-
Notifications
You must be signed in to change notification settings - Fork 0
/
foward_bert.py
259 lines (201 loc) · 13.4 KB
/
foward_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
from torch.nn.functional import gelu, elu
import torch.nn.functional as F
import torch.nn as nn
import math
import torch
from collections import OrderedDict
def functional_bert(fast_weights, config, input_ids=None, attention_mask=None, token_type_ids=None,
position_ids=None, head_mask=None, inputs_embeds=None, encoder_hidden_states=None,
encoder_attention_mask=None, is_train = True):
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
input_shape = input_ids.size()
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
device = input_ids.device if input_ids is not None else inputs_embeds.device
if attention_mask is None:
attention_mask = torch.ones(input_shape, device=device)
if token_type_ids is None:
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
if attention_mask.dim() == 3:
extended_attention_mask = attention_mask[:, None, :, :]
elif attention_mask.dim() == 2:
if config.is_decoder:
batch_size, seq_length = input_shape
seq_ids = torch.arange(seq_length, device=device)
causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
causal_mask = causal_mask.to(torch.long)
extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
else:
extended_attention_mask = attention_mask[:, None, None, :]
else:
raise ValueError("Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(input_shape, attention_mask.shape))
extended_attention_mask = extended_attention_mask.to(dtype=next((p for p in fast_weights.values())).dtype) # fp16 compatibility
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
if config.is_decoder and encoder_hidden_states is not None:
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
if encoder_attention_mask is None:
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
if encoder_attention_mask.dim() == 3:
encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
elif encoder_attention_mask.dim() == 2:
encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
else:
raise ValueError("Wrong shape for encoder_hidden_shape (shape {}) or encoder_attention_mask (shape {})".format(encoder_hidden_shape,
encoder_attention_mask.shape))
encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next((p for p in fast_weights.values())).dtype)
encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
else:
encoder_extended_attention_mask = None
if head_mask is not None:
if head_mask.dim() == 1:
head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
head_mask = head_mask.expand(config.num_hidden_layers, -1, -1, -1, -1)
elif head_mask.dim() == 2:
head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)
head_mask = head_mask.to(dtype=next((p for p in fast_weights.values())).dtype)
else:
head_mask = [None] * config.num_hidden_layers
embedding_output = functional_embeeding(fast_weights, config, input_ids, position_ids,
token_type_ids, inputs_embeds, is_train = is_train)
encoder_outputs = functional_encoder(fast_weights, config, embedding_output,
attention_mask=extended_attention_mask,
head_mask=head_mask, encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_extended_attention_mask, is_train = is_train)
sequence_output = encoder_outputs
outputs = (sequence_output,)
return outputs
def functional_embeeding(fast_weights, config, input_ids, position_ids,
token_type_ids, inputs_embeds = None, is_train = True):
if input_ids is not None:
input_shape = input_ids.size()
else:
input_shape = inputs_embeds.size()[:-1]
seq_length = input_shape[1]
device = input_ids.device if input_ids is not None else inputs_embeds.device
if position_ids is None:
position_ids = torch.arange(seq_length, dtype=torch.long, device=device)
position_ids = position_ids.unsqueeze(0).expand(input_shape)
if token_type_ids is None:
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
if inputs_embeds is None:
inputs_embeds = F.embedding(input_ids, fast_weights['bert.embeddings.word_embeddings.weight'], padding_idx = 0)
position_embeddings = F.embedding(position_ids, fast_weights['bert.embeddings.position_embeddings.weight'])
token_type_embeddings = F.embedding(token_type_ids, fast_weights['bert.embeddings.token_type_embeddings.weight'])
embeddings = inputs_embeds + position_embeddings + token_type_embeddings
embeddings = F.layer_norm(embeddings, [config.hidden_size],
weight=fast_weights['bert.embeddings.LayerNorm.weight'],
bias=fast_weights['bert.embeddings.LayerNorm.bias'],
eps=config.layer_norm_eps)
embeddings = F.dropout(embeddings, p=config.hidden_dropout_prob, training = is_train)
return embeddings
def transpose_for_scores(config, x):
new_x_shape = x.size()[:-1] + (config.num_attention_heads, int(config.hidden_size / config.num_attention_heads))
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def functional_self_attention(fast_weights, config, layer_idx,
hidden_states, attention_mask, head_mask,
encoder_hidden_states, encoder_attention_mask,
is_train = True):
attention_head_size = int(config.hidden_size / config.num_attention_heads)
all_head_size = config.num_attention_heads * attention_head_size
mixed_query_layer = F.linear(hidden_states,
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.query.weight'],
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.query.bias'])
if encoder_hidden_states is not None:
mixed_key_layer = F.linear(encoder_hidden_states,
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.key.weight'],
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.key.bias'])
mixed_value_layer = F.linear(encoder_hidden_states,
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.value.weight'],
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.value.bias'])
attention_mask = encoder_attention_mask
else:
mixed_key_layer = F.linear(hidden_states,
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.key.weight'],
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.key.bias'])
mixed_value_layer = F.linear(hidden_states,
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.value.weight'],
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.self.value.bias'])
query_layer = transpose_for_scores(config, mixed_query_layer)
key_layer = transpose_for_scores(config, mixed_key_layer)
value_layer = transpose_for_scores(config, mixed_value_layer)
