Spaces:
Runtime error
Runtime error
DeepLearning101
commited on
Commit
•
e95b4e9
1
Parent(s):
08f4077
Upload 3 files
Browse files
models/multiple_choice/duma.py
ADDED
@@ -0,0 +1,355 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
# @Time : 2022/4/12 12:12 下午
|
3 |
+
# @Author : JianingWang
|
4 |
+
# @File : duma.py
|
5 |
+
import math
|
6 |
+
|
7 |
+
import torch
|
8 |
+
from torch import nn
|
9 |
+
from torch.nn import CrossEntropyLoss
|
10 |
+
|
11 |
+
from transformers.models.bert.modeling_bert import BertModel, BertPreTrainedModel
|
12 |
+
from transformers.models.roberta.modeling_roberta import RobertaModel, RobertaPreTrainedModel
|
13 |
+
from transformers.models.albert.modeling_albert import AlbertModel, AlbertPreTrainedModel
|
14 |
+
from transformers.models.megatron_bert.modeling_megatron_bert import MegatronBertModel, MegatronBertPreTrainedModel
|
15 |
+
from transformers.modeling_outputs import MultipleChoiceModelOutput
|
16 |
+
|
17 |
+
|
18 |
+
def split_context_query(sequence_output, pq_end_pos, input_ids):
|
19 |
+
context_max_len = sequence_output.size(1)
|
20 |
+
query_max_len = sequence_output.size(1)
|
21 |
+
sep_tok_len = 1 # [SEP]
|
22 |
+
context_sequence_output = sequence_output.new(
|
23 |
+
torch.Size((sequence_output.size(0), context_max_len, sequence_output.size(2)))).zero_()
|
24 |
+
query_sequence_output = sequence_output.new_zeros(
|
25 |
+
(sequence_output.size(0), query_max_len, sequence_output.size(2)))
|
26 |
+
query_attention_mask = sequence_output.new_zeros((sequence_output.size(0), query_max_len))
|
27 |
+
context_attention_mask = sequence_output.new_zeros((sequence_output.size(0), context_max_len))
|
28 |
+
for i in range(0, sequence_output.size(0)):
|
29 |
+
p_end = pq_end_pos[i][0]
|
30 |
+
q_end = pq_end_pos[i][1]
|
31 |
+
context_sequence_output[i, :min(context_max_len, p_end)] = sequence_output[i, 1: 1 + min(context_max_len, p_end)]
|
32 |
+
idx = min(query_max_len, q_end - p_end - sep_tok_len)
|
33 |
+
query_sequence_output[i, :idx] = sequence_output[i, p_end + sep_tok_len + 1: p_end + sep_tok_len + 1 + min(q_end - p_end - sep_tok_len, query_max_len)]
|
34 |
+
query_attention_mask[i, :idx] = sequence_output.new_ones((1, query_max_len))[0, :idx]
|
35 |
+
context_attention_mask[i, : min(context_max_len, p_end)] = sequence_output.new_ones((1, context_max_len))[0, : min(context_max_len, p_end)]
|
36 |
+
return context_sequence_output, query_sequence_output, context_attention_mask, query_attention_mask
|
37 |
+
|
38 |
+
|
39 |
+
class BertCoAttention(nn.Module):
|
40 |
+
def __init__(self, config):
|
41 |
+
super(BertCoAttention, self).__init__()
|
42 |
+
if config.hidden_size % config.num_attention_heads != 0:
|
43 |
+
raise ValueError(
|
44 |
+
"The hidden size (%d) is not a multiple of the number of attention "
|
45 |
+
"heads (%d)" % (config.hidden_size, config.num_attention_heads))
|
46 |
+
self.output_attentions = config.output_attentions
|
47 |
+
|
48 |
+
self.num_attention_heads = config.num_attention_heads
|
49 |
+
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
|
50 |
+
self.all_head_size = self.num_attention_heads * self.attention_head_size
|
51 |
+
|
52 |
+
self.query = nn.Linear(config.hidden_size, self.all_head_size)
|
53 |
+
self.key = nn.Linear(config.hidden_size, self.all_head_size)
|
54 |
+
self.value = nn.Linear(config.hidden_size, self.all_head_size)
|
55 |
+
|
56 |
+
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
|
57 |
+
|
58 |
+
def transpose_for_scores(self, x):
|
59 |
+
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
|
60 |
+
x = x.view(*new_x_shape)
|
61 |
+
return x.permute(0, 2, 1, 3)
|
62 |
+
|
63 |
+
def forward(self, context_states, query_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
|
64 |
+
mixed_query_layer = self.query(query_states)
|
65 |
+
|
66 |
+
extended_attention_mask = attention_mask[:, None, None, :]
