forked from dmlc/gluon-nlp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
transformer.py
1419 lines (1335 loc) · 56.4 KB
/
transformer.py
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
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from abc import ABC
import numpy as np
import mxnet as mx
from mxnet import use_np
from mxnet.gluon import nn, HybridBlock
from typing import Optional, Tuple, List
from ..utils.registry import Registry
from ..attention_cell import MultiHeadAttentionCell, gen_self_attn_mask, gen_mem_attn_mask
from ..layers import PositionalEmbedding, PositionwiseFFN, InitializerType
from ..utils.config import CfgNode as CN
from ..sequence_sampler import BaseStepDecoder
__all__ = ['TransformerEncoderLayer', 'TransformerDecoderLayer',
'TransformerEncoder', 'TransformerDecoder',
'TransformerModel', 'TransformerNMTInference']
transformer_cfg_reg = Registry('transformer_cfg')
@transformer_cfg_reg.register()
def transformer_base():
"""Configuration of Transformer WMT EN-DE Base"""
cfg = CN()
cfg.MODEL = CN()
cfg.MODEL.src_vocab_size = -1
cfg.MODEL.tgt_vocab_size = -1
cfg.MODEL.max_src_length = -1
cfg.MODEL.max_tgt_length = -1
cfg.MODEL.scale_embed = True
cfg.MODEL.pos_embed_type = "sinusoidal"
cfg.MODEL.shared_embed = True
cfg.MODEL.tie_weights = True
cfg.MODEL.attention_dropout = 0.0
cfg.MODEL.activation_dropout = 0.0
cfg.MODEL.dropout = 0.1
cfg.MODEL.layout = 'NT'
cfg.MODEL.dtype = 'float32'
# Parameters for the encoder
cfg.MODEL.ENCODER = CN()
cfg.MODEL.ENCODER.num_layers = 6
cfg.MODEL.ENCODER.units = 512
cfg.MODEL.ENCODER.num_heads = 8
cfg.MODEL.ENCODER.hidden_size = 2048
cfg.MODEL.ENCODER.recurrent = False
cfg.MODEL.ENCODER.activation = 'relu'
cfg.MODEL.ENCODER.pre_norm = False
cfg.MODEL.ENCODER.use_qkv_bias = True
# Parameters for the decoder
cfg.MODEL.DECODER = CN()
cfg.MODEL.DECODER.num_layers = 6
cfg.MODEL.DECODER.units = 512
cfg.MODEL.DECODER.num_heads = 8
cfg.MODEL.DECODER.hidden_size = 2048
cfg.MODEL.DECODER.recurrent = False
cfg.MODEL.DECODER.activation = 'relu'
cfg.MODEL.DECODER.pre_norm = False
cfg.MODEL.DECODER.use_qkv_bias = False
# Parameters for the initializer
cfg.INITIALIZER = CN()
cfg.INITIALIZER.embed = ['xavier', 'gaussian', 'in', 1.0]
cfg.INITIALIZER.weight = ['xavier', 'uniform', 'avg', 3.0]
cfg.INITIALIZER.bias = ['zeros']
cfg.VERSION = 1
cfg.freeze()
return cfg
@transformer_cfg_reg.register()
def transformer_base_prenorm():
cfg = transformer_base()
cfg.defrost()
cfg.MODEL.ENCODER.pre_norm = True
cfg.MODEL.DECODER.pre_norm = True
cfg.freeze()
return cfg
@transformer_cfg_reg.register()
def transformer_iwslt_de_en():
cfg = TransformerModel.get_cfg()
cfg.defrost()
cfg.MODEL.ENCODER.units = 512
cfg.MODEL.ENCODER.hidden_size = 1024
cfg.MODEL.ENCODER.num_heads = 4
cfg.MODEL.ENCODER.num_layers = 6
cfg.MODEL.DECODER.units = 512
cfg.MODEL.DECODER.hidden_size = 1024
cfg.MODEL.DECODER.num_heads = 4
cfg.MODEL.DECODER.num_layers = 6
cfg.freeze()
return cfg
@transformer_cfg_reg.register()
def transformer_wmt_en_de_big():
"""Same wmt_en_de_big architecture as in FairSeq"""
cfg = TransformerModel.get_cfg()
cfg.defrost()
cfg.MODEL.attention_dropout = 0.1
cfg.MODEL.dropout = 0.3
cfg.MODEL.ENCODER.units = 1024
cfg.MODEL.ENCODER.hidden_size = 4096
cfg.MODEL.ENCODER.num_heads = 16
cfg.MODEL.ENCODER.num_layers = 6
cfg.MODEL.DECODER.units = 1024
cfg.MODEL.DECODER.hidden_size = 4096
cfg.MODEL.DECODER.num_heads = 16
cfg.MODEL.DECODER.num_layers = 6
cfg.freeze()
return cfg
@transformer_cfg_reg.register()
