-
Notifications
You must be signed in to change notification settings - Fork 26.7k
/
modeling_flax_bert.py
1719 lines (1432 loc) · 62.1 KB
/
modeling_flax_bert.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
# coding=utf-8
# Copyright 2021 The Google Flax Team Authors and The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Callable, Optional, Tuple
import numpy as np
import flax
import flax.linen as nn
import jax
import jax.numpy as jnp
from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
from flax.linen import combine_masks, make_causal_mask
from flax.linen import partitioning as nn_partitioning
from flax.linen.attention import dot_product_attention_weights
from flax.traverse_util import flatten_dict, unflatten_dict
from jax import lax
from ...modeling_flax_outputs import (
FlaxBaseModelOutputWithPastAndCrossAttentions,
FlaxBaseModelOutputWithPooling,
FlaxBaseModelOutputWithPoolingAndCrossAttentions,
FlaxCausalLMOutputWithCrossAttentions,
FlaxMaskedLMOutput,
FlaxMultipleChoiceModelOutput,
FlaxNextSentencePredictorOutput,
FlaxQuestionAnsweringModelOutput,
FlaxSequenceClassifierOutput,
FlaxTokenClassifierOutput,
)
from ...modeling_flax_utils import (
ACT2FN,
FlaxPreTrainedModel,
append_call_sample_docstring,
append_replace_return_docstrings,
overwrite_call_docstring,
)
from ...utils import ModelOutput, add_start_docstrings, add_start_docstrings_to_model_forward, logging
from .configuration_bert import BertConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "bert-base-uncased"
_CONFIG_FOR_DOC = "BertConfig"
_TOKENIZER_FOR_DOC = "BertTokenizer"
remat = nn_partitioning.remat
@flax.struct.dataclass
class FlaxBertForPreTrainingOutput(ModelOutput):
"""
Output type of [`BertForPreTraining`].
Args:
prediction_logits (`jnp.ndarray` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
seq_relationship_logits (`jnp.ndarray` of shape `(batch_size, 2)`):
Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
before SoftMax).
hidden_states (`tuple(jnp.ndarray)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `jnp.ndarray` (one for the output of the embeddings + one for the output of each layer) of shape
`(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
attentions (`tuple(jnp.ndarray)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `jnp.ndarray` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
prediction_logits: jnp.ndarray = None
seq_relationship_logits: jnp.ndarray = None
hidden_states: Optional[Tuple[jnp.ndarray]] = None
attentions: Optional[Tuple[jnp.ndarray]] = None
BERT_START_DOCSTRING = r"""
This model inherits from [`FlaxPreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading, saving and converting weights from PyTorch models)
This model is also a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
subclass. Use it as a regular Flax linen Module and refer to the Flax documentation for all matter related to
general usage and behavior.
Finally, this model supports inherent JAX features such as:
- [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
- [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
- [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
- [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
Parameters:
config ([`BertConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~FlaxPreTrainedModel.from_pretrained`] method to load the model weights.
dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
`jax.numpy.bfloat16` (on TPUs).
This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
specified all the computation will be performed with the given `dtype`.
**Note that this only specifies the dtype of the computation and does not influence the dtype of model
parameters.**
If you wish to change the dtype of the model parameters, see [`~FlaxPreTrainedModel.to_fp16`] and
[`~FlaxPreTrainedModel.to_bf16`].
dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
`jax.numpy.bfloat16` (on TPUs).
This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
specified all the computation will be performed with the given `dtype`.
**Note that this only specifies the dtype of the computation and does not influence the dtype of model
parameters.**
If you wish to change the dtype of the model parameters, see [`~FlaxPreTrainedModel.to_fp16`] and
[`~FlaxPreTrainedModel.to_bf16`].
"""
BERT_INPUTS_DOCSTRING = r"""
Args:
input_ids (`numpy.ndarray` of shape `({0})`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`BertTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`numpy.ndarray` of shape `({0})`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
token_type_ids (`numpy.ndarray` of shape `({0})`, *optional*):
Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
1]`:
- 0 corresponds to a *sentence A* token,
- 1 corresponds to a *sentence B* token.
