-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
executable file
·192 lines (177 loc) · 9.47 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
import torch
import torch.nn as nn
from torch.autograd import Variable
import torchvision
import torchvision.transforms as transforms
from transformers import RobertaConfig, RobertaModel
from transformers import BigBirdConfig, BigBirdModel
from transformers import AutoConfig, AutoModel, AutoTokenizer
from all_loss_aug import transitivity_loss_H_, transitivity_loss_T_, cross_category_loss_, segment_loss_
from dpn_losses import DirichletKLLoss
import numpy as np
from opt_einsum import contract
'''
HiEve Stats
'''
HierPC_h = 1802.0
HierCP_h = 1846.0
HierCo_h = 758.0
HierNo_h = 63755.0
HierTo_h = HierPC_h + HierCP_h + HierCo_h + HierNo_h # total number of event pairs
hier_weights_h = [0.25*HierTo_h/HierPC_h, 0.25*HierTo_h/HierCP_h, 0.25*HierTo_h/HierCo_h, 0.25*HierTo_h/HierNo_h]
'''
IC Stats
'''
HierPC_i = 2248.0 # before ignoring implicit events: 2257
HierCP_i = 2338.0 # 2354
HierCo_i = 2353.0 # 2358
HierNo_i = 81887.0 # 81857
HierTo_i = HierPC_i + HierCP_i + HierCo_i + HierNo_i # total number of event pairs
hier_weights_i = [0.25*HierTo_i/HierPC_i, 0.25*HierTo_i/HierCP_i, 0.25*HierTo_i/HierCo_i, 0.25*HierTo_i/HierNo_i]
temp_weights = [0.25*818.0/412.0, 0.25*818.0/263.0, 0.25*818.0/30.0, 0.25*818.0/113.0]
# transformers + MLP + Constraints
class transformers_mlp_cons(nn.Module):
def __init__(self, params):
super().__init__()
self.transformers_model = params['transformers_model']
#self.model = AutoModel.from_pretrained(self.transformers_model)
self.model = params['model']
self.cuda = params['cuda']
self.dataset = params['dataset']
self.block_size = params['block_size']
self.add_loss = params['add_loss']
self.dpn = params['dpn']
if self.dpn:
self.target_concentration = 100.0
self.id_criterion = DirichletKLLoss(target_concentration=self.target_concentration,
concentration=1.0,
reverse=True)
self.ood_criterion = DirichletKLLoss(target_concentration=0.0,
concentration=1.0,
reverse=True)
self.out_class = 3
else:
self.hier_class_weights_h = torch.FloatTensor(hier_weights_h).to(self.cuda)
self.hier_class_weights_i = torch.FloatTensor(hier_weights_i).to(self.cuda)
self.temp_class_weights = torch.FloatTensor(temp_weights).cuda()
self.HiEve_anno_loss = nn.CrossEntropyLoss(weight=self.hier_class_weights_h)
self.IC_anno_loss = nn.CrossEntropyLoss(weight=self.hier_class_weights_i)
self.MATRES_anno_loss = nn.CrossEntropyLoss(weight=self.temp_class_weights)
self.transitivity_loss_H = transitivity_loss_H_()
self.transitivity_loss_T = transitivity_loss_T_()
self.cross_category_loss = cross_category_loss_()
self.out_class = 4
self.emb_size = params['emb_size']
self.fc = nn.Linear(2 * self.emb_size, self.emb_size)
self.bilinear = nn.Linear(self.emb_size * self.block_size, self.out_class) # 3 if DPN; else 4
self.fc1 = nn.Linear(3 * self.emb_size, self.emb_size)
self.fc2 = nn.Linear(self.emb_size, self.out_class) # 3 if DPN; else 4
def forward(self, input_ids, attention_mask, event_pos, event_pos_end, event_pair, labels):
""" Encode with Transformer """
output = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
output_attentions=True,
)
seq_output = output[0]
attention = output[-1][-1]
attention_dim = attention.size()[-1]
seq_len = seq_output.size()[1]
if attention_dim != seq_len:
sequence_output = torch.zeros(seq_output.size()[0], attention_dim, seq_output.size()[2]).to(self.cuda)
sequence_output[:, :seq_len, :] = seq_output
else:
sequence_output = seq_output
#print(sequence_output.shape) #[batch_size, doc_len, 768]
#print(attention.shape) # [batch_size, 12, doc_len, doc_len]
""" Get representation for event pairs """
batch_size = input_ids.size(0)
e1_embs_batch = []
e2_embs_batch = []
atts_contract = []
for i in range(batch_size):
event_embs = []
event_atts = []
event_num_in_this_article = len(event_pos[i])
for j in range(event_num_in_this_article):
#e_emb = torch.mean(sequence_output[i, event_pos[i][j]:event_pos_end[i][j], :].unsqueeze(0), dim=1)
