-
Notifications
You must be signed in to change notification settings - Fork 0
/
syntax_lm.py
265 lines (234 loc) · 11.3 KB
/
syntax_lm.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
import math
import os
import warnings
from typing import List, Optional, Tuple, Union
import torch
import torch.utils.checkpoint
import torch.nn.functional as F
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers.models.roberta import RobertaPreTrainedModel, RobertaModel
from transformers.modeling_outputs import SequenceClassifierOutput
class RobertaClassificationHead(nn.Module):
"""Head for sentence-level classification tasks."""
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
classifier_dropout = (
config.classifier_dropout
if config.classifier_dropout is not None
else config.hidden_dropout_prob
)
self.dropout = nn.Dropout(classifier_dropout)
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
def forward(self, features, **kwargs):
# x = features[:, 0, :] # take <s> token (equiv. to [CLS])
x = features
x = self.dropout(x)
x = self.dense(x)
x = torch.tanh(x)
x = self.dropout(x)
x = self.out_proj(x)
return x
class SyntaxLMSequenceClassification(RobertaPreTrainedModel):
_keys_to_ignore_on_load_missing = [r"position_ids"]
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.config = config
# TreeLSTM structure
size = 768
self.size = size # size of input embeddings
# BiLSTM parameters run at the very beginning
# self.forward_bilstm = torch.nn.LSTMCell(size + 25, size, bias=False).to(self.device)
# self.backward_bilstm = torch.nn.LSTMCell(size + 25, size, bias=False).to(self.device)
# LSTM from root to leaves
self.cell_topdown = torch.nn.LSTMCell(size, size, bias=False)
# LSTM over children of the same parent
self.cell_bottomup = torch.nn.LSTMCell(2 * size, size, bias=False)
# "move up" parameters, which takes last vectors of the children sequence to the parent
self.move_up = nn.Linear(size, size, bias=False)
self.move_up_init = nn.Linear(size, size, bias=False)
self.act_move_up = nn.Tanh()
# final LSTM, over different tree embeddings of the same tweet
self.act_root = nn.ReLU()
self.root_to_sent = nn.Linear(size, size, bias=False)
self.sentence_lstm = torch.nn.LSTMCell(size, size, bias=False)
# self.word_emb = torch.cat((w_emb, pad), dim=0).to(self.device) --> from previous version, no more useful
# for POS tagging, 17 one-hot vectors [0,24] + a zero tensor, for padding value
self.one_hot_pos = torch.cat(
(
F.one_hot(torch.Tensor(range(17)).long(), num_classes=17),
torch.Tensor.zero_(torch.Tensor(1, 17)),
)
)
# final linear transformation, before softmax (our MLP)
self.roberta = RobertaModel(config, add_pooling_layer=False)
self.classifier = RobertaClassificationHead(config)
# Initialize weights and apply final processing
self.post_init()
def forward_syntax(self, batch_size, word, visit_order, parent_visit_order):
device = self.roberta.device
# initialize tensor to store final representation
syntax_vector = torch.Tensor.zero_(torch.Tensor(batch_size, self.size)).to(
device
)
# from [tweet, tweet_trees, node] to [trees, node], for each type of data
# tensor to access to different items
std = torch.Tensor(range(len(word))).long().to(device)