[What are token type IDs?](../glossary#token-type-ids)
position_ids (`numpy.ndarray` of shape `({0})`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.max_position_embeddings - 1]`.
head_mask (`numpy.ndarray` of shape `({0})`, `optional):
Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
class FlaxBertEmbeddings(nn.Module):
"""Construct the embeddings from word, position and token_type embeddings."""
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.word_embeddings = nn.Embed(
self.config.vocab_size,
self.config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
)
self.position_embeddings = nn.Embed(
self.config.max_position_embeddings,
self.config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
)
self.token_type_embeddings = nn.Embed(
self.config.type_vocab_size,
self.config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
)
self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
def __call__(self, input_ids, token_type_ids, position_ids, attention_mask, deterministic: bool = True):
# Embed
inputs_embeds = self.word_embeddings(input_ids.astype("i4"))
position_embeds = self.position_embeddings(position_ids.astype("i4"))
token_type_embeddings = self.token_type_embeddings(token_type_ids.astype("i4"))
# Sum all embeddings
hidden_states = inputs_embeds + token_type_embeddings + position_embeds
# Layer Norm
hidden_states = self.LayerNorm(hidden_states)
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
return hidden_states
class FlaxBertSelfAttention(nn.Module):
config: BertConfig
causal: bool = False
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.head_dim = self.config.hidden_size // self.config.num_attention_heads
if self.config.hidden_size % self.config.num_attention_heads != 0:
raise ValueError(
"`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads` "
" : {self.config.num_attention_heads}"
)
self.query = nn.Dense(
self.config.hidden_size,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
)
self.key = nn.Dense(
self.config.hidden_size,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
)
self.value = nn.Dense(
self.config.hidden_size,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
)
if self.causal:
self.causal_mask = make_causal_mask(
jnp.ones((1, self.config.max_position_embeddings), dtype="bool"), dtype="bool"
)
def _split_heads(self, hidden_states):
return hidden_states.reshape(hidden_states.shape[:2] + (self.config.num_attention_heads, self.head_dim))
def _merge_heads(self, hidden_states):
return hidden_states.reshape(hidden_states.shape[:2] + (self.config.hidden_size,))
@nn.compact
# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartAttention._concatenate_to_cache
def _concatenate_to_cache(self, key, value, query, attention_mask):
"""
This function takes projected key, value states from a single input token and concatenates the states to cached
states from previous steps. This function is slighly adapted from the official Flax repository:
https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
"""
# detect if we're initializing by absence of existing cache data.
is_initialized = self.has_variable("cache", "cached_key")
cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
if is_initialized:
*batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
# update key, value caches with our new 1d spatial slices
cur_index = cache_index.value
indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
key = lax.dynamic_update_slice(cached_key.value, key, indices)
value = lax.dynamic_update_slice(cached_value.value, value, indices)
cached_key.value = key
cached_value.value = value
num_updated_cache_vectors = query.shape[1]
cache_index.value = cache_index.value + num_updated_cache_vectors
# causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
pad_mask = jnp.broadcast_to(
jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
)
attention_mask = combine_masks(pad_mask, attention_mask)
return key, value, attention_mask
def __call__(
self,
hidden_states,
attention_mask,
layer_head_mask,
key_value_states: Optional[jnp.array] = None,
init_cache: bool = False,
deterministic=True,
output_attentions: bool = False,
):
# if key_value_states are provided this layer is used as a cross-attention layer
# for the decoder
is_cross_attention = key_value_states is not None
batch_size = hidden_states.shape[0]
# get query proj
query_states = self.query(hidden_states)
# get key, value proj
if is_cross_attention:
# cross_attentions
key_states = self.key(key_value_states)
value_states = self.value(key_value_states)
else:
# self_attention
key_states = self.key(hidden_states)
value_states = self.value(hidden_states)
query_states = self._split_heads(query_states)
key_states = self._split_heads(key_states)
value_states = self._split_heads(value_states)
# handle cache prepare causal attention mask
if self.causal:
query_length, key_length = query_states.shape[1], key_states.shape[1]
if self.has_variable("cache", "cached_key"):
mask_shift = self.variables["cache"]["cache_index"]
max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
causal_mask = lax.dynamic_slice(
self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
)
else:
causal_mask = self.causal_mask[:, :, :query_length, :key_length]
causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
# combine masks if needed
if attention_mask is not None and self.causal:
attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
attention_mask = combine_masks(attention_mask, causal_mask)
elif self.causal:
attention_mask = causal_mask
elif attention_mask is not None:
attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
# During fast autoregressive decoding, we feed one position at a time,
# and cache the keys and values step by step.
if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
key_states, value_states, attention_mask = self._concatenate_to_cache(
key_states, value_states, query_states, attention_mask
)
# Convert the boolean attention mask to an attention bias.
if attention_mask is not None:
# attention mask in the form of attention bias
attention_bias = lax.select(
attention_mask > 0,
jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
)
else:
attention_bias = None
dropout_rng = None
if not deterministic and self.config.attention_probs_dropout_prob > 0.0:
dropout_rng = self.make_rng("dropout")
attn_weights = dot_product_attention_weights(
query_states,
key_states,
bias=attention_bias,
dropout_rng=dropout_rng,
dropout_rate=self.config.attention_probs_dropout_prob,
broadcast_dropout=True,
deterministic=deterministic,
dtype=self.dtype,
precision=None,
)
# Mask heads if we want to
if layer_head_mask is not None:
attn_weights = jnp.einsum("...hqk,h->...hqk", attn_weights, layer_head_mask)
attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
attn_output = attn_output.reshape(attn_output.shape[:2] + (-1,))
outputs = (attn_output, attn_weights) if output_attentions else (attn_output,)
return outputs
class FlaxBertSelfOutput(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.dense = nn.Dense(
self.config.hidden_size,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
dtype=self.dtype,
)
self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
def __call__(self, hidden_states, input_tensor, deterministic: bool = True):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
class FlaxBertAttention(nn.Module):
config: BertConfig
causal: bool = False
dtype: jnp.dtype = jnp.float32
def setup(self):
self.self = FlaxBertSelfAttention(self.config, causal=self.causal, dtype=self.dtype)
self.output = FlaxBertSelfOutput(self.config, dtype=self.dtype)
def __call__(
self,
hidden_states,
attention_mask,
layer_head_mask,
key_value_states=None,
init_cache=False,
deterministic=True,
output_attentions: bool = False,
):
# Attention mask comes in as attention_mask.shape == (*batch_sizes, kv_length)
# FLAX expects: attention_mask.shape == (*batch_sizes, 1, 1, kv_length) such that it is broadcastable
# with attn_weights.shape == (*batch_sizes, num_heads, q_length, kv_length)
attn_outputs = self.self(
hidden_states,
attention_mask,
layer_head_mask=layer_head_mask,
key_value_states=key_value_states,
init_cache=init_cache,
deterministic=deterministic,
output_attentions=output_attentions,
)
attn_output = attn_outputs[0]
hidden_states = self.output(attn_output, hidden_states, deterministic=deterministic)
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_outputs[1],)
return outputs
class FlaxBertIntermediate(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.dense = nn.Dense(
self.config.intermediate_size,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
dtype=self.dtype,
)
self.activation = ACT2FN[self.config.hidden_act]
def __call__(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.activation(hidden_states)
return hidden_states
class FlaxBertOutput(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.dense = nn.Dense(
self.config.hidden_size,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
dtype=self.dtype,
)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
def __call__(self, hidden_states, attention_output, deterministic: bool = True):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
hidden_states = self.LayerNorm(hidden_states + attention_output)
return hidden_states
class FlaxBertLayer(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.attention = FlaxBertAttention(self.config, causal=self.config.is_decoder, dtype=self.dtype)
self.intermediate = FlaxBertIntermediate(self.config, dtype=self.dtype)
self.output = FlaxBertOutput(self.config, dtype=self.dtype)
if self.config.add_cross_attention:
self.crossattention = FlaxBertAttention(self.config, causal=False, dtype=self.dtype)
def __call__(
self,
hidden_states,
attention_mask,
