File size: 19,423 Bytes
ad16788
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
# Copyright 2019 Shigeki Karita
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

"""Decoder definition."""
from typing import Any
from typing import List
from typing import Sequence
from typing import Tuple

import torch
from typeguard import check_argument_types

from espnet.nets.pytorch_backend.nets_utils import make_pad_mask
from espnet.nets.pytorch_backend.transformer.attention import MultiHeadedAttention
from espnet.nets.pytorch_backend.transformer.decoder_layer import DecoderLayer
from espnet.nets.pytorch_backend.transformer.dynamic_conv import DynamicConvolution
from espnet.nets.pytorch_backend.transformer.dynamic_conv2d import DynamicConvolution2D
from espnet.nets.pytorch_backend.transformer.embedding import PositionalEncoding
from espnet.nets.pytorch_backend.transformer.layer_norm import LayerNorm
from espnet.nets.pytorch_backend.transformer.lightconv import LightweightConvolution
from espnet.nets.pytorch_backend.transformer.lightconv2d import LightweightConvolution2D
from espnet.nets.pytorch_backend.transformer.mask import subsequent_mask
from espnet.nets.pytorch_backend.transformer.positionwise_feed_forward import (
    PositionwiseFeedForward,  # noqa: H301
)
from espnet.nets.pytorch_backend.transformer.repeat import repeat
from espnet.nets.scorer_interface import BatchScorerInterface
from espnet2.asr.decoder.abs_decoder import AbsDecoder


class BaseTransformerDecoder(AbsDecoder, BatchScorerInterface):
    """Base class of Transfomer decoder module.

    Args:
        vocab_size: output dim
        encoder_output_size: dimension of attention
        attention_heads: the number of heads of multi head attention
        linear_units: the number of units of position-wise feed forward
        num_blocks: the number of decoder blocks
        dropout_rate: dropout rate
        self_attention_dropout_rate: dropout rate for attention
        input_layer: input layer type
        use_output_layer: whether to use output layer
        pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
        normalize_before: whether to use layer_norm before the first block
        concat_after: whether to concat attention layer's input and output
            if True, additional linear will be applied.
            i.e. x -> x + linear(concat(x, att(x)))
            if False, no additional linear will be applied.
            i.e. x -> x + att(x)
    """

    def __init__(
        self,
        vocab_size: int,
        encoder_output_size: int,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        input_layer: str = "embed",
        use_output_layer: bool = True,
        pos_enc_class=PositionalEncoding,
        normalize_before: bool = True,
    ):
        assert check_argument_types()
        super().__init__()
        attention_dim = encoder_output_size

        if input_layer == "embed":
            self.embed = torch.nn.Sequential(
                torch.nn.Embedding(vocab_size, attention_dim),
                pos_enc_class(attention_dim, positional_dropout_rate),
            )
        elif input_layer == "linear":
            self.embed = torch.nn.Sequential(
                torch.nn.Linear(vocab_size, attention_dim),
                torch.nn.LayerNorm(attention_dim),
                torch.nn.Dropout(dropout_rate),
                torch.nn.ReLU(),
                pos_enc_class(attention_dim, positional_dropout_rate),
            )
        else:
            raise ValueError(f"only 'embed' or 'linear' is supported: {input_layer}")

        self.normalize_before = normalize_before
        if self.normalize_before:
            self.after_norm = LayerNorm(attention_dim)
        if use_output_layer:
            self.output_layer = torch.nn.Linear(attention_dim, vocab_size)
        else:
            self.output_layer = None

        # Must set by the inheritance
        self.decoders = None

    def forward(
        self,
        hs_pad: torch.Tensor,
        hlens: torch.Tensor,
        ys_in_pad: torch.Tensor,
        ys_in_lens: torch.Tensor,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Forward decoder.

