File size: 14,059 Bytes
5548515
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
# 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}
# }
#

"""Multi-Head Attention layer definition."""

import math
from typing import Tuple

import torch
from torch import nn


class MultiHeadedAttention(nn.Module):
    """Multi-Head Attention layer.

    Args:
        n_head (int): The number of heads.
        n_feat (int): The number of features.
        dropout_rate (float): Dropout rate.

    """

    def __init__(self, n_head: int, n_feat: int, dropout_rate: float):
        """Construct an MultiHeadedAttention object."""
        super().__init__()
        assert n_feat % n_head == 0
        # We assume d_v always equals d_k
        self.d_k = n_feat // n_head
        self.h = n_head
        self.linear_q = nn.Linear(n_feat, n_feat)
        self.linear_k = nn.Linear(n_feat, n_feat)
        self.linear_v = nn.Linear(n_feat, n_feat)
        self.linear_out = nn.Linear(n_feat, n_feat)
        self.dropout = nn.Dropout(p=dropout_rate)

    def forward_qkv(
        self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Transform query, key and value.

        Args:
            query (torch.Tensor): Query tensor (#batch, time1, size).
            key (torch.Tensor): Key tensor (#batch, time2, size).
            value (torch.Tensor): Value tensor (#batch, time2, size).

        Returns:
            torch.Tensor: Transformed query tensor, size
                (#batch, n_head, time1, d_k).
            torch.Tensor: Transformed key tensor, size
                (#batch, n_head, time2, d_k).
            torch.Tensor: Transformed value tensor, size
                (#batch, n_head, time2, d_k).

        """
        n_batch = query.size(0)
        q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
        k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
        v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
        q = q.transpose(1, 2)  # (batch, head, time1, d_k)
        k = k.transpose(1, 2)  # (batch, head, time2, d_k)
        v = v.transpose(1, 2)  # (batch, head, time2, d_k)

        return q, k, v

    def forward_attention(
        self,
        value: torch.Tensor,
        scores: torch.Tensor,
        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
    ) -> torch.Tensor:
        """Compute attention context vector.

        Args:
            value (torch.Tensor): Transformed value, size
                (#batch, n_head, time2, d_k).
            scores (torch.Tensor): Attention score, size
                (#batch, n_head, time1, time2).
            mask (torch.Tensor): Mask, size (#batch, 1, time2) or
                (#batch, time1, time2), (0, 0, 0) means fake mask.

        Returns:
            torch.Tensor: Transformed value (#batch, time1, d_model)
                weighted by the attention score (#batch, time1, time2).

        """
        n_batch = value.size(0)
        # NOTE(xcsong): When will `if mask.size(2) > 0` be True?
        #   1. onnx(16/4) [WHY? Because we feed real cache & real mask for the
        #           1st chunk to ease the onnx export.]
        #   2. pytorch training
        if mask.size(2) > 0:  # time2 > 0
            mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)
            # For last chunk, time2 might be larger than scores.size(-1)
            mask = mask[:, :, :, : scores.size(-1)]  # (batch, 1, *, time2)
            scores = scores.masked_fill(mask, -float("inf"))
            attn = torch.softmax(scores, dim=-1).masked_fill(
                mask, 0.0
            )  # (batch, head, time1, time2)
        # NOTE(xcsong): When will `if mask.size(2) > 0` be False?
        #   1. onnx(16/-1, -1/-1, 16/0)
        #   2. jit (16/-1, -1/-1, 16/0, 16/4)
        else:
            attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)

        p_attn = self.dropout(attn)
        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
        x = (
            x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
        )  # (batch, time1, d_model)

        return self.linear_out(x)  # (batch, time1, d_model)

    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
        pos_emb: torch.Tensor = torch.empty(0),
        cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Compute scaled dot product attention.

