-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
197 lines (184 loc) · 9.61 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
import os
import torch
import argparse
from loguru import logger
from datetime import datetime
from torch.nn import DataParallel
from transformers import AdamW
from lib.data_loader import get_data_loader
from lib.models.networks import get_model, get_tokenizer
from lib.training.common import train_common, test_acc
from lib.training.RecAdam import RecAdam
from lib.exp import get_num_labels, seed_everything
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--device', default='0,1,2,3',
type=str, required=False, help='GPU ids')
parser.add_argument('--model', default='roberta-base',help='pretrained model type')
parser.add_argument('--pretrained_model', default=None,
type=str, required=False, help='the path of the model to load')
parser.add_argument('--dataset', default='sst-2', help='training dataset')
parser.add_argument('--epochs', default=5, type=int,
required=False, help='number of training epochs')
parser.add_argument('--batch_size', default=16, type=int,
required=False, help='training batch size')
parser.add_argument('--lr', default=2e-5, type=float,
required=False, help='learning rate')
parser.add_argument("--weight_decay", default=0.0,
type=float, help="Weight decay if we apply some.")
parser.add_argument("--shift_reg", default=0, type=float)
parser.add_argument('--log_step', default=100, type=int,required=False)
parser.add_argument('--max_grad_norm', default=1.0,
type=float, required=False)
parser.add_argument('--output_dir', default='saved_model/',
type=str, required=False, help='save directory')
parser.add_argument('--save_every_epoch', action='store_true',
help='save checkpoint every epoch')
parser.add_argument('--output_name', default='model.pt',
type=str, required=False, help='model save name')
parser.add_argument('--log_file', type=str, default='./log/default.log')
parser.add_argument('--seed', type=int, default=42)
parser.add_argument('--optimizer', type=str,default='adam', choices=['adam', 'recadam'])
parser.add_argument("--recadam_anneal_fun", type=str, default='sigmoid',
choices=["sigmoid", "linear", 'constant'],
help="the type of annealing function in RecAdam. Default sigmoid")
parser.add_argument("--recadam_anneal_k", type=float, default=0.2,
help="k for the annealing function in RecAdam.")
parser.add_argument("--recadam_anneal_t0", type=int, default=1000,
help="t0 for the annealing function in RecAdam.")
parser.add_argument("--recadam_anneal_w", type=float, default=1.0,
help="Weight for the annealing function in RecAdam. Default 1.0.")
parser.add_argument("--recadam_pretrain_cof", type=float, default=5000.0,
help="Coefficient of the quadratic penalty in RecAdam. Default 5000.0.")
parser.add_argument("--loss_type", default='ce',choices=['ce', 'scl', 'margin'])
parser.add_argument("--scl_reg", default=2.0, type=float)
parser.add_argument("--eval_metric", default='acc',type=str, choices=['acc', 'f1'])
parser.add_argument("--save_steps", default=-1, type=int)
args = parser.parse_args()
seed_everything(args.seed)
# args setting
log_file_name = args.log_file
logger.add(log_file_name)
logger.info('args:\n' + args.__repr__())
output_dir = args.output_dir
output_name = args.output_name
if not os.path.exists(output_dir):
os.makedirs(output_dir)
num_labels = get_num_labels(args.dataset)
args.num_labels = num_labels
# model loading
model = get_model(args)
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
model.to(device)
logger.info("{} model loaded".format(args.model))
if args.pretrained_model:
model.load_state_dict(torch.load(args.pretrained_model))
logger.info("model loaded from {}".format(args.pretrained_model))
if torch.cuda.device_count() > 1:
logger.info("Let's use " + str(len(args.device.split(','))) + " GPUs!")
