-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain.py
278 lines (256 loc) · 14 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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
# -*- coding:utf-8 -*-
# @project: BlockShuffleTest
# @filename: train
# @author: 刘聪NLP
# @zhihu: https://www.zhihu.com/people/LiuCongNLP
# @contact: [email protected]
# @time: 2021/9/27 10:51
"""
文件说明:
"""
import torch
import os
import random
import numpy as np
import argparse
import logging
from transformers import BertTokenizer
from data_set import SentimentAnalysisDataSet, collate_func_sentiment_analysis
from model import SentimentAnalysisModel
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from transformers import AdamW, get_linear_schedule_with_warmup
from tqdm import tqdm, trange
from bolck_shuffle_data_loader import BlockShuffleDataLoader
from sklearn.metrics import f1_score, accuracy_score
import json
from tensorboardX import SummaryWriter
import time
logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger = logging.getLogger(__name__)
def train(model, device, tokenizer, args):
tb_write = SummaryWriter()
if args.gradient_accumulation_steps < 1:
raise ValueError("gradient_accumulation_steps参数无效，必须大于等于1")
train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
train_data = SentimentAnalysisDataSet(tokenizer, args.max_len, args.data_dir, "train",
args.train_file_path)
dev_data = SentimentAnalysisDataSet(tokenizer, args.max_len, args.data_dir, "dev",
args.dev_file_path)
if args.is_block_shuffle:
train_data_loader = BlockShuffleDataLoader(train_data, sort_key=lambda x: len(x["input_ids"]),
sort_bs_num=None,
is_shuffle=True, batch_size=train_batch_size,
collate_fn=collate_func_sentiment_analysis)
else:
train_sampler = RandomSampler(train_data)
train_data_loader = DataLoader(train_data, sampler=train_sampler,
batch_size=train_batch_size, collate_fn=collate_func_sentiment_analysis)
total_steps = int(len(train_data_loader) * args.num_train_epochs / args.gradient_accumulation_steps)
logger.info("总训练步数为:{}".format(total_steps))
model.to(device)
no_decay = ["bias", "LayerNorm.weight"]
param_optimizer = list(model.named_parameters())
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay': args.weight_decay, 'lr': args.learning_rate},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,
'lr': args.learning_rate}
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=int(args.warmup_proportion * total_steps),
num_training_steps=total_steps)
torch.cuda.empty_cache()
model.train()
tr_loss, logging_loss, max_acc = 0.0, 0.0, 0.0
global_step = 0
for iepoch in trange(0, int(args.num_train_epochs), desc="Epoch", disable=False):
iter_bar = tqdm(train_data_loader, desc="Iter (loss=X.XXX)", disable=False)
for step, batch in enumerate(iter_bar):
input_ids = batch["input_ids"].to(device)
attention_mask = batch["attention_mask"].to(device)
label = batch["label"].to(device)
outputs = model.forward(input_ids, attention_mask, label)
loss = outputs[0]
tr_loss += loss.item()
iter_bar.set_description("Iter (loss=%5.3f)" % loss.item())
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
scheduler.step()
optimizer.zero_grad()
global_step += 1
if args.logging_steps > 0 and global_step % args.logging_steps == 0:
tb_write.add_scalar("lr", scheduler.get_lr()[0], global_step)
tb_write.add_scalar("train_loss", (tr_loss - logging_loss) /
(args.logging_steps * args.gradient_accumulation_steps), global_step)
logging_loss = tr_loss
if args.save_model_steps > 0 and global_step % args.save_model_steps == 0:
eval_loss, eval_acc, eval_f1 = evaluate(model, device, dev_data, args)
logger.info("eval_loss is {}, eval_acc is {} , eval_f1 is {}".format(eval_loss, eval_acc, eval_f1))
tb_write.add_scalar("eval_loss", eval_loss, global_step)
tb_write.add_scalar("eval_acc", eval_acc, global_step)
tb_write.add_scalar("eval_f1", eval_f1, global_step)
if eval_acc >= max_acc:
max_acc = eval_acc
output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
model_to_save = model.module if hasattr(model, "module") else model
model_to_save.save_pretrained(output_dir)
json_output_dir = os.path.join(output_dir, "json_data.json")
fin = open(json_output_dir, "w", encoding="utf-8")
fin.write(json.dumps(
{"eval_acc": float(eval_acc), "eval_f1": float(eval_f1)},
ensure_ascii=False, indent=4))
fin.close()
model.train()
def evaluate(model, device, dev_data, args):
test_sampler = SequentialSampler(dev_data)
test_data_loader = DataLoader(dev_data, sampler=test_sampler,
batch_size=args.test_batch_size, collate_fn=collate_func_sentiment_analysis)
iter_bar = tqdm(test_data_loader, desc="iter", disable=False)
eval_loss = 0.0
true_label = []
pre_label = []
for step, batch in enumerate(iter_bar):
model.eval()
with torch.no_grad():
input_ids = batch["input_ids"].to(device)
attention_mask = batch["attention_mask"].to(device)
label = batch["label"].to(device)
[loss, predict_label, _] = model.forward(input_ids, attention_mask, label)
eval_loss += loss.item()
true_label.extend(label.cpu().numpy())
pre_label.extend(predict_label.cpu().numpy())
true_label = np.array(true_label)
pre_label = np.array(pre_label)
f1 = f1_score(true_label, pre_label, average='macro')
acc = accuracy_score(true_label, pre_label)
eval_loss = eval_loss / len(test_data_loader)
return eval_loss, acc, f1
def train_ori_time(model, device, tokenizer, args):