# Take the dot product between "query" and "key" to get the raw attention scores.
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
attention_scores = attention_scores / math.sqrt(attention_head_size)
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in BertModel forward() function)
attention_scores = attention_scores + attention_mask
attention_probs = torch.nn.Softmax(dim=-1)(attention_scores)
if is_train:
attention_probs = F.dropout(attention_probs, p= config.attention_probs_dropout_prob)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (all_head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = context_layer
return outputs
def functional_out_attention(fast_weights, config, layer_idx,
hidden_states, input_tensor,
is_train = True):
hidden_states = F.linear(hidden_states,
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.output.dense.weight'],
fast_weights['bert.encoder.layer.'+layer_idx+'.attention.output.dense.bias'])
hidden_states = F.dropout(hidden_states, p=config.hidden_dropout_prob, training = is_train)
hidden_states = F.layer_norm(hidden_states + input_tensor, [config.hidden_size],
weight=fast_weights['bert.encoder.layer.'+layer_idx+'.attention.output.LayerNorm.weight'],
bias=fast_weights['bert.encoder.layer.'+layer_idx+'.attention.output.LayerNorm.bias'],
eps=config.layer_norm_eps)
return hidden_states
def functional_attention(fast_weights, config, layer_idx,
hidden_states, attention_mask=None, head_mask=None,
encoder_hidden_states=None, encoder_attention_mask=None,
is_train = True):
self_outputs = functional_self_attention(fast_weights, config, layer_idx,
hidden_states, attention_mask, head_mask,
encoder_hidden_states, encoder_attention_mask, is_train)
attention_output = functional_out_attention(fast_weights, config, layer_idx,
self_outputs, hidden_states, is_train)
return attention_output
def functional_intermediate(fast_weights, config, layer_idx, hidden_states, is_train = True):
weight_name = 'bert.encoder.layer.' + layer_idx + '.intermediate.dense.weight'
bias_name = 'bert.encoder.layer.' + layer_idx + '.intermediate.dense.bias'
hidden_states = F.linear(hidden_states, fast_weights[weight_name], fast_weights[bias_name])
hidden_states = gelu(hidden_states)
return hidden_states
def functional_output(fast_weights, config, layer_idx, hidden_states, input_tensor, is_train = True):
hidden_states = F.linear(hidden_states,
fast_weights['bert.encoder.layer.'+layer_idx+'.output.dense.weight'],
fast_weights['bert.encoder.layer.'+layer_idx+'.output.dense.bias'])
hidden_states = F.dropout(hidden_states, p=config.hidden_dropout_prob, training = is_train)
hidden_states = F.layer_norm(hidden_states + input_tensor, [config.hidden_size],
weight=fast_weights['bert.encoder.layer.'+layer_idx+'.output.LayerNorm.weight'],
bias=fast_weights['bert.encoder.layer.'+layer_idx+'.output.LayerNorm.bias'],
eps=config.layer_norm_eps)
return hidden_states
def functional_layer(fast_weights, config, layer_idx, hidden_states, attention_mask,
head_mask, encoder_hidden_states, encoder_attention_mask, is_train = True):
self_attention_outputs = functional_attention(fast_weights, config, layer_idx,
hidden_states, attention_mask, head_mask,
encoder_hidden_states, encoder_attention_mask,is_train)
attention_output = self_attention_outputs
intermediate_output = functional_intermediate(fast_weights, config, layer_idx, attention_output, is_train)
layer_output = functional_output(fast_weights, config, layer_idx,
intermediate_output, attention_output, is_train)
return layer_output
def functional_encoder(fast_weights, config , hidden_states, attention_mask,
head_mask, encoder_hidden_states, encoder_attention_mask, is_train = True):
for i in range(0,config.num_hidden_layers):
layer_outputs = functional_layer(fast_weights, config, str(i),
hidden_states, attention_mask, head_mask[i],
encoder_hidden_states, encoder_attention_mask, is_train)
hidden_states = layer_outputs
outputs = hidden_states
return outputs