# TOP DOWN FILTERING
# resulting representation for each word from TOP-DOWN filtering
# 2 silly dimension, useful for PAD elements and "parent of the root" and parent of pad nodes.
# Same purpose in every tensor of embeddings
s_vect = torch.Tensor.zero_(
torch.Tensor(len(word), len(word[0]), self.size)
).to(device)
# tensor to save memory-cells in the RecNN for each word from TOP-DOWN filtering
c_vect = torch.Tensor.zero_(
torch.Tensor(len(word), len(word[0]), self.size)
).to(device)
# follow visit ordering fixed in the dataset, previous state is stored in the parent position (for both s and c)
for i in range(len(visit_order[0])):
vo = visit_order[:, i].long()
pvo = parent_visit_order[:, i].long()
s_vect[std, vo], c_vect[std, vo] = self.cell_topdown(
word[std, vo], (s_vect[std, pvo], c_vect[std, pvo])
)
# initialize vectors for BOTTOM-UP procedure
h_vect = torch.Tensor.zero_(
torch.Tensor(len(word), len(word[0]), self.size)
).to(device)
c_vect = torch.Tensor.zero_(
torch.Tensor(len(word), len(word[0]), self.size)
).to(device)
# first initialization: take vector from top-down procedure and pass all to the "move up" network: in this way,
# leaves are initialized
# probably not useful pad here, need to check
x_init_vect = self.act_move_up(self.move_up_init(s_vect.clone()))
x_vect = x_init_vect.clone()
# BOTTOM UP PROCEDURE
# visit in "reverse" order
for i in reversed(range(len(visit_order[0]) - 1)):
vo = visit_order[:, i + 1].long()
pvo = parent_visit_order[:, i + 1].long()
# take h and c from previous" child in children chain (store in father position), take x from the child in exam
h_vect[std, pvo], c_vect[std, pvo] = self.cell_bottomup(
torch.cat((s_vect[std, pvo], x_vect[std, vo]), dim=1),
(h_vect[std, pvo], c_vect[std, pvo]),
)
# update x every time
x_vect[std, pvo] = self.act_move_up(self.move_up(h_vect[std, pvo]))
# OUTPUT: x vector assigned to the #batch_size roots
syntax_vector = x_vect[std, visit_order[:, 0].long()]
# transform from [trees, vector] to [tweets, tweet_trees, vector]
# Final LSTM!
return syntax_vector
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
token_type_ids: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
map_tokbert_to_tokparse: Optional[torch.LongTensor] = None,
divisors: Optional[torch.FloatTensor] = None,
map_attention: Optional[torch.FloatTensor] = None,
pos: Optional[torch.LongTensor] = None,
visit_order: Optional[torch.LongTensor] = None,
parent_visit_order: Optional[torch.LongTensor] = None,
pad_mask_trees: Optional[torch.FloatTensor] = None,
) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
r"""
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = (
return_dict if return_dict is not None else self.config.use_return_dict
)
outputs = self.roberta(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
device = self.roberta.device
batch_size = input_ids.size()[0]
# number of trees for each tweet
# number of tokens for each tree
word_representation = torch.Tensor.zero_(
torch.Tensor(outputs.last_hidden_state.size())
).to(device)
std = torch.Tensor(range(outputs.last_hidden_state.size()[0])).long().to(device)
for idx in range(outputs.last_hidden_state.size()[1]):
word_representation[std, map_tokbert_to_tokparse[:, idx].long(), :] = (
word_representation[std, map_tokbert_to_tokparse[:, idx].long(), :]
+ outputs.last_hidden_state[:, idx, :]
)
batch_size = input_ids.size()[0]
n_tokens_bert = input_ids.size()[1]
# input_ids = input_ids.reshape(batch_size * n_trees, n_tokens_bert)
map_attention = map_attention.reshape(batch_size, n_tokens_bert, 1)
divisors = divisors.reshape(batch_size, n_tokens_bert, 1)
for idx in range(word_representation.size()[1]):
# print(idx)
word_representation[:, idx, :] = word_representation[
:, idx, :
] * map_attention[:, idx].expand(
map_attention[:, idx].size()[0],
word_representation[:, idx, :].size()[1],
)
word_representation[:, idx, :] = word_representation[:, idx, :] / divisors[
:, idx
].expand(
divisors[:, idx].size()[0], word_representation[:, idx, :].size()[1]
)
# extract tweet "syntactical" embedding and give it in input to final MLP
class_vector = self.forward_syntax(
batch_size, word_representation, visit_order, parent_visit_order
)
# sequence_output = outputs[0]
logits = self.classifier(class_vector)
loss = None
if labels is not None:
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1 and (
labels.dtype == torch.long or labels.dtype == torch.int
):
self.config.problem_type = "single_label_classification"
else:
self.config.problem_type = "multi_label_classification"
if self.config.problem_type == "regression":
loss_fct = MSELoss()
if self.num_labels == 1:
loss = loss_fct(logits.squeeze(), labels.squeeze())
else:
loss = loss_fct(logits, labels)
elif self.config.problem_type == "single_label_classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
elif self.config.problem_type == "multi_label_classification":
loss_fct = BCEWithLogitsLoss()
loss = loss_fct(logits, labels)
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)