File size: 19,697 Bytes
85ba398
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from typing import Dict, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F
from fairseq import utils
from fairseq.incremental_decoding_utils import (
    FairseqIncrementalState,
    with_incremental_state,
)
from fairseq.modules.fairseq_dropout import FairseqDropout
from torch import Tensor

from .unfold import unfold1d


def DynamicConv(
    input_size,
    kernel_size=1,
    padding_l=None,
    num_heads=1,
    weight_dropout=0.0,
    weight_softmax=False,
    renorm_padding=False,
    bias=False,
    conv_bias=False,
    query_size=None,
    in_proj=False,
):
    if torch.cuda.is_available():
        try:
            from fairseq.modules.dynamicconv_layer import DynamicconvLayer

            return DynamicconvLayer(
                input_size,
                kernel_size=kernel_size,
                padding_l=padding_l,
                num_heads=num_heads,
                weight_dropout=weight_dropout,
                weight_softmax=weight_softmax,
                renorm_padding=renorm_padding,
                bias=bias,
                conv_bias=conv_bias,
                query_size=query_size,
            )
        except ImportError as e:
            print(e)
    return DynamicConv1dTBC(
        input_size,
        kernel_size=kernel_size,
        padding_l=padding_l,
        num_heads=num_heads,
        weight_dropout=weight_dropout,
        weight_softmax=weight_softmax,
        renorm_padding=renorm_padding,
        bias=bias,
        conv_bias=conv_bias,
        query_size=query_size,
    )


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, 0.0)
    return m


@with_incremental_state
class DynamicConv1dTBC(nn.Module):
    """Dynamic lightweight convolution taking T x B x C inputs
    Args:
        input_size: # of channels of the input
        kernel_size: convolution channels
        padding_l: padding to the left when using "same" padding
        num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size)
        weight_dropout: the drop rate of the DropConnect to drop the weight
        weight_softmax: normalize the weight with softmax before the convolution
        renorm_padding: re-normalize the filters to ignore the padded part (only the non-padding parts sum up to 1)
        bias: use bias
        conv_bias: bias of the convolution
        query_size: specified when feeding a different input as the query
        in_proj: project the input and generate the filter together

    Shape:
        Input: TxBxC, i.e. (timesteps, batch_size, input_size)
        Output: TxBxC, i.e. (timesteps, batch_size, input_size)

    Attributes:
        weight: the learnable weights of the module of shape
            `(num_heads, 1, kernel_size)`
        bias:   the learnable bias of the module of shape `(input_size)`
    """

    def __init__(
        self,
        input_size,
        kernel_size=1,
        padding_l=None,
        num_heads=1,
        weight_dropout=0.0,
        weight_softmax=False,
        renorm_padding=False,
        bias=False,
        conv_bias=False,
        query_size=None,
        in_proj=False,
    ):
        super().__init__()
        self.input_size = input_size
        self.query_size = input_size if query_size is None else query_size
        self.kernel_size = kernel_size
        self.padding_l = padding_l
        self.num_heads = num_heads
        self.weight_dropout_module = FairseqDropout(
            weight_dropout, module_name=self.__class__.__name__
        )
        self.weight_softmax = weight_softmax
        self.renorm_padding = renorm_padding

        if in_proj:
            self.weight_linear = Linear(
                self.input_size, self.input_size + num_heads * kernel_size * 1
            )
        else:
            self.weight_linear = Linear(
                self.query_size, num_heads * kernel_size * 1, bias=bias
            )
        if conv_bias:
            self.conv_bias = nn.Parameter(torch.Tensor(input_size))
        else:
            self.conv_bias = None
        self.reset_parameters()

