File size: 6,610 Bytes
0883aa1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# 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}
# }
#

"""ConvolutionModule definition."""

from typing import Tuple

import torch
from torch import nn


class ConvolutionModule(nn.Module):
    """ConvolutionModule in Conformer model."""

    def __init__(
        self,
        channels: int,
        kernel_size: int = 15,
        activation: nn.Module = nn.ReLU(),
        norm: str = "batch_norm",
        causal: bool = False,
        bias: bool = True,
        adaptive_scale: bool = False,
        init_weights: bool = False,
    ):
        """Construct an ConvolutionModule object.
        Args:
            channels (int): The number of channels of conv layers.
            kernel_size (int): Kernel size of conv layers.
            causal (int): Whether use causal convolution or not
        """
        super().__init__()
        self.bias = bias
        self.channels = channels
        self.kernel_size = kernel_size
        self.adaptive_scale = adaptive_scale
        self.ada_scale = torch.nn.Parameter(
            torch.ones([1, 1, channels]), requires_grad=adaptive_scale
        )
        self.ada_bias = torch.nn.Parameter(
            torch.zeros([1, 1, channels]), requires_grad=adaptive_scale
        )

        self.pointwise_conv1 = nn.Conv1d(
            channels,
            2 * channels,
            kernel_size=1,
            stride=1,
            padding=0,
            bias=bias,
        )
        # self.lorder is used to distinguish if it's a causal convolution,
        # if self.lorder > 0: it's a causal convolution, the input will be
        #    padded with self.lorder frames on the left in forward.
        # else: it's a symmetrical convolution
        if causal:
            padding = 0
            self.lorder = kernel_size - 1
        else:
            # kernel_size should be an odd number for none causal convolution
            assert (kernel_size - 1) % 2 == 0
            padding = (kernel_size - 1) // 2
            self.lorder = 0
        self.depthwise_conv = nn.Conv1d(
            channels,
            channels,
            kernel_size,
            stride=1,
            padding=padding,
            groups=channels,
            bias=bias,
        )

        assert norm in ["batch_norm", "layer_norm"]
        if norm == "batch_norm":
            self.use_layer_norm = False
            self.norm = nn.BatchNorm1d(channels)
        else:
            self.use_layer_norm = True
            self.norm = nn.LayerNorm(channels)

        self.pointwise_conv2 = nn.Conv1d(
            channels,
            channels,
            kernel_size=1,
            stride=1,
            padding=0,
            bias=bias,
        )
        self.activation = activation
        if init_weights:
            self.init_weights()

    def init_weights(self):
        pw_max = self.channels**-0.5
        dw_max = self.kernel_size**-0.5
        torch.nn.init.uniform_(self.pointwise_conv1.weight.data, -pw_max, pw_max)
        if self.bias:
            torch.nn.init.uniform_(self.pointwise_conv1.bias.data, -pw_max, pw_max)
        torch.nn.init.uniform_(self.depthwise_conv.weight.data, -dw_max, dw_max)
        if self.bias:
            torch.nn.init.uniform_(self.depthwise_conv.bias.data, -dw_max, dw_max)
        torch.nn.init.uniform_(self.pointwise_conv2.weight.data, -pw_max, pw_max)
        if self.bias:
            torch.nn.init.uniform_(self.pointwise_conv2.bias.data, -pw_max, pw_max)

    def forward(
        self,
        x: torch.Tensor,
        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
        cache: torch.Tensor = torch.zeros((0, 0, 0)),
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Compute convolution module.
        Args:
            x (torch.Tensor): Input tensor (#batch, time, channels).
            mask_pad (torch.Tensor): used for batch padding (#batch, 1, time),
                (0, 0, 0) means fake mask.
            cache (torch.Tensor): left context cache, it is only
                used in causal convolution (#batch, channels, cache_t),
                (0, 0, 0) meas fake cache.
        Returns:
            torch.Tensor: Output tensor (#batch, time, channels).
        """
        if self.adaptive_scale:
            x = self.ada_scale * x + self.ada_bias
        # exchange the temporal dimension and the feature dimension
        x = x.transpose(1, 2)  # (#batch, channels, time)
        # mask batch padding
        if mask_pad.size(2) > 0:  # time > 0
            x.masked_fill_(~mask_pad, 0.0)

        if self.lorder > 0:
            if cache.size(2) == 0:  # cache_t == 0
                x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
            else:
                assert cache.size(0) == x.size(0)  # equal batch
                assert cache.size(1) == x.size(1)  # equal channel
                x = torch.cat((cache, x), dim=2)
            assert x.size(2) > self.lorder
            new_cache = x[:, :, -self.lorder :]
        else:
            # It's better we just return None if no cache is required,
            # However, for JIT export, here we just fake one tensor instead of
            # None.
            new_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device)

        # GLU mechanism
        x = self.pointwise_conv1(x)  # (batch, 2*channel, dim)
        x = nn.functional.glu(x, dim=1)  # (batch, channel, dim)

        # 1D Depthwise Conv
        x = self.depthwise_conv(x)
        if self.use_layer_norm:
            x = x.transpose(1, 2)
        x = self.activation(self.norm(x))
        if self.use_layer_norm:
            x = x.transpose(1, 2)
        x = self.pointwise_conv2(x)
        # mask batch padding
        if mask_pad.size(2) > 0:  # time > 0
            x.masked_fill_(~mask_pad, 0.0)

        return x.transpose(1, 2), new_cache