        Args:
            hs_pad: encoded memory, float32  (batch, maxlen_in, feat)
            hlens: (batch)
            ys_in_pad:
                input token ids, int64 (batch, maxlen_out)
                if input_layer == "embed"
                input tensor (batch, maxlen_out, #mels) in the other cases
            ys_in_lens: (batch)
        Returns:
            (tuple): tuple containing:

            x: decoded token score before softmax (batch, maxlen_out, token)
                if use_output_layer is True,
            olens: (batch, )
        """
        tgt = ys_in_pad
        # tgt_mask: (B, 1, L)
        tgt_mask = (~make_pad_mask(ys_in_lens)[:, None, :]).to(tgt.device)
        # m: (1, L, L)
        m = subsequent_mask(tgt_mask.size(-1), device=tgt_mask.device).unsqueeze(0)
        # tgt_mask: (B, L, L)
        tgt_mask = tgt_mask & m

        memory = hs_pad
        memory_mask = (~make_pad_mask(hlens))[:, None, :].to(memory.device)

        x = self.embed(tgt)
        x, tgt_mask, memory, memory_mask = self.decoders(
            x, tgt_mask, memory, memory_mask
        )
        if self.normalize_before:
            x = self.after_norm(x)
        if self.output_layer is not None:
            x = self.output_layer(x)

        olens = tgt_mask.sum(1)
        return x, olens

    def forward_one_step(
        self,
        tgt: torch.Tensor,
        tgt_mask: torch.Tensor,
        memory: torch.Tensor,
        cache: List[torch.Tensor] = None,
    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
        """Forward one step.

        Args:
            tgt: input token ids, int64 (batch, maxlen_out)
            tgt_mask: input token mask,  (batch, maxlen_out)
                      dtype=torch.uint8 in PyTorch 1.2-
                      dtype=torch.bool in PyTorch 1.2+ (include 1.2)
            memory: encoded memory, float32  (batch, maxlen_in, feat)
            cache: cached output list of (batch, max_time_out-1, size)
        Returns:
            y, cache: NN output value and cache per `self.decoders`.
            y.shape` is (batch, maxlen_out, token)
        """
        x = self.embed(tgt)
        if cache is None:
            cache = [None] * len(self.decoders)
        new_cache = []
        for c, decoder in zip(cache, self.decoders):
            x, tgt_mask, memory, memory_mask = decoder(
                x, tgt_mask, memory, None, cache=c
            )
            new_cache.append(x)

        if self.normalize_before:
            y = self.after_norm(x[:, -1])
        else:
            y = x[:, -1]
        if self.output_layer is not None:
            y = torch.log_softmax(self.output_layer(y), dim=-1)

        return y, new_cache

    def score(self, ys, state, x):
        """Score."""
        ys_mask = subsequent_mask(len(ys), device=x.device).unsqueeze(0)
        logp, state = self.forward_one_step(
            ys.unsqueeze(0), ys_mask, x.unsqueeze(0), cache=state
        )
        return logp.squeeze(0), state

    def batch_score(
        self, ys: torch.Tensor, states: List[Any], xs: torch.Tensor
    ) -> Tuple[torch.Tensor, List[Any]]:
        """Score new token batch.

        Args:
            ys (torch.Tensor): torch.int64 prefix tokens (n_batch, ylen).
            states (List[Any]): Scorer states for prefix tokens.
            xs (torch.Tensor):
                The encoder feature that generates ys (n_batch, xlen, n_feat).

        Returns:
            tuple[torch.Tensor, List[Any]]: Tuple of
                batchfied scores for next token with shape of `(n_batch, n_vocab)`
                and next state list for ys.

        """
        # merge states
        n_batch = len(ys)
        n_layers = len(self.decoders)
        if states[0] is None:
            batch_state = None
        else:
            # transpose state of [batch, layer] into [layer, batch]
            batch_state = [
                torch.stack([states[b][i] for b in range(n_batch)])
                for i in range(n_layers)
            ]

        # batch decoding
        ys_mask = subsequent_mask(ys.size(-1), device=xs.device).unsqueeze(0)
        logp, states = self.forward_one_step(ys, ys_mask, xs, cache=batch_state)

        # transpose state of [layer, batch] into [batch, layer]
        state_list = [[states[i][b] for i in range(n_layers)] for b in range(n_batch)]
        return logp, state_list


class TransformerDecoder(BaseTransformerDecoder):
    def __init__(
        self,
        vocab_size: int,
        encoder_output_size: int,
        attention_heads: int = 4,
        linear_units: int = 2048,
        num_blocks: int = 6,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        self_attention_dropout_rate: float = 0.0,
        src_attention_dropout_rate: float = 0.0,
        input_layer: str = "embed",
        use_output_layer: bool = True,
        pos_enc_class=PositionalEncoding,
        normalize_before: bool = True,
        concat_after: bool = False,
    ):
        assert check_argument_types()
        super().__init__(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
            dropout_rate=dropout_rate,
            positional_dropout_rate=positional_dropout_rate,
            input_layer=input_layer,
            use_output_layer=use_output_layer,
            pos_enc_class=pos_enc_class,
            normalize_before=normalize_before,
        )

