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
| def main(args):
cfg=load_json(args.config)
seed=5
random.seed(seed)
torch.manual_seed(seed)
np.random.seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
device = torch.device('cuda')
image_size=cfg['image_size']
batch_size=cfg['batch_size']
train_dataset=SBI_Dataset(phase='train',image_size=image_size)
val_dataset=SBI_Dataset(phase='val',image_size=image_size)
train_loader=torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size//2,
shuffle=True,
collate_fn=train_dataset.collate_fn,
num_workers=4,
pin_memory=True,
drop_last=True,
worker_init_fn=train_dataset.worker_init_fn
)
val_loader=torch.utils.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
collate_fn=val_dataset.collate_fn,
num_workers=4,
pin_memory=True,
worker_init_fn=val_dataset.worker_init_fn
)
model=Detector()
model=model.to('cuda')
iter_loss=[]
train_losses=[]
test_losses=[]
train_accs=[]
test_accs=[]
val_accs=[]
val_losses=[]
n_epoch=cfg['epoch']
lr_scheduler=LinearDecayLR(model.optimizer, n_epoch, int(n_epoch/4*3))
last_loss=99999
now=datetime.now()
save_path='output/{}_'.format(args.session_name)+now.strftime(os.path.splitext(os.path.basename(args.config))[0])+'_'+now.strftime("%m_%d_%H_%M_%S")+'/'
os.mkdir(save_path)
os.mkdir(save_path+'weights/')
os.mkdir(save_path+'logs/')
logger = log(path=save_path+"logs/", file="losses.logs")
criterion=nn.CrossEntropyLoss()
last_auc=0
last_val_auc=0
weight_dict={}
n_weight=5
for epoch in range(n_epoch):
np.random.seed(seed + epoch)
train_loss=0.
train_acc=0.
model.train(mode=True)
for step,data in enumerate(tqdm(train_loader)):
img=data['img'].to(device, non_blocking=True).float()
target=data['label'].to(device, non_blocking=True).long()
output=model.training_step(img, target)
loss=criterion(output,target)
loss_value=loss.item()
iter_loss.append(loss_value)
train_loss+=loss_value
acc=compute_accuray(F.log_softmax(output,dim=1),target)
train_acc+=acc
lr_scheduler.step()
train_losses.append(train_loss/len(train_loader))
train_accs.append(train_acc/len(train_loader))
log_text="Epoch {}/{} | train loss: {:.4f}, train acc: {:.4f}, ".format(
epoch+1,
n_epoch,
train_loss/len(train_loader),
train_acc/len(train_loader),
)
model.train(mode=False)
val_loss=0.
val_acc=0.
output_dict=[]
target_dict=[]
np.random.seed(seed)
for step,data in enumerate(tqdm(val_loader)):
img=data['img'].to(device, non_blocking=True).float()
target=data['label'].to(device, non_blocking=True).long()
with torch.no_grad():
output=model(img)
loss=criterion(output,target)
loss_value=loss.item()
iter_loss.append(loss_value)
val_loss+=loss_value
acc=compute_accuray(F.log_softmax(output,dim=1),target)
val_acc+=acc
output_dict+=output.softmax(1)[:,1].cpu().data.numpy().tolist()
target_dict+=target.cpu().data.numpy().tolist()
val_losses.append(val_loss/len(val_loader))
val_accs.append(val_acc/len(val_loader))
val_auc=roc_auc_score(target_dict,output_dict)
log_text+="val loss: {:.4f}, val acc: {:.4f}, val auc: {:.4f}".format(
val_loss/len(val_loader),
val_acc/len(val_loader),
val_auc
)
if len(weight_dict)<n_weight:
save_model_path=os.path.join(save_path+'weights/',"{}_{:.4f}_val.tar".format(epoch+1,val_auc))
weight_dict[save_model_path]=val_auc
torch.save({
"model":model.state_dict(),
"optimizer":model.optimizer.state_dict(),
"epoch":epoch
},save_model_path)
last_val_auc=min([weight_dict[k] for k in weight_dict])
elif val_auc>=last_val_auc:
save_model_path=os.path.join(save_path+'weights/',"{}_{:.4f}_val.tar".format(epoch+1,val_auc))
for k in weight_dict:
if weight_dict[k]==last_val_auc:
del weight_dict[k]
os.remove(k)
weight_dict[save_model_path]=val_auc
break
torch.save({
"model":model.state_dict(),
"optimizer":model.optimizer.state_dict(),
"epoch":epoch
},save_model_path)
last_val_auc=min([weight_dict[k] for k in weight_dict])
logger.info(log_text)
if __name__=='__main__':
parser=argparse.ArgumentParser()
parser.add_argument(dest='config')
parser.add_argument('-n',dest='session_name')
args=parser.parse_args()
main(args)
|
