File size: 3,830 Bytes
06c138d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
import torch
import torch.nn.functional as F
from torch import nn

from . import layers_123821KB as layers


class BaseASPPNet(nn.Module):
    def __init__(self, nin, ch, dilations=(4, 8, 16)):
        super(BaseASPPNet, self).__init__()
        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)

        self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)

        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)

    def __call__(self, x):
        h, e1 = self.enc1(x)
        h, e2 = self.enc2(h)
        h, e3 = self.enc3(h)
        h, e4 = self.enc4(h)

        h = self.aspp(h)

        h = self.dec4(h, e4)
        h = self.dec3(h, e3)
        h = self.dec2(h, e2)
        h = self.dec1(h, e1)

        return h


class CascadedASPPNet(nn.Module):
    def __init__(self, n_fft):
        super(CascadedASPPNet, self).__init__()
        self.stg1_low_band_net = BaseASPPNet(2, 32)
        self.stg1_high_band_net = BaseASPPNet(2, 32)

        self.stg2_bridge = layers.Conv2DBNActiv(34, 16, 1, 1, 0)
        self.stg2_full_band_net = BaseASPPNet(16, 32)

        self.stg3_bridge = layers.Conv2DBNActiv(66, 32, 1, 1, 0)
        self.stg3_full_band_net = BaseASPPNet(32, 64)

        self.out = nn.Conv2d(64, 2, 1, bias=False)
        self.aux1_out = nn.Conv2d(32, 2, 1, bias=False)
        self.aux2_out = nn.Conv2d(32, 2, 1, bias=False)

        self.max_bin = n_fft // 2
        self.output_bin = n_fft // 2 + 1

        self.offset = 128

    def forward(self, x, aggressiveness=None):
        mix = x.detach()
        x = x.clone()

        x = x[:, :, : self.max_bin]

        bandw = x.size()[2] // 2
        aux1 = torch.cat(
            [
                self.stg1_low_band_net(x[:, :, :bandw]),
                self.stg1_high_band_net(x[:, :, bandw:]),
            ],
            dim=2,
        )

        h = torch.cat([x, aux1], dim=1)
        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))

        h = torch.cat([x, aux1, aux2], dim=1)
        h = self.stg3_full_band_net(self.stg3_bridge(h))

        mask = torch.sigmoid(self.out(h))
        mask = F.pad(
            input=mask,
            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
            mode="replicate",
        )

        if self.training:
            aux1 = torch.sigmoid(self.aux1_out(aux1))
            aux1 = F.pad(
                input=aux1,
                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
                mode="replicate",
            )
            aux2 = torch.sigmoid(self.aux2_out(aux2))
            aux2 = F.pad(
                input=aux2,
                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
                mode="replicate",
            )
            return mask * mix, aux1 * mix, aux2 * mix
        else:
            if aggressiveness:
                mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
                    mask[:, :, : aggressiveness["split_bin"]],
                    1 + aggressiveness["value"] / 3,
                )
                mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
                    mask[:, :, aggressiveness["split_bin"] :],
                    1 + aggressiveness["value"],
                )

            return mask * mix

    def predict(self, x_mag, aggressiveness=None):
        h = self.forward(x_mag, aggressiveness)

        if self.offset > 0:
            h = h[:, :, :, self.offset : -self.offset]
            assert h.size()[3] > 0

        return h