def transformer_wmt_en_de_big_t2t():
"""Parameter used in the T2T implementation"""
cfg = transformer_wmt_en_de_big()
cfg.defrost()
cfg.MODEL.attention_dropout = 0.1
cfg.MODEL.activation_dropout = 0.1
cfg.MODEL.ENCODER.pre_norm = True
cfg.MODEL.DECODER.pre_norm = True
cfg.freeze()
return cfg
@use_np
class TransformerEncoderLayer(HybridBlock):
"""Transformer Encoder Layer"""
def __init__(self,
units: int = 512,
hidden_size: int = 2048,
num_heads: int = 8,
attention_dropout_prob: float = 0.1,
hidden_dropout_prob: float = 0.1,
activation_dropout_prob: float = 0.0,
layer_norm_eps: float = 1e-12,
pre_norm: bool = False,
use_qkv_bias: bool = True,
weight_initializer: Optional[InitializerType] = None,
bias_initializer: Optional[InitializerType] = 'zeros',
activation: str = 'relu',
dtype='float32',
layout='NT'):
"""
Parameters
----------
units
hidden_size
num_heads
attention_dropout_prob
hidden_dropout_prob
activation_dropout_prob
layer_norm_eps
pre_norm
Whether to attach the normalization layer before attention layer
If pre_norm:
norm(data) -> attn -> res(+data) -> ffn
Else:
data -> attn -> norm(res(+data)) -> ffn
use_qkv_bias
Wether to use bias for self attention
weight_initializer
bias_initializer
activation
dtype
layout
"""
super().__init__()
self._units = units
self._hidden_size = hidden_size
self._num_heads = num_heads
self._attention_dropout_prob = attention_dropout_prob
self._hidden_dropout_prob = hidden_dropout_prob
self._activation_dropout_prob = activation_dropout_prob
self._pre_norm = pre_norm
self._dtype = dtype
self._layout = layout
assert layout in ['TN', 'NT'], 'Invalid layout received = {}. ' \
'Only "TN" and "NT" are accepted!'.format(layout)
assert self._units % self._num_heads == 0, 'units must be divisive by the number of heads'
self.dropout_layer = nn.Dropout(hidden_dropout_prob)
self.attn_qkv = nn.Dense(3 * units,
flatten=False,
use_bias=use_qkv_bias,
in_units=units,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
dtype=self._dtype)
self.attention_proj = nn.Dense(units=units,
flatten=False,
in_units=units,
use_bias=True,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
dtype=self._dtype)
attention_layout = 'NTK' if self._layout == 'NT' else 'TNK'
self.attention_cell = \
MultiHeadAttentionCell(
query_units=self._units,
num_heads=self._num_heads,
attention_dropout=self._attention_dropout_prob,
scaled=True,
dtype=self._dtype,
layout=attention_layout
)
self.layer_norm = nn.LayerNorm(epsilon=layer_norm_eps,
in_channels=units)
self.ffn = PositionwiseFFN(units=units,
hidden_size=hidden_size,
dropout=hidden_dropout_prob,
activation_dropout=activation_dropout_prob,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
layer_norm_eps=layer_norm_eps,
activation=activation,
pre_norm=pre_norm,
dtype=self._dtype)
@property
def layout(self) -> str:
return self._layout
def hybrid_forward(self, F, data, attn_mask):
"""
Parameters
----------
F
data :
If layout == 'NT'
Shape (batch_size, seq_length, C_in)
Else
Shape (seq_length, batch_size, C_in)
attn_mask :
Shape (batch_size, seq_length, seq_length)
Returns
-------
out :
If layout == 'NT'
Shape (batch_size, seq_length, C_out)
Else
Shape (seq_length, batch_size, C_out)
attn_weight :
Shape (batch_size, seq_length, seq_length)
"""
if self._pre_norm:
data = self.layer_norm(data)
query, key, value = F.np.split(self.attn_qkv(data), 3, axis=-1)
query = F.npx.reshape(query, (-2, -2, self._num_heads, -1))
key = F.npx.reshape(key, (-2, -2, self._num_heads, -1))
