File size: 23,257 Bytes
67c46fd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
#!/usr/bin/env python3
# -*- coding: utf-8 -*-

# Copyright 2019 Shigeki Karita
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

"""Subsampling layer definition."""
import numpy as np
import torch
import torch.nn.functional as F
from funasr_detach.models.transformer.embedding import PositionalEncoding
import logging
from funasr_detach.models.scama.utils import sequence_mask
from funasr_detach.models.transformer.utils.nets_utils import (
    sub_factor_to_params,
    pad_to_len,
)
from typing import Optional, Tuple, Union
import math


class TooShortUttError(Exception):
    """Raised when the utt is too short for subsampling.

    Args:
        message (str): Message for error catch
        actual_size (int): the short size that cannot pass the subsampling
        limit (int): the limit size for subsampling

    """

    def __init__(self, message, actual_size, limit):
        """Construct a TooShortUttError for error handler."""
        super().__init__(message)
        self.actual_size = actual_size
        self.limit = limit


def check_short_utt(ins, size):
    """Check if the utterance is too short for subsampling."""
    if isinstance(ins, Conv2dSubsampling2) and size < 3:
        return True, 3
    if isinstance(ins, Conv2dSubsampling) and size < 7:
        return True, 7
    if isinstance(ins, Conv2dSubsampling6) and size < 11:
        return True, 11
    if isinstance(ins, Conv2dSubsampling8) and size < 15:
        return True, 15
    return False, -1


class Conv2dSubsampling(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/4 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, idim, odim, dropout_rate, pos_enc=None):
        """Construct an Conv2dSubsampling object."""
        super(Conv2dSubsampling, self).__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 3, 2),
            torch.nn.ReLU(),
        )
        self.out = torch.nn.Sequential(
            torch.nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim),
            pos_enc if pos_enc is not None else PositionalEncoding(odim, dropout_rate),
        )

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 4.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 4.

        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        return x, x_mask[:, :, :-2:2][:, :, :-2:2]

    def __getitem__(self, key):
        """Get item.

        When reset_parameters() is called, if use_scaled_pos_enc is used,
            return the positioning encoding.

        """
        if key != -1:
            raise NotImplementedError("Support only `-1` (for `reset_parameters`).")
        return self.out[key]


class Conv2dSubsamplingPad(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/4 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, idim, odim, dropout_rate, pos_enc=None):
        """Construct an Conv2dSubsampling object."""
        super(Conv2dSubsamplingPad, self).__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, odim, 3, 2, padding=(0, 0)),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 3, 2, padding=(0, 0)),
            torch.nn.ReLU(),
        )
        self.out = torch.nn.Sequential(
            torch.nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim),
            pos_enc if pos_enc is not None else PositionalEncoding(odim, dropout_rate),
        )
        self.pad_fn = torch.nn.ConstantPad1d((0, 4), 0.0)

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 4.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 4.

        """
        x = x.transpose(1, 2)
        x = self.pad_fn(x)
        x = x.transpose(1, 2)
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        x_len = torch.sum(x_mask[:, 0, :], dim=-1)
        x_len = (x_len - 1) // 2 + 1
        x_len = (x_len - 1) // 2 + 1
        mask = sequence_mask(x_len, None, x_len.dtype, x[0].device)
        return x, mask[:, None, :]

    def __getitem__(self, key):
        """Get item.

        When reset_parameters() is called, if use_scaled_pos_enc is used,
            return the positioning encoding.

        """
        if key != -1:
            raise NotImplementedError("Support only `-1` (for `reset_parameters`).")
        return self.out[key]


class Conv2dSubsampling2(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/2 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, idim, odim, dropout_rate, pos_enc=None):
        """Construct an Conv2dSubsampling2 object."""
        super(Conv2dSubsampling2, self).__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 3, 1),
            torch.nn.ReLU(),
        )
        self.out = torch.nn.Sequential(
            torch.nn.Linear(odim * (((idim - 1) // 2 - 2)), odim),
            pos_enc if pos_enc is not None else PositionalEncoding(odim, dropout_rate),
        )

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 2.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 2.

