a-mannion commited on
Commit
c2852a4
1 Parent(s): 0642ab8

Upload 3 files

Browse files

initial commit: library code

Files changed (3) hide show
  1. jargon_configuration.py +88 -0
  2. jargon_model.py +522 -0
  3. linformer.py +740 -0
jargon_configuration.py ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ from transformers.models.roberta.modeling_roberta import RobertaConfig
3
+
4
+
5
+ class JargonConfig(RobertaConfig):
6
+ model_type = "jargon"
7
+
8
+ def __init__(
9
+ self,
10
+ compress_layer= 1,
11
+ shared_layer_kv_compressed=1,
12
+ shared_kv_compressed=0,
13
+ max_positions=512,
14
+ max_position_embeddings=512,
15
+ compressed=4,
16
+ vocab_size=30522,
17
+ freeze_compress=0,
18
+ embed_dim=768,
19
+ num_heads=16,
20
+ dim_feedforward=4096,
21
+ dropout=0.1,
22
+ activation="relu",
23
+ layer_norm_eps=1e-05,
24
+ self_attention=True,
25
+ encoder_decoder_attention=False,
26
+ bias=True,
27
+ q_noise=0,
28
+ qn_block_size=8,
29
+ add_bias_kv=False,
30
+ add_zero_attn=False,
31
+ num_layers=12,
32
+ untie_weights_roberta=False,
33
+ layernorm_embedding=False,
34
+ encoder_normalize_before=False,
35
+ encoder_embed_dim=768,
36
+ encoder_attention_heads=12,
37
+ quant_noise_pq=0.0,
38
+ quant_noise_pq_block_size=8,
39
+ quant_noise_scalar=0,
40
+ encoder_ffn_embed_dim=4096,
41
+ add_pooling_layer=False,
42
+ intermediate_size=4096,
43
+ intermediate_act_fn="relu",
44
+ hidden_act="relu",
45
+ output_hidden_states=False,
46
+ position_embedding_type="learned",
47
+ **kwargs
48
+ ):
49
+ super().__init__(**kwargs)
50
+
51
+ self.add_pooling_layer = add_pooling_layer
52
+ self.compress_layer = compress_layer
53
+ self.shared_layer_kv_compressed = shared_layer_kv_compressed
54
+ self.shared_kv_compressed = shared_kv_compressed
55
+ self.max_positions = max_positions
56
+ self.max_position_embeddings = max_position_embeddings
57
+ self.compressed = compressed
58
+ self.freeze_compress = freeze_compress
59
+ self.embed_dim = embed_dim
60
+ self.num_heads = num_heads
61
+ self.dim_feedforward=dim_feedforward
62
+ self.dropout = dropout
63
+ self.activation= activation
64
+ self.layer_norm_eps = layer_norm_eps
65
+ self.self_attention = self_attention
66
+ self.encoder_decoder_attention = encoder_decoder_attention
67
+ self.bias = bias
68
+ self.q_noise = q_noise
69
+ self.qn_block_size = qn_block_size
70
+ self.add_bias_kv = add_bias_kv
71
+ self.add_zero_attn = add_zero_attn
72
+ self.num_layers = num_layers
73
+ self.untie_weights_roberta = untie_weights_roberta
74
+ self.layernorm_embedding=layernorm_embedding
75
+ self.encoder_embed_dim = encoder_embed_dim
76
+ self.encoder_attention_heads=encoder_attention_heads
77
+ self.quant_noise_pq = quant_noise_pq
78
+ self.quant_noise_pq_block_size=quant_noise_pq_block_size
79
+ self.quant_noise_scalar=quant_noise_scalar
80
+ self.encoder_normalize_before=encoder_normalize_before
81
+ self.encoder_ffn_embed_dim = encoder_ffn_embed_dim
82
+ self.vocab_size = vocab_size
83
+ self.intermediate_size = intermediate_size
84
+ self.intermediate_act_fn = intermediate_act_fn
85
+ self.output_hidden_states = output_hidden_states
86
+ self.hidden_act = hidden_act
87
+ self.position_embedding_type = position_embedding_type
88
+ self.encoder_normalize_before = encoder_normalize_before
jargon_model.py ADDED
@@ -0,0 +1,522 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ from typing import List, Optional, Tuple, Union
3
+
4
+ import torch
5
+ import torch.nn as nn
6
+ import torch.nn.functional as F
7
+ from torch.nn import LayerNorm
8
+ from fairseq.models.roberta import (
9
+ RobertaModel as RobertModel,
10
+ RobertaEncoder as RobertaEncoderFS
11
+ )
12
+ from transformers.models.roberta.modeling_roberta import (
13
+ RobertaEncoder,
14
+ RobertaConfig,
15
+ RobertaModel,
16
+ RobertaLMHead,
17
+ RobertaForMaskedLM,
18
+ RobertaEmbeddings,
19
+ RobertaForTokenClassification,
20
+ RobertaForSequenceClassification
21
+ )
22
+ from transformers.modeling_outputs import (
23
+ MaskedLMOutput,
24
+ BaseModelOutputWithPastAndCrossAttentions,
25
+ BaseModelOutputWithPoolingAndCrossAttentions,
26
+ )
27
+
28
+ from .linformer import LinformerTransformerEncoderLayer
29
+ from .jargon_configuration import JargonConfig
30
+
31
+
32
+ class JargonForSequenceClassification(RobertaForSequenceClassification):
33
+
34
+ config_class = JargonConfig
35
+
36
+ def __init__(self, config, **kwargs):
37
+ base_model_prefix = "jargon"
38
+
39
+ super().__init__(config, **kwargs)
40
+
41
+ self.roberta = JargonModel(config, add_pooling_layer=False)
42
+ self.sbo_head = self.build_sbo_head(config)
43
+
44
+ def build_sbo_head(self, config):
45
+ return SBOHead(
46
+ config,
47
+ embedding_weights=(
48
+ self.roberta.embeddings.word_embeddings.weight
49
+ if not config.untie_weights_roberta
50
+ else None
51
+ )
52
+ )
53
+
54
+
55
+ class JargonForTokenClassification(RobertaForTokenClassification):
56
+
57
+ config_class = JargonConfig
58
+
59
+ def __init__(self, config, **kwargs):
60
+ base_model_prefix = "jargon"
61
+
62
+ super().__init__(config, **kwargs)
63
+
64
+ self.roberta = JargonModel(config, add_pooling_layer=False)
65
+ self.sbo_head = self.build_sbo_head(config)
66
+
67
+ def build_sbo_head(self, config):
68
+ return SBOHead(
69
+ config,
70
+ embedding_weights=(
71
+ self.roberta.embeddings.word_embeddings.weight
72
+ if not config.untie_weights_roberta
73
+ else None
74
+ )
75
+ )
76
+
77
+
78
+ class JargonForMaskedLM(RobertaForMaskedLM):
79
+
80
+ config_class = JargonConfig
81
+
82
+ def __init__(self, config, **kwargs):
83
+ base_model_prefix = "jargon"
84
+
85
+ super().__init__(config, **kwargs)
86
+
87
+ self.roberta = JargonModel(config, add_pooling_layer=False)
88
+ self.sbo_head = self.build_sbo_head(config)
89
+
90
+ def build_sbo_head(self, config):
91
+ return SBOHead(
92
+ config,
93
+ embedding_weights=(
94
+ self.roberta.embeddings.word_embeddings.weight
95
+ if not config.untie_weights_roberta
96
+ else None
97
+ )
98
+ )
99
+
100
+
101
+ class JargonForMaskedLMFS(RobertaForMaskedLM):
102
+
103
+ def __init__(self, config, dictionary, **kwargs):
104
+ config_class = JargonConfig
105
+ base_model_prefix = "jargon"
106
+
107
+ super().__init__(config, **kwargs)
108
+
109
+ self.roberta = FlaubertEncoder(config, dictionary)
110
+
111
+ def build_sbo_head(self, config):
112
+ return SBOHead(
113
+ config,
114
+ embedding_weights=(
115
+ self.roberta.embeddings.word_embeddings.weight
116
+ if not config.untie_weights_roberta
117
+ else None
118
+ )
119
+ )
120
+
121
+
122
+ class JargonEmbeddings(RobertaEmbeddings):
123
+
124
+ def __init__(self, config, **kwargs):
125
+ config_class = JargonConfig
126
+ base_model_prefix = "jargon"
127
+ super().__init__(config, **kwargs)
128
+
129
+ def forward(
130
+ self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
131
+ ):
132
+ if position_ids is None:
133
+ if input_ids is not None:
134
+ # Create the position ids from the input token ids. Any padded tokens remain padded.