        Args:
            query (torch.Tensor): Query tensor (#batch, time1, size).
            key (torch.Tensor): Key tensor (#batch, time2, size).
            value (torch.Tensor): Value tensor (#batch, time2, size).
            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
                (#batch, time1, time2).
                1.When applying cross attention between decoder and encoder,
                the batch padding mask for input is in (#batch, 1, T) shape.
                2.When applying self attention of encoder,
                the mask is in (#batch, T, T)  shape.
                3.When applying self attention of decoder,
                the mask is in (#batch, L, L)  shape.
                4.If the different position in decoder see different block
                of the encoder, such as Mocha, the passed in mask could be
                in (#batch, L, T) shape. But there is no such case in current
                Wenet.
            cache (torch.Tensor): Cache tensor (1, head, cache_t, d_k * 2),
                where `cache_t == chunk_size * num_decoding_left_chunks`
                and `head * d_k == size`


        Returns:
            torch.Tensor: Output tensor (#batch, time1, d_model).
            torch.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2)
                where `cache_t == chunk_size * num_decoding_left_chunks`
                and `head * d_k == size`

        """
        q, k, v = self.forward_qkv(query, key, value)

        # NOTE(xcsong):
        #   when export onnx model, for 1st chunk, we feed
        #       cache(1, head, 0, d_k * 2) (16/-1, -1/-1, 16/0 mode)
        #       or cache(1, head, real_cache_t, d_k * 2) (16/4 mode).
        #       In all modes, `if cache.size(0) > 0` will alwayse be `True`
        #       and we will always do splitting and
        #       concatnation(this will simplify onnx export). Note that
        #       it's OK to concat & split zero-shaped tensors(see code below).
        #   when export jit  model, for 1st chunk, we always feed
        #       cache(0, 0, 0, 0) since jit supports dynamic if-branch.
        # >>> a = torch.ones((1, 2, 0, 4))
        # >>> b = torch.ones((1, 2, 3, 4))
        # >>> c = torch.cat((a, b), dim=2)
        # >>> torch.equal(b, c)        # True
        # >>> d = torch.split(a, 2, dim=-1)
        # >>> torch.equal(d[0], d[1])  # True
        if cache.size(0) > 0:
            key_cache, value_cache = torch.split(cache, cache.size(-1) // 2, dim=-1)
            k = torch.cat([key_cache, k], dim=2)
            v = torch.cat([value_cache, v], dim=2)
        # NOTE(xcsong): We do cache slicing in encoder.forward_chunk, since it's
        #   non-trivial to calculate `next_cache_start` here.
        new_cache = torch.cat((k, v), dim=-1)

        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
        return self.forward_attention(v, scores, mask), new_cache


class RelPositionMultiHeadedAttention(MultiHeadedAttention):
    """Multi-Head Attention layer with relative position encoding.
    Paper: https://arxiv.org/abs/1901.02860
    Args:
        n_head (int): The number of heads.
        n_feat (int): The number of features.
        dropout_rate (float): Dropout rate.
    """

    def __init__(self, n_head, n_feat, dropout_rate):
        """Construct an RelPositionMultiHeadedAttention object."""
        super().__init__(n_head, n_feat, dropout_rate)
        # linear transformation for positional encoding
        self.linear_pos = nn.Linear(n_feat, n_feat, bias=False)
        # these two learnable bias are used in matrix c and matrix d
        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
        self.pos_bias_u = nn.Parameter(torch.Tensor(self.h, self.d_k))
        self.pos_bias_v = nn.Parameter(torch.Tensor(self.h, self.d_k))
        torch.nn.init.xavier_uniform_(self.pos_bias_u)
        torch.nn.init.xavier_uniform_(self.pos_bias_v)

    def rel_shift(self, x, zero_triu: bool = False):
        """Compute relative positinal encoding.
        Args:
            x (torch.Tensor): Input tensor (batch, time, size).
            zero_triu (bool): If true, return the lower triangular part of
                the matrix.
        Returns:
            torch.Tensor: Output tensor.
        """

