
) # B*dim1, dim2, N
|
246 |
+
col_output = self.col_transformer[i](col_input) # B*dim1, dim2, H
|
247 |
+
col_output = (
|
248 |
+
col_output.view(batch_size, dim1, dim2, -1)
|
249 |
+
.permute(0, 3, 1, 2)
|
250 |
+
.contiguous()
|
251 |
+
) # B, N, dim1, dim2
|
252 |
+
output = col_output
|
253 |
+
|
254 |
+
output = self.output(output) # B, output_size, dim1, dim2
|
255 |
+
|
256 |
+
return output
|
257 |
+
|
258 |
+
|
259 |
+
# base module for deep DPT
|
260 |
+
class DPT_base(nn.Module):
|
261 |
+
def __init__(
|
262 |
+
self, input_dim, feature_dim, hidden_dim, num_spk=2, layer=6, segment_size=250
|
263 |
+
):
|
264 |
+
super(DPT_base, self).__init__()
|
265 |
+
|
266 |
+
self.input_dim = input_dim
|
267 |
+
self.feature_dim = feature_dim
|
268 |
+
self.hidden_dim = hidden_dim
|
269 |
+
|
270 |
+
self.layer = layer
|
271 |
+
self.segment_size = segment_size
|
272 |
+
self.num_spk = num_spk
|
273 |
+
|
274 |
+
self.eps = 1e-8
|
275 |
+
|
276 |
+
# bottleneck
|
277 |
+
self.BN = nn.Conv1d(self.input_dim, self.feature_dim, 1, bias=False)
|
278 |
+
|
279 |
+
# DPT model
|
280 |
+
self.DPT = DPT(
|
281 |
+
self.feature_dim,
|
282 |
+
self.hidden_dim,
|
283 |
+
self.feature_dim * self.num_spk,
|
284 |
+
num_layers=layer,
|
285 |
+
)
|
286 |
+
|
287 |
+
def pad_segment(self, input, segment_size):
|
288 |
+
# input is the features: (B, N, T)
|
289 |
+
batch_size, dim, seq_len = input.shape
|
290 |
+
segment_stride = segment_size // 2
|
291 |
+
|
292 |
+
rest = segment_size - (segment_stride + seq_len % segment_size) % segment_size
|
293 |
+
if rest > 0:
|
294 |
+
pad = Variable(torch.zeros(batch_size, dim, rest)).type(input.type())
|
295 |
+
input = torch.cat([input, pad], 2)
|
296 |
+
|
297 |
+
pad_aux = Variable(torch.zeros(batch_size, dim, segment_stride)).type(
|
298 |
+
input.type()
|
299 |
+
)
|
300 |
+
input = torch.cat([pad_aux, input, pad_aux], 2)
|
301 |
+
|
302 |
+
return input, rest
|
303 |
+
|
304 |
+
def split_feature(self, input, segment_size):
|
305 |
+
# split the feature into chunks of segment size
|
306 |
+
# input is the features: (B, N, T)
|
307 |
+
|
308 |
+
input, rest = self.pad_segment(input, segment_size)
|
309 |
+
batch_size, dim, seq_len = input.shape
|
310 |
+
segment_stride = segment_size // 2
|
311 |
+
|
312 |
+
segments1 = (
|
313 |
+
input[:, :, :-segment_stride]
|
314 |
+
.contiguous()
|
315 |
+
.view(batch_size, dim, -1, segment_size)
|
316 |
+
)
|
317 |
+
segments2 = (
|
318 |
+
input[:, :, segment_stride:]
|
319 |
+
.contiguous()
|
320 |
+
.view(batch_size, dim, -1, segment_size)
|
321 |
+
)
|
322 |
+
segments = (
|
323 |
+
torch.cat([segments1, segments2], 3)
|
324 |
+
.view(batch_size, dim, -1, segment_size)
|
325 |
+
.transpose(2, 3)
|
326 |
+
)
|
327 |
+
|
328 |
+
return segments.contiguous(), rest
|
329 |
+
|
330 |
+
def merge_feature(self, input, rest):
|
331 |
+
# merge the splitted features into full utterance
|
332 |
+
# input is the features: (B, N, L, K)
|
333 |
+
|
334 |
+
batch_size, dim, segment_size, _ = input.shape
|
335 |
+
segment_stride = segment_size // 2
|
336 |
+
input = (
|
337 |
+
input.transpose(2, 3)
|
338 |
+
.contiguous()
|
339 |
+