|
67 |
+
# extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
|
68 |
+
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
|
69 |
+
attention_mask = extended_attention_mask
|
70 |
+
|
71 |
+
# If this is instantiated as a cross-attention module, the keys
|
72 |
+
# and values come from an encoder; the attention mask needs to be
|
73 |
+
# such that the encoder"s padding tokens are not attended to.
|
74 |
+
if encoder_hidden_states is not None:
|
75 |
+
mixed_key_layer = self.key(encoder_hidden_states)
|
76 |
+
mixed_value_layer = self.value(encoder_hidden_states)
|
77 |
+
attention_mask = encoder_attention_mask
|
78 |
+
else:
|
79 |
+
mixed_key_layer = self.key(context_states)
|
80 |
+
mixed_value_layer = self.value(context_states)
|
81 |
+
|
82 |
+
query_layer = self.transpose_for_scores(mixed_query_layer)
|
83 |
+
key_layer = self.transpose_for_scores(mixed_key_layer)
|
84 |
+
value_layer = self.transpose_for_scores(mixed_value_layer)
|
85 |
+
|
86 |
+
# Take the dot product between "query" and "key" to get the raw attention scores.
|
87 |
+
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
|
88 |
+
attention_scores = attention_scores / math.sqrt(self.attention_head_size)
|
89 |
+
if attention_mask is not None:
|
90 |
+
# Apply the attention mask is (precomputed for all layers in BertModel forward() function)
|
91 |
+
attention_scores = attention_scores + attention_mask
|
92 |
+
|
93 |
+
# Normalize the attention scores to probabilities.
|
94 |
+
attention_probs = nn.Softmax(dim=-1)(attention_scores)
|
95 |
+
|
96 |
+
# This is actually dropping out entire tokens to attend to, which might
|
97 |
+
# seem a bit unusual, but is taken from the original Transformer paper.
|
98 |
+
attention_probs = self.dropout(attention_probs)
|
99 |
+
|
100 |
+
# Mask heads if we want to
|
101 |
+
if head_mask is not None:
|
102 |
+
attention_probs = attention_probs * head_mask
|
103 |
+
|
104 |
+
context_layer = torch.matmul(attention_probs, value_layer)
|
105 |
+
|
106 |
+
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
|
107 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
|
108 |
+
context_layer = context_layer.view(*new_context_layer_shape)
|
109 |
+
|
110 |
+
# outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)
|
111 |
+
outputs = context_layer
|
112 |
+
return outputs
|
113 |
+
|
114 |
+
|
115 |
+
class BertDUMAForMultipleChoice(BertPreTrainedModel):
|
116 |
+
|
117 |
+
def __init__(self, config):
|
118 |
+
super(BertDUMAForMultipleChoice, self).__init__(config)
|
119 |
+
|
120 |
+
self.bert = BertModel(config)
|
121 |
+
self.classifier_2 = nn.Linear(2 * config.hidden_size, 1)
|
122 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
123 |
+
self.bert_att = BertCoAttention(config)
|
124 |
+
|
125 |
+
self.init_weights()
|
126 |
+
|
127 |
+
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
|
128 |
+
inputs_embeds=None, labels=None, pq_end_pos=None, iter=1):
|
129 |
+
num_choices = input_ids.shape[1]
|
130 |
+
|
131 |
+
flat_input_ids = input_ids.view(-1, input_ids.size(-1))
|
132 |
+
flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
|
133 |
+
flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
134 |
+
flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
135 |
+
flat_head_mask = head_mask.view(-1, head_mask.size(-1)) if head_mask is not None else None
|
136 |
+
flat_inputs_embeds = inputs_embeds.view(-1, inputs_embeds.size(-1)) if inputs_embeds is not None else None
|
137 |
+
|
138 |
+
outputs = self.bert(
|
139 |
+
input_ids=flat_input_ids,
|
140 |
+
attention_mask=flat_attention_mask,
|
141 |
+
token_type_ids=flat_token_type_ids,
|
142 |
+
position_ids=flat_position_ids,
|
143 |
+
head_mask=flat_head_mask,
|
144 |
+
inputs_embeds=flat_inputs_embeds
|
145 |
+
)
|
146 |
+
|
147 |
+
sequence_output = outputs[0]
|
148 |
+
|
149 |
+
pq_end_pos = pq_end_pos.view(-1, pq_end_pos.size(-1))
|
150 |
+
|
151 |
+
context_sequence_output, query_sequence_output, context_attention_mask, query_attention_mask = \
|
152 |
+
split_context_query(sequence_output, pq_end_pos, input_ids)
|
153 |
+
for _ in range(0, iter):
|
154 |
+
cq_biatt_output = self.bert_att(context_sequence_output, query_sequence_output, context_attention_mask)
|
155 |
+
qc_biatt_output = self.bert_att(query_sequence_output, context_sequence_output, query_attention_mask)
|
156 |
+
|
157 |
+
query_sequence_output = cq_biatt_output
|
158 |
+
