/
train.py
211 lines (153 loc) · 8.47 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
from kobert_transformers import get_kobert_model
from kobert_transformers import get_tokenizer
from sklearn.model_selection import train_test_split
from tqdm.auto import tqdm
import torch.nn as nn
import torch
import argparse
import pandas as pd
import os
import datetime
from datasets import(
create_data_loader,
tuplify_with_device,
)
from loss_func import QuoteCSELoss
from models import Encoder
from util import make_pair, AverageMeter, set_seed
from pytorchtools import EarlyStopping
def main():
parser = argparse.ArgumentParser()
# arguments
parser.add_argument("--seed", default=123, type=int, help="set seed")
parser.add_argument("--batch_size", default=8, type=int, help="batch size")
parser.add_argument("--max_len", default=512, type=int, help="max length")
parser.add_argument("--num_workers", default=16, type=int, help="number of workers")
parser.add_argument("--dimension_size", default=768, type=int, help="dimension size")
parser.add_argument("--hidden_size", default=100, type=int, help="hidden size")
parser.add_argument("--learning_rate", default=1e-6, type=float, help="learning rate")
parser.add_argument("--weight_decay", default=1e-7, type=float, help="weight decay")
parser.add_argument("--epochs", default=50, type=int, help="epoch")
parser.add_argument("--static_epochs", default=14, type=int, help="epoch for static")
parser.add_argument("--dynamic_epochs", default=2, type=int, help="epoch for dynamic")
parser.add_argument("--temperature", default=0.05, type=float, help="temperature")
parser.add_argument("--assignment", default='static', type=str, help="assignment type")
parser.add_argument("--MODEL_DIR", default='./model/', type=str, help="where to save the trained model")
parser.add_argument("--MODIFIED_DATA_PATH", default='./data/modified_sample.pkl', type=str, help="data for pretraining")
parser.add_argument("--VERBATIM_DATA_PATH", default='./data/verbatim_sample.pkl', type=str, help="data for pretraining")
args = parser.parse_args()
if not os.path.exists(args.MODEL_DIR):
os.makedirs(args.MODEL_DIR)
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"
os.environ['WANDB_CONSOLE'] = 'off'
set_seed(args.seed)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args.batch_size = args.batch_size * torch.cuda.device_count()
args.backbone_model = get_kobert_model()
args.tokenizer = get_tokenizer()
loss_func = QuoteCSELoss(temperature=args.temperature, batch_size=args.batch_size)
encoder = Encoder(args)
encoder = nn.DataParallel(encoder)
encoder = encoder.to(args.device)
optimizer = torch.optim.Adam(encoder.parameters(), lr=args.learning_rate, weight_decay=args.weight_decay)
optimizer.zero_grad()
print('Making Dataloader')
modified_df = pd.read_pickle(args.MODIFIED_DATA_PATH)
verbatim_df = pd.read_pickle(args.VERBATIM_DATA_PATH)
train_modified_df, test_modified_df = train_test_split(modified_df, test_size=0.2, random_state=args.seed)
valid_modified_df, test_modified_df = train_test_split(test_modified_df, test_size=0.5, random_state=args.seed)
train_verbatim_df, test_verbatim_df = train_test_split(verbatim_df, test_size=0.2, random_state=args.seed)
valid_verbatim_df, test_verbatim_df = train_test_split(test_verbatim_df, test_size=0.5, random_state=args.seed)
train_df = pd.concat([train_modified_df, train_verbatim_df])
valid_df = pd.concat([valid_modified_df, valid_verbatim_df])
test_df = pd.concat([test_modified_df, test_verbatim_df])
train_data_loader = create_data_loader(args,
df = train_df,
shuffle = True,
drop_last = True)
valid_data_loader = create_data_loader(args,
df = valid_df,