.view(batch_size, dim, -1, segment_size * 2)
|
340 |
+
) # B, N, K, L
|
341 |
+
|
342 |
+
input1 = (
|
343 |
+
input[:, :, :, :segment_size]
|
344 |
+
.contiguous()
|
345 |
+
.view(batch_size, dim, -1)[:, :, segment_stride:]
|
346 |
+
)
|
347 |
+
input2 = (
|
348 |
+
input[:, :, :, segment_size:]
|
349 |
+
.contiguous()
|
350 |
+
.view(batch_size, dim, -1)[:, :, :-segment_stride]
|
351 |
+
)
|
352 |
+
|
353 |
+
output = input1 + input2
|
354 |
+
if rest > 0:
|
355 |
+
output = output[:, :, :-rest]
|
356 |
+
|
357 |
+
return output.contiguous() # B, N, T
|
358 |
+
|
359 |
+
def forward(self, input):
|
360 |
+
pass
|
361 |
+
|
362 |
+
|
363 |
+
class DPTSeparation(DPT_base):
|
364 |
+
def __init__(self, *args, **kwargs):
|
365 |
+
super(DPTSeparation, self).__init__(*args, **kwargs)
|
366 |
+
|
367 |
+
# gated output layer
|
368 |
+
self.output = nn.Sequential(
|
369 |
+
nn.Conv1d(self.feature_dim, self.feature_dim, 1), nn.Tanh()
|
370 |
+
)
|
371 |
+
self.output_gate = nn.Sequential(
|
372 |
+
nn.Conv1d(self.feature_dim, self.feature_dim, 1), nn.Sigmoid()
|
373 |
+
)
|
374 |
+
|
375 |
+
def forward(self, input):
|
376 |
+
# input = input.to(device)
|
377 |
+
# input: (B, E, T)
|
378 |
+
batch_size, E, seq_length = input.shape
|
379 |
+
|
380 |
+
enc_feature = self.BN(input) # (B, E, L)-->(B, N, L)
|
381 |
+
# split the encoder output into overlapped, longer segments
|
382 |
+
enc_segments, enc_rest = self.split_feature(
|
383 |
+
enc_feature, self.segment_size
|
384 |
+
) # B, N, L, K: L is the segment_size
|
385 |
+
# print('enc_segments.shape {}'.format(enc_segments.shape))
|
386 |
+
# pass to DPT
|
387 |
+
output = self.DPT(enc_segments).view(
|
388 |
+
batch_size * self.num_spk, self.feature_dim, self.segment_size, -1
|
389 |
+
) # B*nspk, N, L, K
|
390 |
+
|
391 |
+
# overlap-and-add of the outputs
|
392 |
+
output = self.merge_feature(output, enc_rest) # B*nspk, N, T
|
393 |
+
|
394 |
+
# gated output layer for filter generation
|
395 |
+
bf_filter = self.output(output) * self.output_gate(output) # B*nspk, K, T
|
396 |
+
bf_filter = (
|
397 |
+
bf_filter.transpose(1, 2)
|
398 |
+
.contiguous()
|
399 |
+
.view(batch_size, self.num_spk, -1, self.feature_dim)
|
400 |
+
) # B, nspk, T, N
|
401 |
+
|
402 |
+
return bf_filter
|
403 |
+
|
404 |
+
|
405 |
+
class DPTDecoder(nn.Module):
|
406 |
+
def __init__(self, n_filters: int = 64, window_size: int = 2):
|
407 |
+
super().__init__()
|
408 |
+
self.W = window_size
|
409 |
+
self.basis_signals = nn.Linear(n_filters, window_size, bias=False)
|
410 |
+
|
411 |
+
def forward(self, mixture, mask):
|
412 |
+
"""
|
413 |
+
mixture: (batch, n_filters, L)
|
414 |
+
mask: (batch, sources, n_filters, L)
|
415 |
+
"""
|
416 |
+
source_w = torch.unsqueeze(mixture, 1) * mask # [B, C, E, L]
|
417 |
+
source_w = torch.transpose(source_w, 2, 3) # [B, C, L, E]
|
418 |
+
# S = DV
|
419 |
+
est_source = self.basis_signals(source_w) # [B, C, L, W]
|
420 |
+
est_source = overlap_and_add(est_source, self.W // 2) # B x C x T
|
421 |
+
return est_source
|
422 |
+
|
423 |
+
|
424 |
+
def overlap_and_add(signal, frame_step):
|
425 |
+
"""Reconstructs a signal from a framed representation.
