-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
125 lines (115 loc) · 5.61 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
import torch
import os
from torch.utils.data import DataLoader
import argparse
import numpy as np
import datetime
from CSRNet import CSRNet
from data_manager import DataManager
import torchvision.transforms as TF
from torch.utils.tensorboard import SummaryWriter
import shutil
parser = argparse.ArgumentParser('CSRNet training tool')
parser.add_argument('--train',required=True,type=str,help="Path to train folder")
parser.add_argument('-cfg','--model',required=True,default='config.json',type=str,help="path to cfg model CSRNET")
parser.add_argument('--use_pretrain',default=False,action="store_true",help="using pretrain VGG16 for frontend backbone")
parser.add_argument('--density',required=True,type=str,help="Path to file density hdf5 file")
parser.add_argument('--cuda',default=False,action="store_true",help="set flag to use cpu or gpu")
parser.add_argument('--checkpoint',default=False,action="store_true",help="continue train from checkpoint")
parser.add_argument('-i','--espisode',default=50,type=int,help="max iteration to train")
parser.add_argument('-lr','--learning_rate',default=1e-7,type=float,help="learning rate coefficient")
parser.add_argument('-s','--save_point',default=10,type=int,help="define iteration to save checkpoint")
parser.add_argument('-l','--log_path',default='./logs',type=str,help="define logs path to save checkpoint, performace train, parameters train")
parser.add_argument('-bs','--batchsize',default=3,type=int,help="define number of batch size dataset")
parser.add_argument('-wk','--worker',default=3,type=int,help="define number of worker to train")
parser.add_argument('-kck','--keep_checkpoint',default=10,type=int,help="define number of checkpoint can store")
args = parser.parse_args()
def train(args,model,opt,euclidean_dist,writer,device):
model.train()
for i in range(args.espisode):
display_loss = np.array([])
grs = np.array([])
for idx,(inps,labels) in enumerate(loader):
tt = 0
inps = torch.cat(inps,dim=0)
labels = torch.cat(labels,dim=0)
# labels= labels.squeeze(dim=1)
if args.cuda:
inps = inps.to(device)
labels = labels.to(device)
y_pred = model(inps)
loss = euclidean_dist(y_pred,labels)
display_loss = np.append(display_loss,loss.item())
opt.zero_grad()
loss.backward()
for a in model.parameters():
tt = (tt + a.grad.sum()).detach()
grs = np.append(grs,tt.item())
opt.step()
n = grs.shape[0]
for idx in range(n):
writer.add_scalar('train/grads_tt',grs[idx],global_step=i*n + idx)
writer.add_scalar('train/loss',display_loss[idx],global_step=i*n + idx)
print("iteration: {}, grads_tt: {}, loss_avg: {}".format(i,grs.mean(),display_loss.mean()),end='\r',flush=True)
if i % args.save_point == 0:
save(model,opt,args.log_path)
def save(model,optimize,logs_path,keep=10,name='checkpoint'):
current_time = datetime.datetime.now().strftime('%m%d%Y-%H%M%S')
## Save checkpoint model and optimizer
torch.save({
'csrnet':model.state_dict(),
'opt':opt.state_dict()
},'{}/{}-{}.pth'.format(logs_path,name,current_time))
## check %name%.txt existing in %logs_path% if it not exist create one
if not os.path.exists('{}/{}.txt'.format(logs_path,name)):
with open('{}/{}.txt'.format(logs_path,name),'w') as f:
pass
## check number checkpoint over %keep% yet! if we keep %keep% latest checkpoint
# and remove early checkpoints
checkpoints = []
with open('{}/{}.txt'.format(logs_path,name),'r') as f:
_checkpoints = f.read().split('\n')
checkpoints = [x for x in _checkpoints if x != '']
checkpoints.append('{}/{}-{}.pth'.format(logs_path,name,current_time))
if len(checkpoints)>keep:
checkpoints = checkpoints[-keep:]
ckp_rm = checkpoints[:-keep]
for f in ckp_rm:
try:
os.remove(f)
except:
pass
# write file %name%.txt to make sure latest checkpoint will load easy in future
with open('{}/{}.txt'.format(logs_path,name),'w') as f:
for c in checkpoints:
f.write('{}\n'.format(c))
if __name__ == "__main__":
current_time = datetime.datetime.now().strftime('%m%d%Y-%H%M%S')
writer = SummaryWriter(args.log_path+'/CSRnet-train-{}'.format(current_time))
model = CSRNet(args.model,use_pretrain=args.use_pretrain)
opt = torch.optim.SGD(model.parameters(),lr=args.learning_rate,momentum=0.9)
euclidean_dist = torch.nn.MSELoss(reduction='sum')
inp_transform = TF.Compose([
TF.ToTensor(),
TF.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
val_transform = TF.Compose([
TF.ToPILImage(),
TF.Resize(224)
])
manager = DataManager(args.train,args.density,mode='train',transforms=inp_transform)
loader = DataLoader(manager,batch_size=args.batchsize,shuffle=False,num_workers=args.worker)
if args.cuda:
device = torch.device('cuda')
model.cuda()
euclidean_dist.cuda()
else:
device = torch.device('cpu')
if args.checkpoint:
with open('{}/checkpoint.txt'.format(args.log_path)) as f:
path_checkpoint = f.read().split('\n')[-1]
checkpoint = torch.load(path_checkpoint)
model.load_state_dict(checkpoint['csrnet'])
opt.load_state_dict(checkpoint['opt'])
train(args,model,opt,euclidean_dist,writer,device)