-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain_align.py
197 lines (160 loc) · 7.84 KB
/
train_align.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
196
197
import os
import torch
import torch.nn as nn
import numpy as np
from data.dataset import CUHKSZ_align
from models.model import PalmAlignNet
from models.losses import FocalLoss, BinRotLoss
from config import TrainConfig
import utils
def main(config, logger):
logger.info("Logger is set - training start")
# split data to train/validation
data_dir = './data/CUHKSZ/'
train_dataset = CUHKSZ_align(data_dir, 'train')
test_dataset = CUHKSZ_align(data_dir, 'test')
train_loader = torch.utils.data.DataLoader(train_dataset,
shuffle=True,
batch_size=config.batch_size,
num_workers=8,
pin_memory=True,
drop_last=True)
test_loader = torch.utils.data.DataLoader(test_dataset,
shuffle=False,
batch_size=config.batch_size,
num_workers=4,
pin_memory=True)
# model setting
criterion = [FocalLoss(),
nn.L1Loss(),
BinRotLoss(),
nn.CrossEntropyLoss()]
model = PalmAlignNet(config, train_dataset.num_classes).cuda()
# weights optimizer
optimizer_det = torch.optim.SGD(model.model_det.parameters(), config.lr, momentum=config.momentum,
weight_decay=config.weight_decay)
optimizer_cls = torch.optim.SGD(model.model_cls.parameters(), config.lr, momentum=config.momentum,
weight_decay=config.weight_decay)
lr_scheduler_det = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer_det, config.epochs, eta_min=config.lr_min)
lr_scheduler_cls = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer_cls, config.epochs, eta_min=config.lr_min)
# training loop
best_eer = 1
for epoch in range(config.epochs):
# training
train(train_loader, model, optimizer_det, optimizer_cls, criterion, epoch, config)
lr_scheduler_det.step()
lr_scheduler_cls.step()
if (epoch + 1) % 5 == 0:
# validation
eer = validate(test_loader, model, epoch, config)
# save
if best_eer > eer:
best_eer = eer
is_best = True
else:
is_best = False
if (epoch + 1) % config.save_freq == 0:
utils.save_checkpoint(model.state_dict(), config.path, is_best, (epoch + 1))
print("")
torch.cuda.empty_cache()
logger.info("Final best eer {:.4%}".format(eer))
def train(train_loader, model, optimizer1, optimizer2, criterion, epoch, config):
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
losses1 = utils.AverageMeter()
losses2 = utils.AverageMeter()
losses3 = utils.AverageMeter()
losses4 = utils.AverageMeter()
losses = utils.AverageMeter()
cur_lr = optimizer1.param_groups[0]['lr']
logger.info("Epoch {} LR {}".format(epoch + 1, cur_lr))
model.train()
for step, (img, y, hm, disp_gt, wh_gt, bbox_gt, theta_gt) in enumerate(train_loader):
img, y, hm, disp_gt, theta_gt, wh_gt = img.cuda(), y.cuda(), hm.cuda(), disp_gt.cuda(), theta_gt.cuda(), wh_gt.cuda()
N = img.size(0)
hd1, wh_pred, disp_pred, bin_pred, logits, ft = model(img, y)
loss1 = criterion[0](hd1, hm)
loss2_1 = criterion[1](wh_pred, wh_gt)
loss2_2 = criterion[1](disp_pred, disp_gt)
loss2 = loss2_1 + loss2_2
loss3 = criterion[2](bin_pred, theta_gt)
loss4 = criterion[3](logits, y)
if epoch < 15:
loss = config.lw1 * loss1 + config.lw2 * loss2 + config.lw3 * loss3
optimizer1.zero_grad()
loss.backward()