|
426 |
+
Adds potentially overlapping frames of a signal with shape
|
427 |
+
`[..., frames, frame_length]`, offsetting subsequent frames by `frame_step`.
|
428 |
+
The resulting tensor has shape `[..., output_size]` where
|
429 |
+
output_size = (frames - 1) * frame_step + frame_length
|
430 |
+
Args:
|
431 |
+
signal: A [..., frames, frame_length] Tensor.
|
432 |
+
All dimensions may be unknown, and rank must be at least 2.
|
433 |
+
frame_step: An integer denoting overlap offsets. Must be less than or equal to frame_length.
|
434 |
+
Returns:
|
435 |
+
A Tensor with shape [..., output_size] containing the overlap-added frames of signal's
|
436 |
+
inner-most two dimensions.
|
437 |
+
output_size = (frames - 1) * frame_step + frame_length
|
438 |
+
Based on https://github.com/tensorflow/tensorflow/blob/r1.12/tensorflow/contrib/signal/python/ops/reconstruction_ops.py
|
439 |
+
"""
|
440 |
+
outer_dimensions = signal.size()[:-2]
|
441 |
+
frames, frame_length = signal.size()[-2:]
|
442 |
+
|
443 |
+
subframe_length = math.gcd(frame_length, frame_step) # gcd=Greatest Common Divisor
|
444 |
+
subframe_step = frame_step // subframe_length
|
445 |
+
subframes_per_frame = frame_length // subframe_length
|
446 |
+
output_size = frame_step * (frames - 1) + frame_length
|
447 |
+
output_subframes = output_size // subframe_length
|
448 |
+
|
449 |
+
subframe_signal = signal.reshape(*outer_dimensions, -1, subframe_length)
|
450 |
+
|
451 |
+
frame = torch.arange(0, output_subframes).unfold(
|
452 |
+
0, subframes_per_frame, subframe_step
|
453 |
+
)
|
454 |
+
frame = signal.new_tensor(frame).long() # signal may in GPU or CPU
|
455 |
+
frame = frame.contiguous().view(-1)
|
456 |
+
|
457 |
+
result = signal.new_zeros(*outer_dimensions, output_subframes, subframe_length)
|
458 |
+
result.index_add_(-2, frame, subframe_signal)
|
459 |
+
result = result.view(*outer_dimensions, -1)
|
460 |
+
return result
|
remfx/models.py
CHANGED
@@ -1,10 +1,6 @@
|
|
1 |
-
import wandb
|
2 |
import torch
|
3 |
-
import torchaudio
|
4 |
import torchmetrics
|
5 |
import pytorch_lightning as pl
|
6 |
-
import torch.nn.functional as F
|
7 |
-
|
8 |
from torch import Tensor, nn
|
9 |
from einops import rearrange
|
10 |
from torchaudio.models import HDemucs
|
@@ -13,10 +9,12 @@ from auraloss.time import SISDRLoss
|
|
13 |
from auraloss.freq import MultiResolutionSTFTLoss
|
14 |
from umx.openunmix.model import OpenUnmix, Separator
|
15 |
|
16 |
-
from
|
|
|
|
|
17 |
|
18 |
|
19 |
-
class
|
20 |
def __init__(
|
21 |
self,
|
22 |
lr: float,
|
@@ -35,7 +33,7 @@ class RemFXModel(pl.LightningModule):
|
|
35 |
self.lr_weight_decay = lr_weight_decay
|
36 |
self.sample_rate = sample_rate
|
37 |
self.model = network
|
38 |
-
self.metrics =
|
39 |
{
|
40 |
"SISDR": SISDRLoss(),
|
41 |
"STFT": MultiResolutionSTFTLoss(),
|
@@ -94,7 +92,8 @@ class RemFXModel(pl.LightningModule):
|
|
94 |
return loss
|
95 |
|
96 |
def common_step(self, batch, batch_idx, mode: str = "train"):
