-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
190 lines (163 loc) · 9.02 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
#!/usr/bin/env/ python
# coding: utf-8
"""
@File: model.py
@Description: Void
@Copyright: Zhuoxian Wang
@Email: [email protected]
@Date: 2020/3/26
"""
from transformers import XLNetModel
import torch
from torch import nn
import math
from const import PreTrainedXLNetModelPath
class GCN(nn.Module):
def __init__(self, hid_size=1024, concatenate=False):
super(GCN, self).__init__()
self.hid_size = hid_size
# self.W = nn.Parameter(torch.FloatTensor(self.hid_size, self.hid_size // 2).cuda())
# self.b = nn.Parameter(torch.FloatTensor(self.hid_size // 2, ).cuda())
self.concatenate = concatenate
if concatenate:
self.linear = nn.Linear(self.hid_size, self.hid_size // 2, bias=True)
else:
self.linear = nn.Linear(self.hid_size, self.hid_size, bias=True)
# self.init()
def init(self):
stdv = 1 / math.sqrt(self.hid_size // 2)
# self.W.data.uniform_(-stdv, stdv)
# self.b.data.uniform_(-stdv, stdv)
self.linear.weight.data.uniform_(-stdv, stdv)
self.linear.bias.data.uniform_(-stdv, stdv)
def forward(self, inp, adj, is_relu=True, average=False):
# out = torch.matmul(inp, self.W) + self.b
out = self.linear(inp)
out = torch.matmul(adj, out) # [bs, max_seq_len, 1024]
if average:
margin_sum = torch.sum(adj, dim=2)
margin_sum[margin_sum == 0] = 1
out = out / margin_sum.unsqueeze(2)
if is_relu:
# out = nn.functional.relu(out)
out = nn.functional.leaky_relu(out)
return out
def __repr__(self):
return self.__class__.__name__ + '(hid_size=%d)' % self.hid_size
class GraphPreTrain(nn.Module):
def __init__(self, args, drop_out=0.5):
super(GraphPreTrain, self).__init__()
self.xlnet = XLNetModel.from_pretrained(PreTrainedXLNetModelPath)
self.hidden_size = self.xlnet.config.hidden_size
self.rel_feature_size = args.rel_feature_size
self.num_rel = args.num_rel
self.gcn_average = True if args.gcn_average else False
self.gcn_concatenate = True if args.gcn_concatenate else False
# define model architecture and the way of sequence aggregation
if args.model_architecture:
assert args.model_architecture in ["joint", "graph_alone", "sequence_alone"]
self.model_architecture = args.model_architecture
else:
self.model_architecture = "joint"
if self.model_architecture != "graph_alone":
self.linear_s = nn.Linear(self.hidden_size * 2, self.rel_feature_size)
self.batch_norm_s = nn.BatchNorm1d(num_features=2*self.hidden_size)
if self.model_architecture != "sequence_alone":
self.linear_g = nn.Linear(self.hidden_size * 2, self.rel_feature_size)
self.batch_norm_g = nn.BatchNorm1d(num_features=2 * self.hidden_size)
self.linear_rel = nn.Linear(self.rel_feature_size, self.num_rel)
self.entity_position_encoding = args.entity_position_encoding
if args.aggregation_way:
assert args.aggregation_way in ["cls", "entity_start", "max_pool"]
self.aggregation_way = args.aggregation_way
else:
self.aggregation_way = "entity_start"
self.graph_pool_function = "max" if not args.graph_pool_fun else args.graph_pool_fun # max or mean
self.gcn_layer = args.gcn_layer
self.hidden_size = self.xlnet.config.hidden_size
self.gcn_fw = nn.ModuleList([GCN(self.hidden_size, self.gcn_concatenate) for _ in range(self.gcn_layer)])
self.gcn_bw = nn.ModuleList([GCN(self.hidden_size, self.gcn_concatenate) for _ in range(self.gcn_layer)])
self.drop_out = nn.Dropout(drop_out)
def forward(self, input_ids, attn_mask, token_type_ids, dep_fw, dep_bw, en_positions):
"""
token_type_ids == None if entity encoding way is not 'token_type_ids'
for input ['<pad>', '<pad>', 'a', 'e1', 'b', 'e2_0', 'e2_1',...]