layer_head_mask,
encoder_hidden_states: Optional[jnp.ndarray] = None,
encoder_attention_mask: Optional[jnp.ndarray] = None,
init_cache: bool = False,
deterministic: bool = True,
output_attentions: bool = False,
):
# Self Attention
attention_outputs = self.attention(
hidden_states,
attention_mask,
layer_head_mask=layer_head_mask,
init_cache=init_cache,
deterministic=deterministic,
output_attentions=output_attentions,
)
attention_output = attention_outputs[0]
# Cross-Attention Block
if encoder_hidden_states is not None:
cross_attention_outputs = self.crossattention(
attention_output,
attention_mask=encoder_attention_mask,
layer_head_mask=layer_head_mask,
key_value_states=encoder_hidden_states,
deterministic=deterministic,
output_attentions=output_attentions,
)
attention_output = cross_attention_outputs[0]
hidden_states = self.intermediate(attention_output)
hidden_states = self.output(hidden_states, attention_output, deterministic=deterministic)
outputs = (hidden_states,)
if output_attentions:
outputs += (attention_outputs[1],)
if encoder_hidden_states is not None:
outputs += (cross_attention_outputs[1],)
return outputs
class FlaxBertLayerCollection(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
gradient_checkpointing: bool = False
def setup(self):
if self.gradient_checkpointing:
FlaxBertCheckpointLayer = remat(FlaxBertLayer, static_argnums=(5, 6, 7))
self.layers = [
FlaxBertCheckpointLayer(self.config, name=str(i), dtype=self.dtype)
for i in range(self.config.num_hidden_layers)
]
else:
self.layers = [
FlaxBertLayer(self.config, name=str(i), dtype=self.dtype) for i in range(self.config.num_hidden_layers)
]
def __call__(
self,
hidden_states,
attention_mask,
head_mask,
encoder_hidden_states: Optional[jnp.ndarray] = None,
encoder_attention_mask: Optional[jnp.ndarray] = None,
init_cache: bool = False,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
# Check if head_mask has a correct number of layers specified if desired
if head_mask is not None:
if head_mask.shape[0] != (len(self.layers)):
raise ValueError(
f"The head_mask should be specified for {len(self.layers)} layers, but it is for "
f" {head_mask.shape[0]}."
)
for i, layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
layer_outputs = layer(
hidden_states,
attention_mask,
head_mask[i] if head_mask is not None else None,
encoder_hidden_states,
encoder_attention_mask,
init_cache,
deterministic,
output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions += (layer_outputs[1],)
if encoder_hidden_states is not None:
all_cross_attentions += (layer_outputs[2],)
if output_hidden_states:
all_hidden_states += (hidden_states,)
outputs = (hidden_states,)
if not return_dict:
return tuple(v for v in outputs if v is not None)
return FlaxBaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_attentions,
cross_attentions=all_cross_attentions,
)
class FlaxBertEncoder(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
gradient_checkpointing: bool = False
def setup(self):
self.layer = FlaxBertLayerCollection(
self.config,
dtype=self.dtype,
gradient_checkpointing=self.gradient_checkpointing,
)
def __call__(
self,
hidden_states,
attention_mask,
head_mask,
encoder_hidden_states: Optional[jnp.ndarray] = None,
encoder_attention_mask: Optional[jnp.ndarray] = None,
init_cache: bool = False,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
return self.layer(
hidden_states,
attention_mask,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
init_cache=init_cache,
deterministic=deterministic,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
class FlaxBertPooler(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.dense = nn.Dense(
self.config.hidden_size,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
dtype=self.dtype,
)
def __call__(self, hidden_states):
cls_hidden_state = hidden_states[:, 0]
cls_hidden_state = self.dense(cls_hidden_state)
return nn.tanh(cls_hidden_state)
class FlaxBertPredictionHeadTransform(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.dense = nn.Dense(self.config.hidden_size, dtype=self.dtype)
self.activation = ACT2FN[self.config.hidden_act]
self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