|
97 |
-
x, y, _, _ = batch
|
|
|
98 |
loss, output = self.model((x, y))
|
99 |
self.log(f"{mode}_loss", loss)
|
100 |
# Metric logging
|
@@ -201,7 +200,7 @@ class RemFXModel(pl.LightningModule):
|
|
201 |
)
|
202 |
|
203 |
|
204 |
-
class OpenUnmixModel(
|
205 |
def __init__(
|
206 |
self,
|
207 |
n_fft: int = 2048,
|
@@ -234,7 +233,7 @@ class OpenUnmixModel(torch.nn.Module):
|
|
234 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
235 |
n_bins=self.num_bins, sample_rate=self.sample_rate
|
236 |
)
|
237 |
-
self.l1loss =
|
238 |
|
239 |
def forward(self, batch):
|
240 |
x, target = batch
|
@@ -249,7 +248,7 @@ class OpenUnmixModel(torch.nn.Module):
|
|
249 |
return self.separator(x).squeeze(1)
|
250 |
|
251 |
|
252 |
-
class DemucsModel(
|
253 |
def __init__(self, sample_rate, **kwargs) -> None:
|
254 |
super().__init__()
|
255 |
self.model = HDemucs(**kwargs)
|
@@ -257,7 +256,7 @@ class DemucsModel(torch.nn.Module):
|
|
257 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
258 |
n_bins=self.num_bins, sample_rate=sample_rate
|
259 |
)
|
260 |
-
self.l1loss =
|
261 |
|
262 |
def forward(self, batch):
|
263 |
x, target = batch
|
@@ -284,201 +283,42 @@ class DiffusionGenerationModel(nn.Module):
|
|
284 |
return self.model.sample(noise, num_steps=num_steps)
|
285 |
|
286 |
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
max_items: int = 10,
|
294 |
-
):
|
295 |
-
num_items = samples.shape[0]
|
296 |
-
samples = rearrange(samples, "b c t -> b t c")
|
297 |
-
for idx in range(num_items):
|
298 |
-
if idx >= max_items:
|
299 |
-
break
|
300 |
-
logger.experiment.log(
|
301 |
-
{
|
302 |
-
f"{id}_{idx}": wandb.Audio(
|
303 |
-
samples[idx].cpu().numpy(),
|
304 |
-
caption=caption,
|
305 |
-
sample_rate=sampling_rate,
|
306 |
-
)
|
307 |
-
}
|
308 |
-
)
|
309 |
-
|
310 |
-
|
311 |
-
def spectrogram(
|
312 |
-
x: torch.Tensor,
|
313 |
-
window: torch.Tensor,
|
314 |
-
n_fft: int,
|
315 |
-
hop_length: int,
|
316 |
-
alpha: float,
|
317 |
-
) -> torch.Tensor:
|
318 |
-
bs, chs, samp = x.size()
|
319 |
-
x = x.view(bs * chs, -1) # move channels onto batch dim
|
320 |
-
|
321 |
-
X = torch.stft(
|
322 |
-
x,
|
323 |
-
n_fft=n_fft,
|
324 |
-
hop_length=hop_length,
|
325 |
-
window=window,
|
326 |
-
return_complex=True,
|
327 |
-
)
|
328 |
-
|
329 |
-
# move channels back
|
330 |
-
X = X.view(bs, chs, X.shape[-2], X.shape[-1])
|
331 |
-
|
332 |
-
return torch.pow(X.abs() + 1e-8, alpha)
|
333 |
-
|
334 |
-
|
335 |
-
# adapted from https://github.com/qiuqiangkong/audioset_tagging_cnn/blob/master/pytorch/models.py
|
336 |
-
|
337 |
-
|
338 |
-
def init_layer(layer):
|
339 |
-
"""Initialize a Linear or Convolutional layer."""
|
340 |
-
nn.init.xavier_uniform_(layer.weight)
|
341 |
-
|
342 |
-
if hasattr(layer, "bias"):
|
343 |
-
if layer.bias is not None:
|
344 |
-
layer.bias.data.fill_(0.0)
|
345 |
-
|
346 |
-
|
347 |
-
def init_bn(bn):
|
348 |
-
"""Initialize a Batchnorm layer."""