135
+ position_ids = create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length)
136
+ else:
137
+ position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
138
+
139
+ if input_ids is not None:
140
+ input_shape = input_ids.size()
141
+ else:
142
+ input_shape = inputs_embeds.size()[:-1]
143
+
144
+ seq_length = input_shape[1]
145
+
146
+ # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
147
+ # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
148
+ # issue #5664
149
+ if token_type_ids is None:
150
+ if hasattr(self, "token_type_ids"):
151
+ buffered_token_type_ids = self.token_type_ids[:, :seq_length]
152
+ buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
153
+ token_type_ids = buffered_token_type_ids_expanded
154
+ else:
155
+ token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
156
+
157
+ if inputs_embeds is None:
158
+ inputs_embeds = self.word_embeddings(input_ids)
159
+ token_type_embeddings = self.token_type_embeddings(token_type_ids)
160
+
161
+ embeddings = inputs_embeds + token_type_embeddings
162
+ position_embeddings = self.position_embeddings(position_ids)
163
+
164
+ embeddings += position_embeddings
165
+ embeddings = self.dropout(embeddings)
166
+ return embeddings
167
+
168
+
169
+ class JargonEncoder(RobertaEncoder):
170
+
171
+ def __init__(self, args):
172
+ compress_layer = None
173
+ if args.shared_layer_kv_compressed == 1 and compress_layer is None:
174
+ compress_layer = nn.Linear(
175
+ args.max_positions,
176
+ args.max_positions // args.compressed
177
+ )
178
+ # intialize parameters for compressed layer
179
+ nn.init.xavier_uniform_(compress_layer.weight, gain=1 / math.sqrt(2))
180
+ if args.freeze_compress == 1:
181
+ compress_layer.weight.requires_grad = False
182
+ compress_layer = compress_layer
183
+
184
+ super().__init__(args)
185
+
186
+ self.layer = nn.ModuleList([LinformerTransformerEncoderLayer(args, compress_layer) for _ in range(args.num_layers)])
187
+ self.compress_layer = compress_layer
188
+
189
+ if args.encoder_normalize_before:
190
+ self.layer_norm = LayerNorm(args.embed_dim)
191
+ else:
192
+ self.layer_norm = None
193
+
194
+ self.lm_head = None
195
+
196
+ def forward(
197
+ self,
198
+ hidden_states: torch.Tensor,
199
+ attention_mask: Optional[torch.FloatTensor] = None,
200
+ head_mask: Optional[torch.FloatTensor] = None,
201
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
202
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
203
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
204
+ use_cache: Optional[bool] = None,
205
+ output_attentions: Optional[bool] = False,
206
+ output_hidden_states: Optional[bool] = False,
207
+ return_dict: Optional[bool] = True,
208
+ ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
209
+
210
+ x = super().forward(hidden_states=hidden_states,
211
+ attention_mask=attention_mask,
212
+ head_mask=head_mask,
213
+ encoder_hidden_states=encoder_hidden_states,
214
+ encoder_attention_mask=encoder_attention_mask,
215
+ past_key_values=past_key_values,
216
+ use_cache=use_cache,
217
+ output_attentions=output_attentions,
218
+ output_hidden_states=output_hidden_states,
219
+ return_dict=return_dict)
220
+
221
+
222
+ if self.layer_norm is not None:
223
+ x.last_hidden_state = self.layer_norm(x.last_hidden_state)
224
+
225
+ return x
226
+
227
+ def build_encoder(self, args, dictionary, embed_tokens):
228
+ encoder = LinformerTransformerEncoder(args)
229
+ return encoder
230
+ if args.use_linformer:
231
+ encoder = LinformerTransformerEncoder(args, dictionary, embed_tokens)
232
+ elif args.use_fft:
233
+ encoder = FourierTransformerEncoder(args, dictionary, embed_tokens)
234
+ else:
235
+ encoder = TransformerEncoder(args, dictionary, embed_tokens)
236
+
237
+ encoder.apply(init_bert_params)
238
+
239
+ return encoder
240
+
241
+ def output_layer(self, features, masked_tokens=None, pairs=None, **unused):
242
+ lm_out = self.lm_head(features, masked_tokens)
243
+ if pairs is not None:
244
+ sbo_out = self.sbo_head(features, pairs)
245
+ return lm_out, sbo_out
246
+ else:
247
+ return lm_out
248
+
249
+
250
+ class JargonModel(RobertaModel):
251
+ config_class = JargonConfig
252
+ def __init__(self, config, **kwargs):
253
+ config_class = JargonConfig
254
+ base_model_prefix = "jargon"
255
+
256
+ super().__init__(config, **kwargs)
257
+ self.embeddings = JargonEmbeddings(config)
258
+ self.encoder = JargonEncoder(config)
259
+ # Copied from modeling_roberta.py
260
+ # Add transpose of embeddings as implemented in fairseq
261
+ def forward(
262
+ self,
263
+ input_ids: Optional[torch.Tensor] = None,
264
+ attention_mask: Optional[torch.Tensor] = None,
265
+ token_type_ids: Optional[torch.Tensor] = None,
266
+ position_ids: Optional[torch.Tensor] = None,
267
+ head_mask: Optional[torch.Tensor] = None,
268
+ inputs_embeds: Optional[torch.Tensor] = None,
269
+ encoder_hidden_states: Optional[torch.Tensor] = None,
270
+ encoder_attention_mask: Optional[torch.Tensor] = None,
271
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
272
+ use_cache: Optional[bool] = None,
273
+ output_attentions: Optional[bool] = None,
274
+ output_hidden_states: Optional[bool] = None,
275
+ return_dict: Optional[bool] = None,
276
+ ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
277
+ r"""
278
+ encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
279
+ Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
280
+ the model is configured as a decoder.
