File size: 26,506 Bytes
ed1cdd1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
import math
import torch
from torch import nn
from torch.nn import Parameter
import torch.onnx.operators
import torch.nn.functional as F
import utils


class Reshape(nn.Module):
    def __init__(self, *args):
        super(Reshape, self).__init__()
        self.shape = args

    def forward(self, x):
        return x.view(self.shape)


class Permute(nn.Module):
    def __init__(self, *args):
        super(Permute, self).__init__()
        self.args = args

    def forward(self, x):
        return x.permute(self.args)


class LinearNorm(torch.nn.Module):
    def __init__(self, in_dim, out_dim, bias=True, w_init_gain='linear'):
        super(LinearNorm, self).__init__()
        self.linear_layer = torch.nn.Linear(in_dim, out_dim, bias=bias)

        torch.nn.init.xavier_uniform_(
            self.linear_layer.weight,
            gain=torch.nn.init.calculate_gain(w_init_gain))

    def forward(self, x):
        return self.linear_layer(x)


class ConvNorm(torch.nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size=1, stride=1,
                 padding=None, dilation=1, bias=True, w_init_gain='linear'):
        super(ConvNorm, self).__init__()
        if padding is None:
            assert (kernel_size % 2 == 1)
            padding = int(dilation * (kernel_size - 1) / 2)

        self.conv = torch.nn.Conv1d(in_channels, out_channels,
                                    kernel_size=kernel_size, stride=stride,
                                    padding=padding, dilation=dilation,
                                    bias=bias)

        torch.nn.init.xavier_uniform_(
            self.conv.weight, gain=torch.nn.init.calculate_gain(w_init_gain))

    def forward(self, signal):
        conv_signal = self.conv(signal)
        return conv_signal


def Embedding(num_embeddings, embedding_dim, padding_idx=None):
    m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx)
    nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5)
    if padding_idx is not None:
        nn.init.constant_(m.weight[padding_idx], 0)
    return m


def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True, export=False):
    if not export and torch.cuda.is_available():
        try:
            from apex.normalization import FusedLayerNorm
            return FusedLayerNorm(normalized_shape, eps, elementwise_affine)
        except ImportError:
            pass
    return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine)


def Linear(in_features, out_features, bias=True):
    m = nn.Linear(in_features, out_features, bias)
    nn.init.xavier_uniform_(m.weight)
    if bias:
        nn.init.constant_(m.bias, 0.)
    return m


class SinusoidalPositionalEmbedding(nn.Module):
    """This module produces sinusoidal positional embeddings of any length.

    Padding symbols are ignored.
    """

    def __init__(self, embedding_dim, padding_idx, init_size=1024):
        super().__init__()
        self.embedding_dim = embedding_dim
        self.padding_idx = padding_idx
        self.weights = SinusoidalPositionalEmbedding.get_embedding(
            init_size,
            embedding_dim,
            padding_idx,
        )
        self.register_buffer('_float_tensor', torch.FloatTensor(1))

    @staticmethod
    def get_embedding(num_embeddings, embedding_dim, padding_idx=None):
        """Build sinusoidal embeddings.

        This matches the implementation in tensor2tensor, but differs slightly
        from the description in Section 3.5 of "Attention Is All You Need".
        """
        half_dim = embedding_dim // 2
        emb = math.log(10000) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb)
        emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze(1) * emb.unsqueeze(0)
        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1)
        if embedding_dim % 2 == 1:
            # zero pad
            emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1)
        if padding_idx is not None:
            emb[padding_idx, :] = 0
        return emb

    def forward(self, input, incremental_state=None, timestep=None, positions=None, **kwargs):
        """Input is expected to be of size [bsz x seqlen]."""
        bsz, seq_len = input.shape[:2]
        max_pos = self.padding_idx + 1 + seq_len
        if self.weights is None or max_pos > self.weights.size(0):
            # recompute/expand embeddings if needed
            self.weights = SinusoidalPositionalEmbedding.get_embedding(
                max_pos,
                self.embedding_dim,
                self.padding_idx,
            )
        self.weights = self.weights.to(self._float_tensor)

        if incremental_state is not None:
            # positions is the same for every token when decoding a single step
            pos = timestep.view(-1)[0] + 1 if timestep is not None else seq_len
            return self.weights[self.padding_idx + pos, :].expand(bsz, 1, -1)

        positions = utils.make_positions(input, self.padding_idx) if positions is None else positions
        return self.weights.index_select(0, positions.view(-1)).view(bsz, seq_len, -1).detach()