|
349 |
-
bn.bias.data.fill_(0.0)
|
350 |
-
bn.weight.data.fill_(1.0)
|
351 |
-
|
352 |
-
|
353 |
-
class ConvBlock(nn.Module):
|
354 |
-
def __init__(self, in_channels, out_channels):
|
355 |
-
super(ConvBlock, self).__init__()
|
356 |
-
|
357 |
-
self.conv1 = nn.Conv2d(
|
358 |
-
in_channels=in_channels,
|
359 |
-
out_channels=out_channels,
|
360 |
-
kernel_size=(3, 3),
|
361 |
-
stride=(1, 1),
|
362 |
-
padding=(1, 1),
|
363 |
-
bias=False,
|
364 |
-
)
|
365 |
-
|
366 |
-
self.conv2 = nn.Conv2d(
|
367 |
-
in_channels=out_channels,
|
368 |
-
out_channels=out_channels,
|
369 |
-
kernel_size=(3, 3),
|
370 |
-
stride=(1, 1),
|
371 |
-
padding=(1, 1),
|
372 |
-
bias=False,
|
373 |
)
|
|
|
374 |
|
375 |
-
|
376 |
-
|
377 |
-
|
378 |
-
self.
|
379 |
-
|
380 |
-
def init_weight(self):
|
381 |
-
init_layer(self.conv1)
|
382 |
-
init_layer(self.conv2)
|
383 |
-
init_bn(self.bn1)
|
384 |
-
init_bn(self.bn2)
|
385 |
-
|
386 |
-
def forward(self, input, pool_size=(2, 2), pool_type="avg"):
|
387 |
-
x = input
|
388 |
-
x = F.relu_(self.bn1(self.conv1(x)))
|
389 |
-
x = F.relu_(self.bn2(self.conv2(x)))
|
390 |
-
if pool_type == "max":
|
391 |
-
x = F.max_pool2d(x, kernel_size=pool_size)
|
392 |
-
elif pool_type == "avg":
|
393 |
-
x = F.avg_pool2d(x, kernel_size=pool_size)
|
394 |
-
elif pool_type == "avg+max":
|
395 |
-
x1 = F.avg_pool2d(x, kernel_size=pool_size)
|
396 |
-
x2 = F.max_pool2d(x, kernel_size=pool_size)
|
397 |
-
x = x1 + x2
|
398 |
-
else:
|
399 |
-
raise Exception("Incorrect argument!")
|
400 |
|
401 |
-
|
|
|
402 |
|
403 |
|
404 |
-
class
|
405 |
-
def __init__(
|
406 |
-
self,
|
407 |
-
num_classes: int,
|
408 |
-
sample_rate: float,
|
409 |
-
n_fft: int = 2048,
|
410 |
-
hop_length: int = 512,
|
411 |
-
n_mels: int = 128,
|
412 |
-
):
|
413 |
super().__init__()
|
414 |
-
self.
|
415 |
-
self.
|
416 |
-
|
417 |
-
|
418 |
-
window = torch.hann_window(n_fft)
|
419 |
-
self.register_buffer("window", window)
|
420 |
-
|
421 |
-
self.melspec = torchaudio.transforms.MelSpectrogram(
|
422 |
-
sample_rate,
|
423 |
-
n_fft,
|
424 |
-
hop_length=hop_length,
|
425 |
-
n_mels=n_mels,
|
426 |
)
|
|
|
427 |
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
self.