context_sequence_output = qc_biatt_output
|
159 |
+
|
160 |
+
cat_output = torch.cat([torch.mean(qc_biatt_output, 1), torch.mean(cq_biatt_output, 1)], 1)
|
161 |
+
pooled_output = self.dropout(cat_output)
|
162 |
+
logits = self.classifier_2(pooled_output)
|
163 |
+
|
164 |
+
reshaped_logits = logits.view(-1, num_choices)
|
165 |
+
|
166 |
+
outputs = (reshaped_logits,) + outputs[2:] # add hidden states and attention if they are here
|
167 |
+
|
168 |
+
if labels is not None:
|
169 |
+
loss_fct = CrossEntropyLoss()
|
170 |
+
loss = loss_fct(reshaped_logits, labels)
|
171 |
+
outputs = (loss,) + outputs
|
172 |
+
|
173 |
+
return outputs # (loss), reshaped_logits, (hidden_states), (attentions)
|
174 |
+
|
175 |
+
|
176 |
+
class RobertaDUMAForMultipleChoice(RobertaPreTrainedModel):
|
177 |
+
|
178 |
+
def __init__(self, config):
|
179 |
+
super(RobertaDUMAForMultipleChoice, self).__init__(config)
|
180 |
+
|
181 |
+
self.roberta = RobertaModel(config)
|
182 |
+
self.classifier_2 = nn.Linear(2 * config.hidden_size, 1)
|
183 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
184 |
+
self.bert_att = BertCoAttention(config)
|
185 |
+
|
186 |
+
self.init_weights()
|
187 |
+
|
188 |
+
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
|
189 |
+
inputs_embeds=None, labels=None, pq_end_pos=None, iter=1):
|
190 |
+
num_choices = input_ids.shape[1]
|
191 |
+
|
192 |
+
flat_input_ids = input_ids.view(-1, input_ids.size(-1))
|
193 |
+
flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
|
194 |
+
flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
195 |
+
flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
196 |
+
flat_head_mask = head_mask.view(-1, head_mask.size(-1)) if head_mask is not None else None
|
197 |
+
flat_inputs_embeds = inputs_embeds.view(-1, inputs_embeds.size(-1)) if inputs_embeds is not None else None
|
198 |
+
|
199 |
+
outputs = self.roberta(
|
200 |
+
input_ids=flat_input_ids,
|
201 |
+
attention_mask=flat_attention_mask,
|
202 |
+
token_type_ids=flat_token_type_ids,
|
203 |
+
position_ids=flat_position_ids,
|
204 |
+
head_mask=flat_head_mask,
|
205 |
+
inputs_embeds=flat_inputs_embeds
|
206 |
+
)
|
207 |
+
|
208 |
+
sequence_output = outputs[0]
|
209 |
+
|
210 |
+
pq_end_pos = pq_end_pos.view(-1, pq_end_pos.size(-1))
|
211 |
+
|
212 |
+
context_sequence_output, query_sequence_output, context_attention_mask, query_attention_mask = \
|
213 |
+
split_context_query(sequence_output, pq_end_pos, input_ids)
|
214 |
+
for _ in range(0, iter):
|
215 |
+
cq_biatt_output = self.bert_att(context_sequence_output, query_sequence_output, context_attention_mask)
|
216 |
+
qc_biatt_output = self.bert_att(query_sequence_output, context_sequence_output, query_attention_mask)
|
217 |
+
|
218 |
+
query_sequence_output = cq_biatt_output
|
219 |
+
context_sequence_output = qc_biatt_output
|
220 |
+
|
221 |
+
cat_output = torch.cat([torch.mean(qc_biatt_output, 1), torch.mean(cq_biatt_output, 1)], 1)
|
222 |
+
pooled_output = self.dropout(cat_output)
|
223 |
+
logits = self.classifier_2(pooled_output)
|
224 |
+
|
225 |
+
reshaped_logits = logits.view(-1, num_choices)
|
226 |
+
|
227 |
+
outputs = (reshaped_logits,) + outputs[2:] # add hidden states and attention if they are here
|
228 |
+
|
229 |
+
if labels is not None:
|
230 |
+
loss_fct = CrossEntropyLoss()
|
231 |
+
loss = loss_fct(reshaped_logits, labels)
|
232 |
+
outputs = (loss,) + outputs
|
233 |
+
|
234 |
+
return outputs # (loss), reshaped_logits, (hidden_states), (attentions)
|
235 |
+
|
236 |
+
class AlbertDUMAForMultipleChoice(AlbertPreTrainedModel):
|
237 |
+
|
238 |
+
def __init__(self, config):
|
239 |
+
super(AlbertDUMAForMultipleChoice, self).__init__(config)
|
240 |
+
|
241 |
+
self.albert = AlbertModel(config)
|
242 |
+
self.classifier_2 = nn.Linear(2 * config.hidden_size, 1)
|
243 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
244 |
+
self.bert_att = BertCoAttention(config)
|
245 |
+
|
246 |
+
self.init_weights()
|
247 |
+
|
248 |
+
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
|
249 |
+
inputs_embeds=None, labels=None, pq_end_pos=None, iter=1):
|
250 |
+
num_choices = input_ids.shape[1]
|
251 |
+
|
252 |
+
flat_input_ids = input_ids.view(-1, input_ids.size(-1))
|
253 |
+
flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
|
254 |
+
flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
255 |
+
flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
256 |
+