    def max_positions(self):
        """Maximum number of supported positions."""
        return int(1e5)  # an arbitrary large number


class ConvTBC(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, padding=0):
        super(ConvTBC, self).__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.kernel_size = kernel_size
        self.padding = padding

        self.weight = torch.nn.Parameter(torch.Tensor(
            self.kernel_size, in_channels, out_channels))
        self.bias = torch.nn.Parameter(torch.Tensor(out_channels))

    def forward(self, input):
        return torch.conv_tbc(input.contiguous(), self.weight, self.bias, self.padding)


class MultiheadAttention(nn.Module):
    def __init__(self, embed_dim, num_heads, kdim=None, vdim=None, dropout=0., bias=True,
                 add_bias_kv=False, add_zero_attn=False, self_attention=False,
                 encoder_decoder_attention=False):
        super().__init__()
        self.embed_dim = embed_dim
        self.kdim = kdim if kdim is not None else embed_dim
        self.vdim = vdim if vdim is not None else embed_dim
        self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim

        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
        self.scaling = self.head_dim ** -0.5

        self.self_attention = self_attention
        self.encoder_decoder_attention = encoder_decoder_attention

        assert not self.self_attention or self.qkv_same_dim, 'Self-attention requires query, key and ' \
                                                             'value to be of the same size'

        if self.qkv_same_dim:
            self.in_proj_weight = Parameter(torch.Tensor(3 * embed_dim, embed_dim))
        else:
            self.k_proj_weight = Parameter(torch.Tensor(embed_dim, self.kdim))
            self.v_proj_weight = Parameter(torch.Tensor(embed_dim, self.vdim))
            self.q_proj_weight = Parameter(torch.Tensor(embed_dim, embed_dim))

        if bias:
            self.in_proj_bias = Parameter(torch.Tensor(3 * embed_dim))
        else:
            self.register_parameter('in_proj_bias', None)

        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

        if add_bias_kv:
            self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))
            self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))
        else:
            self.bias_k = self.bias_v = None

        self.add_zero_attn = add_zero_attn

        self.reset_parameters()

        self.enable_torch_version = False
        if hasattr(F, "multi_head_attention_forward"):
            self.enable_torch_version = True
        else:
            self.enable_torch_version = False
        self.last_attn_probs = None

    def reset_parameters(self):
        if self.qkv_same_dim:
            nn.init.xavier_uniform_(self.in_proj_weight)
        else:
            nn.init.xavier_uniform_(self.k_proj_weight)
            nn.init.xavier_uniform_(self.v_proj_weight)
            nn.init.xavier_uniform_(self.q_proj_weight)

        nn.init.xavier_uniform_(self.out_proj.weight)
        if self.in_proj_bias is not None:
            nn.init.constant_(self.in_proj_bias, 0.)
            nn.init.constant_(self.out_proj.bias, 0.)
        if self.bias_k is not None:
            nn.init.xavier_normal_(self.bias_k)
        if self.bias_v is not None:
            nn.init.xavier_normal_(self.bias_v)

    def forward(
            self,
            query, key, value,
            key_padding_mask=None,
            incremental_state=None,
            need_weights=True,
            static_kv=False,
            attn_mask=None,
            before_softmax=False,
            need_head_weights=False,
            enc_dec_attn_constraint_mask=None,
            reset_attn_weight=None
    ):
        """Input shape: Time x Batch x Channel

        Args:
            key_padding_mask (ByteTensor, optional): mask to exclude
                keys that are pads, of shape `(batch, src_len)`, where
                padding elements are indicated by 1s.
            need_weights (bool, optional): return the attention weights,
                averaged over heads (default: False).
            attn_mask (ByteTensor, optional): typically used to
                implement causal attention, where the mask prevents the
                attention from looking forward in time (default: None).
            before_softmax (bool, optional): return the raw attention
                weights and values before the attention softmax.
            need_head_weights (bool, optional): return the attention
                weights for each head. Implies *need_weights*. Default:
                return the average attention weights over all heads.
        """
        if need_head_weights:
            need_weights = True

        tgt_len, bsz, embed_dim = query.size()
        assert embed_dim == self.embed_dim
        assert list(query.size()) == [tgt_len, bsz, embed_dim]

