-
Notifications
You must be signed in to change notification settings - Fork 62
/
train_aff.py
143 lines (110 loc) · 5.7 KB
/
train_aff.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
import numpy as np
import torch
from torch.backends import cudnn
from torch.utils.data import DataLoader
cudnn.enabled = True
from torchvision import transforms
import voc12.data
from tool import pyutils, imutils, torchutils
import argparse
import importlib
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--batch_size", default=8, type=int)
parser.add_argument("--max_epoches", default=8, type=int)
parser.add_argument("--network", default="network.vgg16_aff", type=str)
parser.add_argument("--lr", default=0.1, type=float)
parser.add_argument("--num_workers", default=8, type=int)
parser.add_argument("--wt_dec", default=5e-4, type=float)
parser.add_argument("--train_list", default="voc12/train_aug.txt", type=str)
parser.add_argument("--val_list", default="voc12/val.txt", type=str)
parser.add_argument("--session_name", default="vgg_aff", type=str)
parser.add_argument("--crop_size", default=448, type=int)
parser.add_argument("--weights", required=True, type=str)
parser.add_argument("--voc12_root", required=True, type=str)
parser.add_argument("--la_crf_dir", required=True, type=str)
parser.add_argument("--ha_crf_dir", required=True, type=str)
args = parser.parse_args()
pyutils.Logger(args.session_name + '.log')
print(vars(args))
model = getattr(importlib.import_module(args.network), 'Net')()
print(model)
train_dataset = voc12.data.VOC12AffDataset(args.train_list, label_la_dir=args.la_crf_dir, label_ha_dir=args.ha_crf_dir,
voc12_root=args.voc12_root, cropsize=args.crop_size, radius=5,
joint_transform_list=[
None,
None,
imutils.RandomCrop(args.crop_size),
imutils.RandomHorizontalFlip()
],
img_transform_list=[
transforms.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.1),
np.asarray,
model.normalize,
imutils.HWC_to_CHW
],
label_transform_list=[
None,
None,
None,
imutils.AvgPool2d(8)
])
train_data_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers,
pin_memory=True, drop_last=True)
max_step = len(train_dataset) // args.batch_size * args.max_epoches
param_groups = model.get_parameter_groups()
optimizer = torchutils.PolyOptimizer([
{'params': param_groups[0], 'lr': args.lr, 'weight_decay': args.wt_dec},
{'params': param_groups[1], 'lr': 2*args.lr, 'weight_decay': 0},
{'params': param_groups[2], 'lr': 10*args.lr, 'weight_decay': args.wt_dec},
{'params': param_groups[3], 'lr': 20*args.lr, 'weight_decay': 0}
], lr=args.lr, weight_decay=args.wt_dec, max_step=max_step)
if args.weights[-7:] == '.params':
import network.resnet38d
assert args.network == "network.resnet38_aff"
weights_dict = network.resnet38d.convert_mxnet_to_torch(args.weights)
elif args.weights[-11:] == '.caffemodel':
import network.vgg16d
assert args.network == "network.vgg16_aff"
weights_dict = network.vgg16d.convert_caffe_to_torch(args.weights)
else:
weights_dict = torch.load(args.weights)
model.load_state_dict(weights_dict, strict=False)
model = torch.nn.DataParallel(model).cuda()
model.train()
avg_meter = pyutils.AverageMeter('loss', 'bg_loss', 'fg_loss', 'neg_loss', 'bg_cnt', 'fg_cnt', 'neg_cnt')
timer = pyutils.Timer("Session started: ")
for ep in range(args.max_epoches):
for iter, pack in enumerate(train_data_loader):
aff = model.forward(pack[0])
bg_label = pack[1][0].cuda(non_blocking=True)
fg_label = pack[1][1].cuda(non_blocking=True)
neg_label = pack[1][2].cuda(non_blocking=True)
bg_count = torch.sum(bg_label) + 1e-5
fg_count = torch.sum(fg_label) + 1e-5
neg_count = torch.sum(neg_label) + 1e-5
bg_loss = torch.sum(- bg_label * torch.log(aff + 1e-5)) / bg_count
fg_loss = torch.sum(- fg_label * torch.log(aff + 1e-5)) / fg_count
neg_loss = torch.sum(- neg_label * torch.log(1. + 1e-5 - aff)) / neg_count
loss = bg_loss/4 + fg_loss/4 + neg_loss/2
optimizer.zero_grad()
loss.backward()
optimizer.step()
avg_meter.add({
'loss': loss.item(),
'bg_loss': bg_loss.item(), 'fg_loss': fg_loss.item(), 'neg_loss': neg_loss.item(),
'bg_cnt': bg_count.item(), 'fg_cnt': fg_count.item(), 'neg_cnt': neg_count.item()
})
if (optimizer.global_step - 1) % 50 == 0:
timer.update_progress(optimizer.global_step / max_step)
print('Iter:%5d/%5d' % (optimizer.global_step-1, max_step),
'loss:%.4f %.4f %.4f %.4f' % avg_meter.get('loss', 'bg_loss', 'fg_loss', 'neg_loss'),
'cnt:%.0f %.0f %.0f' % avg_meter.get('bg_cnt', 'fg_cnt', 'neg_cnt'),
'imps:%.1f' % ((iter+1) * args.batch_size / timer.get_stage_elapsed()),
'Fin:%s' % (timer.str_est_finish()),
'lr: %.4f' % (optimizer.param_groups[0]['lr']), flush=True)
avg_meter.pop()
else:
print('')
timer.reset_stage()
torch.save(model.module.state_dict(), args.session_name + '.pth')