-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
173 lines (126 loc) · 6.54 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
import torch
import torch.nn as nn
class kobert_Classifier(nn.Module):
def __init__(self, bert, hidden_size=768, num_classes=42, dr_rate=0.0):
super(kobert_Classifier, self).__init__()
self.bert = bert
self.dr_rate = dr_rate
self.classifier = nn.Linear(hidden_size, num_classes)
torch.nn.init.xavier_uniform_(self.classifier.weight)
self.dropout = nn.Dropout(p=dr_rate)
def forward(self, token_ids, attention_mask, segment_ids):
out = self.bert(input_ids=token_ids, attention_mask=attention_mask, token_type_ids=segment_ids)
if self.dr_rate:
out = self.dropout(out)
return self.classifier(out)
class koelectra_Classifier(nn.Module):
def __init__(self, electra, hidden_size=768, num_classes=42, dr_rate=0.0):
super(koelectra_Classifier, self).__init__()
self.electra = electra
self.dr_rate = dr_rate
self.pooler = nn.Linear(hidden_size, hidden_size)
self.classifier = nn.Linear(hidden_size, num_classes)
torch.nn.init.xavier_uniform_(self.pooler.weight)
torch.nn.init.xavier_uniform_(self.classifier.weight)
self.dropout = nn.Dropout(p=dr_rate)
def forward(self, token_ids, attention_mask, segment_ids):
out = self.electra(input_ids=token_ids, attention_mask=attention_mask, token_type_ids=segment_ids)[0]
out = out[:, 0, :] # take <s> token (equiv. to [CLS])
out = self.pooler(out)
out = torch.nn.functional.gelu(out) # although BERT uses tanh here, it seems Electra authors used gelu here
if self.dr_rate:
out = self.dropout(out)
return self.classifier(out)
class roberta_Classifier(nn.Module):
def __init__(self, roberta, hidden_size=1024, num_classes=42, dr_rate=0.0):
super(roberta_Classifier, self).__init__()
self.roberta = roberta
self.dr_rate = dr_rate
self.pooler = nn.Linear(hidden_size, hidden_size//2)
self.classifier = nn.Linear(hidden_size//2, num_classes)
torch.nn.init.xavier_uniform_(self.pooler.weight)
torch.nn.init.xavier_uniform_(self.classifier.weight)
if self.dr_rate:
self.dropout = nn.Dropout(p=dr_rate)
def forward(self, token_ids, attention_mask, segment_ids=None):
out = self.roberta(input_ids=token_ids, attention_mask=attention_mask)[0]
out = out[:, 0, :] # take <s> token (equiv. to [CLS])
out = self.pooler(out)
out = torch.nn.functional.gelu(out)
if self.dr_rate:
out = self.dropout(out)
return self.classifier(out)
'''
Reference:
https://github.com/monologg/R-BERT
'''
class FCLayer(nn.Module):
def __init__(self, input_dim, output_dim, dropout_rate=0.0, use_activation=True):
super(FCLayer, self).__init__()
self.use_activation = use_activation
self.dropout = nn.Dropout(dropout_rate)
self.linear = nn.Linear(input_dim, output_dim)
self.tanh = nn.Tanh()
torch.nn.init.xavier_uniform_(self.linear.weight)
def forward(self, x):
x = self.dropout(x)
if self.use_activation:
x = self.tanh(x)
return self.linear(x)
class r_roberta_Classifier(nn.Module):
def __init__(self, roberta, hidden_size=1024, num_classes=42, dr_rate=0.0):
super(r_roberta_Classifier, self).__init__()
self.roberta = roberta
self.dr_rate = dr_rate
self.cls_fc = FCLayer(hidden_size, hidden_size//2, self.dr_rate)
self.entity_fc = FCLayer(hidden_size, hidden_size//2, self.dr_rate)
self.label_classifier = FCLayer(hidden_size//2 * 3, num_classes, self.dr_rate, False)
def forward(self, token_ids, attention_mask, segment_ids=None):
out = self.roberta(input_ids=token_ids, attention_mask=attention_mask)[0]
entity_end_position = torch.where(token_ids == 2)[1]
entity1_end, entity2_end = entity_end_position[0], entity_end_position[2]
cls_vector = out[:, 0, :] # take <s> token (equiv. to [CLS])
entity1_vector = out[:, 1:entity1_end, :] # Get Entity vector
entity2_vector = out[:, entity1_end+2:entity2_end, :]
entity1_vector = torch.mean(entity1_vector, dim=1) # Average
entity2_vector = torch.mean(entity2_vector, dim=1)
# Dropout -> tanh -> fc_layer (Share FC layer for e1 and e2)
cls_embedding = self.cls_fc(cls_vector)
e1_embedding = self.entity_fc(entity1_vector)
e2_embedding = self.entity_fc(entity2_vector)
# Concat -> fc_layer
concat_embedding = torch.cat([cls_embedding, e1_embedding, e2_embedding], dim=-1)
return self.label_classifier(concat_embedding)
def get_tokenizer(args):
if args.model == 'kobert':
tokenizer = KoBertTokenizer.from_pretrained('monologg/kobert')
elif args.model == 'multi':
tokenizer = AutoTokenizer.from_pretrained("sangrimlee/bert-base-multilingual-cased-korquad")
elif args.models == 'koelectra':
tokenizer = AutoTokenizer.from_pretrained("monologg/koelectra-base-v3-discriminator")
elif args.model == 'r_roberta':
tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-large")
elif args.model == 'roberta':
tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-large")
else:
raise NotImplementedError('Tokenizer & Model not available')
return tokenizer
def get_model(args):
if args.model == 'kobert':
feature_model = BertModel.from_pretrained("monologg/kobert")
model = kobert_Classifier(feature_model, dr_rate=args.dp)
elif args.model == 'multi':
feature_model = BertModel.from_pretrained("sangrimlee/bert-base-multilingual-cased-korquad")
model = kobert_Classifier(feature_model, dr_rate=args.dp)
elif args.models == 'koelectra':
feature_model = ElectraModel.from_pretrained("monologg/koelectra-base-v3-discriminator")
model = koelectra_Classifier(feature_model, dr_rate=args.dp)
elif args.model == 'r_roberta':
feature_model = RobertaModel.from_pretrained("xlm-roberta-large", add_pooling_layer=False)
model = r_roberta_Classifier(feature_model, dr_rate=args.dp)
elif args.model == 'roberta':
feature_model = RobertaModel.from_pretrained("xlm-roberta-large", add_pooling_layer=False)
model = roberta_Classifier(feature_model)
else:
raise NotImplementedError('Tokenizer & Model not available')
return model