def __call__(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.activation(hidden_states)
return self.LayerNorm(hidden_states)
class FlaxBertLMPredictionHead(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32
bias_init: Callable[..., np.ndarray] = jax.nn.initializers.zeros
def setup(self):
self.transform = FlaxBertPredictionHeadTransform(self.config, dtype=self.dtype)
self.decoder = nn.Dense(self.config.vocab_size, dtype=self.dtype, use_bias=False)
self.bias = self.param("bias", self.bias_init, (self.config.vocab_size,))
def __call__(self, hidden_states, shared_embedding=None):
hidden_states = self.transform(hidden_states)
if shared_embedding is not None:
hidden_states = self.decoder.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
else:
hidden_states = self.decoder(hidden_states)
bias = jnp.asarray(self.bias, self.dtype)
hidden_states += bias
return hidden_states
class FlaxBertOnlyMLMHead(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.predictions = FlaxBertLMPredictionHead(self.config, dtype=self.dtype)
def __call__(self, hidden_states, shared_embedding=None):
hidden_states = self.predictions(hidden_states, shared_embedding=shared_embedding)
return hidden_states
class FlaxBertOnlyNSPHead(nn.Module):
dtype: jnp.dtype = jnp.float32
def setup(self):
self.seq_relationship = nn.Dense(2, dtype=self.dtype)
def __call__(self, pooled_output):
return self.seq_relationship(pooled_output)
class FlaxBertPreTrainingHeads(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.predictions = FlaxBertLMPredictionHead(self.config, dtype=self.dtype)
self.seq_relationship = nn.Dense(2, dtype=self.dtype)
def __call__(self, hidden_states, pooled_output, shared_embedding=None):
prediction_scores = self.predictions(hidden_states, shared_embedding=shared_embedding)
seq_relationship_score = self.seq_relationship(pooled_output)
return prediction_scores, seq_relationship_score
class FlaxBertPreTrainedModel(FlaxPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = BertConfig
base_model_prefix = "bert"
module_class: nn.Module = None
def __init__(
self,
config: BertConfig,
input_shape: Tuple = (1, 1),
seed: int = 0,
dtype: jnp.dtype = jnp.float32,
_do_init: bool = True,
gradient_checkpointing: bool = False,
**kwargs
):
module = self.module_class(
config=config,
dtype=dtype,
gradient_checkpointing=gradient_checkpointing,
**kwargs,
)
super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
def enable_gradient_checkpointing(self):
self._module = self.module_class(
config=self.config,
dtype=self.dtype,
gradient_checkpointing=True,
)
def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
# init input tensors
input_ids = jnp.zeros(input_shape, dtype="i4")
token_type_ids = jnp.zeros_like(input_ids)
position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
attention_mask = jnp.ones_like(input_ids)
head_mask = jnp.ones((self.config.num_hidden_layers, self.config.num_attention_heads))
params_rng, dropout_rng = jax.random.split(rng)
rngs = {"params": params_rng, "dropout": dropout_rng}
if self.config.add_cross_attention:
encoder_hidden_states = jnp.zeros(input_shape + (self.config.hidden_size,))
encoder_attention_mask = attention_mask
module_init_outputs = self.module.init(
rngs,
input_ids,
attention_mask,
token_type_ids,
position_ids,
head_mask,
encoder_hidden_states,
encoder_attention_mask,
return_dict=False,
)
else:
module_init_outputs = self.module.init(
rngs, input_ids, attention_mask, token_type_ids, position_ids, head_mask, return_dict=False
)
random_params = module_init_outputs["params"]
if params is not None:
random_params = flatten_dict(unfreeze(random_params))
params = flatten_dict(unfreeze(params))
for missing_key in self._missing_keys:
params[missing_key] = random_params[missing_key]
self._missing_keys = set()
return freeze(unflatten_dict(params))
else:
return random_params
# Copied from transformers.models.bart.modeling_flax_bart.FlaxBartDecoderPreTrainedModel.init_cache
def init_cache(self, batch_size, max_length):
r"""
Args:
batch_size (`int`):
batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
max_length (`int`):
maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
cache.
