hv_pointpillars_secfpn_6x8_160e_kitti-3d-car.py 5.77 KB
Newer Older
zhangwenwei's avatar
zhangwenwei committed
1
# model settings
2
voxel_size = [0.16, 0.16, 4]
zhangwenwei's avatar
zhangwenwei committed
3
4
5
6
point_cloud_range = [0, -39.68, -3, 69.12, 39.68, 1]
model = dict(
    type='VoxelNet',
    voxel_layer=dict(
7
8
9
        max_num_points=64,
        point_cloud_range=point_cloud_range,
        voxel_size=voxel_size,
zhangwenwei's avatar
zhangwenwei committed
10
        max_voxels=(12000, 20000)),
zhangwenwei's avatar
zhangwenwei committed
11
12
    voxel_encoder=dict(
        type='PillarFeatureNet',
zhangwenwei's avatar
zhangwenwei committed
13
14
        in_channels=4,
        feat_channels=[64],
zhangwenwei's avatar
zhangwenwei committed
15
        with_distance=False,
16
        voxel_size=voxel_size,
zhangwenwei's avatar
zhangwenwei committed
17
        point_cloud_range=point_cloud_range),
zhangwenwei's avatar
zhangwenwei committed
18
    middle_encoder=dict(
zhangwenwei's avatar
zhangwenwei committed
19
        type='PointPillarsScatter', in_channels=64, output_shape=[496, 432]),
zhangwenwei's avatar
zhangwenwei committed
20
21
22
23
24
    backbone=dict(
        type='SECOND',
        in_channels=64,
        layer_nums=[3, 5, 5],
        layer_strides=[2, 2, 2],
zhangwenwei's avatar
zhangwenwei committed
25
        out_channels=[64, 128, 256]),
zhangwenwei's avatar
zhangwenwei committed
26
27
28
29
    neck=dict(
        type='SECONDFPN',
        in_channels=[64, 128, 256],
        upsample_strides=[1, 2, 4],
zhangwenwei's avatar
zhangwenwei committed
30
        out_channels=[128, 128, 128]),
zhangwenwei's avatar
zhangwenwei committed
31
    bbox_head=dict(
zhangwenwei's avatar
zhangwenwei committed
32
33
        type='Anchor3DHead',
        num_classes=1,
zhangwenwei's avatar
zhangwenwei committed
34
35
36
        in_channels=384,
        feat_channels=384,
        use_direction_classifier=True,
37
38
39
40
41
42
        anchor_generator=dict(
            type='Anchor3DRangeGenerator',
            ranges=[[0, -39.68, -1.78, 69.12, 39.68, -1.78]],
            sizes=[[1.6, 3.9, 1.56]],
            rotations=[0, 1.57],
            reshape_out=True),
zhangwenwei's avatar
zhangwenwei committed
43
        diff_rad_by_sin=True,
44
        bbox_coder=dict(type='DeltaXYZWLHRBBoxCoder'),
zhangwenwei's avatar
zhangwenwei committed
45
46
47
48
49
50
51
52
        loss_cls=dict(
            type='FocalLoss',
            use_sigmoid=True,
            gamma=2.0,
            alpha=0.25,
            loss_weight=1.0),
        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=2.0),
        loss_dir=dict(
zhangwenwei's avatar
zhangwenwei committed
53
            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.2)))
zhangwenwei's avatar
zhangwenwei committed
54
55
56
57
# model training and testing settings
train_cfg = dict(
    assigner=dict(
        type='MaxIoUAssigner',
zhangwenwei's avatar
zhangwenwei committed
58
        iou_calculator=dict(type='BboxOverlapsNearest3D'),
zhangwenwei's avatar
zhangwenwei committed
59
60
61
62
63
64
65
66
67
68
69
        pos_iou_thr=0.6,
        neg_iou_thr=0.45,
        min_pos_iou=0.45,
        ignore_iof_thr=-1),
    allowed_border=0,
    pos_weight=-1,
    debug=False)
test_cfg = dict(
    use_rotate_nms=True,
    nms_across_levels=False,
    nms_thr=0.01,
70
    score_thr=0.1,
zhangwenwei's avatar
zhangwenwei committed
71
    min_bbox_size=0,
zhangwenwei's avatar
zhangwenwei committed
72
73
    nms_pre=100,
    max_num=50)
zhangwenwei's avatar
zhangwenwei committed
74
75
76
77
78

# dataset settings
dataset_type = 'KittiDataset'
data_root = 'data/kitti/'
class_names = ['Car']
79
input_modality = dict(use_lidar=True, use_camera=False)
zhangwenwei's avatar
zhangwenwei committed
80
db_sampler = dict(
zhangwenwei's avatar
zhangwenwei committed
81
    data_root=data_root,
zhangwenwei's avatar
zhangwenwei committed
82
83
84
    info_path=data_root + 'kitti_dbinfos_train.pkl',
    rate=1.0,
    object_rot_range=[0.0, 0.0],
zhangwenwei's avatar
zhangwenwei committed
85
    prepare=dict(filter_by_difficulty=[-1], filter_by_min_points=dict(Car=5)),
zhangwenwei's avatar
zhangwenwei committed
86
    sample_groups=dict(Car=15),
zhangwenwei's avatar
zhangwenwei committed
87
    classes=class_names)
zhangwenwei's avatar
zhangwenwei committed
88
89