281
+ encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
282
+ Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
283
+ the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
284
+
285
+ - 1 for tokens that are **not masked**,
286
+ - 0 for tokens that are **masked**.
287
+ past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
288
+ Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
289
+
290
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
291
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
292
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
293
+ use_cache (`bool`, *optional*):
294
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
295
+ `past_key_values`).
296
+ """
297
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
298
+ output_hidden_states = (
299
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
300
+ )
301
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
302
+
303
+ if self.config.is_decoder:
304
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
305
+ else:
306
+ use_cache = False
307
+
308
+ if input_ids is not None and inputs_embeds is not None:
309
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
310
+ elif input_ids is not None:
311
+ input_shape = input_ids.size()
312
+ elif inputs_embeds is not None:
313
+ input_shape = inputs_embeds.size()[:-1]
314
+ else:
315
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
316
+
317
+ batch_size, seq_length = input_shape
318
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
319
+
320
+ # past_key_values_length
321
+ past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
322
+
323
+ if attention_mask is None:
324
+ attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
325
+
326
+ if token_type_ids is None:
327
+ if hasattr(self.embeddings, "token_type_ids"):
328
+ buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
329
+ buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
330
+ token_type_ids = buffered_token_type_ids_expanded
331
+ else:
332
+ token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
333
+
334
+ # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
335
+ # ourselves in which case we just need to make it broadcastable to all heads.
336
+ extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
337
+
338
+ # If a 2D or 3D attention mask is provided for the cross-attention
339
+ # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
340
+ if self.config.is_decoder and encoder_hidden_states is not None:
341
+ encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
342
+ encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
343
+ if encoder_attention_mask is None:
344
+ encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
345
+ encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
346
+ else:
347
+ encoder_extended_attention_mask = None
348
+
349
+ # Prepare head mask if needed
350
+ # 1.0 in head_mask indicate we keep the head
351
+ # attention_probs has shape bsz x n_heads x N x N
352
+ # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
353
+ # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
354
+ head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
355
+
356
+ embedding_output = self.embeddings(
357
+ input_ids=input_ids,
358
+ position_ids=position_ids,
359
+ token_type_ids=token_type_ids,
360
+ inputs_embeds=inputs_embeds,
361
+ past_key_values_length=past_key_values_length,
362
+ )
363
+
364
+
365
+ embedding_output = embedding_output.transpose(0,1)
366
+ encoder_outputs = self.encoder(
367
+ embedding_output,
368
+ attention_mask=extended_attention_mask,
369
+ head_mask=head_mask,
370
+ encoder_hidden_states=encoder_hidden_states,
371
+ encoder_attention_mask=encoder_extended_attention_mask,
372
+ past_key_values=past_key_values,
373
+ use_cache=use_cache,
374
+ output_attentions=output_attentions,
375
+ output_hidden_states=output_hidden_states,
376
+ )
377
+
378
+ sequence_output = encoder_outputs[0].transpose(0,1)
379
+
380
+ pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
381
+
382
+ # Fairseq Linformer implementation works with transposed hidden states -> we transpose them back for HF implementation.
383
+ if output_hidden_states:
384
+ encoder_outputs.hidden_states = [h.transpose(0,1) for h in encoder_outputs.hidden_states]
385
+
386
+ if not return_dict:
387
+ return (sequence_output, pooled_output) + encoder_outputs[1:]
388
+
389
+ return BaseModelOutputWithPoolingAndCrossAttentions(
390
+ last_hidden_state=sequence_output,
391
+ pooler_output=pooled_output,
392
+ past_key_values=encoder_outputs.past_key_values,
393
+ hidden_states=encoder_outputs.hidden_states,
394
+ attentions=encoder_outputs.attentions,
395
+ cross_attentions=encoder_outputs.cross_attentions,
396
+ )
397
+
398
+
399
+ class SBOLayer(nn.Module):
400
+
401
+ def __init__(self, input_size, hidden_size, activation, export):
402
+ super().__init__()
403
+ self.layer = nn.Linear(input_size, hidden_size)
404
+ self.activ = get_activation_fn(activation)
405
+ self.norm = LayerNorm(hidden_size)
406
+
407
+ def forward(self, x):
408
+ return self.norm(self.activ(self.layer(x)))
409
+
410
+
411
+ class SBONetwork(nn.Module):
412
+
413
+ def __init__(self, input_size, hidden_size, activation, export):
414
+ super().__init__()
415
+ self.layers = nn.ModuleList([
416
+ self.build_sbo_layer(input_size, hidden_size, activation, export),
417
+ self.build_sbo_layer(hidden_size, hidden_size, activation, export)
418
+ ])
419
+ self.layers = nn.Sequential(*self.layers)
420
+
421
+ def build_sbo_layer(self, input_size, output_size, activation, export):
422
+ return SBOLayer(input_size, output_size, activation, export)
423
+
424
+ def forward(self, x):
425
+ return self.layers(x)
426
+
427
+
428
+ class SBOHead(nn.Module):
429
+
430
+ def __init__(self, args, embedding_weights, max_targets=10, position_embedding_size=200):
431
+ super().__init__()
432
+
433
+ self.position_embeddings = nn.Embedding(max_targets, position_embedding_size)
434
+
435
+ export = getattr(args, "export", False)
436
+ hidden_size = args.embed_dim
437
+ input_size = hidden_size * 2 + position_embedding_size
438
+ activation = getattr(args, "activation_fn", "relu") or "relu"
439
+
440
+ self.mlp_layer_norm = self.build_sbo_network(input_size, hidden_size, activation, export)
441
+
442
+ # The output weights are the same as the input embeddings, but there is
443
+ # an output-only bias for each token.