        attention_dim = encoder_output_size
        self.decoders = repeat(
            num_blocks,
            lambda lnum: DecoderLayer(
                attention_dim,
                MultiHeadedAttention(
                    attention_heads, attention_dim, self_attention_dropout_rate
                ),
                MultiHeadedAttention(
                    attention_heads, attention_dim, src_attention_dropout_rate
                ),
                PositionwiseFeedForward(attention_dim, linear_units, dropout_rate),
                dropout_rate,
                normalize_before,
                concat_after,
            ),
        )


class LightweightConvolutionTransformerDecoder(BaseTransformerDecoder):
    def __init__(
        self,
        vocab_size: int,
        encoder_output_size: int,
        attention_heads: int = 4,
        linear_units: int = 2048,
        num_blocks: int = 6,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        self_attention_dropout_rate: float = 0.0,
        src_attention_dropout_rate: float = 0.0,
        input_layer: str = "embed",
        use_output_layer: bool = True,
        pos_enc_class=PositionalEncoding,
        normalize_before: bool = True,
        concat_after: bool = False,
        conv_wshare: int = 4,
        conv_kernel_length: Sequence[int] = (11, 11, 11, 11, 11, 11),
        conv_usebias: int = False,
    ):
        assert check_argument_types()
        if len(conv_kernel_length) != num_blocks:
            raise ValueError(
                "conv_kernel_length must have equal number of values to num_blocks: "
                f"{len(conv_kernel_length)} != {num_blocks}"
            )
        super().__init__(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
            dropout_rate=dropout_rate,
            positional_dropout_rate=positional_dropout_rate,
            input_layer=input_layer,
            use_output_layer=use_output_layer,
            pos_enc_class=pos_enc_class,
            normalize_before=normalize_before,
        )

        attention_dim = encoder_output_size
        self.decoders = repeat(
            num_blocks,
            lambda lnum: DecoderLayer(
                attention_dim,
                LightweightConvolution(
                    wshare=conv_wshare,
                    n_feat=attention_dim,
                    dropout_rate=self_attention_dropout_rate,
                    kernel_size=conv_kernel_length[lnum],
                    use_kernel_mask=True,
                    use_bias=conv_usebias,
                ),
                MultiHeadedAttention(
                    attention_heads, attention_dim, src_attention_dropout_rate
                ),
                PositionwiseFeedForward(attention_dim, linear_units, dropout_rate),
                dropout_rate,
                normalize_before,
                concat_after,
            ),
        )


class LightweightConvolution2DTransformerDecoder(BaseTransformerDecoder):
    def __init__(
        self,
        vocab_size: int,
        encoder_output_size: int,
        attention_heads: int = 4,
        linear_units: int = 2048,
        num_blocks: int = 6,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        self_attention_dropout_rate: float = 0.0,
        src_attention_dropout_rate: float = 0.0,
        input_layer: str = "embed",
        use_output_layer: bool = True,
        pos_enc_class=PositionalEncoding,
        normalize_before: bool = True,
        concat_after: bool = False,
        conv_wshare: int = 4,
        conv_kernel_length: Sequence[int] = (11, 11, 11, 11, 11, 11),
        conv_usebias: int = False,
    ):
        assert check_argument_types()
        if len(conv_kernel_length) != num_blocks:
            raise ValueError(
                "conv_kernel_length must have equal number of values to num_blocks: "
                f"{len(conv_kernel_length)} != {num_blocks}"
            )
        super().__init__(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
            dropout_rate=dropout_rate,
            positional_dropout_rate=positional_dropout_rate,
            input_layer=input_layer,
            use_output_layer=use_output_layer,
            pos_enc_class=pos_enc_class,
            normalize_before=normalize_before,
        )