train_pipeline = [
zhangwenwei's avatar
zhangwenwei committed
90
91
    dict(type='LoadPointsFromFile', load_dim=4, use_dim=4),
    dict(type='LoadAnnotations3D', with_bbox_3d=True, with_label_3d=True),
zhangwenwei's avatar
zhangwenwei committed
92
93
94
95
96
97
98
    dict(type='ObjectSample', db_sampler=db_sampler),
    dict(
        type='ObjectNoise',
        num_try=100,
        loc_noise_std=[0.25, 0.25, 0.25],
        global_rot_range=[0.0, 0.0],
        rot_uniform_noise=[-0.15707963267, 0.15707963267]),
99
    dict(type='RandomFlip3D', flip_ratio=0.5),
zhangwenwei's avatar
zhangwenwei committed
100
101
102
103
104
105
106
107
    dict(
        type='GlobalRotScale',
        rot_uniform_noise=[-0.78539816, 0.78539816],
        scaling_uniform_noise=[0.95, 1.05]),
    dict(type='PointsRangeFilter', point_cloud_range=point_cloud_range),
    dict(type='ObjectRangeFilter', point_cloud_range=point_cloud_range),
    dict(type='PointShuffle'),
    dict(type='DefaultFormatBundle3D', class_names=class_names),
zhangwenwei's avatar
zhangwenwei committed
108
    dict(type='Collect3D', keys=['points', 'gt_bboxes_3d', 'gt_labels_3d'])
zhangwenwei's avatar
zhangwenwei committed
109
110
]
test_pipeline = [
zhangwenwei's avatar
zhangwenwei committed
111
    dict(type='LoadPointsFromFile', load_dim=4, use_dim=4),
zhangwenwei's avatar
zhangwenwei committed
112
113
114
115
116
    dict(type='PointsRangeFilter', point_cloud_range=point_cloud_range),
    dict(
        type='DefaultFormatBundle3D',
        class_names=class_names,
        with_label=False),
zhangwenwei's avatar
zhangwenwei committed
117
    dict(type='Collect3D', keys=['points'])
zhangwenwei's avatar
zhangwenwei committed
118
119
120
121
122
123
]

data = dict(
    samples_per_gpu=6,
    workers_per_gpu=4,
    train=dict(
124
125
126
127
128
129
130
131
132
133
134
135
        type='RepeatDataset',
        times=2,
        dataset=dict(
            type=dataset_type,
            data_root=data_root,
            ann_file=data_root + 'kitti_infos_train.pkl',
            split='training',
            pts_prefix='velodyne_reduced',
            pipeline=train_pipeline,
            modality=input_modality,
            classes=class_names,
            test_mode=False)),
zhangwenwei's avatar
zhangwenwei committed
136
137
    val=dict(
        type=dataset_type,
zhangwenwei's avatar
zhangwenwei committed
138
        data_root=data_root,
zhangwenwei's avatar
zhangwenwei committed
139
140
        ann_file=data_root + 'kitti_infos_val.pkl',
        split='training',
zhangwenwei's avatar
zhangwenwei committed
141
        pts_prefix='velodyne_reduced',
zhangwenwei's avatar
zhangwenwei committed
142
143
        pipeline=test_pipeline,
        modality=input_modality,
zhangwenwei's avatar
zhangwenwei committed
144
145
        classes=class_names,
        test_mode=True),
zhangwenwei's avatar
zhangwenwei committed
146
147
    test=dict(
        type=dataset_type,
zhangwenwei's avatar
zhangwenwei committed
148
        data_root=data_root,
zhangwenwei's avatar
zhangwenwei committed
149
        ann_file=data_root + 'kitti_infos_val.pkl',
zhangwenwei's avatar
zhangwenwei committed
150
151
        split='training',
        pts_prefix='velodyne_reduced',
zhangwenwei's avatar
zhangwenwei committed
152
153
        pipeline=test_pipeline,
        modality=input_modality,
zhangwenwei's avatar
zhangwenwei committed
154
155
        classes=class_names,
        test_mode=True))
zhangwenwei's avatar
zhangwenwei committed
156
157
158
159
160
161
162
163
164
165
166
# optimizer
lr = 0.001  # max learning rate
optimizer = dict(
    type='AdamW',
    lr=lr,
    betas=(0.95, 0.99),  # the momentum is change during training
    weight_decay=0.01)
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
# learning policy
lr_config = dict(
    policy='cyclic',
zhangwenwei's avatar
zhangwenwei committed
167
    target_ratio=(10, 1e-4),
zhangwenwei's avatar
zhangwenwei committed
168
    cyclic_times=1,
zhangwenwei's avatar
zhangwenwei committed
169
    step_ratio_up=0.4)
zhangwenwei's avatar
zhangwenwei committed
170
171
momentum_config = dict(
    policy='cyclic',
zhangwenwei's avatar
zhangwenwei committed
172
    target_ratio=(0.85 / 0.95, 1),
zhangwenwei's avatar
zhangwenwei committed
173
    cyclic_times=1,
zhangwenwei's avatar
zhangwenwei committed
174
    step_ratio_up=0.4)
zhangwenwei's avatar
zhangwenwei committed
175
checkpoint_config = dict(interval=1)
176
evaluation = dict(interval=1)
zhangwenwei's avatar
zhangwenwei committed
177
178
179
180
181
182
183
184
185
# yapf:disable
log_config = dict(
    interval=50,
    hooks=[
        dict(type='TextLoggerHook'),
        dict(type='TensorboardLoggerHook')
    ])
# yapf:enable
# runtime settings
186
total_epochs = 80
zhangwenwei's avatar
zhangwenwei committed
187
188
189
190
191
192
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/pp_secfpn_80e'
load_from = None
resume_from = None
workflow = [('train', 1)]