-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathfinetune_bert.py
246 lines (226 loc) · 10.2 KB
/
finetune_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
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
#!/usr/bin/env python
# coding: utf-8
from __future__ import absolute_import, division, print_function
import os
import torch
import logging
import argparse
import numpy as np
import pandas as pd
import warnings
from ood_main import load_dataset
warnings.filterwarnings('ignore')
from ood_main import load_extra_dataset
from simpletransformers.classification import ClassificationModel
from simpletransformers.language_modeling import LanguageModelingModel
logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)
seed=42
def set_binary_label(dataframe, indomain, col='labels'):
if indomain:
dataframe[col].values[:] = 1
else:
dataframe[col].values[:] = 0
def train_process(args):
# load data
data_type = args.data_type.strip()
print("data type %s"%( data_type))
train_df, num_classes = load_dataset('clinc150_train', data_type=data_type)
test_df, _ = load_dataset('clinc150_test', data_type=data_type)
eval_df, _ = load_dataset('clinc150_val', data_type=data_type)
if args.model_class == 'bert':
model = ClassificationModel(
'bert',
'bert-base-uncased',
num_labels=num_classes,
use_cuda=True,
cuda_device=int(args.gpu_id),
args={'num_train_epochs': 5,
'fp16':False,
'n_gpu':int(args.n_gpu),
'learning_rate': 4e-5,
'warmup_ratio': 0.10,
'train_batch_size': 32,
'eval_batch_size': 32,
'evaluate_during_training': False,
'evaluate_during_training_steps': 2000,
'do_lower_case': True,
'reprocess_input_data': True,
'overwrite_output_dir': False,
'output_dir': './models/%s_outputs/'%(data_type),
'best_model_dir': "./models/%s_outputs/best_model"%(data_type),
'cache_dir': "./models/%s_cache_dir/"%(data_type)})
else:
raise NotImplementedError
model.train_model(train_df, eval_df=eval_df)
result, model_outputs, wrong_predictions = model.eval_model(test_df)
# pdb.set_trace()
print(result)
print("---------------")
def finetune_bcad(args):
# load data
sed = 42
data_type = args.data_type.strip()
print("data type %s"%( data_type))
if args.load_path:
output_dir = './models/{}_{}_ft_MLM_binary_intent_outputs_{}/'.format(args.model_class, data_type, str(args.neg_sample))
else:
output_dir = './models/{}_{}_ft_binary_intent_outputs_{}/'.format(args.model_class, data_type, str(args.neg_sample))
print (output_dir)
os.makedirs(output_dir, exist_ok=True)
if data_type == "sst":
train_df = load_extra_dataset("./dataset/sst/sst-train.txt")
eval_df = load_extra_dataset("./dataset/sst/sst-dev.txt")
test_df = load_extra_dataset("./dataset/sst/sst-test.txt")
for df in [train_df, test_df, eval_df]:
set_binary_label(df, indomain=True) # set label to one (in domain)
else:
train_df = load_dataset('clinc150_train', data_type=data_type)
test_df = load_dataset('clinc150_test', data_type=data_type)
eval_df = load_dataset('clinc150_val', data_type=data_type)
for df in [train_df, test_df, eval_df]:
set_binary_label(df, indomain=True) # set label to one (in domain)
# neg_df = load_extra_dataset("./dataset/sst-train.txt")
# neg_val_df = load_extra_dataset("./dataset/sst-dev.txt")
# neg_test_df = load_extra_dataset("./dataset/sst-test.txt")
print ("training df size", len(train_df))
book_df = load_extra_dataset("./dataset/bookcorpus/subset_books.txt")
set_binary_label(book_df, indomain=False)
wiki_df = load_extra_dataset("./dataset/wikipedia/squad_train_wiki.txt")
set_binary_label(wiki_df, indomain=False)
# pdb.set_trace()
# train_df = train_df.sample(n=200, random_state=seed)
neg_book_df = book_df.sample(n=args.neg_sample, random_state=seed)
neg_book_val_df = book_df.sample(n=1000, random_state=seed)
neg_book_test_df = book_df.sample(n=2000, random_state=seed)
neg_wiki_df = wiki_df.sample(n=args.neg_sample, random_state=seed)
neg_wiki_val_df = wiki_df.sample(n=1000, random_state=seed)
neg_wiki_test_df = wiki_df.sample(n=2000, random_state=seed)
if data_type != "sst":
neg_book_df['text'] = neg_book_df['text'].apply(lambda x: x.strip(".? "))
neg_book_val_df['text'] = neg_book_val_df['text'].apply(lambda x: x.strip(".? "))
neg_book_test_df['text'] = neg_book_test_df['text'].apply(lambda x: x.strip(".? "))
neg_wiki_df['text'] = neg_wiki_df['text'].apply(lambda x: x.strip(".? "))
neg_wiki_val_df['text'] = neg_wiki_val_df['text'].apply(lambda x: x.strip(".? "))
neg_wiki_test_df['text'] = neg_wiki_test_df['text'].apply(lambda x: x.strip(".? "))