flat_head_mask = head_mask.view(-1, head_mask.size(-1)) if head_mask is not None else None
|
257 |
+
flat_inputs_embeds = inputs_embeds.view(-1, inputs_embeds.size(-1)) if inputs_embeds is not None else None
|
258 |
+
|
259 |
+
outputs = self.albert(
|
260 |
+
input_ids=flat_input_ids,
|
261 |
+
attention_mask=flat_attention_mask,
|
262 |
+
token_type_ids=flat_token_type_ids,
|
263 |
+
position_ids=flat_position_ids,
|
264 |
+
head_mask=flat_head_mask,
|
265 |
+
inputs_embeds=flat_inputs_embeds
|
266 |
+
)
|
267 |
+
|
268 |
+
sequence_output = outputs[0]
|
269 |
+
|
270 |
+
pq_end_pos = pq_end_pos.view(-1, pq_end_pos.size(-1))
|
271 |
+
|
272 |
+
context_sequence_output, query_sequence_output, context_attention_mask, query_attention_mask = \
|
273 |
+
split_context_query(sequence_output, pq_end_pos, input_ids)
|
274 |
+
for _ in range(0, iter):
|
275 |
+
cq_biatt_output = self.bert_att(context_sequence_output, query_sequence_output, context_attention_mask)
|
276 |
+
qc_biatt_output = self.bert_att(query_sequence_output, context_sequence_output, query_attention_mask)
|
277 |
+
|
278 |
+
query_sequence_output = cq_biatt_output
|
279 |
+
context_sequence_output = qc_biatt_output
|
280 |
+
|
281 |
+
cat_output = torch.cat([torch.mean(qc_biatt_output, 1), torch.mean(cq_biatt_output, 1)], 1)
|
282 |
+
pooled_output = self.dropout(cat_output)
|
283 |
+
logits = self.classifier_2(pooled_output)
|
284 |
+
|
285 |
+
reshaped_logits = logits.view(-1, num_choices)
|
286 |
+
|
287 |
+
outputs = (reshaped_logits,) + outputs[2:] # add hidden states and attention if they are here
|
288 |
+
|
289 |
+
if labels is not None:
|
290 |
+
loss_fct = CrossEntropyLoss()
|
291 |
+
loss = loss_fct(reshaped_logits, labels)
|
292 |
+
outputs = (loss,) + outputs
|
293 |
+
|
294 |
+
return outputs # (loss), reshaped_logits, (hidden_states), (attentions)
|
295 |
+
|
296 |
+
|
297 |
+
class MegatronDumaForMultipleChoice(MegatronBertPreTrainedModel):
|
298 |
+
|
299 |
+
def __init__(self, config):
|
300 |
+
super(MegatronDumaForMultipleChoice, self).__init__(config)
|
301 |
+
|
302 |
+
self.bert = MegatronBertModel(config)
|
303 |
+
self.classifier_2 = nn.Linear(2 * config.hidden_size, 1)
|
304 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
305 |
+
self.bert_att = BertCoAttention(config)
|
306 |
+
|
307 |
+
self.init_weights()
|
308 |
+
|
309 |
+
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
|
310 |
+
inputs_embeds=None, labels=None, pq_end_pos=None, iter=1):
|
311 |
+
num_choices = input_ids.shape[1]
|
312 |
+
|
313 |
+
flat_input_ids = input_ids.view(-1, input_ids.size(-1))
|
314 |
+
flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
|
315 |
+
flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
316 |
+
flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
317 |
+
flat_head_mask = head_mask.view(-1, head_mask.size(-1)) if head_mask is not None else None
|
318 |
+
flat_inputs_embeds = inputs_embeds.view(-1, inputs_embeds.size(-1)) if inputs_embeds is not None else None
|
319 |
+
|
320 |
+
outputs = self.bert(
|
321 |
+
input_ids=flat_input_ids,
|
322 |
+
attention_mask=flat_attention_mask,
|
323 |
+
token_type_ids=flat_token_type_ids,
|
324 |
+
position_ids=flat_position_ids,
|
325 |
+
head_mask=flat_head_mask,
|
326 |
+
inputs_embeds=flat_inputs_embeds
|
327 |
+
)
|
328 |
+
|
329 |
+
sequence_output = outputs[0]
|
330 |
+
|
331 |
+
pq_end_pos = pq_end_pos.view(-1, pq_end_pos.size(-1))
|
332 |
+
|
333 |
+
context_sequence_output, query_sequence_output, context_attention_mask, query_attention_mask = \
|
334 |
+
split_context_query(sequence_output, pq_end_pos, input_ids)
|
335 |
+
for _ in range(0, iter):
|
336 |
+
cq_biatt_output = self.bert_att(context_sequence_output, query_sequence_output, context_attention_mask)
|
337 |
+
qc_biatt_output = self.bert_att(query_sequence_output, context_sequence_output, query_attention_mask)
|
338 |
+
|
339 |
+
query_sequence_output = cq_biatt_output
|
340 |
+
context_sequence_output = qc_biatt_output
|
341 |
+
|
342 |
+
cat_output = torch.cat([torch.mean(qc_biatt_output, 1), torch.mean(cq_biatt_output, 1)], 1)
|
343 |
+
pooled_output = self.dropout(cat_output)
|
344 |
+
logits = self.classifier_2(pooled_output)
|
345 |
+
|
346 |
+
reshaped_logits = logits.view(-1, num_choices)
|
347 |
+
|
348 |
+
outputs = (reshaped_logits,) + outputs[2:] # add hidden states and attention if they are here
|
349 |
+
|
350 |
+
if labels is not None:
|