# nn.utils.clip_grad_norm_(model.parameters(), config.grad_clip)
optimizer1.step()
else:
loss = config.lw1 * loss1 + config.lw2 * loss2 + config.lw3 * loss3 + config.lw4 * loss4
optimizer1.zero_grad()
optimizer2.zero_grad()
loss.backward()
# nn.utils.clip_grad_norm_(model.parameters(), config.grad_clip)
optimizer1.step()
optimizer2.step()
losses1.update(loss1.item(), N)
losses2.update(loss2.item(), N)
losses3.update(loss3.item(), N)
losses4.update(loss4.item(), N)
losses.update(loss.item(), N)
prec1, prec5 = utils.accuracy(logits, y, topk=(1, 5))
top1.update(prec1.item(), N)
top5.update(prec5.item(), N)
if (step + 1) % config.print_freq == 0 or step == len(train_loader) - 1:
logger.info("Train: [{:2d}/{}] Step {:03d}/{:03d} Loss {losses.avg:.3f}, Loss1 {losses1.avg:.3f}, "
"Loss2 {losses2.avg:.3f}, Loss3 {losses3.avg:.3f}, Loss4 {losses4.avg:.3f} "
"Prec@(1,5) ({top1.avg:.1%}, {top5.avg:.1%})".format(epoch + 1, config.epochs, step + 1,
len(train_loader), losses=losses, losses1=losses1, losses2=losses2, losses3=losses3,
losses4=losses4, top1=top1, top5=top5))
def validate(valid_loader, model, epoch, config):
IoU = utils.AverageMeter()
dim_feature = 512
num_valid_sample = len(valid_loader.dataset)
feature_map = torch.zeros(num_valid_sample, dim_feature).cuda()
ground_truth = torch.zeros(num_valid_sample).cuda()
index_begin = 0
model.eval()
with torch.no_grad():
for step, (img, y, disp_gt, bbox_gt, theta_gt) in enumerate(valid_loader):
img, y, disp_gt, bbox_gt, theta_gt = img.cuda(), y.cuda(), disp_gt.cuda(), bbox_gt.cuda(), theta_gt.cuda()
N = img.size(0)
bbox_pred, theta_pred, logits, ft = model(img, y)
index_end = index_begin + N
feature_map[index_begin:index_end, :] = ft.view(N, dim_feature)
ground_truth[index_begin:index_end] = y
index_begin = index_end
iou1 = utils.iou(bbox_gt, bbox_pred, theta_gt, theta_pred)
IoU.update(iou1, N)
fm_n = feature_map.norm(p=2, dim=1)
dist = 1 - torch.matmul(feature_map / fm_n.view(num_valid_sample, 1),
(feature_map / fm_n.view(num_valid_sample, 1)).t())
# metrics
eer_a = utils.eer_all2all(ground_truth, dist)
eer_t = utils.eer_test2register(ground_truth, dist, valid_loader.dataset.register_list,
valid_loader.dataset.test_list)
rank1_a = utils.rankn_all2all(ground_truth, dist, 1)
rank1_t = utils.rankn_test2register(ground_truth, dist, valid_loader.dataset.register_list,
valid_loader.dataset.test_list, 1)
logger.info("Valid: [{:2d}/{}] Step {:03d}/{:03d}, EER_a {eer_a:.3%}, EER_t {eer_t:.3%}, Rank-1_a {rank1_a:.3%},"
" Rank-1_t {rank1_t:.3%} IoU {iou:.3%}".format(epoch + 1, config.epochs, step + 1, len(valid_loader),
eer_a=eer_a, eer_t=eer_t, rank1_a=rank1_a, rank1_t=rank1_t, iou=IoU.avg))
return eer_a
def setup_environment():
config = TrainConfig()
# set default gpu device id
torch.cuda.set_device(config.gpus[0])
torch.backends.cudnn.benchmark = True
# set seed
np.random.seed(config.seed)
torch.manual_seed(config.seed)
torch.cuda.manual_seed_all(config.seed)
# logging
if not os.path.isdir(config.path):
os.mkdir(config.path)
logger = utils.get_logger(os.path.join(config.path, "{}.log".format(config.name)))
config.print_params(logger.info)
return config, logger
if __name__ == "__main__":
config, logger = setup_environment()
main(config, logger)