value = F.npx.reshape(value, (-2, -2, self._num_heads, -1))
out, [_, attn_weight] = self.attention_cell(query, key, value, attn_mask)
out = self.attention_proj(out)
out = self.dropout_layer(out)
out = out + data
if not self._pre_norm:
out = self.layer_norm(out)
out = self.ffn(out)
return out, attn_weight
@use_np
class TransformerEncoder(HybridBlock):
def __init__(self, num_layers=6, recurrent=False,
units=512, hidden_size=2048, num_heads=8,
activation_dropout=0.0, dropout=0.1, use_qkv_bias=True,
attention_dropout=0.1, layer_norm_eps=1E-5, data_norm=False,
pre_norm=False, weight_initializer=None, bias_initializer='zeros',
activation='relu', dtype='float32', layout='NT'):
"""
Parameters
----------
num_layers :
The number of layers
recurrent : bool
Whether the layers share weights or not
units
hidden_size
num_heads
dropout
layer_norm_eps
data_norm
Whether to apply LayerNorm to the data
pre_norm
Whether to apply LayerNorm before the attention layer.
weight_initializer
bias_initializer
activation
dtype
layout
"""
super().__init__()
self._dtype = dtype
self.num_layers = num_layers
self._recurrent = recurrent
self._data_norm = data_norm
self._pre_norm = pre_norm
self._layout = layout
assert layout in ['TN', 'NT'], 'Invalid layout received = {}. ' \
'Only "TN" and "NT" are accepted!'.format(layout)
self.dropout_layer = nn.Dropout(dropout)
if self._pre_norm:
self.ln_final = nn.LayerNorm(epsilon=layer_norm_eps,
in_channels=units)
if self._data_norm:
self.ln_data = nn.LayerNorm(epsilon=layer_norm_eps,
in_channels=units)
# Construct the intermediate layers
self.layers = nn.HybridSequential()
real_num_layers = 1 if recurrent else num_layers
for i in range(real_num_layers):
self.layers.add(TransformerEncoderLayer(
units=units,
hidden_size=hidden_size,
num_heads=num_heads,
hidden_dropout_prob=dropout,
attention_dropout_prob=attention_dropout,
activation_dropout_prob=activation_dropout,
use_qkv_bias=use_qkv_bias,
layer_norm_eps=layer_norm_eps,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
pre_norm=pre_norm,
activation=activation,
layout=self._layout,
dtype=dtype))
@property
def layout(self) -> str:
return self._layout
def hybrid_forward(self, F, data, valid_length):
"""
Parameters
----------
F
data :
- layout = 'NT'
Shape (batch_size, seq_length, C)
- layout = 'TN'
Shape (seq_length, batch_size, C)
valid_length :
Shape (batch_size,)
Returns
-------
out :
- layout = 'NT'
Shape (batch_size, seq_length, C_out)
- layout = 'TN'
Shape (seq_length, batch_size, C_out)
"""
# 1. Embed the data
attn_mask = gen_self_attn_mask(F, data, valid_length,
dtype=self._dtype,
layout=self.layout,
attn_type='full')
out = self.dropout_layer(data)
if self._data_norm:
out = self.ln_data(out)
for i in range(self.num_layers):
if self._recurrent:
layer = self.layers[0]
else:
layer = self.layers[i]
out, _ = layer(out, attn_mask)
if self._pre_norm:
out = self.ln_final(out)
return out
@use_np
class TransformerDecoderLayer(HybridBlock):
def __init__(self, units: int = 512,
mem_units: Optional[int] = None,
hidden_size: int = 2048,
num_heads: int = 8,
activation_dropout: float = 0.0,
dropout: float = 0.1,
attention_dropout: float = 0.1,
layer_norm_eps: float = 1E-5,
activation: str = 'relu',
pre_norm: bool = False,
use_qkv_bias: bool = True,
weight_initializer=None,
bias_initializer='zeros',
dtype='float32',
layout='NT'):
"""
Parameters
----------
units
mem_units
The number of units in the memory. By default, it is initialized to be the
same as the units.