|
432 |
-
|
433 |
-
self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
|
434 |
-
self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
|
435 |
-
self.conv_block6 = ConvBlock(in_channels=1024, out_channels=2048)
|
436 |
-
|
437 |
-
self.fc1 = nn.Linear(2048, 2048, bias=True)
|
438 |
-
self.fc_audioset = nn.Linear(2048, num_classes, bias=True)
|
439 |
-
|
440 |
-
self.init_weight()
|
441 |
-
|
442 |
-
def init_weight(self):
|
443 |
-
init_bn(self.bn0)
|
444 |
-
init_layer(self.fc1)
|
445 |
-
init_layer(self.fc_audioset)
|
446 |
|
447 |
-
def
|
448 |
-
|
449 |
-
Input: (batch_size, data_length)"""
|
450 |
-
|
451 |
-
x = self.melspec(x)
|
452 |
-
x = x.permute(0, 2, 1, 3)
|
453 |
-
x = self.bn0(x)
|
454 |
-
x = x.permute(0, 2, 1, 3)
|
455 |
-
|
456 |
-
if self.training:
|
457 |
-
pass
|
458 |
-
# x = self.spec_augmenter(x)
|
459 |
-
|
460 |
-
x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
|
461 |
-
x = F.dropout(x, p=0.2, training=self.training)
|
462 |
-
x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
|
463 |
-
x = F.dropout(x, p=0.2, training=self.training)
|
464 |
-
x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
|
465 |
-
x = F.dropout(x, p=0.2, training=self.training)
|
466 |
-
x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
|
467 |
-
x = F.dropout(x, p=0.2, training=self.training)
|
468 |
-
x = self.conv_block5(x, pool_size=(2, 2), pool_type="avg")
|
469 |
-
x = F.dropout(x, p=0.2, training=self.training)
|
470 |
-
x = self.conv_block6(x, pool_size=(1, 1), pool_type="avg")
|
471 |
-
x = F.dropout(x, p=0.2, training=self.training)
|
472 |
-
x = torch.mean(x, dim=3)
|
473 |
-
|
474 |
-
(x1, _) = torch.max(x, dim=2)
|
475 |
-
x2 = torch.mean(x, dim=2)
|
476 |
-
x = x1 + x2
|
477 |
-
x = F.dropout(x, p=0.5, training=self.training)
|
478 |
-
x = F.relu_(self.fc1(x))
|
479 |
-
clipwise_output = self.fc_audioset(x)
|
480 |
-
|
481 |
-
return clipwise_output
|
482 |
|
483 |
|
484 |
class FXClassifier(pl.LightningModule):
|
@@ -501,7 +341,7 @@ class FXClassifier(pl.LightningModule):
|
|
501 |
def common_step(self, batch, batch_idx, mode: str = "train"):
|
502 |
x, y, dry_label, wet_label = batch
|
503 |
pred_label = self.network(x)
|
504 |
-
loss =
|
505 |
self.log(
|
506 |
f"{mode}_loss",
|
507 |
loss,
|
|
|
|
|
1 |
import torch
|
|
|
2 |
import torchmetrics
|
3 |
import pytorch_lightning as pl
|
|
|
|
|
4 |
from torch import Tensor, nn
|
5 |
from einops import rearrange
|
6 |
from torchaudio.models import HDemucs
|
|
|
9 |
from auraloss.freq import MultiResolutionSTFTLoss
|
10 |
from umx.openunmix.model import OpenUnmix, Separator
|
11 |
|
12 |
+
from utils import FADLoss, spectrogram, log_wandb_audio_batch
|
13 |
+
from dptnet import DPTNet_base
|
14 |
+
from dcunet import RefineSpectrogramUnet
|
15 |
|
16 |
|
17 |
+
class RemFX(pl.LightningModule):
|
18 |
def __init__(
|
19 |
self,
|
20 |
lr: float,
|
|
|
33 |
self.lr_weight_decay = lr_weight_decay
|
34 |
self.sample_rate = sample_rate
|
35 |
self.model = network
|
36 |
+
self.metrics = nn.ModuleDict(
|
37 |
{
|
38 |
"SISDR": SISDRLoss(),
|
39 |
"STFT": MultiResolutionSTFTLoss(),
|
|
|
92 |
return loss
|
93 |
|
94 |
def common_step(self, batch, batch_idx, mode: str = "train"):
|
95 |
+
x, y, _, _ = batch # x, y = (B, C, T), (B, C, T)
|
96 |
+
|
97 |
loss, output = self.model((x, y))
|
98 |
self.log(f"{mode}_loss", loss)
|
99 |