        if self.enable_torch_version and incremental_state is None and not static_kv and reset_attn_weight is None:
            if self.qkv_same_dim:
                return F.multi_head_attention_forward(query, key, value,
                                                      self.embed_dim, self.num_heads,
                                                      self.in_proj_weight,
                                                      self.in_proj_bias, self.bias_k, self.bias_v,
                                                      self.add_zero_attn, self.dropout,
                                                      self.out_proj.weight, self.out_proj.bias,
                                                      self.training, key_padding_mask, need_weights,
                                                      attn_mask)
            else:
                return F.multi_head_attention_forward(query, key, value,
                                                      self.embed_dim, self.num_heads,
                                                      torch.empty([0]),
                                                      self.in_proj_bias, self.bias_k, self.bias_v,
                                                      self.add_zero_attn, self.dropout,
                                                      self.out_proj.weight, self.out_proj.bias,
                                                      self.training, key_padding_mask, need_weights,
                                                      attn_mask, use_separate_proj_weight=True,
                                                      q_proj_weight=self.q_proj_weight,
                                                      k_proj_weight=self.k_proj_weight,
                                                      v_proj_weight=self.v_proj_weight)

        if incremental_state is not None:
            print('Not implemented error.')
            exit()
        else:
            saved_state = None

        if self.self_attention:
            # self-attention
            q, k, v = self.in_proj_qkv(query)
        elif self.encoder_decoder_attention:
            # encoder-decoder attention
            q = self.in_proj_q(query)
            if key is None:
                assert value is None
                k = v = None
            else:
                k = self.in_proj_k(key)
                v = self.in_proj_v(key)

        else:
            q = self.in_proj_q(query)
            k = self.in_proj_k(key)
            v = self.in_proj_v(value)
        q *= self.scaling

        if self.bias_k is not None:
            assert self.bias_v is not None
            k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
            v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
            if attn_mask is not None:
                attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1)
            if key_padding_mask is not None:
                key_padding_mask = torch.cat(
                    [key_padding_mask, key_padding_mask.new_zeros(key_padding_mask.size(0), 1)], dim=1)

        q = q.contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim).transpose(0, 1)
        if k is not None:
            k = k.contiguous().view(-1, bsz * self.num_heads, self.head_dim).transpose(0, 1)
        if v is not None:
            v = v.contiguous().view(-1, bsz * self.num_heads, self.head_dim).transpose(0, 1)

        if saved_state is not None:
            print('Not implemented error.')
            exit()

        src_len = k.size(1)

        # This is part of a workaround to get around fork/join parallelism
        # not supporting Optional types.
        if key_padding_mask is not None and key_padding_mask.shape == torch.Size([]):
            key_padding_mask = None

        if key_padding_mask is not None:
            assert key_padding_mask.size(0) == bsz
            assert key_padding_mask.size(1) == src_len

        if self.add_zero_attn:
            src_len += 1
            k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)
            v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)
            if attn_mask is not None:
                attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1)
            if key_padding_mask is not None:
                key_padding_mask = torch.cat(
                    [key_padding_mask, torch.zeros(key_padding_mask.size(0), 1).type_as(key_padding_mask)], dim=1)

        attn_weights = torch.bmm(q, k.transpose(1, 2))
        attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)

        assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]

        if attn_mask is not None:
            if len(attn_mask.shape) == 2:
                attn_mask = attn_mask.unsqueeze(0)
            elif len(attn_mask.shape) == 3:
                attn_mask = attn_mask[:, None].repeat([1, self.num_heads, 1, 1]).reshape(
                    bsz * self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights + attn_mask

        if enc_dec_attn_constraint_mask is not None:  # bs x head x L_kv
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights.masked_fill(
                enc_dec_attn_constraint_mask.unsqueeze(2).bool(),
                -1e9,
            )
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        if key_padding_mask is not None:
            # don't attend to padding symbols
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights.masked_fill(
                key_padding_mask.unsqueeze(1).unsqueeze(2),
                -1e9,
            )
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        attn_logits = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)

        if before_softmax:
            return attn_weights, v

        attn_weights_float = utils.softmax(attn_weights, dim=-1)
        attn_weights = attn_weights_float.type_as(attn_weights)
        attn_probs = F.dropout(attn_weights_float.type_as(attn_weights), p=self.dropout, training=self.training)

        if reset_attn_weight is not None:
            if reset_attn_weight:
                self.last_attn_probs = attn_probs.detach()
            else:
                assert self.last_attn_probs is not None
                attn_probs = self.last_attn_probs
        attn = torch.bmm(attn_probs, v)
        assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
        attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
        attn = self.out_proj(attn)

        if need_weights:
            attn_weights = attn_weights_float.view(bsz, self.num_heads, tgt_len, src_len).transpose(1, 0)
            if not need_head_weights:
                # average attention weights over heads
                attn_weights = attn_weights.mean(dim=0)
        else:
            attn_weights = None