hidden_size
num_heads
activation_dropout
dropout
attention_dropout
layer_norm_eps
activation
pre_norm
Whether to apply normalization before the attention layer
use_qkv_bias
Wether to use bias for both self attention and contextual attention
weight_initializer
bias_initializer
dtype
Data type
layout
Layout of the input
"""
super().__init__()
self._dtype = dtype
self._units = units
if mem_units is None:
mem_units = units
self._mem_units = mem_units
self._pre_norm = pre_norm
self._num_heads = num_heads
self._attention_dropout = attention_dropout
self._dtype = dtype
self._layout = layout
assert layout in ['TN', 'NT'], 'Invalid layout received = {}. ' \
'Only "TN" and "NT" are accepted!'.format(layout)
attention_layout = 'NTK' if layout == 'NT' else 'TNK'
self.dropout_layer = nn.Dropout(dropout)
if units % num_heads:
raise ValueError('In Transformer, units should be divided exactly by the number of '
'heads. Received units={}, num_heads={}'.format(units, num_heads))
self.attn_in_qkv = nn.Dense(3 * units, in_units=units,
use_bias=use_qkv_bias,
flatten=False,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
dtype=dtype)
self.self_attention = MultiHeadAttentionCell(query_units=units,
num_heads=num_heads,
attention_dropout=self._attention_dropout,
dtype=dtype,
layout=attention_layout)
self.proj_in = nn.Dense(units=units, in_units=units, flatten=False, use_bias=True,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
dtype=dtype)
self.attn_inter_q = nn.Dense(units,
in_units=units,
use_bias=use_qkv_bias,
flatten=False,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
dtype=dtype)
self.attn_inter_k = nn.Dense(units, in_units=mem_units,
use_bias=use_qkv_bias,
flatten=False,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
dtype=dtype)
self.attn_inter_v = nn.Dense(units, in_units=mem_units,
use_bias=use_qkv_bias,
flatten=False,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
dtype=dtype)
self.inter_attention = MultiHeadAttentionCell(query_units=units,
num_heads=num_heads,
attention_dropout=self._attention_dropout,
dtype=dtype,
layout=attention_layout)
self.proj_inter = nn.Dense(units=units, in_units=units,
flatten=False, use_bias=True,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
dtype=dtype)
# TODO(sxjscience) Add DType to LayerNorm
self.ln_in = nn.LayerNorm(epsilon=layer_norm_eps,
in_channels=units)
self.ln_inter = nn.LayerNorm(epsilon=layer_norm_eps,
in_channels=units)
self.ffn = PositionwiseFFN(units=units,
hidden_size=hidden_size,
dropout=dropout,
activation_dropout=activation_dropout,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
layer_norm_eps=layer_norm_eps,
activation=activation,
pre_norm=pre_norm,
dtype=dtype)
@property
def layout(self) -> str:
return self._layout
def hybrid_forward(self, F, data, mem, self_causal_mask, mem_attn_mask):
"""
Parameters
----------
F
data :
- layout = 'NT'
Shape (batch_size, seq_length, C_in)
- layout = 'TN'
Shape (seq_length, batch_size, C_in)
mem :
- layout = 'NT'
Shape (batch_size, mem_length, C_mem)
- layout = 'TN'
Shape (mem_length, batch_size, C_mem)
self_causal_mask :
Shape (batch_size, seq_length, seq_length)
Mask for the causal self-attention.
self_causal_mask[i, j, :] masks the elements that token `j` attends to.
