-
Notifications
You must be signed in to change notification settings - Fork 16
/
Copy pathtrain.py
195 lines (161 loc) · 7.2 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
from __future__ import print_function
import os
import argparse
import torch
import torch.optim as optim
from torch.optim import lr_scheduler
import torchvision.transforms as transforms
from loss import FocalLoss
from sstdnet import load_sstdnet
from datagen import ListDataset
from torch.autograd import Variable
parser = argparse.ArgumentParser(description='PyTorch SSTDnet Training')
parser.add_argument('--lr', default=0.0001, type=float, help='learning rate')
parser.add_argument('--resume', '-r', action='store_true', default=False, help='resume from checkpoint')
parser.add_argument('--resume_path', type=str, default="checkpoint/ckpt-0.pth", help='checkpoint path')
parser.add_argument('--num_classes', '-classes', type=int, help='num. of classes')
parser.add_argument('--optimizer', '-op', type=str, default='SGD', help='SGD or Adam')
parser.add_argument('--num_crops', '-nc', type=int, default=2, help='How many crops')
parser.add_argument('--batch_size', '-batch', type=int, default=8, help='Batch size')
args = parser.parse_args()
lr = args.lr
num_classes = args.num_classes
selected_optim = args.optimizer
n_crops = args.num_crops
batch_size = args.batch_size
using_gpu = torch.cuda.is_available()
checkpoint_dir = 'checkpoint'
start_epoch = 0 # start from epoch 0 or last epoch
# Data
print('==> Preparing data..')
transform = transforms.Compose([
transforms.ToTensor()
])
trainset = ListDataset(root="../train", gt_extension=".txt",
labelmap_path="class_label_map.xlsx", is_train=True, transform=transform, input_image_size=512,
num_crops=n_crops, original_img_size=512)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=8, collate_fn=trainset.collate_fn)
validset = ListDataset(root="../valid", gt_extension=".txt",
labelmap_path="class_label_map.xlsx", is_train=False, transform=transform, input_image_size=512,
num_crops=5, original_img_size=512)
validloader = torch.utils.data.DataLoader(validset, batch_size=batch_size, shuffle=False, num_workers=8, collate_fn=validset.collate_fn)
print("lr : " + str(lr))
print("num. of classes : " + str(num_classes))
print("optimizer : " + selected_optim)
print("Using cuda : " + str(using_gpu))
print("Num. of crops : " + str(n_crops))
print("Size of batch : " + str(batch_size))
print("num. train data : " + str(trainset.__len__()))
raw_input("Press any key to continue..")
parameters_out = open("params.txt", 'w')
parameters_out.write("lr: "+str(lr)+"\n")
parameters_out.write("num. classes: "+str(num_classes)+"\n")
parameters_out.write("optimizer: "+selected_optim+"\n")
parameters_out.write("num. crops: "+str(n_crops)+"\n")
parameters_out.write("batch size: "+str(batch_size)+"\n")
parameters_out.write("num. training sample: "+str(trainset.__len__())+"\n")
parameters_out.close()
# Model
net = load_sstdnet(num_classes=num_classes, using_pretrained=True)
num_parameters = 0.
for param in net.parameters():
sizes = param.size()
num_layer_param = 1.
for size in sizes:
num_layer_param *= size
num_parameters += num_layer_param
print(net)
print("num. of parameters : " + str(num_parameters))
if args.resume:
print('==> Resuming from checkpoint..')
checkpoint = torch.load(args.resume_path)
net.load_state_dict(checkpoint['net'])
best_loss = checkpoint['loss']
start_epoch = checkpoint['epoch']
criterion = FocalLoss(num_classes=net.num_classes)
if torch.cuda.device_count() > 1:
net = torch.nn.DataParallel(net)
if using_gpu is True:
net.cuda()
if selected_optim == 'SGD':
optimizer = optim.SGD(net.parameters(), lr=lr, momentum=0.9, weight_decay=1e-4)
elif selected_optim == 'Adam':
optimizer = optim.Adam(net.parameters(), lr=lr)
else:
print("not supported optimizer")
# step_exp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.1)
# Training
def train(epoch):
print('\nEpoch: %d' % epoch)
net.train()
train_loss = 0
avg_matched_anchor = 0.
# step_exp_lr_scheduler.step()
for batch_idx, (inputs, loc_targets, cls_targets, mask_targets) in enumerate(trainloader):
if using_gpu is True:
inputs = Variable(inputs.cuda())
loc_targets = Variable(loc_targets.cuda())
cls_targets = Variable(cls_targets.cuda())
mask_targets = Variable(mask_targets.cuda())
else:
inputs = Variable(inputs)
loc_targets = Variable(loc_targets)
cls_targets = Variable(cls_targets)
mask_targets = Variable(mask_targets)
optimizer.zero_grad()
loc_preds, cls_preds, mask_preds = net(inputs)
loc_loss, cls_loss, mask_loss, num_matched_anchors = \
criterion(loc_preds, loc_targets, cls_preds, cls_targets, mask_preds, mask_targets)
loss = ((loc_loss + cls_loss) / num_matched_anchors) + mask_loss
loss.backward()
optimizer.step()
train_loss += loss.data[0]
avg_matched_anchor += float(num_matched_anchors)
print('epoch: %3d | iter: %4d | loc_loss: %.3f | cls_loss: %.3f | mask_loss: %.3f | train_loss: %.3f | avg_loss: %.3f | avg_num. matched: %d'
% (epoch, batch_idx, loc_loss.data[0] / num_matched_anchors, cls_loss.data[0] / num_matched_anchors,
mask_loss.data[0], loss.data[0], train_loss / (batch_idx + 1), avg_matched_anchor / (batch_idx + 1)))
# Test
def valid(epoch):
print('\nValid')
net.eval()
valid_loss = 0
for batch_idx, (inputs, loc_targets, cls_targets, mask_targets) in enumerate(validloader):
if using_gpu is True:
inputs = Variable(inputs.cuda(), volatile=True)
loc_targets = Variable(loc_targets.cuda())
cls_targets = Variable(cls_targets.cuda())
mask_targets = Variable(mask_targets.cuda())
else:
inputs = Variable(inputs, volatile=True)
loc_targets = Variable(loc_targets)
cls_targets = Variable(cls_targets)
mask_targets = Variable(mask_targets)
loc_preds, cls_preds, mask_preds = net(inputs)
loc_loss, cls_loss, mask_loss, num_matched_anchors = \
criterion(loc_preds, loc_targets, cls_preds, cls_targets, mask_preds, mask_targets)
loss = ((loc_loss + cls_loss) / num_matched_anchors) + mask_loss
valid_loss += loss.data[0]
print('loc_loss: %.3f | cls_loss: %.3f | valid_loss: %.3f | avg_loss: %.3f'
% (loc_loss.data[0] / num_matched_anchors, cls_loss.data[0] / num_matched_anchors,
loss.data[0], valid_loss / (batch_idx + 1)))
# Save checkpoint
# Every checkpoints are stored to analyze how is going training
# Model is selected by low-validation error.
valid_loss /= len(validloader)
print('Saving..')
state = {
'net': net.module.state_dict(),
'loss': valid_loss,
'epoch': epoch,
'num_classes': num_classes,
'lr': lr,
'batch': batch_size,
'crops': n_crops,
'op': selected_optim
}
if not os.path.isdir(checkpoint_dir):
os.mkdir(checkpoint_dir)
torch.save(state, checkpoint_dir+'/ckpt-'+str(epoch)+'.pth')
for epoch in range(start_epoch, start_epoch+200):
train(epoch)
valid(epoch)