File size: 8,366 Bytes
c968fc3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
# This module is from [WeNet](https://github.com/wenet-e2e/wenet).

# ## Citations

# ```bibtex
# @inproceedings{yao2021wenet,
#   title={WeNet: Production oriented Streaming and Non-streaming End-to-End Speech Recognition Toolkit},
#   author={Yao, Zhuoyuan and Wu, Di and Wang, Xiong and Zhang, Binbin and Yu, Fan and Yang, Chao and Peng, Zhendong and Chen, Xiaoyu and Xie, Lei and Lei, Xin},
#   booktitle={Proc. Interspeech},
#   year={2021},
#   address={Brno, Czech Republic },
#   organization={IEEE}
# }

# @article{zhang2022wenet,
#   title={WeNet 2.0: More Productive End-to-End Speech Recognition Toolkit},
#   author={Zhang, Binbin and Wu, Di and Peng, Zhendong and Song, Xingchen and Yao, Zhuoyuan and Lv, Hang and Xie, Lei and Yang, Chao and Pan, Fuping and Niu, Jianwei},
#   journal={arXiv preprint arXiv:2203.15455},
#   year={2022}
# }
#


"""Subsampling layer definition."""

from typing import Tuple, Union

import torch


class BaseSubsampling(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.right_context = 0
        self.subsampling_rate = 1

    def position_encoding(
        self, offset: Union[int, torch.Tensor], size: int
    ) -> torch.Tensor:
        return self.pos_enc.position_encoding(offset, size)


class LinearNoSubsampling(BaseSubsampling):
    """Linear transform the input without subsampling

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.

    """

    def __init__(
        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
    ):
        """Construct an linear object."""
        super().__init__()
        self.out = torch.nn.Sequential(
            torch.nn.Linear(idim, odim),
            torch.nn.LayerNorm(odim, eps=1e-5),
            torch.nn.Dropout(dropout_rate),
        )
        self.pos_enc = pos_enc_class
        self.right_context = 0
        self.subsampling_rate = 1

    def forward(
        self,
        x: torch.Tensor,
        x_mask: torch.Tensor,
        offset: Union[int, torch.Tensor] = 0,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Input x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: linear input tensor (#batch, time', odim),
                where time' = time .
            torch.Tensor: linear input mask (#batch, 1, time'),
                where time' = time .

        """
        x = self.out(x)
        x, pos_emb = self.pos_enc(x, offset)
        return x, pos_emb, x_mask


class Conv2dSubsampling4(BaseSubsampling):
    """Convolutional 2D subsampling (to 1/4 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.

    """

    def __init__(
        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
    ):
        """Construct an Conv2dSubsampling4 object."""
        super().__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 3, 2),
            torch.nn.ReLU(),
        )
        self.out = torch.nn.Sequential(
            torch.nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
        )
        self.pos_enc = pos_enc_class
        # The right context for every conv layer is computed by:
        # (kernel_size - 1) * frame_rate_of_this_layer
        self.subsampling_rate = 4
        # 6 = (3 - 1) * 1 + (3 - 1) * 2
        self.right_context = 6

    def forward(
        self,
        x: torch.Tensor,
        x_mask: torch.Tensor,
        offset: Union[int, torch.Tensor] = 0,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 4.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 4.
            torch.Tensor: positional encoding

        """
        x = x.unsqueeze(1)  # (b, c=1, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        x, pos_emb = self.pos_enc(x, offset)
        return x, pos_emb, x_mask[:, :, 2::2][:, :, 2::2]


class Conv2dSubsampling6(BaseSubsampling):
    """Convolutional 2D subsampling (to 1/6 length).
    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.
    """

    def __init__(
        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
    ):
        """Construct an Conv2dSubsampling6 object."""
        super().__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 5, 3),
            torch.nn.ReLU(),
        )
        self.linear = torch.nn.Linear(odim * (((idim - 1) // 2 - 2) // 3), odim)
        self.pos_enc = pos_enc_class
        # 10 = (3 - 1) * 1 + (5 - 1) * 2
        self.subsampling_rate = 6
        self.right_context = 10

    def forward(
        self,
        x: torch.Tensor,
        x_mask: torch.Tensor,
        offset: Union[int, torch.Tensor] = 0,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Subsample x.
        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 6.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 6.
            torch.Tensor: positional encoding
        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.linear(x.transpose(1, 2).contiguous().view(b, t, c * f))
        x, pos_emb = self.pos_enc(x, offset)
        return x, pos_emb, x_mask[:, :, 2::2][:, :, 4::3]


class Conv2dSubsampling8(BaseSubsampling):
    """Convolutional 2D subsampling (to 1/8 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.

    """

    def __init__(
        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
    ):
        """Construct an Conv2dSubsampling8 object."""
        super().__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 3, 2),
            torch.nn.ReLU(),
        )
        self.linear = torch.nn.Linear(
            odim * ((((idim - 1) // 2 - 1) // 2 - 1) // 2), odim
        )
        self.pos_enc = pos_enc_class
        self.subsampling_rate = 8
        # 14 = (3 - 1) * 1 + (3 - 1) * 2 + (3 - 1) * 4
        self.right_context = 14

    def forward(
        self,
        x: torch.Tensor,
        x_mask: torch.Tensor,
        offset: Union[int, torch.Tensor] = 0,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 8.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 8.
            torch.Tensor: positional encoding
        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.linear(x.transpose(1, 2).contiguous().view(b, t, c * f))
        x, pos_emb = self.pos_enc(x, offset)
        return x, pos_emb, x_mask[:, :, 2::2][:, :, 2::2][:, :, 2::2]