To understand the self-causal attention mask, we can look at the following example:
['I', 'can', 'now', 'use', 'numpy', 'in', 'Gluon@@', 'NLP']
'I': 1, 0, 0, 0, 0, 0, 0, 0
'can': 1, 1, 0, 0, 0, 0, 0, 0
'now': 1, 1, 1, 0, 0, 0, 0, 0
'use': 1, 1, 1, 1, 0, 0, 0, 0
'numpy': 1, 1, 1, 1, 1, 0, 0, 0
'in': 1, 1, 1, 1, 1, 1, 0, 0
'Gluon@@': 1, 1, 1, 1, 1, 1, 1, 0
'NLP': 1, 1, 1, 1, 1, 1, 1, 1
mem_attn_mask :
Shape (batch_size, seq_length, mem_length)
Mask between the decoding input and the memory.
['numpy', 'in', 'Gluon@@', 'NLP']
'I': 1, 1, 1, 1
'can': 1, 1, 1, 1
'now': 1, 1, 1, 1
'use': 1, 1, 1, 1
Returns
-------
out :
- layout = 'NT'
Shape (batch_size, seq_length, C_out)
- layout = 'TN'
Shape (seq_length, batch_size, C_out)
"""
# 1. Get the causal self-attention value
if self._pre_norm:
data = self.ln_in(data)
self_query, self_key, self_value = F.np.split(self.attn_in_qkv(data), 3, axis=-1)
out, [_, self_attn_weight] = self.self_attention(
F.npx.reshape(self_query, (-2, -2, self._num_heads, -1)),
F.npx.reshape(self_key, (-2, -2, self._num_heads, -1)),
F.npx.reshape(self_value, (-2, -2, self._num_heads, -1)),
self_causal_mask)
out = self.proj_in(out)
out = self.dropout_layer(out)
out = out + data
if not self._pre_norm:
out = self.ln_in(out)
# 2. Attend to the contextual memory
data = out
if self._pre_norm:
data = self.ln_inter(data)
out, [_, context_attn_weight] = self.inter_attention(
F.npx.reshape(self.attn_inter_q(data), (-2, -2, self._num_heads, -1)),
F.npx.reshape(self.attn_inter_k(mem), (-2, -2, self._num_heads, -1)),
F.npx.reshape(self.attn_inter_v(mem), (-2, -2, self._num_heads, -1)),
mem_attn_mask)
out = self.proj_inter(out)
out = self.dropout_layer(out)
out = out + data
if not self._pre_norm:
out = self.ln_inter(out)
# 3. Encode the output via an FFN layer
out = self.ffn(out)
return out
@property
def state_batch_axis(self):
if self.layout == 'NT':
return 0, 0
else:
return 1, 1
def init_states(self, batch_size, ctx, dtype='float32'):
"""Initialize the states required for incremental decoding
Returns
-------
init_key
- layout = 'NT'
Shape (batch_size, 0, N, C_key)
- layout = 'TN'
Shape (0, batch_size, N, C_key)
init_value :
- layout = 'NT'
Shape (batch_size, 0, N, C_value)
- layout = 'TN'
Shape (0, batch_size, N, C_value)
"""
if self.layout == 'NT':
init_key = mx.np.zeros(shape=(batch_size, 0, self._num_heads,
self._units // self._num_heads), ctx=ctx, dtype=dtype)
init_value = mx.np.zeros(shape=(batch_size, 0, self._num_heads,
self._units // self._num_heads), ctx=ctx, dtype=dtype)
else:
init_key = mx.np.zeros(shape=(0, batch_size, self._num_heads,
self._units // self._num_heads), ctx=ctx, dtype=dtype)
init_value = mx.np.zeros(shape=(0, batch_size, self._num_heads,
self._units // self._num_heads), ctx=ctx, dtype=dtype)
return init_key, init_value
def incremental_decode(self, F, data, states, mem, mem_valid_length, mem_attn_mask=None):
"""Incrementally generate the output given the decoder input.