444
+ self.decoder = nn.Linear(
445
+ embedding_weights.size(1),
446
+ embedding_weights.size(0),
447
+ bias=False
448
+ )
449
+ if embedding_weights is not None:
450
+ self.decoder.weight = embedding_weights
451
+
452
+ self.bias = nn.Parameter(torch.zeros(embedding_weights.size(0)))
453
+ self.max_targets = max_targets
454
+
455
+ def build_sbo_network(self, input_size, hidden_size, activation, export):
456
+ return SBONetwork(input_size, hidden_size, activation, export)
457
+
458
+ def forward(self, hidden_states, pairs):
459
+ bs, num_pairs, _ = pairs.size()
460
+ bs, seq_len, dim = hidden_states.size()
461
+ # pair indices: (bs, num_pairs)
462
+ left, right = pairs[:,:, 0], pairs[:, :, 1]
463
+ # (bs, num_pairs, dim)
464
+ left_hidden = torch.gather(hidden_states, 1, left.unsqueeze(2).repeat(1, 1, dim))
465
+ # pair states: bs * num_pairs, max_targets, dim
466
+ left_hidden = left_hidden.contiguous().view(bs * num_pairs, dim).unsqueeze(1).repeat(1, self.max_targets, 1)
467
+
468
+ right_hidden = torch.gather(hidden_states, 1, right.unsqueeze(2).repeat(1, 1, dim))
469
+ # bs * num_pairs, max_targets, dim
470
+ right_hidden = right_hidden.contiguous().view(bs * num_pairs, dim).unsqueeze(1).repeat(1, self.max_targets, 1)
471
+
472
+ # (max_targets, dim)
473
+ position_embeddings = self.position_embeddings.weight
474
+
475
+ z = torch.cat((left_hidden, right_hidden, position_embeddings.unsqueeze(0).repeat(bs * num_pairs, 1, 1)), -1)
476
+
477
+ hidden_states = self.mlp_layer_norm(torch.cat((left_hidden, right_hidden, position_embeddings.unsqueeze(0).repeat(bs * num_pairs, 1, 1)), -1))
478
+ # target scores : bs * num_pairs, max_targets, vocab_size
479
+ target_scores = self.decoder(hidden_states) + self.bias
480
+ return target_scores
481
+
482
+
483
+ def get_activation_fn(activation):
484
+ """Returns the activation function corresponding to `activation`"""
485
+
486
+ if activation == "relu":
487
+ return F.relu
488
+ elif activation == "relu_squared":
489
+ return F.relu_squared
490
+ elif activation == "gelu":
491
+ return F.gelu
492
+ elif activation == "gelu_fast":
493
+ deprecation_warning(
494
+ "--activation-fn=gelu_fast has been renamed to gelu_accurate"
495
+ )
496
+ return F.gelu_accurate
497
+ elif activation == "gelu_accurate":
498
+ return F.gelu_accurate
499
+ elif activation == "tanh":
500
+ return torch.tanh
501
+ elif activation == "linear":
502
+ return lambda x: x
503
+ elif activation == "swish":
504
+ return torch.nn.SiLU
505
+ else:
506
+ raise RuntimeError("--activation-fn {} not supported".format(activation))
507
+
508
+
509
+ def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0):
510
+ """
511
+ Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
512
+ are ignored. This is modified from fairseq's `utils.make_positions`.
513
+
514
+ Args:
515
+ x: torch.Tensor x:
516
+
517
+ Returns: torch.Tensor
518
+ """
519
+ # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
520
+ mask = input_ids.ne(padding_idx).int()
521
+ incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
522
+ return incremental_indices.long() + padding_idx
linformer.py ADDED
@@ -0,0 +1,740 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+ #
3
+ # This source code is licensed under the MIT license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ import math
7
+ import inspect
8
+ from typing import Callable, Dict, List, Optional, Set, Tuple, Union
9
+
10
+ import torch
11
+ import torch.nn as nn
12
+ import torch.nn.functional as F
13
+ from torch.nn import TransformerEncoder, TransformerEncoderLayer
14
+ from fairseq import utils
15
+ from fairseq.models.transformer import *
16
+ from fairseq.incremental_decoding_utils import with_incremental_state
17
+ from fairseq.modules.quant_noise import quant_noise
18
+ from transformers.models.roberta.modeling_roberta import (
19
+ RobertaEncoder,
20
+ RobertaConfig,
21
+ RobertaModel,
22
+ RobertaLMHead,
23
+ RobertaForMaskedLM,
24
+ RobertaLayer
25
+ )
26
+
27
+ # from .multihead_linear_attention import MultiheadLinearAttention
28
+
29
+
30
+ class LinformerTransformerEncoderLayer(RobertaLayer):
31
+ """
32
+ Implements a Linformer Encoder Layer used in BERT/XLM style pre-trained
33
+ models.
34
+ """
35
+
36
+ def __init__(self, config, shared_compress_layer):
37
+ # wrap in a list so it's not automatically registered by PyTorch
38
+ self.shared_compress_layer = [shared_compress_layer]
39
+ d_model=config.embed_dim
40
+ nhead=config.num_heads
41
+ dim_feedforward=config.dim_feedforward
42
+ dropout=config.dropout
43
+ activation=config.activation
44
+ layer_norm_eps=config.layer_norm_eps
45
+
46
+ super().__init__(config)
47
+ self.attention = self.build_self_attention(config.embed_dim, config)
48
+ self.attn_layer_norm = nn.LayerNorm(config.hidden_size, eps=1e-5)
49
+ self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=1e-5)
50
+ self.output = RobertaOutput(config)
51
+
52
+ def build_self_attention(self, embed_dim, args):
53
+
54
+ attn = MultiheadLinearAttention(
55
+ embed_dim,
56
+ args.encoder_attention_heads,
57
+ dropout=args.dropout,
58
+ self_attention=True,
59
+ q_noise=args.quant_noise_pq,
60
+ qn_block_size=args.quant_noise_pq_block_size,
61
+ compressed=args.compressed,
62
+ max_seq_len=args.max_positions,
63
+ shared_kv_compressed=args.shared_kv_compressed,
64
+ shared_compress_layer=self.shared_compress_layer[0],
65
+ freeze_compress=args.freeze_compress,
66
+ )
67
+ return attn
68
+
69
+ def feed_forward_chunk(self, attention_output):
70
+ residual = attention_output
71
+ x = self.intermediate(attention_output)
72
+ layer_output = self.output(x, residual)
73
+ return layer_output
74
+
75
+ def forward(
76
+ self,
77
+ hidden_states: torch.Tensor,
78
+ attention_mask: Optional[torch.FloatTensor] = None,
79
+ head_mask: Optional[torch.FloatTensor] = None,
80
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
81
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
82
+ past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
83
+ output_attentions: Optional[bool] = False,
84
+ ) -> Tuple[torch.Tensor]:
85
+
86
+ residual = hidden_states
87
+
88
+ if self.attn_layer_norm is not None:
89
+ hidden_states = self.attn_layer_norm(hidden_states)
90
+
91
+ # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
92
+ self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
93
+ self_attention_outputs = self.attention(
94
+ hidden_states,
95
+ attention_mask,
96
+ head_mask,
97
+ output_attentions=output_attentions,
98
+ past_key_value=self_attn_past_key_value,
99
+ )
100
+ attention_output = self_attention_outputs[0]
101
+
102
+ # if decoder, the last output is tuple of self-attn cache
103
+ if self.is_decoder:
104
+ outputs = self_attention_outputs[1:-1]
105
+ present_key_value = self_attention_outputs[-1]
106
+ else:
107
+ outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
108
+
109
+ cross_attn_present_key_value = None
110
+ if self.is_decoder and encoder_hidden_states is not None:
111
+ if not hasattr(self, "crossattention"):
112
+ raise ValueError(
113
+ f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
114
+ " by setting `config.add_cross_attention=True`"
115
+ )
116
+
117
+ # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
118
+ cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
119
+ cross_attention_outputs = self.crossattention(
120
+ attention_output,
121
+ attention_mask,
122
+ head_mask,
123
+ encoder_hidden_states,
124
+ encoder_attention_mask,
125
+ cross_attn_past_key_value,
126
+ output_attentions,
127
+ )
128
+ attention_output = cross_attention_outputs[0]
129
+ outputs = outputs + cross_attention_outputs[1:-1] # add cross attentions if we output attention weights
130
+
131
+ # add cross-attn cache to positions 3,4 of present_key_value tuple
132
+ cross_attn_present_key_value = cross_attention_outputs[-1]
133
+ present_key_value = present_key_value + cross_attn_present_key_value
134
+
135
+ attention_output = attention_output + residual
136
+ residual = attention_output
137
+ attention_output = self.final_layer_norm(attention_output)
138
+ layer_output = apply_chunking_to_forward(
139
+ self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
140
+ )
141
+ layer_output = layer_output + residual
142
+
143
+ outputs = (layer_output,) + outputs
144
+
145
+ # if decoder, return the attn key/values as the last output
146
+ if self.is_decoder:
147
+ outputs = outputs + (present_key_value,)
148
+
149
+ return outputs
150
+
151
+
152
+ class RobertaOutput(nn.Module):
153
+ def __init__(self, config):
154
+ super().__init__()
155
+ self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
156
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
157
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
158
+
159
+ def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
160
+ hidden_states = self.dense(hidden_states)
161
+ return hidden_states
162
+
163
+
164
+ class LinformerTransformerEncoder(RobertaEncoder):
165
+ """
166
+ Implementation for a Bi-directional Linformer based Sentence Encoder used
167
+ in BERT/XLM style pre-trained models.
168
+
169
+ This first computes the token embedding using the token embedding matrix,
170
+ position embeddings (if specified) and segment embeddings
171
+ (if specified). After applying the specified number of
172
+ LinformerEncoderLayers, it outputs all the internal states of the
173
+ encoder as well as the final representation associated with the first
174
+ token (usually CLS token).
175
+
176
+ Input:
177
+ - tokens: B x T matrix representing sentences
178
+ - segment_labels: B x T matrix representing segment label for tokens
179
+
180
+ Output:
181
+ - a tuple of the following:
182
+ - a list of internal model states used to compute the
183
+ predictions where each tensor has shape T x B x C
184
+ - sentence representation associated with first input token
185
+ in format B x C.
186
+ """
187
+
188
+ def __init__(self, config,**kwargs):
189
+ compress_layer = None
190
+ if config.shared_layer_kv_compressed == 1 and compress_layer is None:
191
+ compress_layer = nn.Linear(
192
+ config.max_positions,
193
+ config.max_positions // config.compressed
194
+ )
195
+ # intialize parameters for compressed layer
196
+ nn.init.xavier_uniform_(compress_layer.weight, gain=1 / math.sqrt(2))
197
+ if config.freeze_compress == 1:
198
+ compress_layer.weight.requires_grad = False
199
+ compress_layer = compress_layer
200
+ #encoder_layer = LinformerTransformerEncoderLayer(config, compress_layer)
201
+
202
+ super().__init__(config)
203
+
204
+ self.layer = nn.ModuleList([LinformerTransformerEncoderLayer(config, compress_layer) for _ in range(config.num_layers)])
205
+ self.compress_layer = compress_layer
206
+ self.layer_norm = nn.LayerNorm(config.embed_dim)
207
+
208
+
209
+ @with_incremental_state
210
+ class MultiheadLinearAttention(nn.Module):
211
+ def __init__(
212
+ self,
213
+ embed_dim,
214
+ num_heads,
215
+ kdim=None,
216
+ vdim=None,
217
+ dropout=0.0,
218
+ bias=True,
219
+ add_bias_kv=False,
220
+ add_zero_attn=False,
221
+ self_attention=False,
222
+ encoder_decoder_attention=False,
223
+ q_noise=0.0,
224
+ qn_block_size=8,
225
+ compressed=1,
226
+ max_seq_len=256,
227
+ shared_kv_compressed=0,
228
+ shared_compress_layer=None,
229
+ freeze_compress=0,
230
+ ):
231
+ super().__init__()
232
+ self.embed_dim = embed_dim
233
+ self.kdim = kdim if kdim is not None else embed_dim
234
+ self.vdim = vdim if vdim is not None else embed_dim
235
+ self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim
236
+ self.num_heads = num_heads
237
+ self.dropout = dropout
238
+ self.head_dim = embed_dim // num_heads
239
+ assert (
240
+ self.head_dim * num_heads == self.embed_dim
241
+ ), "embed_dim must be divisible by num_heads"
242
+ self.scaling = self.head_dim ** -0.5
243
+
244
+ self.self_attention = self_attention
245
+ self.encoder_decoder_attention = encoder_decoder_attention
246
+ assert not self.self_attention or self.qkv_same_dim, (
247
+ "Self-attention requires query, key and " "value to be of the same size"
248
+ )
249
+
250
+ self.k_proj = quant_noise(
251
+ nn.Linear(self.kdim, embed_dim, bias=bias), q_noise, qn_block_size
252
+ )
253
+ self.v_proj = quant_noise(
254
+ nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size
255
+ )
256
+ self.q_proj = quant_noise(
