-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_supervised_bert.py
149 lines (129 loc) · 5.15 KB
/
train_supervised_bert.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
# coding:utf-8
from cmath import log
from curses import A_REVERSE
from re import L
import sys
sys.path.append('../')
import torch
import numpy as np
import json
import opennre
from opennre import encoder, model, framework
import os
import argparse
import logging
import random
def set_seed(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
parser = argparse.ArgumentParser()
parser.add_argument('--pretrain_path', default='roberta-large',
help='Pre-trained ckpt path / model name (hugginface)')
parser.add_argument('--ckpt', default='',
help='Checkpoint name')
parser.add_argument('--pooler', default='cls', choices=['cls', 'entity'],
help='Sentence representation pooler')
parser.add_argument('--only_test', action='store_true',
help='Only run test')
parser.add_argument('--mask_entity', action='store_true',
help='Mask entity mentions')
parser.add_argument('--labeling', default="False", choices=['True','False'],
help='Generate soft labels')
parser.add_argument("--stutrain",default="False",choices=['True','False'],
help='Teacher training or student training')
# Data
parser.add_argument('--metric', default='micro_f1', choices=['micro_f1', 'acc', 'macro_f1'],
help='Metric for picking up best checkpoint')
parser.add_argument('--dataset', default='semeval', required=True,
help='Dataset. If not none, the following args can be ignored')
parser.add_argument('--train_file', default='', type=str,
help='Training data file')
# parser.add_argument('--val_file', default='', type=str,
# help='Validation data file')
parser.add_argument('--test_file', default='', type=str,
help='Test data file')
parser.add_argument('--rel2id_file', default='', type=str,
help='Relation to ID file')
# Hyper-parameters
parser.add_argument('--batch_size', default=64, type=int,
help='Batch size')
parser.add_argument('--lr', default=2e-5, type=float,
help='Learning rate')
parser.add_argument('--max_length', default=128, type=int,
help='Maximum sentence length')
parser.add_argument('--max_epoch', default=10, type=int,
help='Max number of training epochs')
parser.add_argument('--lambda_u',default=0.2,type=float,
help="weight of soft loss")
parser.add_argument('--use_loss',default="nn.CrossEntropyLoss",type=str,
choices=["nn.CrossEntropyLoss","DiceLoss","MultiDSCLoss","MultiFocalLoss","GHMC_Loss","LDAMLoss","CBLoss"],
help='Loss function')
# Seed
parser.add_argument('--seed', default=42, type=int,
help='Seed')
args = parser.parse_args()
# Set random seed
set_seed(args.seed)
# Some basic settings
root_path = '.'
sys.path.append(root_path)
if not os.path.exists('ckpt'):
os.mkdir('ckpt')
if len(args.ckpt) == 0:
args.ckpt = '{}_{}_{}'.format(args.dataset, args.pretrain_path.split('/')[-1], args.pooler)
ckpt = 'ckpt/{}.pth.tar'.format(args.ckpt)
opennre.download(args.dataset, root_path=root_path)
args.train_file = os.path.join(root_path, 'benchmark', args.dataset, 'train.json')
if args.labeling=="True":
args.test_file = os.path.join(root_path, 'benchmark', args.dataset, 'label.json')
else:
args.test_file = os.path.join(root_path, 'benchmark', args.dataset, 'test.json')
args.rel2id_file = os.path.join(root_path, 'benchmark', args.dataset, 'rel2id.json')
logging.info('Arguments:')
for arg in vars(args):
logging.info(' {}: {}'.format(arg, getattr(args, arg)))
rel2id = json.load(open(args.rel2id_file))
sentence_encoder = opennre.encoder.BERTEncoder(
max_length=args.max_length,
pretrain_path=args.pretrain_path,
mask_entity=args.mask_entity
)
# Define the model
model = opennre.model.SoftmaxNN(sentence_encoder, len(rel2id), rel2id)
# Define the whole training framework
framework = opennre.framework.SentenceRE(
train_path=args.train_file,
test_path=args.test_file,
model=model,
ckpt=ckpt,
batch_size=args.batch_size,
max_epoch=args.max_epoch,
lr=args.lr,
opt='adamw',
stutrain = args.stutrain,
lambda_u=args.lambda_u,
use_loss=args.use_loss
)
# Train the model
if args.labeling=="False" and not args.only_test:
framework.train_model(args.metric)
# Test
framework.load_state_dict(torch.load(ckpt)['state_dict'])
if args.labeling=="False":
result = framework.eval_model(framework.test_loader)
# Print the result
logging.info('Test set results:')
logging.info('Accuracy: {}'.format(result['acc']))
logging.info('Micro precision: {}'.format(result['micro_p']))
logging.info('Micro recall: {}'.format(result['micro_r']))
logging.info('Micro F1: {}'.format(result['micro_f1']))
logging.info('Macro precision: {}'.format(result['macro_p']))
logging.info('Macro recall: {}'.format(result['macro_r']))
logging.info('Macro F1: {}'.format(result['macro_f1']))
# logging.info("Predicted Labels: {}".format(result['pred_labels']))
logging.info('F1 per Relation: {}'.format(result['f1_per_relation']))
logging.info('Classification Report: {}'.format(result['report']))
else:
result = framework.test_model(framework.test_loader)