    @property
    def in_proj(self):
        return (
            self.weight_linear.out_features
            == self.input_size + self.num_heads * self.kernel_size
        )

    def reset_parameters(self):
        self.weight_linear.reset_parameters()
        if self.conv_bias is not None:
            nn.init.constant_(self.conv_bias, 0.0)

    def forward(self, x, incremental_state=None, query=None, unfold=None):
        """Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C
        args:
            x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size)
            incremental_state: A dict to keep the state
            unfold: unfold the input or not. If not, we use the matrix trick instead
            query: use the specified query to predict the conv filters
        """
        unfold = (
            x.size(0) > 512 if unfold is None else unfold
        )  # use unfold mode as default for long sequence to save memory
        unfold = unfold or (incremental_state is not None)
        assert query is None or not self.in_proj

        if query is None:
            query = x
        if unfold:
            output = self._forward_unfolded(x, incremental_state, query)
        else:
            output = self._forward_expanded(x, incremental_state, query)

        if self.conv_bias is not None:
            output = output + self.conv_bias.view(1, 1, -1)
        return output

    def _forward_unfolded(self, x, incremental_state, query):
        """The conventional implementation of convolutions.
        Unfolding the input by having a window shifting to the right."""
        T, B, C = x.size()
        K, H = self.kernel_size, self.num_heads
        R = C // H
        assert R * H == C == self.input_size

        if self.in_proj:
            proj = self.weight_linear(x)
            x = proj.narrow(2, 0, self.input_size).contiguous()
            weight = (
                proj.narrow(2, self.input_size, H * K).contiguous().view(T * B * H, -1)
            )
        else:
            weight = self.weight_linear(query).view(T * B * H, -1)

        # renorm_padding is only implemented in _forward_expanded
        assert not self.renorm_padding or incremental_state is not None

        if incremental_state is not None:
            input_buffer = self._get_input_buffer(incremental_state)
            if input_buffer is None:
                input_buffer = x.new()
            x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3)
            if self.kernel_size > 1:
                self._set_input_buffer(
                    incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :]
                )
            x_unfold = x_unfold.view(T * B * H, R, -1)
        else:
            padding_l = self.padding_l
            if K > T and padding_l == K - 1:
                weight = weight.narrow(1, K - T, T)
                K, padding_l = T, T - 1
            # unfold the input: T x B x C --> T' x B x C x K
            x_unfold = unfold1d(x, K, padding_l, 0)
            x_unfold = x_unfold.view(T * B * H, R, K)

        if self.weight_softmax and not self.renorm_padding:
            weight = F.softmax(weight, dim=1)
        weight = weight.narrow(1, 0, K)

        if incremental_state is not None:
            weight = weight[:, -x_unfold.size(2) :]
            K = weight.size(1)

        if self.weight_softmax and self.renorm_padding:
            weight = F.softmax(weight, dim=1)

        weight = self.weight_dropout_module(weight, inplace=False)

        output = torch.bmm(x_unfold, weight.unsqueeze(2))  # T*B*H x R x 1
        output = output.view(T, B, C)
        return output

    def _forward_expanded(self, x, incremental_stat, query):
        """Turn the convolution filters into band matrices and do matrix multiplication.
        This is faster when the sequence is short, but less memory efficient.
        This is not used in the decoder during inference.
        """
        T, B, C = x.size()
        K, H = self.kernel_size, self.num_heads
        R = C // H
        assert R * H == C == self.input_size
        if self.in_proj:
            proj = self.weight_linear(x)
            x = proj.narrow(2, 0, self.input_size).contiguous()
            weight = (
                proj.narrow(2, self.input_size, H * K).contiguous().view(T * B * H, -1)
            )
        else:
            weight = self.weight_linear(query).view(T * B * H, -1)

        if not self.renorm_padding:
            if self.weight_softmax:
                weight = F.softmax(weight, dim=1)
            weight = self.weight_dropout_module(weight, inplace=False)
        weight = weight.narrow(1, 0, K).contiguous()
        weight = weight.view(T, B * H, K).transpose(0, 1)