257
+ nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size
258
+ )
259
+
260
+ # used for compress sequence to subsequence
261
+ if shared_compress_layer is None:
262
+ self.compress_seq_len = max_seq_len // compressed
263
+ self.compress_k = nn.Linear(max_seq_len, self.compress_seq_len, bias=False)
264
+ if shared_kv_compressed == 0:
265
+ self.compress_v = nn.Linear(
266
+ max_seq_len, self.compress_seq_len, bias=False
267
+ )
268
+ self.layerwise_sharing = False
269
+ else:
270
+ self.compress_k = shared_compress_layer
271
+ if shared_kv_compressed == 0:
272
+ self.compress_v = shared_compress_layer
273
+ self.layerwise_sharing = True
274
+ self.shared_kv_compressed = shared_kv_compressed
275
+
276
+ self.out_proj = quant_noise(
277
+ nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size)
278
+
279
+ if add_bias_kv:
280
+ self.bias_k = nn.Parameter(torch.Tensor(1, 1, embed_dim))
281
+ self.bias_v = nn.Parameter(torch.Tensor(1, 1, embed_dim))
282
+ else:
283
+ self.bias_k = self.bias_v = None
284
+
285
+ self.add_zero_attn = add_zero_attn
286
+
287
+ self.reset_parameters()
288
+
289
+ if freeze_compress == 1:
290
+ self.compress_k.weight.requires_grad = False
291
+ if shared_kv_compressed == 0:
292
+ self.compress_v.weight.requires_grad = False
293
+
294
+ self.onnx_trace = False
295
+ def reset_parameters(self):
296
+ if self.qkv_same_dim:
297
+ # Empirically observed the convergence to be much better with
298
+ # the scaled initialization
299
+ nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))
300
+ nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))
301
+ nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))
302
+ if (
303
+ not self.layerwise_sharing
304
+ ): # otherwise, we already initialize the parameters
305
+ nn.init.xavier_uniform_(self.compress_k.weight, gain=1 / math.sqrt(2))
306
+ if self.shared_kv_compressed == 0:
307
+ nn.init.xavier_uniform_(
308
+ self.compress_v.weight, gain=1 / math.sqrt(2)
309
+ )
310
+ else:
311
+ nn.init.xavier_uniform_(self.k_proj.weight)
312
+ nn.init.xavier_uniform_(self.v_proj.weight)
313
+ nn.init.xavier_uniform_(self.q_proj.weight)
314
+ if (
315
+ not self.layerwise_sharing
316
+ ): # otherwise, we already initialize the parameters
317
+ nn.init.xavier_uniform_(self.compress_k.weight)
318
+ if self.shared_kv_compressed == 0:
319
+ nn.init.xavier_uniform_(self.compress_v.weight)
320
+
321
+ nn.init.xavier_uniform_(self.out_proj.weight)
322
+ if self.out_proj.bias is not None:
323
+ nn.init.constant_(self.out_proj.bias, 0.0)
324
+ if self.bias_k is not None:
325
+ nn.init.xavier_normal_(self.bias_k)
326
+ if self.bias_v is not None:
327
+ nn.init.xavier_normal_(self.bias_v)
328
+
329
+ def prepare_for_onnx_export_(self):
330
+ self.onnx_trace = True
331
+
332
+ def forward(
333
+ self,
334
+ query,
335
+ key: Optional[torch.Tensor],
336
+ value: Optional[torch.Tensor],
337
+ key_padding_mask: Optional[torch.Tensor] = None,
338
+ incremental_state: Optional[Dict[str, Dict[str, Optional[torch.Tensor]]]] = None,
339
+ output_attentions: bool = True,
340
+ need_weights: bool = True,
341
+ static_kv: bool = False,
342
+ attn_mask: Optional[torch.Tensor] = None,
343
+ before_softmax: bool = False,
344
+ need_head_weights: bool = False,
345
+ past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
346
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
347
+ """Input shape: Time x Batch x Channel
348
+
349
+ Args:
350
+ key_padding_mask (ByteTensor, optional): mask to exclude
351
+ keys that are pads, of shape `(batch, src_len)`, where
352
+ padding elements are indicated by 1s.
353
+ need_weights (bool, optional): return the attention weights,
354
+ averaged over heads (default: False).
355
+ attn_mask (ByteTensor, optional): typically used to
356
+ implement causal attention, where the mask prevents the
357
+ attention from looking forward in time (default: None).
358
+ before_softmax (bool, optional): return the raw attention
359
+ weights and values before the attention softmax.
360
+ need_head_weights (bool, optional): return the attention
361
+ weights for each head. Implies *need_weights*. Default:
362
+ return the average attention weights over all heads.
363
+ """
364
+
365
+ if need_head_weights:
366
+ need_weights = True
367
+
368
+ tgt_len, bsz, embed_dim = query.size()
369
+ assert embed_dim == self.embed_dim
370
+ assert list(query.size()) == [tgt_len, bsz, embed_dim]
371
+
372
+ if incremental_state is not None:
373
+ saved_state = self._get_input_buffer(incremental_state)
374
+ if saved_state is not None and "prev_key" in saved_state:
375
+ # previous time steps are cached - no need to recompute
376
+ # key and value if they are static
377
+ if static_kv:
378
+ assert self.encoder_decoder_attention and not self.self_attention
379
+ key = value = None
380
+ else:
381
+ saved_state = None
382
+
383
+ if self.self_attention:
384
+ q = self.q_proj(query)
385
+
386
+ k_input = query.permute(1, 2, 0).contiguous() # B * C * T
387
+ k_input = (
388
+ F.linear(k_input, self.compress_k.weight[:, 0:tgt_len])
389
+ .permute(2, 0, 1)
390
+ .contiguous()
391
+ )
392
+ k = self.k_proj(k_input)
393
+
394
+ v_input = query.permute(1, 2, 0).contiguous() # B * C * T
395
+ if self.shared_kv_compressed == 0:
396
+ v_input = (
397
+ F.linear(v_input, self.compress_v.weight[:, 0:tgt_len])
398
+ .permute(2, 0, 1)
399
+ .contiguous()
400
+ )
401
+ if self.shared_kv_compressed == 1: # use shared kv compressed linear layer
402
+ v_input = (
403
+ F.linear(v_input, self.compress_k.weight[:, 0:tgt_len])
404
+ .permute(2, 0, 1)
405
+ .contiguous()
406
+ )
407
+ v = self.v_proj(v_input)
408
+ elif self.encoder_decoder_attention:
409
+ # encoder-decoder attention
410
+ q = self.q_proj(query)
411
+ if key is None:
412
+ assert value is None
413
+ k = v = None
414
+ else:
415
+ k = self.k_proj(key)
416
+ v = self.v_proj(key)
417
+
418
+ else:
419
+ assert key is not None and value is not None
420
+ q = self.q_proj(query)
421
+ k = self.k_proj(key)
422
+ v = self.v_proj(value)
423
+ q *= self.scaling
424
+
425
+ if self.bias_k is not None:
426
+ assert self.bias_v is not None
427
