-
Notifications
You must be signed in to change notification settings - Fork 1.1k
/
bag_re.py
184 lines (173 loc) · 6.66 KB
/
bag_re.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
import torch
from torch import nn, optim
import json
from .data_loader import SentenceRELoader, BagRELoader
from .utils import AverageMeter
from tqdm import tqdm
import os
class BagRE(nn.Module):
def __init__(self,
model,
train_path,
val_path,
test_path,
ckpt,
batch_size=32,
max_epoch=100,
lr=0.1,
weight_decay=1e-5,
opt='sgd',
bag_size=0,
loss_weight=False):
super().__init__()
self.max_epoch = max_epoch
self.bag_size = bag_size
# Load data
if train_path != None:
self.train_loader = BagRELoader(
train_path,
model.rel2id,
model.sentence_encoder.tokenize,
batch_size,
True,
bag_size=bag_size,
entpair_as_bag=False)
if val_path != None:
self.val_loader = BagRELoader(
val_path,
model.rel2id,
model.sentence_encoder.tokenize,
batch_size,
False,
bag_size=bag_size,
entpair_as_bag=True)
if test_path != None:
self.test_loader = BagRELoader(
test_path,
model.rel2id,
model.sentence_encoder.tokenize,
batch_size,
False,
bag_size=bag_size,
entpair_as_bag=True
)
# Model
self.model = nn.DataParallel(model)
# Criterion
if loss_weight:
self.criterion = nn.CrossEntropyLoss(weight=self.train_loader.dataset.weight)
else:
self.criterion = nn.CrossEntropyLoss()
# Params and optimizer
params = self.model.parameters()
self.lr = lr
if opt == 'sgd':
self.optimizer = optim.SGD(params, lr, weight_decay=weight_decay)
elif opt == 'adam':
self.optimizer = optim.Adam(params, lr, weight_decay=weight_decay)
elif opt == 'adamw':
from transformers import AdamW
params = list(self.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
grouped_params = [
{
'params': [p for n, p in params if not any(nd in n for nd in no_decay)],
'weight_decay': 0.01,
'lr': lr,
'ori_lr': lr
},
{
'params': [p for n, p in params if any(nd in n for nd in no_decay)],
'weight_decay': 0.0,
'lr': lr,
'ori_lr': lr
}
]
self.optimizer = AdamW(grouped_params, correct_bias=False)
else:
raise Exception("Invalid optimizer. Must be 'sgd' or 'adam' or 'bert_adam'.")
# Cuda
if torch.cuda.is_available():
self.cuda()
# Ckpt
self.ckpt = ckpt
def train_model(self, metric='auc'):
best_metric = 0
for epoch in range(self.max_epoch):
# Train
self.train()
print("=== Epoch %d train ===" % epoch)
avg_loss = AverageMeter()
avg_acc = AverageMeter()
avg_pos_acc = AverageMeter()
t = tqdm(self.train_loader)
for iter, data in enumerate(t):
if torch.cuda.is_available():
for i in range(len(data)):
try:
data[i] = data[i].cuda()
except:
pass
label = data[0]
bag_name = data[1]
scope = data[2]
args = data[3:]
logits = self.model(label, scope, *args, bag_size=self.bag_size)
loss = self.criterion(logits, label)
score, pred = logits.max(-1) # (B)
acc = float((pred == label).long().sum()) / label.size(0)
pos_total = (label != 0).long().sum()
pos_correct = ((pred == label).long() * (label != 0).long()).sum()
if pos_total > 0:
pos_acc = float(pos_correct) / float(pos_total)
else:
pos_acc = 0
# Log
avg_loss.update(loss.item(), 1)
avg_acc.update(acc, 1)
avg_pos_acc.update(pos_acc, 1)
t.set_postfix(loss=avg_loss.avg, acc=avg_acc.avg, pos_acc=avg_pos_acc.avg)
# Optimize
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
# Val
print("=== Epoch %d val ===" % epoch)
result = self.eval_model(self.val_loader)
print("AUC: %.4f" % result['auc'])
print("Micro F1: %.4f" % (result['max_micro_f1']))
if result[metric] > best_metric:
print("Best ckpt and saved.")
torch.save({'state_dict': self.model.module.state_dict()}, self.ckpt)
best_metric = result[metric]
print("Best %s on val set: %f" % (metric, best_metric))
def eval_model(self, eval_loader):
self.model.eval()
with torch.no_grad():
t = tqdm(eval_loader)
pred_result = []
for iter, data in enumerate(t):
if torch.cuda.is_available():
for i in range(len(data)):
try:
data[i] = data[i].cuda()
except:
pass
label = data[0]
bag_name = data[1]
scope = data[2]
args = data[3:]
logits = self.model(None, scope, *args, train=False, bag_size=self.bag_size) # results after softmax
logits = logits.cpu().numpy()
for i in range(len(logits)):
for relid in range(self.model.module.num_class):
if self.model.module.id2rel[relid] != 'NA':
pred_result.append({
'entpair': bag_name[i][:2],
'relation': self.model.module.id2rel[relid],
'score': logits[i][relid]
})
result = eval_loader.dataset.eval(pred_result)
return result
def load_state_dict(self, state_dict):
self.model.module.load_state_dict(state_dict)