"""
# init input variables to retrieve cache
input_ids = jnp.ones((batch_size, max_length), dtype="i4")
attention_mask = jnp.ones_like(input_ids, dtype="i4")
position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
init_variables = self.module.init(
jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
)
return unfreeze(init_variables["cache"])
@add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
def __call__(
self,
input_ids,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
params: dict = None,
dropout_rng: jax.random.PRNGKey = None,
train: bool = False,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
past_key_values: dict = None,
):
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.return_dict
# init input tensors if not passed
if token_type_ids is None:
token_type_ids = jnp.zeros_like(input_ids)
if position_ids is None:
position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
if attention_mask is None:
attention_mask = jnp.ones_like(input_ids)
if head_mask is None:
head_mask = jnp.ones((self.config.num_hidden_layers, self.config.num_attention_heads))
# Handle any PRNG if needed
rngs = {}
if dropout_rng is not None:
rngs["dropout"] = dropout_rng
inputs = {"params": params or self.params}
if self.config.add_cross_attention:
# if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed
# down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be
# changed by FlaxBertAttention module
if past_key_values:
inputs["cache"] = past_key_values
mutable = ["cache"]
else:
mutable = False
outputs = self.module.apply(
inputs,
jnp.array(input_ids, dtype="i4"),
jnp.array(attention_mask, dtype="i4"),
token_type_ids=jnp.array(token_type_ids, dtype="i4"),
position_ids=jnp.array(position_ids, dtype="i4"),
head_mask=jnp.array(head_mask, dtype="i4"),
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
deterministic=not train,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
rngs=rngs,
mutable=mutable,
)
# add updated cache to model output
if past_key_values is not None and return_dict:
outputs, past_key_values = outputs
outputs["past_key_values"] = unfreeze(past_key_values["cache"])
return outputs
elif past_key_values is not None and not return_dict:
outputs, past_key_values = outputs
outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
else:
outputs = self.module.apply(
inputs,
jnp.array(input_ids, dtype="i4"),
jnp.array(attention_mask, dtype="i4"),
token_type_ids=jnp.array(token_type_ids, dtype="i4"),
position_ids=jnp.array(position_ids, dtype="i4"),
head_mask=jnp.array(head_mask, dtype="i4"),
deterministic=not train,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
rngs=rngs,
)
return outputs
class FlaxBertModule(nn.Module):
config: BertConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
add_pooling_layer: bool = True
gradient_checkpointing: bool = False
def setup(self):
self.embeddings = FlaxBertEmbeddings(self.config, dtype=self.dtype)
self.encoder = FlaxBertEncoder(
self.config,
dtype=self.dtype,
gradient_checkpointing=self.gradient_checkpointing,
)
self.pooler = FlaxBertPooler(self.config, dtype=self.dtype)
def __call__(
self,
input_ids,
attention_mask,
token_type_ids: Optional[jnp.ndarray] = None,
position_ids: Optional[jnp.ndarray] = None,
head_mask: Optional[jnp.ndarray] = None,
encoder_hidden_states: Optional[jnp.ndarray] = None,
encoder_attention_mask: Optional[jnp.ndarray] = None,
init_cache: bool = False,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
# make sure `token_type_ids` is correctly initialized when not passed
if token_type_ids is None:
token_type_ids = jnp.zeros_like(input_ids)
# make sure `position_ids` is correctly initialized when not passed
if position_ids is None:
position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
hidden_states = self.embeddings(
input_ids, token_type_ids, position_ids, attention_mask, deterministic=deterministic
)
outputs = self.encoder(
hidden_states,
attention_mask,
head_mask=head_mask,
deterministic=deterministic,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,