-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
272 lines (234 loc) · 11.5 KB
/
model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
import torch
import torch.nn as nn
from transformers import BertModel, BertTokenizer
import ipdb
from config import args, device
import torch.nn.functional as F
import random
from config import args
class REModel(nn.Module):
def __init__(self, hidden_size=768):
super(REModel, self).__init__()
tokenizer = BertTokenizer.from_pretrained(args.model)
tokenizer.add_tokens(["[s1]", "[s2]"])
self.bert = BertModel.from_pretrained(args.model)
self.bert.resize_token_embeddings(len(tokenizer))
self.proj_relation = nn.Linear(2 * hidden_size, 37)
self.proj_trigger = nn.Linear(hidden_size, 2)
self.proj_binary = nn.Linear(hidden_size, 2)
def reset(self, class_cnt, hidden_size=768):
self.proj_relation = nn.Linear(2 * hidden_size, class_cnt)
def forward(self, inputs):
# Phase 1: feed inputs to self.bert and extract the hidden states
# Phase 2-1: keep the context part unmasked, and apply start & end prediction
# Phase 2-2: concatenate hid_trigger, x, and y,
# then pass this concatenated tensor to self.proj_relation
last_hidden_states = self.bert(inputs['input_ids'], inputs['attention_mask'], inputs['token_type_ids'])["last_hidden_state"]
start_end_logit = self.proj_trigger(last_hidden_states) # shape: (batch_size, seq_len, 2)
ids = []
for x_idx in inputs['x_idx']:
ids.append([0, x_idx[0]-1])
masked_start_end_logit = self.get_masked(
start_end_logit,
torch.tensor(ids),
mask_val=float('-inf')
)
ids = self.get_triggers_ids(masked_start_end_logit)
trigger = self.attention(last_hidden_states, torch.tensor(ids))
x = []
for b_idx in range(len(last_hidden_states)):
x.append(last_hidden_states[b_idx][inputs['x_idx'][b_idx][1]+1, :])
x = torch.vstack(x)
concat_hid = torch.hstack((trigger, last_hidden_states[:, 0, :]))
relation_logit = self.proj_relation(concat_hid)
binary_logit = self.proj_binary(x)
return relation_logit, start_end_logit, binary_logit#, distributions#, tri_len_logit
def get_triggers_ids(self, masked_start_end_logit, tri_len=None):
ids = []
if tri_len is None:
for batch_idx, sample in enumerate(masked_start_end_logit):
start = sample[:,0] # start: shape (512)
end = sample[:,1]
start_candidates = torch.topk(start, k=30)
end_candidates = torch.topk(end, k=30)
ans_candidates = [(0, 1)]
scores = [-100]
start_logits = F.softmax(start_candidates[0])
end_logits = F.softmax(end_candidates[0])
for i, s in enumerate(start_candidates[1]):
for j, e in enumerate(end_candidates[1]):
if s == 0:
ans_candidates.append((s, s+1))
# scores.append(start_candidates[0][i]+end_candidates[0][j])
scores.append(start_logits[i] * end_logits[j])
if s<e and e-s <= 10:
ans_candidates.append((s, e))
# scores.append(start_candidates[0][i]+end_candidates[0][j])
scores.append(start_logits[i] * end_logits[j])
results = list(zip(scores, ans_candidates))
results.sort()
results.reverse()
ids.append([int(results[0][1][0]), int(results[0][1][1])])
return ids
else:
for batch_idx, sample in enumerate(masked_start_end_logit):
start = sample[:,0] # start: shape (512)
end = sample[:,1]