351 |
+
loss_fct = CrossEntropyLoss()
|
352 |
+
loss = loss_fct(reshaped_logits, labels)
|
353 |
+
outputs = (loss,) + outputs
|
354 |
+
|
355 |
+
return outputs # (loss), reshaped_logits, (hidden_states), (attentions)
|
models/multiple_choice/multiple_choice.py
ADDED
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
# @Time : 2022/4/16 12:10 下午
|
3 |
+
# @Author : JianingWang
|
4 |
+
# @File : multiple_choice.py
|
5 |
+
import torch
|
6 |
+
from torch import nn
|
7 |
+
from torch.nn import CrossEntropyLoss
|
8 |
+
import torch.nn.functional as F
|
9 |
+
# from transformers import MegatronBertPreTrainedModel, MegatronBertModel
|
10 |
+
from transformers.models.megatron_bert import MegatronBertPreTrainedModel, MegatronBertModel
|
11 |
+
from transformers.modeling_outputs import MultipleChoiceModelOutput
|
12 |
+
|
13 |
+
|
14 |
+
class MegatronBertForMultipleChoice(MegatronBertPreTrainedModel):
|
15 |
+
def __init__(self, config):
|
16 |
+
super().__init__(config)
|
17 |
+
|
18 |
+
self.bert = MegatronBertModel(config)
|
19 |
+
# classifier_dropout = (
|
20 |
+
# config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
|
21 |
+
# )
|
22 |
+
classifier_dropout = 0.2
|
23 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
24 |
+
self.classifier = nn.Linear(config.hidden_size, 1)
|
25 |
+
|
26 |
+
# Initialize weights and apply final processing
|
27 |
+
self.post_init()
|
28 |
+
|
29 |
+
def forward(
|
30 |
+
self,
|
31 |
+
input_ids=None,
|
32 |
+
attention_mask=None,
|
33 |
+
token_type_ids=None,
|
34 |
+
position_ids=None,
|
35 |
+
head_mask=None,
|
36 |
+
inputs_embeds=None,
|
37 |
+
labels=None,
|
38 |
+
output_attentions=None,
|
39 |
+
output_hidden_states=None,
|
40 |
+
return_dict=None,
|
41 |
+
pseudo=None
|
42 |
+
):
|
43 |
+
r"""
|
44 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
45 |
+
Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
|
46 |
+
num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
|
47 |
+
`input_ids` above)
|
48 |
+
"""
|
49 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
50 |
+
num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
|
51 |
+
|
52 |
+
input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
|
53 |
+
attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
54 |
+
token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
55 |
+
position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
|
56 |
+
inputs_embeds = (
|
57 |
+
inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
|
58 |
+
if inputs_embeds is not None
|
59 |
+
else None
|
60 |
+
)
|
61 |
+
|
62 |
+
outputs = self.bert(
|
63 |
+
input_ids,
|
64 |
+
attention_mask=attention_mask,
|
65 |
+
token_type_ids=token_type_ids,
|
66 |
+
position_ids=position_ids,
|
67 |
+
head_mask=head_mask,
|
68 |
+
inputs_embeds=inputs_embeds,
|
69 |
+
output_attentions=output_attentions,
|
70 |
+
output_hidden_states=output_hidden_states,
|
71 |
+
return_dict=return_dict,
|
72 |
+
)
|
73 |
+
|
74 |
+
pooled_output = outputs[1] # [batch_size, num_choices, hidden_dim]
|
75 |
+
|
76 |
+
pooled_output = self.dropout(pooled_output)
|
77 |
+
logits = self.classifier(pooled_output) # [batch_size, num_choices, 1]
|
78 |
+
reshaped_logits = logits.view(-1, num_choices) # [batch_size, num_choices]
|
79 |
+
|
80 |
+
loss = None
|
81 |
+
if labels is not None:
|
82 |
+
if pseudo is None:
|
83 |
+
loss_fct = CrossEntropyLoss()
|
84 |
+
loss = loss_fct(reshaped_logits, labels)
|
85 |
+
else:
|
86 |
+
loss_fct = CrossEntropyLoss(reduction="none")
|
87 |
+
loss = loss_fct(reshaped_logits, labels)
|
88 |
+
weight = 1 - pseudo * 0.9
|
89 |
+
loss *= weight
|
90 |
+
loss = loss.mean()
|
91 |
+
|
92 |
+
if not return_dict:
|
93 |
+
output = (reshaped_logits,) + outputs[2:]
|
94 |
+
return ((loss,) + output) if loss is not None else output
|
95 |
+
|
96 |
+
return MultipleChoiceModelOutput(
|
97 |
+
loss=loss,
|
98 |
+
logits=reshaped_logits,
|
99 |
+
hidden_states=outputs.hidden_states,
|
100 |
+
attentions=outputs.attentions,
|
101 |
+
)
|
102 |
+
|
103 |
+
|
104 |
+
class MegatronBertRDropForMultipleChoice(MegatronBertPreTrainedModel):
|
105 |
+
def __init__(self, config):
|
106 |
+
super().__init__(config)
|
107 |
+
|
108 |
+
self.bert = MegatronBertModel(config)
|
109 |
+
# classifier_dropout = (
|
110 |
+
# config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
|
111 |
+
# )
|
112 |
+
classifier_dropout = 0.2
|
113 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