Parameters
----------
F
data
Shape (batch_size, C_in)
states
The previous states, contains
1. layout = 'NT':
- prev_multi_key
Shape (batch_size, prev_seq_length, num_heads, C_key)
- prev_multi_value
Shape (batch_size, prev_seq_length, num_heads, C_value)
2. layout = 'TN'
- prev_multi_key
Shape (prev_seq_length, batch_size, num_heads, C_key)
- prev_multi_value
Shape (prev_seq_length, batch_size, num_heads, C_value)
mem
The memory
1. layout = 'NT':
Shape (batch_size, mem_length, C_mem)
2. layout = 'TN'
Shape (mem_length, batch_size, C_mem)
mem_valid_length
Valid length of the memory
Shape (batch_size,)
mem_attn_mask
The attention mask between data and the memory
Has shape (batch_size, 1, mem_length)
Returns
-------
out
Shape (batch_size, C_out)
updated_states
- new_key
Shape (batch_size, prev_seq_length + 1, num_heads, C_key)
- new_value
Shape (batch_size, prev_seq_length + 1, num_heads, C_value)
"""
if self._pre_norm:
data = self.ln_in(data)
if self.layout == 'NT':
time_axis = 1
else:
time_axis = 0
data = F.np.expand_dims(data, axis=time_axis)
# Shape (B, prev_L, #Head, C_K), (B, prev_L, #Head, C_V)
# or (prev_L, B, #Head, C_K), (prev_L, B, #Head, C_V)
prev_key, prev_value = states
if mem_attn_mask is None:
mem_attn_mask = gen_mem_attn_mask(F, mem, mem_valid_length, data, None,
dtype=self._dtype, layout=self.layout)
# 1. Get the causal self-attention value, we need to attend to both the current data
# and the previous stored key/values
# Shape (B, 1, 3 * num_heads * C_key)
# or (1, B, 3 * num_heads * C_key)
step_qkv = self.attn_in_qkv(data)
step_query, step_key, step_value = F.np.split(step_qkv, 3, axis=-1)
step_query = F.npx.reshape(step_query, (-2, -2, self._num_heads, -1))
step_key = F.npx.reshape(step_key, (-2, -2, self._num_heads, -1))
step_value = F.npx.reshape(step_value, (-2, -2, self._num_heads, -1))
new_key = F.np.concatenate([prev_key, step_key], axis=time_axis)
new_value = F.np.concatenate([prev_value, step_value], axis=time_axis)
out, [_, attn_weight] = self.self_attention(step_query, new_key, new_value, None)
out = self.proj_in(out)
out = self.dropout_layer(out)
out = out + data
if not self._pre_norm:
out = self.ln_in(out)
# 2. Attend to the contextual memory
data = out
if self._pre_norm:
data = self.ln_inter(data)
out, _ = self.inter_attention(F.npx.reshape(self.attn_inter_q(data),
(-2, -2, self._num_heads, -1)),
F.npx.reshape(self.attn_inter_k(mem),
(-2, -2, self._num_heads, -1)),
F.npx.reshape(self.attn_inter_v(mem),
(-2, -2, self._num_heads, -1)),
mem_attn_mask)
out = self.proj_inter(out)
out = self.dropout_layer(out)
out = out + data
if not self._pre_norm:
out = self.ln_inter(out)
# 3. Encode the output via an FFN layer
out = self.ffn(out)
out = F.npx.reshape(out, (-5, -1))
return out, (new_key, new_value)
@use_np
class TransformerDecoder(HybridBlock):
def __init__(self, num_layers=6, recurrent=False,
units=512, mem_units=None, hidden_size=2048, use_qkv_bias=True,
num_heads=8, max_shift=None, activation_dropout=0.0,
dropout=0.1, attention_dropout=0.1, layer_norm_eps=1E-5, data_norm=False,
pre_norm=False, weight_initializer=None, bias_initializer=None,
activation='relu', dtype='float32',