e_emb = sequence_output[i, event_pos[i][j], :] # [768]
event_embs.append(e_emb)
e_att = attention[i, :, event_pos[i][j]] # [12, doc_len]
event_atts.append(e_att)
event_embs = torch.squeeze(torch.stack(event_embs, dim = 0))
event_atts = torch.squeeze(torch.stack(event_atts, dim = 0))
e1_embs = torch.index_select(event_embs, 0, torch.add(torch.tensor(event_pair[i]).to(event_embs.device)[:, 0], -1))
e2_embs = torch.index_select(event_embs, 0, torch.add(torch.tensor(event_pair[i]).to(event_embs.device)[:, 1], -1))
e1_embs_batch.append(e1_embs)
e2_embs_batch.append(e2_embs)
e1_atts = torch.index_select(event_atts, 0, torch.add(torch.tensor(event_pair[i]).to(event_embs.device)[:, 0], -1)) # torch.Size([780, 12, 428])
e2_atts = torch.index_select(event_atts, 0, torch.add(torch.tensor(event_pair[i]).to(event_embs.device)[:, 1], -1))
event_pair_att = (e1_atts * e2_atts).mean(1)
event_pair_att = event_pair_att / (event_pair_att.sum(1, keepdim=True) + 1e-5) # torch.Size([780, 428])
event_pair_contract = contract("ld,rl->rd", sequence_output[i], event_pair_att) # torch.Size([780, 768])
atts_contract.append(event_pair_contract)
e1_embs_batch = torch.cat(e1_embs_batch, dim = 0)
e2_embs_batch = torch.cat(e2_embs_batch, dim = 0)
#print("e1_embs_batch.shape:", e1_embs_batch.shape) # [pairs_num, 768]
atts_contract = torch.cat(atts_contract, dim = 0)
assert self.emb_size == e1_embs_batch.size(1)
""" Calculating loss """
wenxuan = True
if wenxuan:
e1_representation = torch.tanh(self.fc(torch.cat([e1_embs_batch, atts_contract], dim = 1)))
e2_representation = torch.tanh(self.fc(torch.cat([e2_embs_batch, atts_contract], dim = 1)))
gb1 = e1_representation.view(-1, self.emb_size // self.block_size, self.block_size) # group bilinear
gb2 = e2_representation.view(-1, self.emb_size // self.block_size, self.block_size) # group bilinear
bl = (gb1.unsqueeze(3) * gb2.unsqueeze(2)).view(-1, self.emb_size * self.block_size) # ?
logits = self.bilinear(bl)
else:
mul = torch.mul(e1_embs_batch, e2_embs_batch)
logits = self.fc2(torch.tanh(self.fc1(torch.cat((e1_embs_batch, e2_embs_batch, mul), 1))))
loss = 0.0
labels = [torch.tensor(label) for label in labels]
labels = torch.cat(labels, dim=0).to(logits).long()
#print(labels.shape)
#print(logits.shape)
"""
if self.dataset == 'HiEve':
loss += self.HiEve_anno_loss(logits, labels)
elif self.dataset == 'IC':
loss += self.IC_anno_loss(logits, labels)
elif self.dataset == "MATRES":
loss += self.MATRES_anno_loss(logits, labels)
else:
raise Exception('Dataset error!')
"""
# Updated on Apr 13, 2022
if self.dpn:
for i in range(labels.shape[0]):
if labels[i] != 3:
loss += temp_weights[int(labels[i])] * self.id_criterion(torch.unsqueeze(logits[i], 0), torch.unsqueeze(labels[i], 0)) / self.target_concentration
else:
loss += temp_weights[int(labels[i])] * self.ood_criterion(torch.unsqueeze(logits[i], 0), None)
else:
loss += self.MATRES_anno_loss(logits, labels)
if self.add_loss:
print("adding loss...")
start_index = 0
for i in range(batch_size):
event_num = len(event_pos[i])
triples = []
for a in range(0, event_num):
for b in range(a+1, event_num):
for c in range(b+1, event_num):
triples.append([a, b, c])
alpha_logits = torch.index_select(logits[start_index:start_index+len(event_pair[i])], 0, torch.tensor(triples).to(event_embs.device)[:, 0])
beta_logits = torch.index_select(logits[start_index:start_index+len(event_pair[i])], 0, torch.tensor(triples).to(event_embs.device)[:, 1])
gamma_logits = torch.index_select(logits[start_index:start_index+len(event_pair[i])], 0, torch.tensor(triples).to(event_embs.device)[:, 2])
if self.dataset in ['HiEve', 'IC']:
loss += self.add_loss * self.transitivity_loss_H(alpha_logits, beta_logits, gamma_logits).sum()
else:
loss += self.add_loss * self.transitivity_loss_T(alpha_logits, beta_logits, gamma_logits).sum()
start_index += len(event_pair[i])
#print(start_index)
#print(len(logits))
#print(logits.shape)
assert start_index == len(logits)
return logits, loss