en_position = (3, 3), (5, 6)
"""
batch_size = input_ids.shape[0]
transformers_out = self.xlnet(
input_ids=input_ids,
attention_mask=attn_mask,
token_type_ids=token_type_ids # may be None
)[0]
if self.aggregation_way == 'cls':
sequence_features = torch.cat((transformers_out[:, -1, :], transformers_out[:, -1, :]), dim=1)
elif self.aggregation_way == "max_pool":
en1_seq_feature_list = list()
en2_seq_feature_list = list()
for batch_id in range(batch_size):
en1_position = en_positions[batch_id][0]
en2_position = en_positions[batch_id][1]
en1_seq_feature = transformers_out[batch_id, en1_position[0]:en1_position[1] + 1].max(dim=0, keepdim=True)[0]
en1_seq_feature_list.append(en1_seq_feature)
en2_seq_feature = transformers_out[batch_id, en2_position[0]:en2_position[1] + 1].max(dim=0, keepdim=True)[0]
en2_seq_feature_list.append(en2_seq_feature)
en1_seq_features = torch.cat(en1_seq_feature_list, dim=0)
en2_seq_features = torch.cat(en2_seq_feature_list, dim=0)
sequence_features = torch.cat((en1_seq_features, en2_seq_features), dim=1) # [bs, 2 * 1024]
elif self.aggregation_way == "entity_start":
en1_seq_feature_list = list()
en2_seq_feature_list = list()
for batch_id in range(batch_size):
en1_position = en_positions[batch_id][0]
en2_position = en_positions[batch_id][1]
entity_start_pos1 = en1_position[0] - 1 if self.entity_position_encoding == "entity_tag" else \
en1_position[0]
en1_seq_feature = transformers_out[batch_id, entity_start_pos1].unsqueeze(0)
en1_seq_feature_list.append(en1_seq_feature)
entity_start_pos2 = en2_position[0] - 1 if self.entity_position_encoding == "entity_tag" else \
en2_position[0]
en2_seq_feature = transformers_out[batch_id, entity_start_pos2].unsqueeze(0)
en2_seq_feature_list.append(en2_seq_feature)
en1_seq_features = torch.cat(en1_seq_feature_list, dim=0)
en2_seq_features = torch.cat(en2_seq_feature_list, dim=0)
sequence_features = torch.cat((en1_seq_features, en2_seq_features), dim=1) # [bs, 2 * 1024]
else:
raise ValueError("Aggregation way error!")
if self.model_architecture != "sequence_alone":
out = transformers_out
for i in range(self.gcn_layer):
out_fw = self.gcn_fw[i](out, dep_fw, average=self.gcn_average)
out_bw = self.gcn_bw[i](out, dep_bw, average=self.gcn_average)
if self.gcn_concatenate:
out = torch.cat([out_fw, out_bw], dim=2)
else:
out = out_fw + out_bw
out = self.drop_out(out) # [bs, max_len, 1024]
en1_graph_feature_list = list()
en2_graph_feature_list = list()
for batch_id in range(batch_size):
en1_position = en_positions[batch_id][0]
en2_position = en_positions[batch_id][1]
if self.graph_pool_function == "max":
en1_graph_feature = out[batch_id, en1_position[0]:en1_position[1]+1].max(dim=0, keepdim=True)[0]
en1_graph_feature_list.append(en1_graph_feature)
en2_graph_feature = out[batch_id, en2_position[0]:en2_position[1]+1].max(dim=0, keepdim=True)[0]
en2_graph_feature_list.append(en2_graph_feature)
else:
en1_graph_feature = out[batch_id, en1_position[0]:en1_position[1]+1].mean(dim=0, keepdim=True)
en1_graph_feature_list.append(en1_graph_feature)
en2_graph_feature = out[batch_id, en2_position[0]:en2_position[1]+1].mean(dim=0, keepdim=True)
en2_graph_feature_list.append(en2_graph_feature)
en1_graph_features = torch.cat(en1_graph_feature_list, dim=0)
en2_graph_features = torch.cat(en2_graph_feature_list, dim=0)
graph_features = torch.cat((en1_graph_features, en2_graph_features), dim=1) # [bs, 2 * 1024]
if self.model_architecture == "sequence_alone":
# sequence alone
relation_features = self.linear_s(self.batch_norm_s(sequence_features))
elif self.model_architecture == "graph_alone":
relation_features = self.linear_g(self.batch_norm_g(graph_features))
else:
relation_features = self.linear_s(self.batch_norm_s(sequence_features)) + self.linear_g(self.batch_norm_g(graph_features))
scores = self.linear_rel(nn.functional.leaky_relu(relation_features))
return scores