-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
154 lines (126 loc) · 5.37 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
# coding:utf-8
import torch
import torch.nn as nn
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
import time
import argparse
import os
import tqdm
from torch.utils.data import DataLoader
from torch.autograd import Variable
from tensorboardX import SummaryWriter
from senet import se_resnet50
from dataset import custom_Dataset
from config import cfg
from loss import CELoss
from eval import eval
from utils.Tools import *
def get_args():
parses = argparse.ArgumentParser(description = 'Train config')
#parses.add_argument('--gpus' ,type=str,default='0,1,2,3')
parses.add_argument('--model_path',type=str,default=None)
parses.add_argument('--start_epoch',type=int,default=0)
args = parses.parse_args()
return args
def train(model,optimizer,scheduler,cfg,args):
trainsets = custom_Dataset(cfg,phase = 'train')
trainloader = DataLoader(trainsets, num_workers=4,batch_size=cfg['batch_size'],shuffle=True)
valsets = custom_Dataset(cfg, phase='val')
valloader = DataLoader(valsets, num_workers=4, batch_size=cfg['batch_size'], shuffle=True)
out_dir = '{}_{}_{}'.format(cfg['model_name'], time.strftime("%Y%m%d"),time.strftime("%H%M%S"))
criterion = CELoss()
save_dir = os.path.join(cfg['checkpoint_dir'],out_dir)
writer = SummaryWriter(save_dir)
model_path = os.path.join(save_dir, '{}_epoch{}.pth')
create_dir(save_dir)
model.train(True)
model.apply(weight_init)
model.cuda()
model = torch.nn.DataParallel(model, device_ids=[0,1,2,3])
#断点重训
if args.model_path != None:
if args.start_epoch == None:
print('input epoch !')
exit(0)
model.load_state_dict(torch.load(args.model_path))
lr = cfg['base_lr'] * (cfg['gamma'] ** int(args.start_epoch))
print('change base lr {} to {} at start epoch {}'.format(cfg['base_lr'],lr,args.start_epoch))
for param_group in optimizer.param_groups:
param_group["lr"] = lr
device = "cuda" if torch.cuda.is_available() else "cpu"
step = 0
best_score = 0
best_epoch = -1
min_loss = 100.0
tic_batch = time.time()
for epoch in range(args.start_epoch,cfg['epochs']):
train_loss = 0.0
train_acc = 0.0
for idx, (imgs, labels) in enumerate(trainloader):
#print(imgs.shape)
optimizer.zero_grad()
if device == 'cuda':
imgs = Variable(imgs.cuda())
labels = Variable(labels.cuda())
else:
inputs, labels = Variable(imgs), Variable(labels)
preds = model(imgs)
loss = criterion(preds, labels)
output = torch.argmax(preds,dim=1)
loss.backward()
optimizer.step()
train_loss += loss.item()
#print(output,labels)
cur_correctrs = torch.sum(output == labels.data)
train_acc += cur_correctrs.float()
batch_acc = (cur_correctrs.float()) / (len(output))
if idx % cfg['print_freq'] == 0:
print('[Epoch {}/{}]-[batch:{}/{}] lr={:.6f} Loss: {:.6f} Acc: {:.4f} Time: {:.4f}batch/sec'.format(
epoch+1, cfg['epochs'], idx + 1, len(trainloader), scheduler.get_lr()[0],loss.item(), batch_acc, \
cfg['print_freq']/(time.time()-tic_batch)))
tic_batch = time.time()
train_loss /= len(trainloader)
train_acc /= len(trainloader) * cfg['batch_size']
print("Train Epoch {} : mean Accu {:.4f} --- mean Loss {:.6f}".format(epoch + 1,train_acc, train_loss))
if epoch !=0 and epoch % cfg['checkpoint_freq'] == 0:
torch.save(model.state_dict(), model_path.format(cfg['model_name'],epoch))
if True: # eval:
print("Evaluate at epoch {}".format(epoch + 1))
model.eval()
eval_acc,eval_loss = eval(model,valloader,criterion,device,cfg)
model.train()
if best_score < eval_acc:
best_score = eval_acc
if min_loss > eval_loss:
best_model_path = os.path.join(save_dir, 'best.pth')
torch.save(model.state_dict(), best_model_path)
best_epoch = epoch + 1
min_loss = eval_loss
print("Val Epoch {} : Accu {:.4f} , best Accu: {:.4f} --- mean Loss {:.6f} , min Loss {:.6f} , best at epoch {}".format(epoch + 1,eval_acc, best_score,eval_loss,min_loss,best_epoch))
writer.add_scalars('loss', {
'train': train_loss,
'val': eval_loss
}, epoch + 1)
writer.add_scalars('accu', {
'train': train_acc,
'val': eval_acc
}, epoch + 1)
train_loss = 0.0
train_acc = 0.0
scheduler.step()
if __name__ == '__main__':
create_dir(cfg['checkpoint_dir'])
# if not os.path.exists(cfg['val_path']):
# split_dataset(cfg)
#img_mean_std(cfg)
args = get_args()
model = se_resnet50(9,None)
optimizer = optim.SGD(
filter(lambda p: p.requires_grad, model.parameters()),
momentum=0.9,lr = cfg['base_lr'], weight_decay=0.001
)
scheduler = lr_scheduler.ExponentialLR(optimizer, gamma=cfg['gamma'])
#scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,mode='min',factor=0.2,patience=3,verbose=True,)
#summary(model,(3,224,224))
train(model,optimizer,scheduler,cfg,args)