114 |
+
self.classifier = nn.Linear(config.hidden_size, 1)
|
115 |
+
|
116 |
+
# Initialize weights and apply final processing
|
117 |
+
self.post_init()
|
118 |
+
|
119 |
+
def forward(
|
120 |
+
self,
|
121 |
+
input_ids=None,
|
122 |
+
attention_mask=None,
|
123 |
+
token_type_ids=None,
|
124 |
+
position_ids=None,
|
125 |
+
head_mask=None,
|
126 |
+
inputs_embeds=None,
|
127 |
+
labels=None,
|
128 |
+
output_attentions=None,
|
129 |
+
output_hidden_states=None,
|
130 |
+
return_dict=None,
|
131 |
+
|
132 |
+
):
|
133 |
+
r"""
|
134 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
135 |
+
Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
|
136 |
+
num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
|
137 |
+
`input_ids` above)
|
138 |
+
"""
|
139 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
140 |
+
num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
|
141 |
+
|
142 |
+
input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
|
143 |
+
attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
144 |
+
token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
145 |
+
position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
|
146 |
+
inputs_embeds = (
|
147 |
+
inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
|
148 |
+
if inputs_embeds is not None
|
149 |
+
else None
|
150 |
+
)
|
151 |
+
|
152 |
+
logits_list = []
|
153 |
+
for i in range(2):
|
154 |
+
outputs = self.bert(
|
155 |
+
input_ids,
|
156 |
+
attention_mask=attention_mask,
|
157 |
+
token_type_ids=token_type_ids,
|
158 |
+
position_ids=position_ids,
|
159 |
+
head_mask=head_mask,
|
160 |
+
inputs_embeds=inputs_embeds,
|
161 |
+
output_attentions=output_attentions,
|
162 |
+
output_hidden_states=output_hidden_states,
|
163 |
+
return_dict=return_dict,
|
164 |
+
)
|
165 |
+
pooled_output = outputs[1]
|
166 |
+
pooled_output = self.dropout(pooled_output)
|
167 |
+
logits = self.classifier(pooled_output)
|
168 |
+
logits_list.append(logits.view(-1, num_choices))
|
169 |
+
|
170 |
+
loss = None
|
171 |
+
alpha = 1.0
|
172 |
+
for logits in logits_list:
|
173 |
+
if labels is not None:
|
174 |
+
loss_fct = CrossEntropyLoss()
|
175 |
+
l = loss_fct(logits, labels)
|
176 |
+
if loss:
|
177 |
+
loss += alpha * l
|
178 |
+
else:
|
179 |
+
loss = alpha * l
|
180 |
+
|
181 |
+
if loss is not None:
|
182 |
+
p = torch.log_softmax(logits_list[0], dim=-1)
|
183 |
+
p_tec = torch.exp(p)
|
184 |
+
q = torch.log_softmax(logits_list[-1], dim=-1)
|
185 |
+
q_tec = torch.exp(q)
|
186 |
+
|
187 |
+
kl_loss = F.kl_div(p, q_tec, reduction="none").sum()
|
188 |
+
reverse_kl_loss = F.kl_div(q, p_tec, reduction="none").sum()
|
189 |
+
loss += 0.5 * (kl_loss + reverse_kl_loss) / 2.
|
190 |
+
|
191 |
+
return MultipleChoiceModelOutput(
|
192 |
+
loss=loss,
|
193 |
+
logits=logits_list[0],
|
194 |
+
hidden_states=None,
|
195 |
+
attentions=None
|
196 |
+
)
|
models/multiple_choice/multiple_choice_tag.py
ADDED
@@ -0,0 +1,271 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
# @Time : 2022/3/3 7:59 下午
|
3 |
+
# @Author : JianingWang
|
4 |
+
# @File : multiple_choice.py
|
5 |
+
import torch
|
6 |
+
from roformer import RoFormerPreTrainedModel, RoFormerModel
|
7 |
+
from torch import nn
|
8 |
+
from torch.nn import CrossEntropyLoss
|
9 |
+
|
10 |
+
from transformers import MegatronBertPreTrainedModel, MegatronBertModel
|
11 |
+
from transformers.modeling_outputs import MultipleChoiceModelOutput
|
12 |
+
from transformers.models.bert import BertPreTrainedModel, BertModel
|
13 |
+
|
14 |
+
|
15 |
+
class BertForTagMultipleChoice(BertPreTrainedModel):
|
16 |
+
def __init__(self, config):
|
17 |
+
super().__init__(config)
|
18 |
+
|
19 |
+
self.bert = BertModel(config)
|
20 |
+
classifier_dropout = (
|
21 |
+
config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
|
22 |
+
)
|
23 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
24 |
+
self.classifier = nn.Linear(config.hidden_size * 2, 1)
|
25 |
+
|
26 |
+
# Initialize weights and apply final processing
|
27 |
+
self.post_init()
|
28 |
+
|
29 |
+
def forward(
|
30 |
+
self,
|
31 |
+
input_ids=None,
|
32 |
+
attention_mask=None,
|
33 |
+
token_type_ids=None,
|
34 |
+
position_ids=None,
|
35 |
+
head_mask=None,
|
36 |
+
inputs_embeds=None,
|
37 |
+
labels=None,
|
38 |
+
output_attentions=None,
|
39 |
+
output_hidden_states=None,
|
40 |
+
return_dict=None,
|
41 |
+
pseudo=None
|
42 |
+
):
|
43 |
+
