File size: 4,225 Bytes
528df8b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
import torch
import torch.nn.functional as F
from torch import nn

from . import layers_new


class BaseNet(nn.Module):
    def __init__(
        self, nin, nout, nin_lstm, nout_lstm, dilations=((4, 2), (8, 4), (12, 6))
    ):
        super(BaseNet, self).__init__()
        self.enc1 = layers_new.Conv2DBNActiv(nin, nout, 3, 1, 1)
        self.enc2 = layers_new.Encoder(nout, nout * 2, 3, 2, 1)
        self.enc3 = layers_new.Encoder(nout * 2, nout * 4, 3, 2, 1)
        self.enc4 = layers_new.Encoder(nout * 4, nout * 6, 3, 2, 1)
        self.enc5 = layers_new.Encoder(nout * 6, nout * 8, 3, 2, 1)

        self.aspp = layers_new.ASPPModule(nout * 8, nout * 8, dilations, dropout=True)

        self.dec4 = layers_new.Decoder(nout * (6 + 8), nout * 6, 3, 1, 1)
        self.dec3 = layers_new.Decoder(nout * (4 + 6), nout * 4, 3, 1, 1)
        self.dec2 = layers_new.Decoder(nout * (2 + 4), nout * 2, 3, 1, 1)
        self.lstm_dec2 = layers_new.LSTMModule(nout * 2, nin_lstm, nout_lstm)
        self.dec1 = layers_new.Decoder(nout * (1 + 2) + 1, nout * 1, 3, 1, 1)

    def __call__(self, x):
        e1 = self.enc1(x)
        e2 = self.enc2(e1)
        e3 = self.enc3(e2)
        e4 = self.enc4(e3)
        e5 = self.enc5(e4)

        h = self.aspp(e5)

        h = self.dec4(h, e4)
        h = self.dec3(h, e3)
        h = self.dec2(h, e2)
        h = torch.cat([h, self.lstm_dec2(h)], dim=1)
        h = self.dec1(h, e1)

        return h


class CascadedNet(nn.Module):
    def __init__(self, n_fft, nout=32, nout_lstm=128):
        super(CascadedNet, self).__init__()

        self.max_bin = n_fft // 2
        self.output_bin = n_fft // 2 + 1
        self.nin_lstm = self.max_bin // 2
        self.offset = 64

        self.stg1_low_band_net = nn.Sequential(
            BaseNet(2, nout // 2, self.nin_lstm // 2, nout_lstm),
            layers_new.Conv2DBNActiv(nout // 2, nout // 4, 1, 1, 0),
        )

        self.stg1_high_band_net = BaseNet(
            2, nout // 4, self.nin_lstm // 2, nout_lstm // 2
        )

        self.stg2_low_band_net = nn.Sequential(
            BaseNet(nout // 4 + 2, nout, self.nin_lstm // 2, nout_lstm),
            layers_new.Conv2DBNActiv(nout, nout // 2, 1, 1, 0),
        )
        self.stg2_high_band_net = BaseNet(
            nout // 4 + 2, nout // 2, self.nin_lstm // 2, nout_lstm // 2
        )

        self.stg3_full_band_net = BaseNet(
            3 * nout // 4 + 2, nout, self.nin_lstm, nout_lstm
        )

        self.out = nn.Conv2d(nout, 2, 1, bias=False)
        self.aux_out = nn.Conv2d(3 * nout // 4, 2, 1, bias=False)

    def forward(self, x):
        x = x[:, :, : self.max_bin]

        bandw = x.size()[2] // 2
        l1_in = x[:, :, :bandw]
        h1_in = x[:, :, bandw:]
        l1 = self.stg1_low_band_net(l1_in)
        h1 = self.stg1_high_band_net(h1_in)
        aux1 = torch.cat([l1, h1], dim=2)

        l2_in = torch.cat([l1_in, l1], dim=1)
        h2_in = torch.cat([h1_in, h1], dim=1)
        l2 = self.stg2_low_band_net(l2_in)
        h2 = self.stg2_high_band_net(h2_in)
        aux2 = torch.cat([l2, h2], dim=2)

        f3_in = torch.cat([x, aux1, aux2], dim=1)
        f3 = self.stg3_full_band_net(f3_in)

        mask = torch.sigmoid(self.out(f3))
        mask = F.pad(
            input=mask,
            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
            mode="replicate",
        )

        if self.training:
            aux = torch.cat([aux1, aux2], dim=1)
            aux = torch.sigmoid(self.aux_out(aux))
            aux = F.pad(
                input=aux,
                pad=(0, 0, 0, self.output_bin - aux.size()[2]),
                mode="replicate",
            )
            return mask, aux
        else:
            return mask

    def predict_mask(self, x):
        mask = self.forward(x)

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

        return mask

    def predict(self, x, aggressiveness=None):
        mask = self.forward(x)
        pred_mag = x * mask

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

        return pred_mag