-
Notifications
You must be signed in to change notification settings - Fork 605
/
super_dimp_simple.py
132 lines (108 loc) · 7.86 KB
/
super_dimp_simple.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
import torch.optim as optim
from ltr.dataset import Lasot, Got10k, TrackingNet, MSCOCOSeq
from ltr.data import processing, sampler, LTRLoader
from ltr.models.tracking import dimpnet
import ltr.models.loss as ltr_losses
import ltr.models.loss.kl_regression as klreg_losses
import ltr.actors.tracking as tracking_actors
from ltr.trainers import LTRTrainer
import ltr.data.transforms as tfm
from ltr import MultiGPU
def run(settings):
settings.description = 'SuperDiMPHinge: Combines the DiMP classifier with the PrDiMP bounding box regressor and better' \
'training settings (larger batch size, inside_major cropping, and flipping augmentation.' \
'Gives results significantly better than both DiMP-50 and PrDiMP-50. Compared to SuperDiMP' \
'it omits the learned discriminative learning loss and used a gaussian label maps instead of' \
'bounding boxes for supervision of the target classifier.'
settings.batch_size = 20
settings.num_workers = 8
settings.multi_gpu = False
settings.print_interval = 1
settings.normalize_mean = [0.485, 0.456, 0.406]
settings.normalize_std = [0.229, 0.224, 0.225]
settings.search_area_factor = 6.0
settings.output_sigma_factor = 1/4
settings.target_filter_sz = 4
settings.feature_sz = 22
settings.output_sz = settings.feature_sz * 16
settings.center_jitter_factor = {'train': 3, 'test': 5.5}
settings.scale_jitter_factor = {'train': 0.25, 'test': 0.5}
settings.hinge_threshold = 0.05
# settings.print_stats = ['Loss/total', 'Loss/iou', 'ClfTrain/init_loss', 'ClfTrain/test_loss']
# Train datasets
lasot_train = Lasot(settings.env.lasot_dir, split='train')
got10k_train = Got10k(settings.env.got10k_dir, split='vottrain')
trackingnet_train = TrackingNet(settings.env.trackingnet_dir, set_ids=list(range(4)))
coco_train = MSCOCOSeq(settings.env.coco_dir)
# Validation datasets
got10k_val = Got10k(settings.env.got10k_dir, split='votval')
# Data transform
transform_joint = tfm.Transform(tfm.ToGrayscale(probability=0.05),
tfm.RandomHorizontalFlip(probability=0.5))
transform_train = tfm.Transform(tfm.ToTensorAndJitter(0.2),
tfm.RandomHorizontalFlip(probability=0.5),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
transform_val = tfm.Transform(tfm.ToTensor(),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
# The tracking pairs processing module
output_sigma = settings.output_sigma_factor / settings.search_area_factor
proposal_params = {'boxes_per_frame': 128, 'gt_sigma': (0.05, 0.05), 'proposal_sigma': [(0.05, 0.05), (0.5, 0.5)]}
label_params = {'feature_sz': settings.feature_sz, 'sigma_factor': output_sigma, 'kernel_sz': settings.target_filter_sz}
label_density_params = {'feature_sz': settings.feature_sz, 'sigma_factor': output_sigma, 'kernel_sz': settings.target_filter_sz}
data_processing_train = processing.KLDiMPProcessing(search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
crop_type='inside_major',
max_scale_change=1.5,
mode='sequence',
proposal_params=proposal_params,
label_function_params=label_params,
label_density_params=label_density_params,
transform=transform_train,
joint_transform=transform_joint)
data_processing_val = processing.KLDiMPProcessing(search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
crop_type='inside_major',
max_scale_change=1.5,
mode='sequence',
proposal_params=proposal_params,
label_function_params=label_params,
label_density_params=label_density_params,
transform=transform_val,
joint_transform=transform_joint)
# Train sampler and loader
dataset_train = sampler.DiMPSampler([lasot_train, got10k_train, trackingnet_train, coco_train], [1,1,1,1],
samples_per_epoch=40000, max_gap=200, num_test_frames=3, num_train_frames=3,
processing=data_processing_train)
loader_train = LTRLoader('train', dataset_train, training=True, batch_size=settings.batch_size, num_workers=settings.num_workers,
shuffle=True, drop_last=True, stack_dim=1)
# Validation samplers and loaders
dataset_val = sampler.DiMPSampler([got10k_val], [1], samples_per_epoch=10000, max_gap=200,
num_test_frames=3, num_train_frames=3,
processing=data_processing_val)
loader_val = LTRLoader('val', dataset_val, training=False, batch_size=settings.batch_size, num_workers=settings.num_workers,
shuffle=False, drop_last=True, epoch_interval=5, stack_dim=1)
# Create network and actor
net = dimpnet.dimpnet50_simple(filter_size=settings.target_filter_sz, backbone_pretrained=True, optim_iter=5,
clf_feat_norm=True, clf_feat_blocks=0, final_conv=True, out_feature_dim=512,
optim_init_reg=0.1, score_act='relu', hinge_threshold=settings.hinge_threshold,
activation_leak=0.1, frozen_backbone_layers=['conv1', 'bn1', 'layer1', 'layer2'])
# Wrap the network for multi GPU training
if settings.multi_gpu:
net = MultiGPU(net, dim=1)
objective = {'bb_ce': klreg_losses.KLRegression(), 'test_clf': ltr_losses.LBHinge(threshold=settings.hinge_threshold)}
loss_weight = {'bb_ce': 0.01, 'test_clf': 100, 'test_init_clf': 100, 'test_iter_clf': 400}
actor = tracking_actors.DiMPSimpleActor(net=net, objective=objective, loss_weight=loss_weight)
# Optimizer
optimizer = optim.Adam([{'params': actor.net.classifier.filter_initializer.parameters(), 'lr': 5e-5},
{'params': actor.net.classifier.filter_optimizer.parameters(), 'lr': 5e-4},
{'params': actor.net.classifier.feature_extractor.parameters(), 'lr': 5e-5},
{'params': actor.net.bb_regressor.parameters(), 'lr': 1e-3},
{'params': actor.net.feature_extractor.layer3.parameters(), 'lr': 2e-5}],
lr=2e-4)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=15, gamma=0.2)
trainer = LTRTrainer(actor, [loader_train, loader_val], optimizer, settings, lr_scheduler)
trainer.train(50, load_latest=True, fail_safe=True)