layout='NT'):
super().__init__()
self._dtype = dtype
self._units = units
self._mem_units = mem_units
self.num_layers = num_layers
self.recurrent = recurrent
self.max_shift = max_shift
self._data_norm = data_norm
self._pre_norm = pre_norm
self._layout = layout
assert layout in ['TN', 'NT'], 'Invalid layout received = {}. ' \
'Only "TN" and "NT" are accepted!'.format(layout)
self.dropout_layer = nn.Dropout(dropout)
if self._data_norm:
self.ln_data = nn.LayerNorm(epsilon=layer_norm_eps,
in_channels=units)
if self._pre_norm:
self.ln_final = nn.LayerNorm(epsilon=layer_norm_eps,
in_channels=units)
# Construct the intermediate layers
self.layers = nn.HybridSequential()
real_num_layers = 1 if recurrent else num_layers
for i in range(real_num_layers):
self.layers.add(TransformerDecoderLayer(units=units,
mem_units=mem_units,
hidden_size=hidden_size,
num_heads=num_heads,
activation_dropout=activation_dropout,
use_qkv_bias=use_qkv_bias,
dropout=dropout,
attention_dropout=attention_dropout,
layer_norm_eps=layer_norm_eps,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
activation=activation,
pre_norm=pre_norm,
layout=layout,
dtype=dtype))
@property
def layout(self) -> str:
return self._layout
def hybrid_forward(self, F, data, valid_length, mem_data, mem_valid_length):
"""
Parameters
----------
F
data
- layout = 'NT'
Shape (batch_size, seq_length, C_in)
- layout = 'TN'
Shape (seq_length, batch_size, C_in)
valid_length
Shape (batch_size,)
mem_data
- layout = 'NT'
Shape (batch_size, mem_length, C_mem)
- layout = 'TN'
Shape (mem_length, batch_size, C_mem)
mem_valid_length
Shape (batch_size,)
Returns
-------
out
- layout = 'NT'
Shape (batch_size, seq_length, C_out)
- layout = 'TN'
Shape (seq_length, batch_size, C_out)
"""
# 1. Embed the data
out = self.dropout_layer(data)
if self._data_norm:
out = self.ln_data(out)
self_causal_mask = gen_self_attn_mask(F, data, valid_length,
dtype=self._dtype,
attn_type='causal',
layout=self._layout)
mem_attn_mask = gen_mem_attn_mask(F, mem_data, mem_valid_length, data, valid_length,
dtype=self._dtype,
layout=self._layout)
for i in range(self.num_layers):
if self.recurrent:
layer = self.layers[0]
else:
layer = self.layers[i]
out = layer(out, mem_data, self_causal_mask, mem_attn_mask)
if self._pre_norm:
out = self.ln_final(out)
return out
@property
def state_batch_axis(self):
ret = []
for i in range(self.num_layers):
if self.recurrent:
layer = self.layers[0]
else:
layer = self.layers[i]
ret.append(layer.state_batch_axis)
return ret
def init_states(self, batch_size, ctx, dtype='float32'):
"""Initialize the states required for incremental decoding
Returns
-------
states
A list of states, each includes:
- init_key :
layout = 'NT':
Shape (batch_size, 0, N, C_key)
- init_value :
layout = 'TN':
Shape (0, batch_size, N, C_value)
"""
states = []
for i in range(self.num_layers):
if self.recurrent:
layer = self.layers[0]
else:
layer = self.layers[i]
states.append(layer.init_states(batch_size=batch_size,
ctx=ctx,
dtype=dtype))
return states
def incremental_decode(self, F, data, states, mem, mem_valid_length):
"""Incrementally generate the output given the decoder input.
