File size: 4,755 Bytes
d596074
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import torch
import torch.nn as nn
import torch.nn.functional as F
from scaling import Balancer


class Decoder(nn.Module):
    """This class modifies the stateless decoder from the following paper:

        RNN-transducer with stateless prediction network
        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419

    It removes the recurrent connection from the decoder, i.e., the prediction
    network. Different from the above paper, it adds an extra Conv1d
    right after the embedding layer.

    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
    """

    def __init__(
        self,
        vocab_size: int,
        decoder_dim: int,
        blank_id: int,
        context_size: int,
    ):
        """
        Args:
          vocab_size:
            Number of tokens of the modeling unit including blank.
          decoder_dim:
            Dimension of the input embedding, and of the decoder output.
          blank_id:
            The ID of the blank symbol.
          context_size:
            Number of previous words to use to predict the next word.
            1 means bigram; 2 means trigram. n means (n+1)-gram.
        """
        super().__init__()

        self.embedding = nn.Embedding(
            num_embeddings=vocab_size,
            embedding_dim=decoder_dim,
        )
        # the balancers are to avoid any drift in the magnitude of the
        # embeddings, which would interact badly with parameter averaging.
        self.balancer = Balancer(
            decoder_dim,
            channel_dim=-1,
            min_positive=0.0,
            max_positive=1.0,
            min_abs=0.5,
            max_abs=1.0,
            prob=0.05,
        )

        self.blank_id = blank_id

        assert context_size >= 1, context_size
        self.context_size = context_size
        self.vocab_size = vocab_size

        if context_size > 1:
            self.conv = nn.Conv1d(
                in_channels=decoder_dim,
                out_channels=decoder_dim,
                kernel_size=context_size,
                padding=0,
                groups=decoder_dim // 4,  # group size == 4
                bias=False,
            )
            self.balancer2 = Balancer(
                decoder_dim,
                channel_dim=-1,
                min_positive=0.0,
                max_positive=1.0,
                min_abs=0.5,
                max_abs=1.0,
                prob=0.05,
            )
        else:
            # To avoid `RuntimeError: Module 'Decoder' has no attribute 'conv'`
            # when inference with torch.jit.script and context_size == 1
            self.conv = nn.Identity()
            self.balancer2 = nn.Identity()

    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
        """
        Args:
          y:
            A 2-D tensor of shape (N, U).
          need_pad:
            True to left pad the input. Should be True during training.
            False to not pad the input. Should be False during inference.
        Returns:
          Return a tensor of shape (N, U, decoder_dim).
        """
        y = y.to(torch.int64)
        # this stuff about clamp() is a temporary fix for a mismatch
        # at utterance start, we use negative ids in beam_search.py
        embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)

        embedding_out = self.balancer(embedding_out)

        if self.context_size > 1:
            embedding_out = embedding_out.permute(0, 2, 1)
            if need_pad is True:
                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
            else:
                # During inference time, there is no need to do extra padding
                # as we only need one output
                assert embedding_out.size(-1) == self.context_size
            embedding_out = self.conv(embedding_out)
            embedding_out = embedding_out.permute(0, 2, 1)
            embedding_out = F.relu(embedding_out)
            embedding_out = self.balancer2(embedding_out)

        return embedding_out