# Metric logging
|
|
|
200 |
)
|
201 |
|
202 |
|
203 |
+
class OpenUnmixModel(nn.Module):
|
204 |
def __init__(
|
205 |
self,
|
206 |
n_fft: int = 2048,
|
|
|
233 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
234 |
n_bins=self.num_bins, sample_rate=self.sample_rate
|
235 |
)
|
236 |
+
self.l1loss = nn.L1Loss()
|
237 |
|
238 |
def forward(self, batch):
|
239 |
x, target = batch
|
|
|
248 |
return self.separator(x).squeeze(1)
|
249 |
|
250 |
|
251 |
+
class DemucsModel(nn.Module):
|
252 |
def __init__(self, sample_rate, **kwargs) -> None:
|
253 |
super().__init__()
|
254 |
self.model = HDemucs(**kwargs)
|
|
|
256 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
257 |
n_bins=self.num_bins, sample_rate=sample_rate
|
258 |
)
|
259 |
+
self.l1loss = nn.L1Loss()
|
260 |
|
261 |
def forward(self, batch):
|
262 |
x, target = batch
|
|
|
283 |
return self.model.sample(noise, num_steps=num_steps)
|
284 |
|
285 |
|
286 |
+
class DPTNetModel(nn.Module):
|
287 |
+
def __init__(self, sample_rate, **kwargs):
|
288 |
+
super().__init__()
|
289 |
+
self.model = DPTNet_base(**kwargs)
|
290 |
+
self.mrstftloss = MultiResolutionSTFTLoss(
|
291 |
+
n_bins=self.num_bins, sample_rate=sample_rate
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
292 |
)
|
293 |
+
self.l1loss = nn.L1Loss()
|
294 |
|
295 |
+
def forward(self, batch):
|
296 |
+
x, target = batch
|
297 |
+
output = self.model(x).squeeze(1)
|
298 |
+
loss = self.mrstftloss(output, target) + self.l1loss(output, target) * 100
|
299 |
+
return loss, output
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
300 |
|
301 |
+
def sample(self, x: Tensor) -> Tensor:
|
302 |
+
return self.model.sample(x)
|
303 |
|
304 |
|
305 |
+
class DCUNetModel(nn.Module):
|
306 |
+
def __init__(self, sample_rate, **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
307 |
super().__init__()
|
308 |
+
self.model = RefineSpectrogramUnet(**kwargs)
|
309 |
+
self.mrstftloss = MultiResolutionSTFTLoss(
|
310 |
+
n_bins=self.num_bins, sample_rate=sample_rate
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
311 |
)
|
312 |
+
self.l1loss = nn.L1Loss()
|
313 |
|
314 |
+
def forward(self, batch):
|
315 |
+
x, target = batch
|
316 |
+
output = self.model(x).squeeze(1)
|
317 |
+
loss = self.mrstftloss(output, target) + self.l1loss(output, target) * 100
|
318 |
+
return loss, output
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
319 |
|
320 |
+
def sample(self, x: Tensor) -> Tensor:
|
321 |
+
return self.model.sample(x)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
322 |
|
323 |
|
324 |
class FXClassifier(pl.LightningModule):
|
|
|
341 |
def common_step(self, batch, batch_idx, mode: str = "train"):
|
342 |
x, y, dry_label, wet_label = batch
|
343 |
pred_label = self.network(x)
|
344 |
+
loss = nn.functional.cross_entropy(pred_label, dry_label)
|
345 |
self.log(
|
346 |
f"{mode}_loss",
|
347 |
loss,
|
remfx/utils.py
CHANGED
@@ -7,6 +7,10 @@ from frechet_audio_distance import FrechetAudioDistance
|
|
7 |
import numpy as np
|
8 |
import torch
|
9 |
import torchaudio
|
|
|
|
|
|
|
|
|
10 |
|
11 |
|
12 |
def get_logger(name=__name__) -> logging.Logger:
|
@@ -138,3 +142,91 @@ def create_sequential_chunks(
|
|
138 |
break
|
139 |
chunks.append(audio[:, start : start + chunk_size])
|
140 |
return chunks, sr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
import numpy as np
|
8 |
import torch
|