+ k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
428
+ v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
429
+ if attn_mask is not None:
430
+ attn_mask = torch.cat(
431
+ [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
432
+ )
433
+ if key_padding_mask is not None:
434
+ key_padding_mask = torch.cat(
435
+ [
436
+ key_padding_mask,
437
+ key_padding_mask.new_zeros(key_padding_mask.size(0), 1),
438
+ ],
439
+ dim=1,
440
+ )
441
+
442
+ q = (
443
+ q.contiguous()
444
+ .view(tgt_len, bsz * self.num_heads, self.head_dim)
445
+ .transpose(0, 1)
446
+ )
447
+ if k is not None:
448
+ k = (
449
+ k.contiguous()
450
+ .view(-1, bsz * self.num_heads, self.head_dim)
451
+ .transpose(0, 1)
452
+ )
453
+ if v is not None:
454
+ v = (
455
+ v.contiguous()
456
+ .view(-1, bsz * self.num_heads, self.head_dim)
457
+ .transpose(0, 1)
458
+ )
459
+
460
+ if saved_state is not None:
461
+ # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)
462
+ if "prev_key" in saved_state:
463
+ _prev_key = saved_state["prev_key"]
464
+ assert _prev_key is not None
465
+ prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim)
466
+ if static_kv:
467
+ k = prev_key
468
+ else:
469
+ assert k is not None
470
+ k = torch.cat([prev_key, k], dim=1)
471
+ if "prev_value" in saved_state:
472
+ _prev_value = saved_state["prev_value"]
473
+ assert _prev_value is not None
474
+ prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim)
475
+ if static_kv:
476
+ v = prev_value
477
+ else:
478
+ assert v is not None
479
+ v = torch.cat([prev_value, v], dim=1)
480
+ prev_key_padding_mask: Optional[torch.Tensor] = None
481
+ if "prev_key_padding_mask" in saved_state:
482
+ prev_key_padding_mask = saved_state["prev_key_padding_mask"]
483
+ assert k is not None and v is not None
484
+ key_padding_mask = MultiheadLinearAttention._append_prev_key_padding_mask(
485
+ key_padding_mask=key_padding_mask,
486
+ prev_key_padding_mask=prev_key_padding_mask,
487
+ batch_size=bsz,
488
+ src_len=k.size(1),
489
+ static_kv=static_kv,
490
+ )
491
+
492
+ saved_state["prev_key"] = k.view(bsz, self.num_heads, -1, self.head_dim)
493
+ saved_state["prev_value"] = v.view(bsz, self.num_heads, -1, self.head_dim)
494
+ saved_state["prev_key_padding_mask"] = key_padding_mask
495
+ # In this branch incremental_state is never None
496
+ assert incremental_state is not None
497
+ incremental_state = self._set_input_buffer(incremental_state, saved_state)
498
+ assert k is not None
499
+ src_len = k.size(1)
500
+
501
+ if self.add_zero_attn:
502
+ assert v is not None
503
+ src_len += 1
504
+ k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)
505
+ v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)
506
+ if attn_mask is not None:
507
+ attn_mask = torch.cat(
508
+ [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
509
+ )
510
+
511
+ attn_weights = torch.bmm(q, k.transpose(1, 2))
512
+ attn_weights = MultiheadLinearAttention.apply_sparse_mask(
513
+ attn_weights, tgt_len, src_len, bsz
514
+ )
515
+
516
+ assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]
517
+
518
+ if attn_mask is not None:
519
+ attn_mask = attn_mask.unsqueeze(0)
520
+ if self.onnx_trace:
521
+ attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1)
522
+ attn_weights += attn_mask
523
+
524
+ if before_softmax:
525
+ return attn_weights, v
526
+
527
+ attn_weights_float = utils.softmax(
528
+ attn_weights, dim=-1, onnx_trace=self.onnx_trace
529
+ )
530
+ attn_weights = attn_weights_float.type_as(attn_weights)
531
+ attn_probs = F.dropout(
532
+ attn_weights,
533
+ p=self.dropout,
534
+ training=self.training,
535
+ )
536
+ assert v is not None
537
+ attn = torch.bmm(attn_probs, v)
538
+ assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
539
+ if self.onnx_trace and attn.size(1) == 1:
540
+ # when ONNX tracing a single decoder step (sequence length == 1)
541
+ # the transpose is a no-op copy before view, thus unnecessary
542
+ attn = attn.contiguous().view(tgt_len, bsz, embed_dim)
543
+ else:
544
+ attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
545
+ attn = self.out_proj(attn)
546
+ attn_weights: Optional[torch.Tensor] = None
547
+ if output_attentions:
548
+ attn_weights = attn_weights_float.view(
549
+ bsz, self.num_heads, tgt_len, src_len
550
+ ).transpose(1, 0)
551
+ if not need_head_weights:
552
+ # average attention weights over heads
553
+ attn_weights = attn_weights.mean(dim=0)
554
+
555
+
556
+ return attn, attn_weights
557
+
558
+ @staticmethod
559
+ def _append_prev_key_padding_mask(
560
+ key_padding_mask: Optional[torch.Tensor],
561
+ prev_key_padding_mask: Optional[torch.Tensor],
562
+ batch_size: int,
563
+ src_len: int,
564
+ static_kv: bool,
565
+ ) -> Optional[torch.Tensor]:
566
+ # saved key padding masks have shape (bsz, seq_len)
567
+ if prev_key_padding_mask is not None and static_kv:
568
+ new_key_padding_mask = prev_key_padding_mask
569
+ elif prev_key_padding_mask is not None and key_padding_mask is not None:
570
+ new_key_padding_mask = torch.cat(
571
+ [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1
572
+ )
573
+ # During incremental decoding, as the padding token enters and
574
+ # leaves the frame, there will be a time when prev or current
575
+ # is None
576
+ elif prev_key_padding_mask is not None:
577
+ filler = torch.zeros(
578
+ (batch_size, src_len - prev_key_padding_mask.size(1)),
579
+ device=prev_key_padding_mask.device,
580
+ )
581
+ new_key_padding_mask = torch.cat(
582
+ [prev_key_padding_mask.float(), filler.float()], dim=1
583
+ )
584
+ elif key_padding_mask is not None:
585
+ filler = torch.zeros(
586
+ (batch_size, src_len - key_padding_mask.size(1)),
587
+ device=key_padding_mask.device,
588
+ )
589
+ new_key_padding_mask = torch.cat(
590
+ [filler.float(), key_padding_mask.float()], dim=1
591
+ )
592
+ else:
593
+ new_key_padding_mask = prev_key_padding_mask
594
+ return new_key_padding_mask
595
+
596
+ @torch.jit.export
597
+ def reorder_incremental_state(
598
+ self,
599
+ incremental_state: Dict[str, Dict[str, Optional[torch.Tensor]]],
600
+ new_order: torch.Tensor,
601
+ ):
602
+ """Reorder buffered internal state (for incremental generation)."""