        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        return x, x_mask[:, :, :-2:2][:, :, :-2:1]

    def __getitem__(self, key):
        """Get item.

        When reset_parameters() is called, if use_scaled_pos_enc is used,
            return the positioning encoding.

        """
        if key != -1:
            raise NotImplementedError("Support only `-1` (for `reset_parameters`).")
        return self.out[key]


class Conv2dSubsampling6(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/6 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, idim, odim, dropout_rate, pos_enc=None):
        """Construct an Conv2dSubsampling6 object."""
        super(Conv2dSubsampling6, self).__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 5, 3),
            torch.nn.ReLU(),
        )
        self.out = torch.nn.Sequential(
            torch.nn.Linear(odim * (((idim - 1) // 2 - 2) // 3), odim),
            pos_enc if pos_enc is not None else PositionalEncoding(odim, dropout_rate),
        )

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 6.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 6.

        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        return x, x_mask[:, :, :-2:2][:, :, :-4:3]


class Conv2dSubsampling8(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/8 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, idim, odim, dropout_rate, pos_enc=None):
        """Construct an Conv2dSubsampling8 object."""
        super(Conv2dSubsampling8, self).__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 3, 2),
            torch.nn.ReLU(),
            torch.nn.Conv2d(odim, odim, 3, 2),
            torch.nn.ReLU(),
        )
        self.out = torch.nn.Sequential(
            torch.nn.Linear(odim * ((((idim - 1) // 2 - 1) // 2 - 1) // 2), odim),
            pos_enc if pos_enc is not None else PositionalEncoding(odim, dropout_rate),
        )

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 8.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 8.

        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        return x, x_mask[:, :, :-2:2][:, :, :-2:2][:, :, :-2:2]


class Conv1dSubsampling(torch.nn.Module):
    """Convolutional 1D subsampling (to 1/2 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(
        self,
        idim,
        odim,
        kernel_size,
        stride,
        pad,
        tf2torch_tensor_name_prefix_torch: str = "stride_conv",
        tf2torch_tensor_name_prefix_tf: str = "seq2seq/proj_encoder/downsampling",
    ):
        super(Conv1dSubsampling, self).__init__()
        self.conv = torch.nn.Conv1d(idim, odim, kernel_size, stride)
        self.pad_fn = torch.nn.ConstantPad1d(pad, 0.0)
        self.stride = stride
        self.odim = odim
        self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch
        self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf

    def output_size(self) -> int:
        return self.odim

    def forward(self, x, x_len):
        """Subsample x."""
        x = x.transpose(1, 2)  # (b, d ,t)
        x = self.pad_fn(x)
        # x = F.relu(self.conv(x))
        x = F.leaky_relu(self.conv(x), negative_slope=0.0)
        x = x.transpose(1, 2)  # (b, t ,d)

        if x_len is None:

            return x, None
        x_len = (x_len - 1) // self.stride + 1
        return x, x_len

    def gen_tf2torch_map_dict(self):
        tensor_name_prefix_torch = self.tf2torch_tensor_name_prefix_torch
        tensor_name_prefix_tf = self.tf2torch_tensor_name_prefix_tf
        map_dict_local = {
            ## predictor
            "{}.conv.weight".format(tensor_name_prefix_torch): {
                "name": "{}/conv1d/kernel".format(tensor_name_prefix_tf),
                "squeeze": None,
                "transpose": (2, 1, 0),
            },  # (256,256,3),(3,256,256)
            "{}.conv.bias".format(tensor_name_prefix_torch): {
                "name": "{}/conv1d/bias".format(tensor_name_prefix_tf),
                "squeeze": None,
                "transpose": None,
            },  # (256,),(256,)
        }
        return map_dict_local

    def convert_tf2torch(
        self,
        var_dict_tf,
        var_dict_torch,
    ):

        map_dict = self.gen_tf2torch_map_dict()