Parameters
----------
F
data
Shape (batch_size, C_in)
states
The previous states, contain a list of
1. layout = 'NT'
- prev_multi_key
Shape (batch_size, prev_seq_length, num_heads, C_key)
- prev_multi_value
Shape (batch_size, prev_seq_length, num_heads, C_value)
2. layout = 'TN'
- prev_multi_key
Shape (prev_seq_length, batch_size, num_heads, C_key)
- prev_multi_value
Shape (prev_seq_length, batch_size, num_heads, C_value)
mem
The memory
1. layout = 'NT'
Shape (batch_size, mem_length, C_mem)
2. layout = 'TN'
Shape (mem_length, batch_size, C_mem)
mem_valid_length
Valid length of the memory
Shape (batch_size,)
Returns
-------
out
Shape (batch_size, C_out)
new_states
The updated states, contain a list of
1. layout = 'NT'
- new_key
Shape (batch_size, prev_seq_length + 1, num_heads, C_key)
2. layout = 'TN'
- new_value
Shape (prev_seq_length + 1, batch_size, num_heads, C_value)
"""
# 1. Embed the data
out = self.dropout_layer(data)
if self._data_norm:
out = self.ln_data(out)
time_axis = 0 if self.layout == 'TN' else 1
# Generate the mem_attn_mask
time_steps = F.npx.arange_like(mem, axis=time_axis) # (mem_length,)
mem_attn_mask = F.np.reshape(time_steps, (1, 1, -1))\
< F.np.reshape(mem_valid_length, (-1, 1, 1))
# TODO(sxjscience) Try with boolean masking
mem_attn_mask = mem_attn_mask.astype(self._dtype)
new_states = []
for i in range(self.num_layers):
if self.recurrent:
layer = self.layers[0]
else:
layer = self.layers[i]
out, new_state = layer.incremental_decode(F, out, states[i],
mem, mem_valid_length, mem_attn_mask)
new_states.append(new_state)
if self._pre_norm:
out = self.ln_final(out)
return out, new_states
@use_np
class TransformerModel(HybridBlock):
def __init__(self, src_vocab_size: int,
tgt_vocab_size: int,
max_src_length: Optional[int] = None,
max_tgt_length: Optional[int] = None,
scale_embed: bool = True,
pos_embed_type="sinusoidal",
shared_embed: bool = True,
tie_weights: bool = True,
activation_dropout: float = 0.0,
dropout: float = 0.1,
attention_dropout: float = 0.1,
layer_norm_eps: float = 1E-5,
data_norm: bool = False,
enc_units: int = 512,
enc_hidden_size: int = 2048,
enc_num_heads: int = 8,
enc_num_layers: int = 6,
enc_recurrent: bool = False,
enc_activation='relu',
enc_pre_norm: bool = False,
enc_use_qkv_bias: bool = True,
dec_units: int = 512,
dec_hidden_size: int = 2048,
dec_num_heads: int = 8,
dec_num_layers: int = 6,
dec_recurrent: bool = False,
dec_activation='relu',
dec_pre_norm: bool = False,
dec_use_qkv_bias: bool = True,
embed_initializer=mx.init.Xavier('gaussian', 'in', 1),
weight_initializer=mx.init.Xavier('uniform', 'avg', 3),
bias_initializer='zeros',
dtype='float32',
layout='NT'):
"""
Parameters
----------
src_vocab_size
The vocabulary size of the source language
tgt_vocab_size
The vocabulary size of the target language
max_src_length
The maximal length of the source sequence.
If it's negative, we will use treat it as not set.
max_tgt_length
The maximal length of the target sequence.
If it's negative, we will use treat it as not set.
scale_embed
Whether to multiply the src and dst embeddings by sqrt(units)
pos_embed_type
Type of the positional embedding
shared_embed
Whether to share the embedding of the src and tgt language
tie_weights
Whether to tie the weights of input + output.
activation_dropout
The ratio of the activation dropout in FFN
dropout
The default dropout ratio
attention_dropout
The ratio of the attention dropout
layer_norm_eps
The epsilon of the layer normalization
data_norm
Whether to add layer normalization layer after the input.
enc_units
Units of the encoder
enc_hidden_size
Hidden size of the encoder
enc_num_heads
Number of heads of the encoder
enc_num_layers
Number of layers of the encoder
enc_recurrent
Whether to use recurrent encoder (share weights)
enc_activation
Activation of the encoder layer
enc_pre_norm
Whether to add layer_norm before self-attention in the encoder