shuffle = False,
drop_last = True)
early_stopping = EarlyStopping(patience = 3, verbose = True, path=args.MODEL_DIR + 'checkpoint_static_dynamic_early.bin')
# train
print('Start Training')
loss_data = []
stop = False
encoder.train()
for epoch in range(args.epochs):
if epoch >= args.static_epochs:
args.assignment = 'dynamic'
if epoch >= args.dynamic_epochs + args.static_epochs:
break
losses = AverageMeter()
valid_loss = []
tbar1 = tqdm(train_data_loader)
tbar2 = tqdm(valid_data_loader)
for title, body, body_len, pos_idx, neg_idx in tbar1:
title_id, title_at = title['input_ids'].to(args.device).long(), title['attention_mask'].to(args.device).long()
b_ids = []
b_atts = []
for b in range(len(body_len)):
i = body_len[b]
b_id, b_at = body['input_ids'][b][:i].to(args.device).long(), body['attention_mask'][b][:i].to(args.device).long()
b_ids.append(b_id)
b_atts.append(b_at)
body_ids = torch.cat(b_ids, dim=0)
body_atts = torch.cat(b_atts, dim=0)
if args.assignment == 'static':
pos_body_ids, neg_body_ids, pos_body_atts, neg_body_atts = make_pair(args, body, title_id, title_at, body_ids, body_atts, body_len, encoder, pos_idx, neg_idx)
elif args.assignment == 'dynamic':
pos_body_ids, neg_body_ids, pos_body_atts, neg_body_atts = make_pair(args, body, title_id, title_at, body_ids, body_atts, body_len, encoder)
del body_ids, body_atts, body_len
outputs = encoder(
input_ids = torch.cat([title_id, pos_body_ids, neg_body_ids]),
attention_mask = torch.cat([title_at, pos_body_atts, neg_body_atts]),
)
loss = loss_func(outputs)
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses.update(loss.item(), args.batch_size)
tbar1.set_description("loss: {0:.6f}".format(losses.avg), refresh=True)
del title_id, pos_body_ids, neg_body_ids, title_at, pos_body_atts, neg_body_atts, outputs, loss
ts = datetime.datetime.now().timestamp()
loss_data.append([epoch, losses.avg, 'Train', ts])
# valid
with torch.no_grad():
for title, body, body_len, pos_idx, neg_idx in tbar2:
title_id, title_at = title['input_ids'].to(args.device).long(), title['attention_mask'].to(args.device).long()
b_ids = []
b_atts = []
for b in range(len(body_len)):
i = body_len[b]
b_id, b_at = body['input_ids'][b][:i].to(args.device).long(), body['attention_mask'][b][:i].to(args.device).long()
b_ids.append(b_id)
b_atts.append(b_at)
body_ids = torch.cat(b_ids, dim=0)
body_atts = torch.cat(b_atts, dim=0)
if args.assignment == 'static':
pos_body_ids, neg_body_ids, pos_body_atts, neg_body_atts = make_pair(args, body, title_id, title_at, body_ids, body_atts, body_len, encoder, pos_idx, neg_idx)
elif args.assignment == 'dynamic':
pos_body_ids, neg_body_ids, pos_body_atts, neg_body_atts = make_pair(args, body, title_id, title_at, body_ids, body_atts, body_len, encoder)
del body_ids, body_atts, body_len
outputs = encoder(
input_ids = torch.cat([title_id, pos_body_ids, neg_body_ids]),
attention_mask = torch.cat([title_at, pos_body_atts, neg_body_atts]),
)
loss = loss_func(outputs)
valid_loss.append(loss.item())
del title_id, pos_body_ids, neg_body_ids, title_at, pos_body_atts, neg_body_atts, outputs, loss
avg_valid_loss = sum(valid_loss) / len(valid_loss)
ts = datetime.datetime.now().timestamp()
loss_data.append([epoch, avg_valid_loss, 'Valid', ts])
print(str(epoch), 'th epoch, Avg Valid Loss: ', str(avg_valid_loss))
early_stopping(avg_valid_loss, encoder)
if early_stopping.early_stop:
break
torch.save(encoder.state_dict(), args.MODEL_DIR + 'checkpoint.bin')
# save loss
df_loss = pd.DataFrame(loss_data, columns=('Epoch', 'Loss', 'Type', 'Time'))
df_loss.to_csv(args.MODEL_DIR + 'loss.csv', sep=',', index=False)
if __name__ == "__main__":
main()