        var_dict_torch_update = dict()
        for name in sorted(var_dict_torch.keys(), reverse=False):
            names = name.split(".")
            if names[0] == self.tf2torch_tensor_name_prefix_torch:
                name_tf = map_dict[name]["name"]
                data_tf = var_dict_tf[name_tf]
                if map_dict[name]["squeeze"] is not None:
                    data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"])
                if map_dict[name]["transpose"] is not None:
                    data_tf = np.transpose(data_tf, map_dict[name]["transpose"])
                data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")

                var_dict_torch_update[name] = data_tf

                logging.info(
                    "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(
                        name, data_tf.size(), name_tf, var_dict_tf[name_tf].shape
                    )
                )
        return var_dict_torch_update


class StreamingConvInput(torch.nn.Module):
    """Streaming ConvInput module definition.
    Args:
        input_size: Input size.
        conv_size: Convolution size.
        subsampling_factor: Subsampling factor.
        vgg_like: Whether to use a VGG-like network.
        output_size: Block output dimension.
    """

    def __init__(
        self,
        input_size: int,
        conv_size: Union[int, Tuple],
        subsampling_factor: int = 4,
        vgg_like: bool = True,
        conv_kernel_size: int = 3,
        output_size: Optional[int] = None,
    ) -> None:
        """Construct a ConvInput object."""
        super().__init__()
        if vgg_like:
            if subsampling_factor == 1:
                conv_size1, conv_size2 = conv_size

                self.conv = torch.nn.Sequential(
                    torch.nn.Conv2d(
                        1,
                        conv_size1,
                        conv_kernel_size,
                        stride=1,
                        padding=(conv_kernel_size - 1) // 2,
                    ),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(
                        conv_size1,
                        conv_size1,
                        conv_kernel_size,
                        stride=1,
                        padding=(conv_kernel_size - 1) // 2,
                    ),
                    torch.nn.ReLU(),
                    torch.nn.MaxPool2d((1, 2)),
                    torch.nn.Conv2d(
                        conv_size1,
                        conv_size2,
                        conv_kernel_size,
                        stride=1,
                        padding=(conv_kernel_size - 1) // 2,
                    ),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(
                        conv_size2,
                        conv_size2,
                        conv_kernel_size,
                        stride=1,
                        padding=(conv_kernel_size - 1) // 2,
                    ),
                    torch.nn.ReLU(),
                    torch.nn.MaxPool2d((1, 2)),
                )

                output_proj = conv_size2 * ((input_size // 2) // 2)

                self.subsampling_factor = 1

                self.stride_1 = 1

                self.create_new_mask = self.create_new_vgg_mask

            else:
                conv_size1, conv_size2 = conv_size

                kernel_1 = int(subsampling_factor / 2)

                self.conv = torch.nn.Sequential(
                    torch.nn.Conv2d(
                        1,
                        conv_size1,
                        conv_kernel_size,
                        stride=1,
                        padding=(conv_kernel_size - 1) // 2,
                    ),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(
                        conv_size1,
                        conv_size1,
                        conv_kernel_size,
                        stride=1,
                        padding=(conv_kernel_size - 1) // 2,
                    ),
                    torch.nn.ReLU(),
                    torch.nn.MaxPool2d((kernel_1, 2)),
                    torch.nn.Conv2d(
                        conv_size1,
                        conv_size2,
                        conv_kernel_size,
                        stride=1,
                        padding=(conv_kernel_size - 1) // 2,
                    ),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(
                        conv_size2,
                        conv_size2,
                        conv_kernel_size,
                        stride=1,
                        padding=(conv_kernel_size - 1) // 2,
                    ),
                    torch.nn.ReLU(),
                    torch.nn.MaxPool2d((2, 2)),
                )

                output_proj = conv_size2 * ((input_size // 2) // 2)

                self.subsampling_factor = subsampling_factor

                self.create_new_mask = self.create_new_vgg_mask

                self.stride_1 = kernel_1

        else:
            if subsampling_factor == 1:
                self.conv = torch.nn.Sequential(
                    torch.nn.Conv2d(1, conv_size, 3, [1, 2], [1, 0]),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(
                        conv_size, conv_size, conv_kernel_size, [1, 2], [1, 0]
                    ),
                    torch.nn.ReLU(),
                )

                output_proj = conv_size * (((input_size - 1) // 2 - 1) // 2)