start_logits = F.softmax(start)
end_logits = F.softmax(end)
# start_candidates = torch.topk(start, k=30)
# end_candidates = torch.topk(end, k=30)
max_score = float('-inf')
cand = None
for i in range(len(start_logits)-tri_len[batch_idx]):
# for j in range(i+1, len(end_logits)):
# if j - i <= 14:
cur_score = start_logits[i] + end_logits[i+tri_len[batch_idx]]
if cur_score > max_score:
max_score = cur_score
cand = [i, i+tri_len[batch_idx]]
ids.append(cand)
return ids
def infer(self, inputs):
last_hidden_states = self.bert(inputs['input_ids'], inputs['attention_mask'], inputs['token_type_ids'])["last_hidden_state"]
start_end_logit = self.proj_trigger(last_hidden_states) # shape: (batch_size, seq_len, 2)
ids = []
for x_idx in inputs['x_idx']:
ids.append([0, x_idx[0]-1])
masked_start_end_logit = self.get_masked(
start_end_logit,
torch.tensor(ids),
mask_val=float('-inf')
)
# y = []
# for b_idx in range(len(last_hidden_states)):
# y.append(last_hidden_states[b_idx][inputs['x_idx'][b_idx][0]-1, :])
# y = torch.vstack(y)
# tri_len_logit = self.proj_tri_len(y)
# tri_len = torch.argmax(tri_len_logit, dim=1)
# ids = self.get_triggers_ids(masked_start_end_logit, tri_len)
ids = self.get_triggers_ids(masked_start_end_logit)
tokenizer = BertTokenizer.from_pretrained(args.model)
tokenizer.add_tokens(["[s1]", "[s2]"])
p_trigs, gt_trigs = [], []
for i in range(len(inputs['input_ids'])):
p_trig = tokenizer.decode(inputs['input_ids'][i][ids[i][0] : ids[i][1]])
p_trigs.append(p_trig)
gt_trig = tokenizer.decode(inputs['input_ids'][i]\
[inputs['t_idx'][i][0] : inputs['t_idx'][i][1]])
gt_trigs.append(gt_trig)
# trigger = self.attention(last_hidden_states, inputs['t_idx'])
trigger = self.attention(last_hidden_states, torch.tensor(ids))
# trigger, inv_lengths = self.get_trigger_and_lengths(last_hidden_states, torch.tensor(ids))
# trigger, inv_lengths = self.get_trigger_and_lengths(last_hidden_states, inputs['t_idx'])
# trigger = torch.mul(trigger.sum(dim=1), inv_lengths.to(device))
# if len(trigger.shape) == 2:
# trigger = trigger.unsqueeze(dim=1)
# cls_and_trig = torch.hstack((last_hidden_states[:, 0, :].unsqueeze(dim=1), trigger))
# concat_hid = torch.mul(
# cls_and_trig.sum(dim=1),
# (1 / (1 + (1 / inv_lengths))).to(device)
# )
# trigger = self.get_masked(last_hidden_states, torch.tensor(ids)).mean(dim=1)
# trigger = self.get_masked(last_hidden_states, inputs['t_idx']).mean(dim=1)
# _, trigger = self.rnn(
# self.get_trigger(last_hidden_states, torch.tensor(ids))
# )
# trigger = trigger.view(trigger.shape[1], -1)
# _, trigger = self.rnn(
# self.get_trigger(last_hidden_states, inputs['t_idx'])
# )
# trigger = trigger.view(trigger.shape[1], -1)
# x = self.get_masked(last_hidden_states, inputs['x_idx']).mean(dim=1)
# y = self.get_masked(last_hidden_states, inputs['y_idx']).mean(dim=1)
# concat_hid = torch.hstack((trigger, x, y))
# trigger = self.get_masked(last_hidden_states, torch.tensor(ids)).mean(dim=1)
# no_trig = [i for i in range(len(ids)) if ids[i][0] == 0]
# no_trig = [i for i in range(len(inputs['t_idx'])) if inputs['t_idx'][i][0] == 0]
# for i in no_trig:
# trigger[i, :] = self.uni_trigger[:]
x = []