        zero_pad = torch.zeros(
            (x.size()[0], x.size()[1], x.size()[2], 1), device=x.device, dtype=x.dtype
        )
        x_padded = torch.cat([zero_pad, x], dim=-1)

        x_padded = x_padded.view(x.size()[0], x.size()[1], x.size(3) + 1, x.size(2))
        x = x_padded[:, :, 1:].view_as(x)

        if zero_triu:
            ones = torch.ones((x.size(2), x.size(3)))
            x = x * torch.tril(ones, x.size(3) - x.size(2))[None, None, :, :]

        return x

    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
        pos_emb: torch.Tensor = torch.empty(0),
        cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Compute 'Scaled Dot Product Attention' with rel. positional encoding.
        Args:
            query (torch.Tensor): Query tensor (#batch, time1, size).
            key (torch.Tensor): Key tensor (#batch, time2, size).
            value (torch.Tensor): Value tensor (#batch, time2, size).
            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
                (#batch, time1, time2), (0, 0, 0) means fake mask.
            pos_emb (torch.Tensor): Positional embedding tensor
                (#batch, time2, size).
            cache (torch.Tensor): Cache tensor (1, head, cache_t, d_k * 2),
                where `cache_t == chunk_size * num_decoding_left_chunks`
                and `head * d_k == size`
        Returns:
            torch.Tensor: Output tensor (#batch, time1, d_model).
            torch.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2)
                where `cache_t == chunk_size * num_decoding_left_chunks`
                and `head * d_k == size`
        """
        q, k, v = self.forward_qkv(query, key, value)
        q = q.transpose(1, 2)  # (batch, time1, head, d_k)

        # NOTE(xcsong):
        #   when export onnx model, for 1st chunk, we feed
        #       cache(1, head, 0, d_k * 2) (16/-1, -1/-1, 16/0 mode)
        #       or cache(1, head, real_cache_t, d_k * 2) (16/4 mode).
        #       In all modes, `if cache.size(0) > 0` will alwayse be `True`
        #       and we will always do splitting and
        #       concatnation(this will simplify onnx export). Note that
        #       it's OK to concat & split zero-shaped tensors(see code below).
        #   when export jit  model, for 1st chunk, we always feed
        #       cache(0, 0, 0, 0) since jit supports dynamic if-branch.
        # >>> a = torch.ones((1, 2, 0, 4))
        # >>> b = torch.ones((1, 2, 3, 4))
        # >>> c = torch.cat((a, b), dim=2)
        # >>> torch.equal(b, c)        # True
        # >>> d = torch.split(a, 2, dim=-1)
        # >>> torch.equal(d[0], d[1])  # True
        if cache.size(0) > 0:
            key_cache, value_cache = torch.split(cache, cache.size(-1) // 2, dim=-1)
            k = torch.cat([key_cache, k], dim=2)
            v = torch.cat([value_cache, v], dim=2)
        # NOTE(xcsong): We do cache slicing in encoder.forward_chunk, since it's
        #   non-trivial to calculate `next_cache_start` here.
        new_cache = torch.cat((k, v), dim=-1)

        n_batch_pos = pos_emb.size(0)
        p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
        p = p.transpose(1, 2)  # (batch, head, time1, d_k)

        # (batch, head, time1, d_k)
        q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
        # (batch, head, time1, d_k)
        q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)

        # compute attention score
        # first compute matrix a and matrix c
        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
        # (batch, head, time1, time2)
        matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))

        # compute matrix b and matrix d
        # (batch, head, time1, time2)
        matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
        # Remove rel_shift since it is useless in speech recognition,
        # and it requires special attention for streaming.
        # matrix_bd = self.rel_shift(matrix_bd)

        scores = (matrix_ac + matrix_bd) / math.sqrt(
            self.d_k
        )  # (batch, head, time1, time2)

        return self.forward_attention(v, scores, mask), new_cache