603
+ input_buffer = self._get_input_buffer(incremental_state)
604
+ if input_buffer is not None:
605
+ for k in input_buffer.keys():
606
+ input_buffer_k = input_buffer[k]
607
+ if input_buffer_k is not None:
608
+ if self.encoder_decoder_attention and input_buffer_k.size(
609
+ 0
610
+ ) == new_order.size(0):
611
+ break
612
+ input_buffer[k] = input_buffer_k.index_select(0, new_order)
613
+ incremental_state = self._set_input_buffer(incremental_state, input_buffer)
614
+ return incremental_state
615
+
616
+ def _get_input_buffer(
617
+ self, incremental_state: Optional[Dict[str, Dict[str, Optional[torch.Tensor]]]]
618
+ ) -> Dict[str, Optional[torch.Tensor]]:
619
+ result = self.get_incremental_state(incremental_state, "attn_state")
620
+ if result is not None:
621
+ return result
622
+ else:
623
+ empty_result: Dict[str, Optional[torch.Tensor]] = {}
624
+ return empty_result
625
+
626
+ def _set_input_buffer(
627
+ self,
628
+ incremental_state: Dict[str, Dict[str, Optional[torch.Tensor]]],
629
+ buffer: Dict[str, Optional[torch.Tensor]],
630
+ ):
631
+ return self.set_incremental_state(incremental_state, "attn_state", buffer)
632
+
633
+ def apply_sparse_mask(attn_weights, tgt_len: int, src_len: int, bsz: int):
634
+ return attn_weights
635
+
636
+ def upgrade_state_dict_named(self, state_dict, name):
637
+ prefix = name + "." if name != "" else ""
638
+ items_to_add = {}
639
+ keys_to_remove = []
640
+ for k in state_dict.keys():
641
+ if k.endswith(prefix + "in_proj_weight"):
642
+ # in_proj_weight used to be q + k + v with same dimensions
643
+ dim = int(state_dict[k].shape[0] / 3)
644
+ items_to_add[prefix + "q_proj.weight"] = state_dict[k][:dim]
645
+ items_to_add[prefix + "k_proj.weight"] = state_dict[k][dim : 2 * dim]
646
+ items_to_add[prefix + "v_proj.weight"] = state_dict[k][2 * dim :]
647
+
648
+ keys_to_remove.append(k)
649
+
650
+ k_bias = prefix + "in_proj_bias"
651
+ if k_bias in state_dict.keys():
652
+ dim = int(state_dict[k].shape[0] / 3)
653
+ items_to_add[prefix + "q_proj.bias"] = state_dict[k_bias][:dim]
654
+ items_to_add[prefix + "k_proj.bias"] = state_dict[k_bias][
655
+ dim : 2 * dim
656
+ ]
657
+ items_to_add[prefix + "v_proj.bias"] = state_dict[k_bias][2 * dim :]
658
+
659
+ keys_to_remove.append(prefix + "in_proj_bias")
660
+
661
+ for k in keys_to_remove:
662
+ del state_dict[k]
663
+
664
+ for key, value in items_to_add.items():
665
+ state_dict[key] = value
666
+
667
+
668
+
669
+ def apply_chunking_to_forward(
670
+ forward_fn: Callable[..., torch.Tensor], chunk_size: int, chunk_dim: int, *input_tensors
671
+ ) -> torch.Tensor:
672
+ """
673
+ This function chunks the `input_tensors` into smaller input tensor parts of size `chunk_size` over the dimension
674
+ `chunk_dim`. It then applies a layer `forward_fn` to each chunk independently to save memory.
675
+
676
+ If the `forward_fn` is independent across the `chunk_dim` this function will yield the same result as directly
677
+ applying `forward_fn` to `input_tensors`.
678
+
679
+ Args:
680
+ forward_fn (`Callable[..., torch.Tensor]`):
681
+ The forward function of the model.
682
+ chunk_size (`int`):
683
+ The chunk size of a chunked tensor: `num_chunks = len(input_tensors[0]) / chunk_size`.
684
+ chunk_dim (`int`):
685
+ The dimension over which the `input_tensors` should be chunked.
686
+ input_tensors (`Tuple[torch.Tensor]`):
687
+ The input tensors of `forward_fn` which will be chunked
688
+
689
+ Returns:
690
+ `torch.Tensor`: A tensor with the same shape as the `forward_fn` would have given if applied`.
691
+
692
+
693
+ Examples:
694
+
695
+ ```python
696
+ # rename the usual forward() fn to forward_chunk()
697
+ def forward_chunk(self, hidden_states):
698
+ hidden_states = self.decoder(hidden_states)
699
+ return hidden_states
700
+
701
+
702
+ # implement a chunked forward function
703
+ def forward(self, hidden_states):
704
+ return apply_chunking_to_forward(self.forward_chunk, self.chunk_size_lm_head, self.seq_len_dim, hidden_states)
705
+ ```"""
706
+ assert len(input_tensors) > 0, f"{input_tensors} has to be a tuple/list of tensors"
707
+
708
+ # inspect.signature exist since python 3.5 and is a python method -> no problem with backward compatibility
709
+ num_args_in_forward_chunk_fn = len(inspect.signature(forward_fn).parameters)
710
+ if num_args_in_forward_chunk_fn != len(input_tensors):
711
+ raise ValueError(
712
+ f"forward_chunk_fn expects {num_args_in_forward_chunk_fn} arguments, but only {len(input_tensors)} input "
713
+ "tensors are given"
714
+ )
715
+
716
+ if chunk_size > 0:
717
+ tensor_shape = input_tensors[0].shape[chunk_dim]
718
+ for input_tensor in input_tensors:
719
+ if input_tensor.shape[chunk_dim] != tensor_shape:
720
+ raise ValueError(
721
+ f"All input tenors have to be of the same shape: {tensor_shape}, "
722
+ f"found shape {input_tensor.shape[chunk_dim]}"
723
+ )
724
+
725
+ if input_tensors[0].shape[chunk_dim] % chunk_size != 0:
726
+ raise ValueError(
727
+ f"The dimension to be chunked {input_tensors[0].shape[chunk_dim]} has to be a multiple of the chunk "
728
+ f"size {chunk_size}"
729
+ )
730
+
731
+ num_chunks = input_tensors[0].shape[chunk_dim] // chunk_size
732
+
733
+ # chunk input tensor into tuples
734
+ input_tensors_chunks = tuple(input_tensor.chunk(num_chunks, dim=chunk_dim) for input_tensor in input_tensors)
735
+ # apply forward fn to every tuple
736
+ output_chunks = tuple(forward_fn(*input_tensors_chunk) for input_tensors_chunk in zip(*input_tensors_chunks))
737
+ # concatenate output at same dimension
738
+ return torch.cat(output_chunks, dim=chunk_dim)
739
+
740
+ return forward_fn(*input_tensors)