        return attn, (attn_weights, attn_logits)

    def in_proj_qkv(self, query):
        return self._in_proj(query).chunk(3, dim=-1)

    def in_proj_q(self, query):
        if self.qkv_same_dim:
            return self._in_proj(query, end=self.embed_dim)
        else:
            bias = self.in_proj_bias
            if bias is not None:
                bias = bias[:self.embed_dim]
            return F.linear(query, self.q_proj_weight, bias)

    def in_proj_k(self, key):
        if self.qkv_same_dim:
            return self._in_proj(key, start=self.embed_dim, end=2 * self.embed_dim)
        else:
            weight = self.k_proj_weight
            bias = self.in_proj_bias
            if bias is not None:
                bias = bias[self.embed_dim:2 * self.embed_dim]
            return F.linear(key, weight, bias)

    def in_proj_v(self, value):
        if self.qkv_same_dim:
            return self._in_proj(value, start=2 * self.embed_dim)
        else:
            weight = self.v_proj_weight
            bias = self.in_proj_bias
            if bias is not None:
                bias = bias[2 * self.embed_dim:]
            return F.linear(value, weight, bias)

    def _in_proj(self, input, start=0, end=None):
        weight = self.in_proj_weight
        bias = self.in_proj_bias
        weight = weight[start:end, :]
        if bias is not None:
            bias = bias[start:end]
        return F.linear(input, weight, bias)


    def apply_sparse_mask(self, attn_weights, tgt_len, src_len, bsz):
        return attn_weights


class Swish(torch.autograd.Function):
    @staticmethod
    def forward(ctx, i):
        result = i * torch.sigmoid(i)
        ctx.save_for_backward(i)
        return result

    @staticmethod
    def backward(ctx, grad_output):
        i = ctx.saved_variables[0]
        sigmoid_i = torch.sigmoid(i)
        return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i)))


class CustomSwish(nn.Module):
    def forward(self, input_tensor):
        return Swish.apply(input_tensor)
        
class Mish(nn.Module):
    def forward(self, x):
        return x * torch.tanh(F.softplus(x))

class TransformerFFNLayer(nn.Module):
    def __init__(self, hidden_size, filter_size, padding="SAME", kernel_size=1, dropout=0., act='gelu'):
        super().__init__()
        self.kernel_size = kernel_size
        self.dropout = dropout
        self.act = act
        if padding == 'SAME':
            self.ffn_1 = nn.Conv1d(hidden_size, filter_size, kernel_size, padding=kernel_size // 2)
        elif padding == 'LEFT':
            self.ffn_1 = nn.Sequential(
                nn.ConstantPad1d((kernel_size - 1, 0), 0.0),
                nn.Conv1d(hidden_size, filter_size, kernel_size)
            )
        self.ffn_2 = Linear(filter_size, hidden_size)
        if self.act == 'swish':
            self.swish_fn = CustomSwish()

    def forward(self, x, incremental_state=None):
        # x: T x B x C
        if incremental_state is not None:
            assert incremental_state is None, 'Nar-generation does not allow this.'
            exit(1)

        x = self.ffn_1(x.permute(1, 2, 0)).permute(2, 0, 1)
        x = x * self.kernel_size ** -0.5

        if incremental_state is not None:
            x = x[-1:]
        if self.act == 'gelu':
            x = F.gelu(x)
        if self.act == 'relu':
            x = F.relu(x)
        if self.act == 'swish':
            x = self.swish_fn(x)
        x = F.dropout(x, self.dropout, training=self.training)
        x = self.ffn_2(x)
        return x


class BatchNorm1dTBC(nn.Module):
    def __init__(self, c):
        super(BatchNorm1dTBC, self).__init__()
        self.bn = nn.BatchNorm1d(c)

    def forward(self, x):
        """

        :param x: [T, B, C]
        :return: [T, B, C]
        """
        x = x.permute(1, 2, 0)  # [B, C, T]
        x = self.bn(x)  # [B, C, T]
        x = x.permute(2, 0, 1)  # [T, B, C]
        return x