for b_idx in range(len(last_hidden_states)):
x.append(last_hidden_states[b_idx][inputs['x_idx'][b_idx][1]+1, :])
x = torch.vstack(x)
# concat_hid = torch.hstack((trigger, x))
binary_logit = self.proj_binary(x)
bin_pred = torch.argmax(binary_logit, dim=1)
# cls = []
for batch_idx in range(len(bin_pred)):
if bin_pred[batch_idx] == 0:
# trigger[batch_idx][:] = last_hidden_states[batch_idx, 0, :]
# cls.append(last_hidden_states[batch_idx, 0, :])
# if inputs['has_trigger'][batch_idx] == 0:
trigger[batch_idx][:] = torch.zeros(len(last_hidden_states[batch_idx, 0, :]))
# else:
# cls.append(torch.zeros(len(last_hidden_states[batch_idx, 0, :]), device=device))
# cls = torch.vstack(cls)
# For bert baseline (only use bert and no other modules)
#for batch_idx in range(len(inputs['has_trigger'])):
# trigger[batch_idx][:] = torch.zeros(len(last_hidden_states[batch_idx, 0, :]))
concat_hid = torch.hstack((trigger, last_hidden_states[:, 0, :]))
# concat_hid = torch.hstack((trigger, cls))
# concat_hid = (trigger + cls) / 2
# concat_hid = (trigger + last_hidden_states[:, 0, :]) / 2
# relation_logit = self.proj_relation(
# torch.sigmoid(self.proj_relation0(concat_hid))
# )
relation_logit = self.proj_relation(concat_hid)
# return torch.argmax(relation_logit, dim=1), p_trigs, gt_trigs
argmax = torch.argmax(relation_logit, dim=1)
has_trigger = torch.argmax(binary_logit, dim=1)
return argmax, has_trigger, ids, p_trigs, gt_trigs
def get_masked(self, mat, ids, mask_val=0):
batch_size, seq_len, cls = mat.shape
mask = torch.ones(batch_size, seq_len, cls)
for i in range(batch_size):
mask[i, ids[i][0]:ids[i][1], :] = 0
mask = mask.bool()
return mat.masked_fill(mask.to(device), mask_val)
def get_trigger(self, mat, ids, mask_val=0, length=15):
batch_size, seq_len, cls = mat.shape
triggers = []
for b_id in range(batch_size):
# self.total += 1
trigger = mat[b_id][ids[b_id][0] : ids[b_id][1]][:]
if len(trigger) < length:
padding = torch.zeros(length - len(trigger), cls)
padding = padding.to(device)
trigger = torch.vstack((trigger, padding))
else:
# self.long_trig += 1
trigger = trigger[:length]
try:
assert len(trigger) == length
except:
ipdb.set_trace()
triggers.append(trigger)
return torch.vstack(triggers).view(batch_size, -1, cls)
def attention(self, mat, ids):
triggers = []
batch_size, _, _ = mat.shape
for b_id in range(batch_size):
trigger = mat[b_id][ids[b_id][0] : ids[b_id][1]][:]
score = []
cls = mat[b_id, 0, :]
# score.append(torch.dot(cls, cls))
for j in range(len(trigger)):
score.append(torch.dot(cls, trigger[j]))
score = torch.tensor(score, device=device)
score = F.softmax(score)
# trigger = torch.vstack((cls, trigger))
triggers.append(torch.matmul(trigger.T, score))
return torch.vstack(triggers)
def get_trigger_and_lengths(self, mat, ids, mask_val=0):
lengths = []
batch_size, seq_len, cls = mat.shape
mask = torch.ones(batch_size, seq_len, cls)
for i in range(batch_size):
mask[i, ids[i][0]:ids[i][1], :] = 0
mask = mask.bool()
for b_id in range(batch_size):
# self.total += 1
lengths.append(ids[b_id][1] - ids[b_id][0])
return mat.masked_fill(mask.to(device), mask_val), \
1 / torch.vstack(lengths)