        x = x.view(T, B * H, R).transpose(0, 1)
        if self.weight_softmax and self.renorm_padding:
            # turn the convolution filters into band matrices
            weight_expanded = weight.new(B * H, T, T + K - 1).fill_(float("-inf"))
            weight_expanded.as_strided(
                (B * H, T, K), (T * (T + K - 1), T + K, 1)
            ).copy_(weight)
            weight_expanded = weight_expanded.narrow(2, self.padding_l, T)
            # normalize the weight over valid positions like self-attention
            weight_expanded = F.softmax(weight_expanded, dim=2)
            weight_expanded = self.weight_dropout_module(weight_expanded, inplace=False)
        else:
            P = self.padding_l
            # For efficiency, we cut the kernel size and reduce the padding when the kernel is larger than the length
            if K > T and P == K - 1:
                weight = weight.narrow(2, K - T, T)
                K, P = T, T - 1
            # turn the convolution filters into band matrices
            weight_expanded = weight.new_zeros(B * H, T, T + K - 1, requires_grad=False)
            weight_expanded.as_strided(
                (B * H, T, K), (T * (T + K - 1), T + K, 1)
            ).copy_(weight)
            weight_expanded = weight_expanded.narrow(2, P, T)  # B*H x T x T
        output = torch.bmm(weight_expanded, x)
        output = output.transpose(0, 1).contiguous().view(T, B, C)
        return output

    def reorder_incremental_state(self, incremental_state, new_order):
        input_buffer = self._get_input_buffer(incremental_state)
        if input_buffer is not None:
            input_buffer = input_buffer.index_select(1, new_order)
            self._set_input_buffer(incremental_state, input_buffer)

    def _get_input_buffer(self, incremental_state):
        return utils.get_incremental_state(self, incremental_state, "input_buffer")

    def _set_input_buffer(self, incremental_state, new_buffer):
        return utils.set_incremental_state(
            self, incremental_state, "input_buffer", new_buffer
        )

    def extra_repr(self):
        s = "{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, conv_bias={}, renorm_padding={}, in_proj={}".format(
            self.input_size,
            self.kernel_size,
            self.padding_l,
            self.num_heads,
            self.weight_softmax,
            self.conv_bias is not None,
            self.renorm_padding,
            self.in_proj,
        )

        if self.query_size != self.input_size:
            s += ", query_size={}".format(self.query_size)
        if self.weight_dropout_module.p > 0.0:
            s += ", weight_dropout={}".format(self.weight_dropout_module.p)
        return s


class DynamicConv_scripatable(nn.Module, FairseqIncrementalState):
    """Dynamic lightweight convolution taking T x B x C inputs
    Args:
        input_size: # of channels of the input
        kernel_size: convolution channels
        padding_l: padding to the left when using "same" padding
        num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size)
        weight_dropout: the drop rate of the DropConnect to drop the weight
        weight_softmax: normalize the weight with softmax before the convolution
        renorm_padding: re-normalize the filters to ignore the padded part (only the non-padding parts sum up to 1)
        bias: use bias
        conv_bias: bias of the convolution
        query_size: specified when feeding a different input as the query
        in_proj: project the input and generate the filter together

    Shape:
        Input: TxBxC, i.e. (timesteps, batch_size, input_size)
        Output: TxBxC, i.e. (timesteps, batch_size, input_size)

    Attributes:
        weight: the learnable weights of the module of shape
            `(num_heads, 1, kernel_size)`
        bias:   the learnable bias of the module of shape `(input_size)`
    """

    def __init__(
        self,
        input_size,
        kernel_size=1,
        padding_l=None,
        num_heads=1,
        weight_dropout=0.0,
        weight_softmax=False,
        renorm_padding=False,
        bias=False,
        conv_bias=False,
        query_size=None,
        in_proj=False,
    ):
        super().__init__()
        self.input_size = input_size
        self.query_size = input_size if query_size is None else query_size
        self.kernel_size = kernel_size
        self.padding_l = padding_l
        self.num_heads = num_heads
        self.weight_dropout_module = FairseqDropout(
            weight_dropout, module_name=self.__class__.__name__
        )
        self.weight_softmax = weight_softmax
        self.renorm_padding = renorm_padding

        if in_proj:
            self.weight_linear = Linear(
                self.input_size, self.input_size + num_heads * kernel_size * 1
            )
        else:
            self.weight_linear = Linear(
                self.query_size, num_heads * kernel_size * 1, bias=bias
            )
        self.in_proj = (
            self.weight_linear.out_features
            == self.input_size + self.num_heads * self.kernel_size
        )
        self.has_conv_bias = conv_bias
        self.conv_bias = nn.Parameter(torch.Tensor(input_size).view(1, 1, -1))
        self.init_incremental_state()