r"""
|
44 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
45 |
+
Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
|
46 |
+
num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
|
47 |
+
`input_ids` above)
|
48 |
+
"""
|
49 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
50 |
+
num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
|
51 |
+
|
52 |
+
input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
|
53 |
+
attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
54 |
+
token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
55 |
+
position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
|
56 |
+
inputs_embeds = (
|
57 |
+
inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
|
58 |
+
if inputs_embeds is not None
|
59 |
+
else None
|
60 |
+
)
|
61 |
+
|
62 |
+
outputs = self.bert(
|
63 |
+
input_ids,
|
64 |
+
attention_mask=attention_mask,
|
65 |
+
token_type_ids=token_type_ids,
|
66 |
+
position_ids=position_ids,
|
67 |
+
head_mask=head_mask,
|
68 |
+
inputs_embeds=inputs_embeds,
|
69 |
+
output_attentions=output_attentions,
|
70 |
+
output_hidden_states=output_hidden_states,
|
71 |
+
return_dict=return_dict,
|
72 |
+
)
|
73 |
+
|
74 |
+
w = torch.logical_and(input_ids >= min(self.config.start_token_ids), input_ids <= max(self.config.start_token_ids))
|
75 |
+
start_index = w.nonzero()[:, 1].view(-1, 2)
|
76 |
+
# <start_entity> + <end_entity> 进分类
|
77 |
+
pooled_output = torch.cat([torch.cat([x[y[0], :], x[y[1], :]]).unsqueeze(0) for x, y in zip(outputs.last_hidden_state, start_index)])
|
78 |
+
|
79 |
+
pooled_output = self.dropout(pooled_output)
|
80 |
+
logits = self.classifier(pooled_output)
|
81 |
+
reshaped_logits = logits.view(-1, num_choices)
|
82 |
+
|
83 |
+
loss = None
|
84 |
+
if labels is not None:
|
85 |
+
if pseudo is None:
|
86 |
+
loss_fct = CrossEntropyLoss()
|
87 |
+
loss = loss_fct(reshaped_logits, labels)
|
88 |
+
else:
|
89 |
+
loss_fct = CrossEntropyLoss(reduction="none")
|
90 |
+
loss = loss_fct(reshaped_logits, labels)
|
91 |
+
weight = 1 - pseudo * 0.9
|
92 |
+
loss *= weight
|
93 |
+
loss = loss.mean()
|
94 |
+
|
95 |
+
if not return_dict:
|
96 |
+
output = (reshaped_logits,) + outputs[2:]
|
97 |
+
return ((loss,) + output) if loss is not None else output
|
98 |
+
|
99 |
+
return MultipleChoiceModelOutput(
|
100 |
+
loss=loss,
|
101 |
+
logits=reshaped_logits,
|
102 |
+
hidden_states=outputs.hidden_states,
|
103 |
+
attentions=outputs.attentions,
|
104 |
+
)
|
105 |
+
|
106 |
+
|
107 |
+
class RoFormerForTagMultipleChoice(RoFormerPreTrainedModel):
|
108 |
+
def __init__(self, config):
|
109 |
+
super().__init__(config)
|
110 |
+
|
111 |
+
self.roformer = RoFormerModel(config, add_pooling_layer=False)
|
112 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
113 |
+
self.classifier = nn.Linear(config.hidden_size * 2, 1)
|
114 |
+
|
115 |
+
# Initialize weights and apply final processing
|
116 |
+
self.post_init()
|
117 |
+
|
118 |
+
def forward(
|
119 |
+
self,
|
120 |
+
input_ids=None,
|
121 |
+
attention_mask=None,
|
122 |
+
token_type_ids=None,
|
123 |
+
head_mask=None,
|
124 |
+
inputs_embeds=None,
|
125 |
+
labels=None,
|
126 |
+
output_attentions=None,
|
127 |
+
output_hidden_states=None,
|
128 |
+
return_dict=None,
|
129 |
+
):
|
130 |
+
r"""
|
131 |
+
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
|
132 |
+
Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
|
133 |
+
num_choices-1]`` where :obj:`num_choices` is the size of the second dimension of the input tensors. (See
|
134 |
+
:obj:`input_ids` above)
|
135 |
+
"""
|
136 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
137 |
+
num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
|
138 |
+
|
139 |
+
input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
|
140 |
+
attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
141 |
+
token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
142 |
+
|
143 |
+
inputs_embeds = (
|
144 |
+
inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
|
145 |
+
if inputs_embeds is not None
|
146 |
+
else None
|
147 |
+
)
|
148 |
+
|
149 |
+
outputs = self.roformer(
|
150 |
+
input_ids,
|
151 |
+
attention_mask=attention_mask,
|
152 |
+
token_type_ids=token_type_ids,
|
153 |
+
head_mask=head_mask,
|
154 |
+
inputs_embeds=inputs_embeds,
|
155 |
+
output_attentions=output_attentions,
|
156 |
+
output_hidden_states=output_hidden_states,
|
157 |
+
return_dict=return_dict,