9 |
import torchaudio
|
10 |
+
from torch import Tensor, nn
|
11 |
+
import wandb
|
12 |
+
from einops import rearrange
|
13 |
+
from torch._six import container_abcs
|
14 |
|
15 |
|
16 |
def get_logger(name=__name__) -> logging.Logger:
|
|
|
142 |
break
|
143 |
chunks.append(audio[:, start : start + chunk_size])
|
144 |
return chunks, sr
|
145 |
+
|
146 |
+
|
147 |
+
def log_wandb_audio_batch(
|
148 |
+
logger: pl.loggers.WandbLogger,
|
149 |
+
id: str,
|
150 |
+
samples: Tensor,
|
151 |
+
sampling_rate: int,
|
152 |
+
caption: str = "",
|
153 |
+
max_items: int = 10,
|
154 |
+
):
|
155 |
+
num_items = samples.shape[0]
|
156 |
+
samples = rearrange(samples, "b c t -> b t c")
|
157 |
+
for idx in range(num_items):
|
158 |
+
if idx >= max_items:
|
159 |
+
break
|
160 |
+
logger.experiment.log(
|
161 |
+
{
|
162 |
+
f"{id}_{idx}": wandb.Audio(
|
163 |
+
samples[idx].cpu().numpy(),
|
164 |
+
caption=caption,
|
165 |
+
sample_rate=sampling_rate,
|
166 |
+
)
|
167 |
+
}
|
168 |
+
)
|
169 |
+
|
170 |
+
|
171 |
+
def spectrogram(
|
172 |
+
x: torch.Tensor,
|
173 |
+
window: torch.Tensor,
|
174 |
+
n_fft: int,
|
175 |
+
hop_length: int,
|
176 |
+
alpha: float,
|
177 |
+
) -> torch.Tensor:
|
178 |
+
bs, chs, samp = x.size()
|
179 |
+
x = x.view(bs * chs, -1) # move channels onto batch dim
|
180 |
+
|
181 |
+
X = torch.stft(
|
182 |
+
x,
|
183 |
+
n_fft=n_fft,
|
184 |
+
hop_length=hop_length,
|
185 |
+
window=window,
|
186 |
+
return_complex=True,
|
187 |
+
)
|
188 |
+
|
189 |
+
# move channels back
|
190 |
+
X = X.view(bs, chs, X.shape[-2], X.shape[-1])
|
191 |
+
|
192 |
+
return torch.pow(X.abs() + 1e-8, alpha)
|
193 |
+
|
194 |
+
|
195 |
+
def init_layer(layer):
|
196 |
+
"""Initialize a Linear or Convolutional layer."""
|
197 |
+
nn.init.xavier_uniform_(layer.weight)
|
198 |
+
|
199 |
+
if hasattr(layer, "bias"):
|
200 |
+
if layer.bias is not None:
|
201 |
+
layer.bias.data.fill_(0.0)
|
202 |
+
|
203 |
+
|
204 |
+
def init_bn(bn):
|
205 |
+
"""Initialize a Batchnorm layer."""
|
206 |
+
bn.bias.data.fill_(0.0)
|
207 |
+
bn.weight.data.fill_(1.0)
|
208 |
+
|
209 |
+
|
210 |
+
def _ntuple(n: int):
|
211 |
+
def parse(x):
|
212 |
+
if isinstance(x, container_abcs.Iterable):
|
213 |
+
return x
|
214 |
+
return tuple([x] * n)
|
215 |
+
|
216 |
+
return parse
|
217 |
+
|
218 |
+
|
219 |
+
single = _ntuple(1)
|
220 |
+
|
221 |
+
|
222 |
+
def concat_complex(a: torch.tensor, b: torch.tensor, dim: int = 1) -> torch.tensor:
|
223 |
+
"""
|
224 |
+
Concatenate two complex tensors in same dimension concept
|
225 |
+
:param a: complex tensor
|
226 |
+
:param b: another complex tensor
|
227 |
+
:param dim: target dimension
|
228 |
+
:return: concatenated tensor
|
229 |
+
"""
|
230 |
+
a_real, a_img = a.chunk(2, dim)
|
231 |
+
b_real, b_img = b.chunk(2, dim)
|
232 |
+
return torch.cat([a_real, b_real, a_img, b_img], dim=dim)
|
scripts/test.py
CHANGED
@@ -3,7 +3,6 @@ import hydra
|
|
3 |
from omegaconf import DictConfig
|
4 |
import remfx.utils as utils
|
5 |
from pytorch_lightning.utilities.model_summary import ModelSummary
|
6 |
-
from remfx.models import RemFXModel
|
7 |
import torch
|
8 |
|
9 |
log = utils.get_logger(__name__)
|
|
|
3 |
from omegaconf import DictConfig
|
4 |
import remfx.utils as utils
|
5 |
from pytorch_lightning.utilities.model_summary import ModelSummary
|
|
|
6 |
import torch
|
7 |
|
8 |
log = utils.get_logger(__name__)
|