-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
167 lines (153 loc) · 7.66 KB
/
train.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
# -*- coding: utf-8 -*-
"""
@Version: 0.1
@Author: Charles
@Time: 2022/11/3 14:55
@File: train.py
@Desc:
"""
import json
import os
import shutil
import numpy as np
import torch
import torch.nn.functional as F
from sklearn import metrics
import time
from utils import get_time_dif, WarmupPolyLR, setup_logger
from tensorboardX import SummaryWriter
from dataloader import MyDataset, dataset_collect
from transformers import BertTokenizer, BertConfig
from torch.utils.data import DataLoader
import transformers
from models import build_model
from config.config import Config
transformers.logging.set_verbosity_error()
def train(config):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # 设备
writer = SummaryWriter(log_dir=config.output_dir)
logger = setup_logger(os.path.join(config.output_dir, 'train.log'))
model_save_dir = os.path.join(config.output_dir, 'checkpoint')
os.makedirs(model_save_dir, exist_ok=True)
# 训练配置
logger.info(json.dumps(vars(config), ensure_ascii=False, indent=2))
with open(os.path.join(config.output_dir, 'model_config.json'), 'w', encoding='utf-8') as f:
f.write(json.dumps(vars(config), ensure_ascii=False, indent=2))
tokenizer = BertTokenizer.from_pretrained('bert-base-chinese')
train_dataset = MyDataset(config.train_path, tokenizer, config)
test_dataset = MyDataset(config.test_path, tokenizer, config)
train_dataloader = DataLoader(train_dataset, batch_size=config.batch_size, shuffle=True, collate_fn=dataset_collect)
test_dataloader = DataLoader(test_dataset, batch_size=config.batch_size, shuffle=False, collate_fn=dataset_collect)
# 保存vocab.txt
tokenizer.save_pretrained(model_save_dir)
model = build_model(config.model_name, config.pretrain_dir)
model.to(device)
start_time = time.time()
model.train()
optimizer = torch.optim.Adam(model.parameters(), lr=config.learning_rate)
if config.warmup:
warmup_iters = config.warmup_epoch * len(train_dataloader)
scheduler = WarmupPolyLR(optimizer, max_iters=config.num_epochs * len(train_dataloader), warmup_iters=warmup_iters,
warmup_epoch=config.warmup_epoch, last_epoch=-1)
total_batch = 0 # 记录进行到多少batch
test_best_acc = 0
last_improve = 0 # 记录上次验证集loss下降的batch数
flag = False # 记录是否很久没有效果提升
logger.info('train dataset has {} samples, {} in dataloader, validate dataset has {} samples, {} in dataloader'.
format(len(train_dataloader.dataset), len(train_dataloader), len(test_dataloader.dataset), len(test_dataloader)))
for epoch in range(config.num_epochs):
lr = optimizer.param_groups[0]['lr']
for i, batch in enumerate(train_dataloader):
lr = optimizer.param_groups[0]['lr']
encodings, labels = batch
for idx, encoding in enumerate(encodings):
encodings[idx] = {k: v.to(device) for k, v in encoding.items()}
labels = labels.to(device)
outputs = model(encodings)
model.zero_grad()
loss = F.cross_entropy(outputs, labels)
loss.backward()
optimizer.step()
if config.warmup:
scheduler.step()
if total_batch % config.log_iter == 0:
# 每多少轮输出在训练集和验证集上的效果
true = labels.data.cpu()
predict = torch.max(outputs, 1)[1].cpu()
# train_acc = metrics.accuracy_score(true, predict)
train_acc, train_rec, train_f1, _ = metrics.precision_recall_fscore_support(true, predict,
average='macro',
zero_division=0)
test_acc, test_rec, test_f1, test_loss = evaluate(config, model, test_dataloader)
if test_acc > test_best_acc:
test_best_acc = test_acc
torch.save(model.state_dict(), os.path.join(model_save_dir, f'best_{epoch+1}.ckpt'))
torch.save(model.state_dict(), os.path.join(model_save_dir, f'best.ckpt'))
model.save_pretrained(model_save_dir)
improve = '*'
last_improve = total_batch
else:
improve = ''
time_dif = get_time_dif(start_time)
msg = 'Epoch [{}/{}], Iter: {:>6}, Train Loss: {:>5.2f}, Train Acc: {:>7.2%}, Test Loss: {:>5.2f}, Test Acc: {:>6.2%}, LR: {:>7.6f},Time: {} {}'
logger.info(msg.format(epoch + 1, config.num_epochs, total_batch, loss.item(), train_acc, test_loss, test_acc, lr, time_dif, improve))
writer.add_scalar("loss/train", loss.item(), total_batch)
writer.add_scalar("loss/test", test_loss, total_batch)
writer.add_scalar("f1/train", train_f1, total_batch)
writer.add_scalar("f1/test", test_f1, total_batch)
writer.add_scalar("train/lr", lr, total_batch)
model.train()
total_batch += 1
if total_batch - last_improve > config.require_improvement:
# 验证集loss超过1000batch没下降，结束训练
logger.info("No optimization for a long time, auto-stopping...")
flag = True
break
if flag:
break
writer.close()
test(config, model, test_dataloader, logger, os.path.join(model_save_dir, f'best.ckpt'))
def test(config, model, test_iter, logger, best_model_path=''):
# test
if best_model_path:
model.load_state_dict(torch.load(best_model_path))
model.eval()
start_time = time.time()
test_acc, test_recall, test_f1, test_loss, test_report, test_confusion = evaluate(config, model, test_iter, test=True)
msg = 'Test Loss: {0:>5.2}, Test Acc: {1:>6.2%}, Test Rec: {2:>6.2%}, Test F1: {3:>6.2%}'
logger.info(msg.format(test_loss, test_acc, test_recall, test_f1))
logger.info("\nPrecision, Recall and F1-Score...")
logger.info('\n{}'.format(test_report))
logger.info("Confusion Matrix...\n")
print(test_confusion)
time_dif = get_time_dif(start_time)
logger.info("Time usage:".format(time_dif))
def evaluate(config, model, data_iter, test=False):
model.eval()
loss_total = 0
predict_all = np.array([], dtype=int)
labels_all = np.array([], dtype=int)
with torch.no_grad():
for batch in data_iter:
encodings, labels = batch
for idx, encoding in enumerate(encodings):
encodings[idx] = {k: v.cuda() for k, v in encoding.items()}
labels = labels.cuda()
outputs = model(encodings)
loss = F.cross_entropy(outputs, labels)
loss_total += loss
labels = labels.data.cpu().numpy()
predict = torch.max(outputs, 1)[1].cpu().numpy()
labels_all = np.append(labels_all, labels)
predict_all = np.append(predict_all, predict)
# acc = metrics.accuracy_score(labels_all, predict_all)
acc, recall, f1, _ = metrics.precision_recall_fscore_support(labels_all, predict_all, average='macro', zero_division=0)
if test:
report = metrics.classification_report(labels_all, predict_all, digits=4)
confusion = metrics.confusion_matrix(labels_all, predict_all)
return acc, recall, f1, loss_total / len(data_iter), report, confusion
return acc, recall, f1, loss_total / len(data_iter)
if __name__ == '__main__':
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
config = Config()
train(config)