train_batch_size = int(args.train_batch_size)
train_data = SentimentAnalysisDataSet(tokenizer, args.max_len, args.data_dir, "train", args.train_file_path)
train_sampler = RandomSampler(train_data)
train_data_loader = DataLoader(train_data, sampler=train_sampler,
batch_size=train_batch_size, collate_fn=collate_func_sentiment_analysis)
total_steps = int(len(train_data_loader) * args.num_train_epochs)
logger.info("总训练步数为:{}".format(total_steps))
model.to(device)
no_decay = ["bias", "LayerNorm.weight"]
param_optimizer = list(model.named_parameters())
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay': args.weight_decay, 'lr': args.learning_rate},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,
'lr': args.learning_rate}
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=int(args.warmup_proportion * total_steps),
num_training_steps=total_steps)
torch.cuda.empty_cache()
model.train()
T1 = time.time()
for iepoch in trange(0, 2, desc="Epoch", disable=False):
iter_bar = tqdm(train_data_loader, desc="Iter", disable=False)
for step, batch in enumerate(iter_bar):
input_ids = batch["input_ids"].to(device)
attention_mask = batch["attention_mask"].to(device)
label = batch["label"].to(device)
outputs = model.forward(input_ids, attention_mask, label)
loss = outputs[0]
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
T2 = time.time()
print("原始DataLoader，运行2个epoch时间为{}秒".format(T2 - T1))
def train_block_shuffle_time(model, device, tokenizer, args):
train_batch_size = int(args.train_batch_size)
train_data = SentimentAnalysisDataSet(tokenizer, args.max_len, args.data_dir, "train", args.train_file_path)
train_data_loader = BlockShuffleDataLoader(train_data, sort_key=lambda x: len(x["input_ids"]),
sort_bs_num=None,
is_shuffle=True, batch_size=train_batch_size,
collate_fn=collate_func_sentiment_analysis)
total_steps = int(len(train_data_loader) * args.num_train_epochs)
logger.info("总训练步数为:{}".format(total_steps))
model.to(device)
no_decay = ["bias", "LayerNorm.weight"]
param_optimizer = list(model.named_parameters())
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay': args.weight_decay, 'lr': args.learning_rate},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,
'lr': args.learning_rate}
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=int(args.warmup_proportion * total_steps),
num_training_steps=total_steps)
torch.cuda.empty_cache()
model.train()
T1 = time.time()
for iepoch in trange(0, 2, desc="Epoch", disable=False):
iter_bar = tqdm(train_data_loader, desc="Iter", disable=False)
for step, batch in enumerate(iter_bar):
input_ids = batch["input_ids"].to(device)
attention_mask = batch["attention_mask"].to(device)
label = batch["label"].to(device)
outputs = model.forward(input_ids, attention_mask, label)
loss = outputs[0]
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
T2 = time.time()
print("BlockShuffleDataLoader，运行2个epoch时间为{}秒".format(T2 - T1))
def set_args():
parser = argparse.ArgumentParser()
parser.add_argument('--device', default='0', type=str, help='')
parser.add_argument('--train_file_path', default='data/train.json', type=str, help='')
parser.add_argument('--dev_file_path', default='data/test.json', type=str, help='')
parser.add_argument('--vocab_path', default="chinese_L-12_H-768_A-12/vocab.txt",
type=str, help='')
parser.add_argument('--pretrained_model_path', default="chinese_L-12_H-768_A-12",
type=str, help='')
parser.add_argument('--data_dir', default='data/', type=str, help='')
parser.add_argument('--num_train_epochs', default=2, type=int, help='')
parser.add_argument('--train_batch_size', default=4, type=int, help='')
parser.add_argument('--test_batch_size', default=4, type=int, help='')
parser.add_argument('--learning_rate', default=5e-5, type=float, help='')
parser.add_argument('--warmup_proportion', default=0.1, type=float, help='')
parser.add_argument("--weight_decay", default=0.01, type=float, help="")
parser.add_argument('--adam_epsilon', default=1e-8, type=float, help='')
parser.add_argument('--save_model_steps', default=12, type=int, help='')
parser.add_argument('--logging_steps', default=5, type=int, help='')
parser.add_argument('--gradient_accumulation_steps', default=1, type=int, help='')
parser.add_argument('--max_grad_norm', default=1.0, type=float, help='')
parser.add_argument('--output_dir', default='output_dir', type=str,
help='')
parser.add_argument('--is_block_shuffle', type=bool, default=True, help='')
parser.add_argument('--seed', type=int, default=2020, help='')
parser.add_argument('--max_len', type=int, default=256, help='')
parser.add_argument('--num_labels', type=int, default=6, help='')
return parser.parse_args()
def main():
args = set_args()
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.device
device = torch.device("cuda" if torch.cuda.is_available() and int(args.device) >= 0 else "cpu")
if args.seed:
torch.manual_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
model = SentimentAnalysisModel.from_pretrained(args.pretrained_model_path, num_labels=args.num_labels)
tokenizer = BertTokenizer.from_pretrained(args.vocab_path, do_lower_case=True)
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
train_ori_time(model, device, tokenizer, args)
train_block_shuffle_time(model, device, tokenizer, args)
train(model, device, tokenizer, args)
if __name__ == '__main__':
main()