        self.reset_parameters()

    def reset_parameters(self):
        self.weight_linear.reset_parameters()
        if self.has_conv_bias:
            nn.init.constant_(self.conv_bias, 0.0)

    def forward(
        self,
        x,
        incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
        query: Optional[Tensor] = None,
    ):
        """Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C
        args:
            x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size)
            incremental_state: A dict to keep the state
            unfold: unfold the input or not. If not, we use the matrix trick instead
            query: use the specified query to predict the conv filters
        """
        assert query is None or not self.in_proj

        if query is None:
            query = x

        output = self._forward_unfolded(x, incremental_state, query)

        if self.has_conv_bias:
            output = output + self.conv_bias
        return output

    def _forward_unfolded(
        self,
        x,
        incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]],
        query,
    ):
        """The conventional implementation of convolutions.
        Unfolding the input by having a window shifting to the right."""
        T, B, C = x.size()
        K, H = self.kernel_size, self.num_heads
        R = C // H
        assert R * H == C == self.input_size

        TxBxH = T * B * H

        if self.in_proj:
            proj = self.weight_linear(x)
            x = proj.narrow(2, 0, self.input_size).contiguous()
            weight = proj.narrow(2, self.input_size, H * K).contiguous().view(TxBxH, -1)
        else:
            weight = self.weight_linear(query).view(TxBxH, -1)

        # renorm_padding is only implemented in _forward_expanded
        assert not self.renorm_padding or incremental_state is not None

        if incremental_state is not None:
            input_buffer = self._get_input_buffer(incremental_state)
            if input_buffer is not None:
                x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3)
            else:
                x_unfold = x.unsqueeze(3).clone()
            if self.kernel_size > 1:
                self._set_input_buffer(
                    incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :]
                )
            x_unfold = x_unfold.view(TxBxH, R, -1)
        else:
            padding_l = self.padding_l
            if K > T and padding_l == K - 1:
                weight = weight.narrow(1, K - T, T)
                K, padding_l = T, T - 1
            # unfold the input: T x B x C --> T' x B x C x K
            x_unfold = unfold1d(x, K, padding_l, 0.0)
            x_unfold = x_unfold.view(TxBxH, R, K)

        if self.weight_softmax and not self.renorm_padding:
            weight = F.softmax(weight, dim=1)
        weight = weight.narrow(1, 0, K)

        if incremental_state is not None:
            weight = weight[:, -(x_unfold.size(2)) :]
            K = weight.size(1)

        if self.weight_softmax and self.renorm_padding:
            weight = F.softmax(weight, dim=1)

        weight = self.weight_dropout_module(weight, inplace=False)

        output = torch.bmm(x_unfold, weight.unsqueeze(2))  # T x B x H x R x 1
        output = output.view(T, B, C)
        return output

    def reorder_incremental_state(
        self,
        incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]],
        new_order: Tensor,
    ):
        input_buffer = self._get_input_buffer(incremental_state)
        if input_buffer is not None:
            input_buffer = input_buffer.index_select(1, new_order)
            self._set_input_buffer(incremental_state, input_buffer)

    def _get_input_buffer(
        self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]
    ):
        result = self.get_incremental_state(incremental_state, "input_buffer")
        if result is not None and "input_buffer" in result:
            return result["input_buffer"]
        else:
            return None

    def _set_input_buffer(
        self,
        incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]],
        new_buffer: Optional[Tensor],
    ):
        result = self.set_incremental_state(
            incremental_state, "input_buffer", {"input_buffer": new_buffer}
        )
        if result is not None:
            incremental_state = result
        return incremental_state

    def extra_repr(self):
        s = "{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, conv_bias={}, renorm_padding={}, in_proj={}".format(  # noqa
            self.input_size,
            self.kernel_size,
            self.padding_l,
            self.num_heads,
            self.weight_softmax,
            self.conv_bias is not None,
            self.renorm_padding,
            self.in_proj,
        )

        if self.query_size != self.input_size:
            s += ", query_size={}".format(self.query_size)
        if self.weight_dropout_module.p > 0.0:
            s += ", weight_dropout={}".format(self.weight_dropout_module.p)
        return s