|
158 |
+
)
|
159 |
+
|
160 |
+
w = torch.logical_and(input_ids >= min(self.config.start_token_ids), input_ids <= max(self.config.start_token_ids))
|
161 |
+
start_index = w.nonzero()[:, 1].view(-1, 2)
|
162 |
+
# <start_entity> + <end_entity> 进分类
|
163 |
+
pooled_output = torch.cat([torch.cat([x[y[0], :], x[y[1], :]]).unsqueeze(0) for x, y in zip(outputs.last_hidden_state, start_index)])
|
164 |
+
|
165 |
+
pooled_output = self.dropout(pooled_output)
|
166 |
+
logits = self.classifier(pooled_output)
|
167 |
+
reshaped_logits = logits.view(-1, num_choices)
|
168 |
+
|
169 |
+
loss = None
|
170 |
+
if labels is not None:
|
171 |
+
loss_fct = CrossEntropyLoss()
|
172 |
+
loss = loss_fct(reshaped_logits, labels)
|
173 |
+
|
174 |
+
if not return_dict:
|
175 |
+
output = (reshaped_logits,) + outputs[2:]
|
176 |
+
return ((loss,) + output) if loss is not None else output
|
177 |
+
|
178 |
+
return MultipleChoiceModelOutput(
|
179 |
+
loss=loss,
|
180 |
+
logits=reshaped_logits,
|
181 |
+
hidden_states=outputs.hidden_states,
|
182 |
+
attentions=outputs.attentions,
|
183 |
+
)
|
184 |
+
|
185 |
+
|
186 |
+
class MegatronBertForTagMultipleChoice(MegatronBertPreTrainedModel):
|
187 |
+
def __init__(self, config):
|
188 |
+
super().__init__(config)
|
189 |
+
|
190 |
+
self.bert = MegatronBertModel(config)
|
191 |
+
self.dropout = nn.Dropout(0.2)
|
192 |
+
self.classifier = nn.Linear(config.hidden_size * 2, 1)
|
193 |
+
|
194 |
+
# Initialize weights and apply final processing
|
195 |
+
self.post_init()
|
196 |
+
|
197 |
+
def forward(
|
198 |
+
self,
|
199 |
+
input_ids=None,
|
200 |
+
attention_mask=None,
|
201 |
+
token_type_ids=None,
|
202 |
+
position_ids=None,
|
203 |
+
head_mask=None,
|
204 |
+
inputs_embeds=None,
|
205 |
+
labels=None,
|
206 |
+
output_attentions=None,
|
207 |
+
output_hidden_states=None,
|
208 |
+
return_dict=None,
|
209 |
+
pseudo=None
|
210 |
+
):
|
211 |
+
r"""
|
212 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
213 |
+
Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
|
214 |
+
num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
|
215 |
+
`input_ids` above)
|
216 |
+
"""
|
217 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
218 |
+
num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
|
219 |
+
input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
|
220 |
+
attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
221 |
+
token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
222 |
+
position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
|
223 |
+
inputs_embeds = (
|
224 |
+
inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
|
225 |
+
if inputs_embeds is not None
|
226 |
+
else None
|
227 |
+
)
|
228 |
+
|
229 |
+
outputs = self.bert(
|
230 |
+
input_ids,
|
231 |
+
attention_mask=attention_mask,
|
232 |
+
token_type_ids=token_type_ids,
|
233 |
+
position_ids=position_ids,
|
234 |
+
head_mask=head_mask,
|
235 |
+
inputs_embeds=inputs_embeds,
|
236 |
+
output_attentions=output_attentions,
|
237 |
+
output_hidden_states=output_hidden_states,
|
238 |
+
return_dict=return_dict,
|
239 |
+
)
|
240 |
+
|
241 |
+
w = torch.logical_and(input_ids >= min(self.config.start_token_ids), input_ids <= max(self.config.start_token_ids))
|
242 |
+
start_index = w.nonzero()[:, 1].view(-1, 2)
|
243 |
+
# <start_entity> + <end_entity> 进分类
|
244 |
+
pooled_output = torch.cat([torch.cat([x[y[0], :], x[y[1], :]]).unsqueeze(0) for x, y in zip(outputs.last_hidden_state, start_index)])
|
245 |
+
|
246 |
+
pooled_output = self.dropout(pooled_output)
|
247 |
+
logits = self.classifier(pooled_output)
|
248 |
+
reshaped_logits = logits.view(-1, num_choices)
|
249 |
+
|
250 |
+
loss = None
|
251 |
+
if labels is not None:
|
252 |
+
if pseudo is None:
|
253 |
+
loss_fct = CrossEntropyLoss()
|
254 |
+
loss = loss_fct(reshaped_logits, labels)
|
255 |
+
else:
|
256 |
+
loss_fct = CrossEntropyLoss(reduction="none")
|
257 |
+
loss = loss_fct(reshaped_logits, labels)
|
258 |
+
weight = 1 - pseudo*0.9
|
259 |
+
loss *= weight
|
260 |
+
loss = loss.mean()
|
261 |
+
|
262 |
+
if not return_dict:
|
263 |
+
output = (reshaped_logits,) + outputs[2:]
|
264 |
+
return ((loss,) + output) if loss is not None else output
|
265 |
+
|
266 |
+
return MultipleChoiceModelOutput(
|
267 |
+
loss=loss,
|
268 |
+
logits=reshaped_logits,
|
269 |
+
hidden_states=outputs.hidden_states,
|
270 |
+
attentions=outputs.attentions,
|
271 |
+
)
|