model = DataParallel(model, device_ids=[int(i) for i in args.device.split(',')])
tokenizer = get_tokenizer(args.model)
logger.info("{} tokenizer loaded".format(args.model))
if tokenizer.pad_token_id is None:
tokenizer.pad_token = tokenizer.eos_token
# data loading
train_loader = get_data_loader(args.dataset, 'train', tokenizer, args.batch_size, shuffle=True)
val_loader = get_data_loader(
args.dataset, 'dev', tokenizer, args.batch_size)
logger.info("dataset {} loaded".format(args.dataset))
logger.info("num_labels: {}".format(num_labels))
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
model.train()
model.to(device)
pretrained_model = get_model(args)
pretrained_model.to(device)
if torch.cuda.device_count() > 1:
pretrained_model = DataParallel(pretrained_model, device_ids=[int(i) for i in args.device.split(',')])
pretrained_model.eval()
no_decay = ["bias", "LayerNorm.weight"]
if args.optimizer == 'adam':
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)
elif args.optimizer == 'recadam':
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if
not any(nd in n for nd in no_decay) and 'bert' in n.lower()],
"weight_decay": args.weight_decay,
"anneal_w": args.recadam_anneal_w,
"pretrain_params": [p_p for p_n, p_p in pretrained_model.named_parameters() if
not any(nd in p_n for nd in no_decay) and 'bert' in p_n.lower()]
},
{
"params": [p for n, p in model.named_parameters() if
not any(nd in n for nd in no_decay) and 'bert' not in n.lower()],
"weight_decay": args.weight_decay,
"anneal_w": 0.0,
"pretrain_params": [p_p for p_n, p_p in pretrained_model.named_parameters() if
not any(nd in p_n for nd in no_decay) and 'bert' not in p_n.lower()]
},
{
"params": [p for n, p in model.named_parameters() if
any(nd in n for nd in no_decay) and 'bert' in n.lower()],
"weight_decay": 0.0,
"anneal_w": args.recadam_anneal_w,
"pretrain_params": [p_p for p_n, p_p in pretrained_model.named_parameters() if
any(nd in p_n for nd in no_decay) and 'bert' in p_n.lower()]
},
{
"params": [p for n, p in model.named_parameters() if
any(nd in n for nd in no_decay) and 'bert' not in n.lower()],
"weight_decay": 0.0,
"anneal_w": 0.0,
"pretrain_params": [p_p for p_n, p_p in pretrained_model.named_parameters() if
any(nd in p_n for nd in no_decay) and 'bert' not in p_n.lower()]
}
]
optimizer = RecAdam(optimizer_grouped_parameters, lr=args.lr,
anneal_fun=args.recadam_anneal_fun, anneal_k=args.recadam_anneal_k,
anneal_t0=args.recadam_anneal_t0, pretrain_cof=args.recadam_pretrain_cof)
logger.info('starting training')
best_acc = 0
step_counter = 0
for epoch in range(args.epochs):
model.train()
pretrained_model.eval()
logger.info("epoch {} start".format(epoch))
start_time = datetime.now()
step_counter = train_common(model, optimizer, train_loader,
epoch, log_steps=args.log_step, pre_model=pretrained_model, shift_reg=args.shift_reg,
loss_type=args.loss_type, scl_reg=args.scl_reg,
save_steps=args.save_steps, save_dir=output_dir, step_counter=step_counter)
acc = test_acc(model, val_loader, args.eval_metric)
logger.info("epoch {} validation {}: {:.4f} ".format(epoch, args.eval_metric, acc))
if acc > best_acc:
best_acc = acc
logger.info("best validation {} improved to {:.4f}".format(
args.eval_metric, best_acc))
model_to_save = model.module if hasattr(model, 'module') else model
torch.save(model_to_save.state_dict(),'{}/{}'.format(output_dir, output_name))
logger.info("model saved to {}/{}".format(output_dir, output_name))
if args.save_every_epoch:
model_to_save = model.module if hasattr(model, 'module') else model
torch.save(model_to_save.state_dict(),'{}/epoch{}_{}'.format(output_dir, epoch, output_name))
logger.info("model saved to {}/epoch{}_{}".format(output_dir, epoch, output_name))
end_time = datetime.now()
logger.info('time for one epoch: {}'.format(end_time - start_time))
logger.info("training finished")
test_loader = get_data_loader(args.dataset, 'test', tokenizer, args.batch_size)
model = get_model(args)
model.to(device)
model.load_state_dict(torch.load('{}/{}'.format(output_dir, output_name)))
logger.info("best model loaded")
logger.info("test {}: {:.4f}".format(args.eval_metric,test_acc(model, test_loader, args.eval_metric)))
if __name__ == '__main__':
main()