        attention_dim = encoder_output_size
        self.decoders = repeat(
            num_blocks,
            lambda lnum: DecoderLayer(
                attention_dim,
                LightweightConvolution2D(
                    wshare=conv_wshare,
                    n_feat=attention_dim,
                    dropout_rate=self_attention_dropout_rate,
                    kernel_size=conv_kernel_length[lnum],
                    use_kernel_mask=True,
                    use_bias=conv_usebias,
                ),
                MultiHeadedAttention(
                    attention_heads, attention_dim, src_attention_dropout_rate
                ),
                PositionwiseFeedForward(attention_dim, linear_units, dropout_rate),
                dropout_rate,
                normalize_before,
                concat_after,
            ),
        )


class DynamicConvolutionTransformerDecoder(BaseTransformerDecoder):
    def __init__(
        self,
        vocab_size: int,
        encoder_output_size: int,
        attention_heads: int = 4,
        linear_units: int = 2048,
        num_blocks: int = 6,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        self_attention_dropout_rate: float = 0.0,
        src_attention_dropout_rate: float = 0.0,
        input_layer: str = "embed",
        use_output_layer: bool = True,
        pos_enc_class=PositionalEncoding,
        normalize_before: bool = True,
        concat_after: bool = False,
        conv_wshare: int = 4,
        conv_kernel_length: Sequence[int] = (11, 11, 11, 11, 11, 11),
        conv_usebias: int = False,
    ):
        assert check_argument_types()
        if len(conv_kernel_length) != num_blocks:
            raise ValueError(
                "conv_kernel_length must have equal number of values to num_blocks: "
                f"{len(conv_kernel_length)} != {num_blocks}"
            )
        super().__init__(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
            dropout_rate=dropout_rate,
            positional_dropout_rate=positional_dropout_rate,
            input_layer=input_layer,
            use_output_layer=use_output_layer,
            pos_enc_class=pos_enc_class,
            normalize_before=normalize_before,
        )
        attention_dim = encoder_output_size

        self.decoders = repeat(
            num_blocks,
            lambda lnum: DecoderLayer(
                attention_dim,
                DynamicConvolution(
                    wshare=conv_wshare,
                    n_feat=attention_dim,
                    dropout_rate=self_attention_dropout_rate,
                    kernel_size=conv_kernel_length[lnum],
                    use_kernel_mask=True,
                    use_bias=conv_usebias,
                ),
                MultiHeadedAttention(
                    attention_heads, attention_dim, src_attention_dropout_rate
                ),
                PositionwiseFeedForward(attention_dim, linear_units, dropout_rate),
                dropout_rate,
                normalize_before,
                concat_after,
            ),
        )


class DynamicConvolution2DTransformerDecoder(BaseTransformerDecoder):
    def __init__(
        self,
        vocab_size: int,
        encoder_output_size: int,
        attention_heads: int = 4,
        linear_units: int = 2048,
        num_blocks: int = 6,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        self_attention_dropout_rate: float = 0.0,
        src_attention_dropout_rate: float = 0.0,
        input_layer: str = "embed",
        use_output_layer: bool = True,
        pos_enc_class=PositionalEncoding,
        normalize_before: bool = True,
        concat_after: bool = False,
        conv_wshare: int = 4,
        conv_kernel_length: Sequence[int] = (11, 11, 11, 11, 11, 11),
        conv_usebias: int = False,
    ):
        assert check_argument_types()
        if len(conv_kernel_length) != num_blocks:
            raise ValueError(
                "conv_kernel_length must have equal number of values to num_blocks: "
                f"{len(conv_kernel_length)} != {num_blocks}"
            )
        super().__init__(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
            dropout_rate=dropout_rate,
            positional_dropout_rate=positional_dropout_rate,
            input_layer=input_layer,
            use_output_layer=use_output_layer,
            pos_enc_class=pos_enc_class,
            normalize_before=normalize_before,
        )
        attention_dim = encoder_output_size

        self.decoders = repeat(
            num_blocks,
            lambda lnum: DecoderLayer(
                attention_dim,
                DynamicConvolution2D(
                    wshare=conv_wshare,
                    n_feat=attention_dim,
                    dropout_rate=self_attention_dropout_rate,
                    kernel_size=conv_kernel_length[lnum],
                    use_kernel_mask=True,
                    use_bias=conv_usebias,
                ),
                MultiHeadedAttention(
                    attention_heads, attention_dim, src_attention_dropout_rate
                ),
                PositionwiseFeedForward(attention_dim, linear_units, dropout_rate),
                dropout_rate,
                normalize_before,
                concat_after,
            ),
        )