-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
185 lines (167 loc) · 8.41 KB
/
main.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
#coding:utf8
from __future__ import absolute_import, division, print_function, unicode_literals
import argparse
from shutil import copyfile
import random
from io import open
import os
from torch.autograd import Variable
import numpy as np
import argparse
import torch
from torch import nn,optim
from torch.utils.data import DataLoader, Dataset, RandomSampler
from torch.utils.data.distributed import DistributedSampler
from torch.autograd import detect_anomaly
from tqdm import tqdm, trange
from modeling import BilingualModel
from preprocessing import DataProvider
from modeling.optimization import BertAdam, warmup_linear
from torch.utils.data import Dataset
import random
from utils import *
from config import config
import logging
import time
from datetime import datetime
from torch.cuda import amp
import pytorch_warmup as warmup
from thop import profile
logging.basicConfig(format='%(levelname)s - %(asctime)s - %(message)s',
datefmt='%d %H:%M:%S',
level=logging.INFO)
logger = logging.getLogger(__name__)
def evaluate(val_loader, model, epoch, device):
val_progressor = ProgressBar(mode="Vali",\
epoch=epoch,\
total_epoch=config.num_train_epochs,\
model_name=config.model_name,\
total=len(val_loader))
model.to(device)
losses = AverageMeter()
avg_recall = AverageMeter()
with torch.no_grad():
for i, batch in enumerate(val_loader):
val_progressor.current = i
batch = tuple(t.to(device) for t in batch)
sent1_ids, sent1_len, sent2_ids, sent2_len, lm_label_id, masked_la = batch
loss, recall = model(sent1_ids, sent1_len, sent2_ids, sent2_len, lm_label_id, masked_la)
losses.update(torch.mean(loss).item(), sent1_ids.size(0))
avg_recall.update(torch.mean(recall).item(), sent1_ids.size(0))
val_progressor.current_loss = losses.val
val_progressor.avg_loss = losses.avg
val_progressor.cur_time = str(datetime.now().strftime('%d %H:%M:%S'))
if i%100==0:
val_progressor()
val_progressor.done()
return losses.avg
def main():
if not config.resume:
if not os.path.exists(config.output_dir):
os.mkdir(config.output_dir)
if os.path.exists(os.path.join(config.output_dir, config.model_name)):
if os.path.exists(os.path.join(config.output_dir, config.model_name, config.best_models)) and os.listdir(os.path.join(config.output_dir, config.model_name, config.best_models)):
logger.error('Exists best checkpoints in {}. Please Check.'.format(os.path.join(config.output_dir, config.model_name, config.best_models)))
exit(0)
if os.path.exists(os.path.join(config.output_dir, config.model_name)):
shutil.rmtree(os.path.join(config.output_dir , config.model_name))
os.mkdir(os.path.join(config.output_dir , config.model_name))
os.mkdir(os.path.join(config.output_dir , config.model_name, config.best_models))
shutil.copy(config.config_path, os.path.join(config.output_dir, config.model_name, 'config_'+config.model_name +'.py'))
logger.info('Copying config.py to {}'.format(os.path.join(config.output_dir,config.model_name, 'config_'+config.model_name)))
device = torch.device("cuda:{}".format(torch.cuda.current_device()) if torch.cuda.is_available() else "cpu")
model = BilingualModel(config.vocab_size, config)
if torch.cuda.device_count()>1:
model = model.cuda()
device_ids = list(range(torch.cuda.device_count()))
model = nn.DataParallel(model, device_ids=device_ids)
model.to(device)
optimizer = optim.Adam(model.parameters(), lr=config.learning_rate, weight_decay=config.weight_decay, amsgrad=True)
start_epoch = 0
best_loss = np.inf
best_loss_save = np.inf
resume = config.resume
if resume:
if not os.path.exists(os.path.join(config.output_dir, config.model_name)):
logger.error('No model found in path: {}'.format(os.path.join(config.output_dir, config.model_name)))
checkpoint = torch.load(os.path.join(config.output_dir, config.model_name, config.best_models, 'model_best.pth.tar'))
old_state = checkpoint['state_dict']
start_epoch = checkpoint["epoch"]
best_loss = checkpoint["best_loss"]
model.load_state_dict(old_state)
optimizer.load_state_dict(checkpoint["optimizer"])
dataloader = DataProvider(config, True, 'train')
train_data, validation_data = dataloader.data_loader, dataloader.vali_loader
num_train_optimization_steps = int(dataloader.dataset.num_samples / config.train_batch_size) * config.num_train_epochs
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[config.lr_decay_from], gamma=0.1)
warmup_scheduler = warmup.UntunedLinearWarmup(optimizer)
warmup_scheduler.last_step = -1
global_step = 0
logger.info("***** Running training *****")
logger.info(" Model Name = %s", config.model_name)
logger.info(" bpe_model_path = %s", config.bpe_path)
logger.info(" Num examples = %d", train_data.dataset.num_samples)
logger.info(" Batch size = %d", config.train_batch_size)
logger.info(" Num steps per epoch = %d", num_train_optimization_steps/config.num_train_epochs + 1)
logger.info(" Has FC: %s", str(config.has_FC))
logger.info(" Has sentence-alignment loss: %s", str(config.has_sentence_loss))
logger.info(" Has sentence-similarity loss: %s", str(config.has_sentence_similarity_loss))
model.train()
scaler = amp.GradScaler()
for epoch in range(start_epoch, int(config.num_train_epochs)):
logger.info('Start new epoch {}'.format(str(epoch)))
tr_loss = AverageMeter()
train_progressor = ProgressBar(mode="Train",\
epoch=epoch,\
total_epoch=int(config.num_train_epochs),\
model_name=config.model_name,\
total=len(train_data))
for step, batch in enumerate(train_data):
lr_scheduler.step(epoch)
warmup_scheduler.dampen()
batch = tuple(t.to(device) for t in batch)
sent1_ids, sent1_len, sent2_ids, sent2_len, lm_label_id, masked_la = batch
with amp.autocast():
loss, recall = model(sent1_ids, sent1_len, sent2_ids, sent2_len, lm_label_id, masked_la)
tr_loss.update(torch.mean(loss).item(), batch[0].size(0))
train_progressor.current = step + 1
train_progressor.current_loss = tr_loss.val
train_progressor.avg_loss = tr_loss.avg
train_progressor.cur_time = str(datetime.now().strftime('%d %H:%M:%S'))
if step % 10000 == 0:
train_progressor()
logger.info("Token Classification Acc: Not Defined")
optimizer.zero_grad()
scaler.scale(torch.mean(loss)).backward()
scaler.step(optimizer)
global_step += 1
scaler.update()
train_progressor.done()
eval_loss = loss
if config.has_validation:
valid_loss = evaluate(validation_data, model, epoch, device)
eval_loss = valid_loss
is_best = eval_loss <= best_loss
best_loss = min(eval_loss, best_loss)
try:
best_loss_save = best_loss.cpu().data.numpy()
except:
pass
logger.info('Trying to save model.')
save_checkpoint({
"epoch": epoch + 1,
"model_name": config.model_name,
"state_dict": model.state_dict(),
"best_loss": best_loss,
"optimizer": optimizer.state_dict(),
"valid_loss": eval_loss,
},is_best,epoch)
if __name__ =="__main__":
parser = argparse.ArgumentParser()
parser.add_argument('-m', dest='modelName', default='', help='modelName')
parser.add_argument('-r', dest='resume', default=False, help='resume training')
parser.add_argument('-t', dest='is_train', default=True, help='is train?')
parser.add_argument('-la', dest='target_la', default='fr', help='target language')
args = parser.parse_args()
config.set_config(args.modelName, args.resume, args.is_train)
main()