                self.subsampling_factor = subsampling_factor
                self.kernel_2 = conv_kernel_size
                self.stride_2 = 1

                self.create_new_mask = self.create_new_conv2d_mask

            else:
                kernel_2, stride_2, conv_2_output_size = sub_factor_to_params(
                    subsampling_factor,
                    input_size,
                )

                self.conv = torch.nn.Sequential(
                    torch.nn.Conv2d(1, conv_size, 3, 2, [1, 0]),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(
                        conv_size,
                        conv_size,
                        kernel_2,
                        stride_2,
                        [(kernel_2 - 1) // 2, 0],
                    ),
                    torch.nn.ReLU(),
                )

                output_proj = conv_size * conv_2_output_size

                self.subsampling_factor = subsampling_factor
                self.kernel_2 = kernel_2
                self.stride_2 = stride_2

                self.create_new_mask = self.create_new_conv2d_mask

        self.vgg_like = vgg_like
        self.min_frame_length = 7

        if output_size is not None:
            self.output = torch.nn.Linear(output_proj, output_size)
            self.output_size = output_size
        else:
            self.output = None
            self.output_size = output_proj

    def forward(
        self,
        x: torch.Tensor,
        mask: Optional[torch.Tensor],
        chunk_size: Optional[torch.Tensor],
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Encode input sequences.
        Args:
            x: ConvInput input sequences. (B, T, D_feats)
            mask: Mask of input sequences. (B, 1, T)
        Returns:
            x: ConvInput output sequences. (B, sub(T), D_out)
            mask: Mask of output sequences. (B, 1, sub(T))
        """
        if mask is not None:
            mask = self.create_new_mask(mask)
            olens = max(mask.eq(0).sum(1))

        b, t, f = x.size()
        x = x.unsqueeze(1)  # (b. 1. t. f)

        if chunk_size is not None:
            max_input_length = int(
                chunk_size
                * self.subsampling_factor
                * (math.ceil(float(t) / (chunk_size * self.subsampling_factor)))
            )
            x = map(lambda inputs: pad_to_len(inputs, max_input_length, 1), x)
            x = list(x)
            x = torch.stack(x, dim=0)
            N_chunks = max_input_length // (chunk_size * self.subsampling_factor)
            x = x.view(b * N_chunks, 1, chunk_size * self.subsampling_factor, f)

        x = self.conv(x)

        _, c, _, f = x.size()
        if chunk_size is not None:
            x = x.transpose(1, 2).contiguous().view(b, -1, c * f)[:, :olens, :]
        else:
            x = x.transpose(1, 2).contiguous().view(b, -1, c * f)

        if self.output is not None:
            x = self.output(x)

        return x, mask[:, :olens][:, : x.size(1)]

    def create_new_vgg_mask(self, mask: torch.Tensor) -> torch.Tensor:
        """Create a new mask for VGG output sequences.
        Args:
            mask: Mask of input sequences. (B, T)
        Returns:
            mask: Mask of output sequences. (B, sub(T))
        """
        if self.subsampling_factor > 1:
            vgg1_t_len = mask.size(1) - (mask.size(1) % (self.subsampling_factor // 2))
            mask = mask[:, :vgg1_t_len][:, :: self.subsampling_factor // 2]

            vgg2_t_len = mask.size(1) - (mask.size(1) % 2)
            mask = mask[:, :vgg2_t_len][:, ::2]
        else:
            mask = mask

        return mask

    def create_new_conv2d_mask(self, mask: torch.Tensor) -> torch.Tensor:
        """Create new conformer mask for Conv2d output sequences.
        Args:
            mask: Mask of input sequences. (B, T)
        Returns:
            mask: Mask of output sequences. (B, sub(T))
        """
        if self.subsampling_factor > 1:
            return mask[:, ::2][:, :: self.stride_2]
        else:
            return mask

    def get_size_before_subsampling(self, size: int) -> int:
        """Return the original size before subsampling for a given size.
        Args:
            size: Number of frames after subsampling.
        Returns:
            : Number of frames before subsampling.
        """
        return size * self.subsampling_factor