-
Notifications
You must be signed in to change notification settings - Fork 254
/
Copy pathtext_classification.py
196 lines (179 loc) · 8.11 KB
/
text_classification.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
r"""
文本分类任务之多分类教程，全流程直观展示，适合NLP入门人员
"""
import os.path
import sys
sys.path.append('..')
import argparse
from torch import nn
import torch.nn.functional as F
from torch.optim import AdamW
from torch.utils.data import DataLoader
from prettytable import PrettyTable
from transformers import get_linear_schedule_with_warmup
from transformers import BertModel, BertTokenizer, DataCollatorWithPadding
from codes.nlper.modules.mlp import MLP
from codes.nlper.utils import DatasetCLF
from codes.nlper.modules.trainer import Trainer
from codes.nlper.modules.metrics import Metrics, PMetric, RMetric, F1Metric
from codes.nlper.utils.format_convert import smp2020_ewect_convert
from codes.nlper.utils import seed_everything, set_devices, Dict2Obj, dataset_names
class BertCLF(nn.Module):
def __init__(self, args):
super(BertCLF, self).__init__()
self.bert = BertModel.from_pretrained(args.pretrained_model)
self.dropout = nn.Dropout(self.bert.config.hidden_dropout_prob)
self.clf = MLP([self.bert.config.hidden_size, args.num_class],
'tanh',
dropout=args.dropout)
def forward(self, input_ids, attention_mask, token_type_ids, **kwargs):
outputs = self.bert(input_ids, attention_mask, token_type_ids)
logits = self.clf(outputs[1])
return logits
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# general
parser.add_argument('--device_ids', default=[0],
help='the GPU ID for running model and we only support single gpu now, [-1] means using CPU')
parser.add_argument('--seed', default=1000,
help='random seed for reproduction')
# dataset
parser.add_argument('--dataset_name', default='text_clf/smp2020-ewect-usual',
help="if train/val/test can't be found, automatic download")
parser.add_argument('--dataset_cache_dir', default='../data/smp2020-ewect-usual')
parser.add_argument('--num_class', default=6,
help='the number of class')
parser.add_argument('--train_file', default='train.tsv',
help='relative path for dataset_cache_dir')
parser.add_argument('--val_file', default='dev.tsv',
help='relative path for dataset_cache_dir')
parser.add_argument('--test_file', default='test.tsv',
help='relative path for dataset_cache_dir')
# train
parser.add_argument('--is_train', default=True,
help='whether to train')
parser.add_argument('--is_test', default=True,
help='whether to test')
parser.add_argument('--pretrained_model', default='bert-base-chinese',
help='the name or path of pretrained language model')
parser.add_argument('--checkpoint', default='',
help='continue to train from checkpoint')
parser.add_argument('--num_epochs', default=2,
help='the number of training epochs')
parser.add_argument('--max_len', default=512,
help='the upper bound of input sentence')
parser.add_argument('--train_batch_size', default=8)
parser.add_argument('--test_batch_size', default=8)
parser.add_argument('--lr', default=3e-5)
parser.add_argument("--warmup_steps", type=int, default=200,
help="the number of steps to warm up optimizer")
parser.add_argument("--weight_decay", type=float, default=0.01,
help="l2 reg")
parser.add_argument('--dropout', default=0.1)
# save
parser.add_argument('--best_model_path', default='saved/model.bin',
help='the path to save model with the highest performance')
# parser.add_argument('--model_saved',
# default='saved/model.bin',
# help='the path of saved model')
parser.add_argument('--pred_saved', default='saved/pred.txt',
help='the path to save prediction of test')
# print arguments
args = parser.parse_args()
seed_everything(args.seed)
device = set_devices(args.device_ids)
table = PrettyTable(['Param', 'Value'])
for item in vars(args):
table.add_row([item, vars(args)[item]])
table.align['Param'] = 'l'
table.align['Value'] = 'l'
print(table)
# load data
print('load data')
tokenizer = BertTokenizer.from_pretrained(args.pretrained_model)
args.train_file = os.path.join(args.dataset_cache_dir, args.train_file)
args.val_file = os.path.join(args.dataset_cache_dir, args.val_file)
args.test_file = os.path.join(args.dataset_cache_dir, args.test_file)
# 如果三者同时不存在（适用于第一次运行该代码）
if not (
os.path.isfile(args.train_file)
or os.path.isfile(args.val_file)
or os.path.isfile(args.test_file)
):
# 自动下载数据集
corpus = dataset_names[args.dataset_name](cache_dir=args.dataset_cache_dir)
is_over = corpus.prepare_data()
if not is_over:
print(f'please download dataset manually, and make sure data file path is correct')
exit()
train_data = smp2020_ewect_convert(args.train_file)
val_data = smp2020_ewect_convert(args.val_file)
test_data = smp2020_ewect_convert(args.test_file)
print(f'sentence number of train: {len(train_data)}')
print(f'sentence number of dev: {len(val_data)}')
print(f'sentence number of test: {len(test_data)}')
# padding
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
# dataset
train_dataset = DatasetCLF(train_data, tokenizer, max_len=args.max_len)
val_dataset = DatasetCLF(val_data, tokenizer, max_len=args.max_len)
test_dataset = DatasetCLF(test_data, tokenizer, max_len=args.max_len)
# dataloader
train_loader = DataLoader(train_dataset, batch_size=args.train_batch_size, shuffle=True, collate_fn=data_collator)
val_loader = DataLoader(val_dataset, batch_size=args.test_batch_size, shuffle=False, collate_fn=data_collator)
test_loader = DataLoader(test_dataset, batch_size=args.test_batch_size, shuffle=False, collate_fn=data_collator)
print(f'batch number of train: {len(train_loader)}')
print(f'batch number of dev: {len(val_loader)}')
print(f'batch number of test: {len(test_loader)}')
print('----------------------------------')
# create model
print('create model')
model = BertCLF(args).to(device)
print(model)
print('----------------------------------')
# create optimizer
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
'weight_decay': args.weight_decay},
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
'weight_decay': 0.0}
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.lr)
# create scheduler
scheduler = get_linear_schedule_with_warmup(
optimizer,
args.warmup_steps,
args.num_epochs * len(train_loader)
)
# create evaluation metric
precision_metric = PMetric(average='macro')
recall_metric = RMetric(average='macro')
f1_metric = F1Metric(average='macro')
metrics = Metrics(
metrics={
precision_metric.name: precision_metric,
recall_metric.name: recall_metric,
f1_metric.name: f1_metric
},
target_metric=f1_metric.name # used for early stop and model saving
)
# load config
trainer_config = Dict2Obj(vars(args))
trainer_config.update({
'task_name': 'text_clf',
'train_loader': train_loader,
'val_loader': val_loader,
'test_loader': test_loader,
'device': device,
'loss_fn': F.cross_entropy,
'optimizer': optimizer,
'scheduler': scheduler,
'metrics': metrics
})
# running
trainer = Trainer(model, trainer_config)
if args.is_train:
trainer.train()
if args.is_test:
trainer.test()