class EncSALayer(nn.Module):
    def __init__(self, c, num_heads, dropout, attention_dropout=0.1,
                 relu_dropout=0.1, kernel_size=9, padding='SAME', norm='ln', act='gelu'):
        super().__init__()
        self.c = c
        self.dropout = dropout
        self.num_heads = num_heads
        if num_heads > 0:
            if norm == 'ln':
                self.layer_norm1 = LayerNorm(c)
            elif norm == 'bn':
                self.layer_norm1 = BatchNorm1dTBC(c)
            self.self_attn = MultiheadAttention(
                self.c, num_heads, self_attention=True, dropout=attention_dropout, bias=False,
            )
        if norm == 'ln':
            self.layer_norm2 = LayerNorm(c)
        elif norm == 'bn':
            self.layer_norm2 = BatchNorm1dTBC(c)
        self.ffn = TransformerFFNLayer(
            c, 4 * c, kernel_size=kernel_size, dropout=relu_dropout, padding=padding, act=act)

    def forward(self, x, encoder_padding_mask=None, **kwargs):
        layer_norm_training = kwargs.get('layer_norm_training', None)
        if layer_norm_training is not None:
            self.layer_norm1.training = layer_norm_training
            self.layer_norm2.training = layer_norm_training
        if self.num_heads > 0:
            residual = x
            x = self.layer_norm1(x)
            x, _, = self.self_attn(
                query=x,
                key=x,
                value=x,
                key_padding_mask=encoder_padding_mask
            )
            x = F.dropout(x, self.dropout, training=self.training)
            x = residual + x
            x = x * (1 - encoder_padding_mask.float()).transpose(0, 1)[..., None]

        residual = x
        x = self.layer_norm2(x)
        x = self.ffn(x)
        x = F.dropout(x, self.dropout, training=self.training)
        x = residual + x
        x = x * (1 - encoder_padding_mask.float()).transpose(0, 1)[..., None]
        return x


class DecSALayer(nn.Module):
    def __init__(self, c, num_heads, dropout, attention_dropout=0.1, relu_dropout=0.1, kernel_size=9, act='gelu'):
        super().__init__()
        self.c = c
        self.dropout = dropout
        self.layer_norm1 = LayerNorm(c)
        self.self_attn = MultiheadAttention(
            c, num_heads, self_attention=True, dropout=attention_dropout, bias=False
        )
        self.layer_norm2 = LayerNorm(c)
        self.encoder_attn = MultiheadAttention(
            c, num_heads, encoder_decoder_attention=True, dropout=attention_dropout, bias=False,
        )
        self.layer_norm3 = LayerNorm(c)
        self.ffn = TransformerFFNLayer(
            c, 4 * c, padding='LEFT', kernel_size=kernel_size, dropout=relu_dropout, act=act)

    def forward(
            self,
            x,
            encoder_out=None,
            encoder_padding_mask=None,
            incremental_state=None,
            self_attn_mask=None,
            self_attn_padding_mask=None,
            attn_out=None,
            reset_attn_weight=None,
            **kwargs,
    ):
        layer_norm_training = kwargs.get('layer_norm_training', None)
        if layer_norm_training is not None:
            self.layer_norm1.training = layer_norm_training
            self.layer_norm2.training = layer_norm_training
            self.layer_norm3.training = layer_norm_training
        residual = x
        x = self.layer_norm1(x)
        x, _ = self.self_attn(
            query=x,
            key=x,
            value=x,
            key_padding_mask=self_attn_padding_mask,
            incremental_state=incremental_state,
            attn_mask=self_attn_mask
        )
        x = F.dropout(x, self.dropout, training=self.training)
        x = residual + x

        residual = x
        x = self.layer_norm2(x)
        if encoder_out is not None:
            x, attn = self.encoder_attn(
                query=x,
                key=encoder_out,
                value=encoder_out,
                key_padding_mask=encoder_padding_mask,
                incremental_state=incremental_state,
                static_kv=True,
                enc_dec_attn_constraint_mask=None, #utils.get_incremental_state(self, incremental_state, 'enc_dec_attn_constraint_mask'),
                reset_attn_weight=reset_attn_weight
            )
            attn_logits = attn[1]
        else:
            assert attn_out is not None
            x = self.encoder_attn.in_proj_v(attn_out.transpose(0, 1))
            attn_logits = None
        x = F.dropout(x, self.dropout, training=self.training)
        x = residual + x

        residual = x
        x = self.layer_norm3(x)
        x = self.ffn(x, incremental_state=incremental_state)
        x = F.dropout(x, self.dropout, training=self.training)
        x = residual + x
        # if len(attn_logits.size()) > 3:
        #    indices = attn_logits.softmax(-1).max(-1).values.sum(-1).argmax(-1)
        #    attn_logits = attn_logits.gather(1,
        #        indices[:, None, None, None].repeat(1, 1, attn_logits.size(-2), attn_logits.size(-1))).squeeze(1)
        return x, attn_logits