train_df = pd.concat([train_df, neg_book_df, neg_wiki_df])
val_df = pd.concat([eval_df, neg_book_val_df, neg_wiki_val_df])
print ("train_df", len(train_df['labels']), train_df['labels'][10:20], train_df['text'][0:10])
if args.load_path:
load_path = args.load_path
else:
load_path = 'bert-base-uncased' if args.model_class == 'bert' else 'roberta-base'
if args.model_class == 'bert' or args.model_class == 'roberta':
model = ClassificationModel(
args.model_class,
load_path,
num_labels=2,
use_cuda=True,
cuda_device=int(args.gpu_id),
args={'num_train_epochs': 2,
'fp16':False,
'n_gpu':int(args.n_gpu),
'learning_rate': 4e-5,
'warmup_ratio': 0.10,
'do_lower_case': True,
"max_seq_length": 256,
"train_batch_size": 16,
'reprocess_input_data': True,
'overwrite_output_dir': True,
'save_model_every_epoch': False,
'evaluate_during_training': True,
'evaluate_during_training_verbose': True,
'evaluate_during_training_steps': 2000,
'output_dir': output_dir,
'best_model_dir': "{}/best_model".format(output_dir),
'cache_dir': output_dir.replace("outputs", "cache_dir")})
else:
raise NotImplementedError
# model.train_model(pd.concat([train_df, neg_df]), eval_df=pd.concat([eval_df, neg_val_df]))
model.train_model(train_df, eval_df=eval_df)
# result, model_outputs, wrong_predictions = model.eval_model(pd.concat([test_df, neg_test_df]))
result, model_outputs, wrong_predictions = model.eval_model(eval_df)
# pdb.set_trace()
print(result)
print (output_dir)
print("---------------")
def finetune_imlm(args):
# load data
data_type = args.data_type.strip()
print("data type %s"%( data_type))
output_dir = './models/{}_{}_mlm_ft_outputs/'.format(args.model_class, data_type)
os.makedirs(output_dir, exist_ok=True)
train_file = os.path.join(output_dir, 'train.txt')
eval_file = os.path.join(output_dir, 'eval.txt')
test_file = os.path.join(output_dir, 'test.txt')
if not (os.path.exists(train_file) and os.path.exists(eval_file)):
if data_type == "sst":
train_df = load_extra_dataset("./dataset/sst/sst-train.txt", label=1)
eval_df = load_extra_dataset("./dataset/sst/sst-dev.txt", label=1)
# test_df = load_extra_dataset("./dataset/sst-test.txt", label=1)
else:
train_df = load_dataset('clinc150_train', data_type=data_type)
# test_df = load_dataset('clinc150_test', data_type=data_type)
eval_df = load_dataset('clinc150_val', data_type=data_type)
# pdb.set_trace()
print ("number of training instances", len(train_df['text']))
with open(train_file, "w") as f:
for line in train_df['text']:
f.write(line + "\n")
print ("number of eval instances", len(eval_df['text']))
with open(eval_file, "w") as f:
for line in eval_df['text']:
f.write(line + "\n")
if args.model_class == 'bert' or args.model_class == 'roberta':
model = LanguageModelingModel(
args.model_class,
'bert-base-uncased' if args.model_class == 'bert' else 'roberta-base',
use_cuda=True,
cuda_device=int(args.gpu_id),
args={'num_train_epochs': 10,
'fp16':False,
'n_gpu':int(args.n_gpu),
'do_lower_case': True,
"max_seq_length": 128,
"train_batch_size": 4,
'reprocess_input_data': True,
'overwrite_output_dir': True,
'save_model_every_epoch': False,
'evaluate_during_training': True,
'output_dir': output_dir,
'best_model_dir': "{}/best_model".format(output_dir),
'cache_dir': output_dir.replace("outputs", "cache_dir")})
else:
raise NotImplementedError
model.train_model(train_file, eval_file=eval_file)
model.eval_model(eval_file)
# print(result)
print("---------------")
def main(args):
run_type = args.type
# set seed
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
torch.cuda.set_device(args.gpu_id)
if run_type == 'train_classifier':
train_process(args)
elif run_type == 'finetune_bcad':
finetune_bcad(args)
elif run_type == 'finetune_imlm':
finetune_imlm(args)
else:
raise NotImplementedError
if __name__ == '__main__':
parser = argparse.ArgumentParser("Bert Model OOD Fine-tuning")
parser.add_argument('--type', default='finetune_bcad', choices=['train_classifier', 'finetune_bcad', 'finetune_imlm'])
parser.add_argument('--model_class', default='bert', type=str)
parser.add_argument('--gpu_id', default=0, type=int)
parser.add_argument('--n_rep', default=1, type=int)
parser.add_argument('--n_gpu', default=1, type=int)
parser.add_argument('--load_path', default=None, type=str)
parser.add_argument('--data_type', default='full', choices = ['clinic', 'sst'])
parser.add_argument('--neg_sample', default=7500, type=int)
args = parser.parse_args()
main(args)