Run bevdet4d without CBGS

[thom966]

Hello, is there a way to run bevdet4d without CBGS strategy?

[HuangJunJie2017]

@thom966 please refer to https://github.com/HuangJunJie2017/BEVDet/blob/dev2.0/configs/bevdet/bevdet-r50.py for example of training bevdet series without CBGS strategy.

[thom966]

I have changed the data dictionary part in the config file as used in bevdet-r50.py like below: `data = dict( samples_per_gpu=8, workers_per_gpu=4, train=dict( data_root=data_root, ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl', pipeline=train_pipeline, classes=class_names, test_mode=False, use_valid_flag=True,

we use box_type_3d='LiDAR' in kitti and nuscenes dataset

    # and box_type_3d='Depth' in sunrgbd and scannet dataset.
    box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)`

When I run the file bevdet4d-r50.py, it throws the following error:

Traceback (most recent call last): File "tools/train.py", line 264, in main() File "tools/train.py", line 119, in main cfg = Config.fromfile(args.config) File "/home/anaconda3/envs/bevdet_new/lib/python3.8/site-packages/mmcv/utils/config.py", line 340, in fromfile cfg_dict, cfg_text = Config._file2dict(filename, File "/home/anaconda3/envs/bevdet_new/lib/python3.8/site-packages/mmcv/utils/config.py", line 208, in _file2dict mod = import_module(temp_module_name) File "/home/anaconda3/envs/bevdet_new/lib/python3.8/importlib/init.py", line 127, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "", line 1014, in _gcd_import File "", line 991, in _find_and_load File "", line 975, in _find_and_load_unlocked File "", line 671, in _load_unlocked File "", line 843, in exec_module File "", line 219, in _call_with_frames_removed File "/tmp/tmpd7sqaj29/tmp41gfpq8e.py", line 264, in KeyError: 'dataset' Exception ignored in: <function _TemporaryFileCloser.del at 0x7f7cd6e48040> Traceback (most recent call last): File "/home/anaconda3/envs/bevdet_new/lib/python3.8/tempfile.py", line 440, in del self.close() File "/home/anaconda3/envs/bevdet_new/lib/python3.8/tempfile.py", line 436, in close unlink(self.name) FileNotFoundError: [Errno 2] No such file or directory: '/tmp/tmpd7sqaj29/tmp41gfpq8e.py'

[HuangJunJie2017]

https://github.com/HuangJunJie2017/BEVDet/blob/cf9d62d42906c76ddd9cfe67324c0ddc4d8c3b2d/configs/bevdet/bevdet-r50.py#L229
L229 and L230 in configs/bevdet/bevdet-r50.py should be adjusted accordingly as well. Besides, the schedule should be adjusted as well. Please compare configs/bevdet/bevdet-r50.py with configs/bevdet/bevdet-r50-cbgs.py carefully and check all modified details.

[thom966]

I have modified the above two sections in config file. I also have checked all modified details and I got the result below:

mAP: 0.0953 mATE: 1.0440 mASE: 0.3154 mAOE: 1.0378 mAVE: 0.7831 mAAE: 0.2801 NDS: 0.2098

Is the result reasonable? mAP seems weird.

Below is the config file:

_base_ = ['../_base_/datasets/nus-3d.py', '../_base_/default_runtime.py']

point_cloud_range = [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]

class_names = [
    'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
    'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]

data_config = {
    'cams': [
        'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
        'CAM_BACK', 'CAM_BACK_RIGHT'
    ],
    'Ncams':
    6,
    'input_size': (256, 704),
    'src_size': (900, 1600),

    # Augmentation
    'resize': (-0.06, 0.11),
    'rot': (-5.4, 5.4),
    'flip': True,
    'crop_h': (0.0, 0.0),
    'resize_test': 0.00,
}

# Model
grid_config = {
    'x': [-51.2, 51.2, 0.8],
    'y': [-51.2, 51.2, 0.8],
    'z': [-5, 3, 8],
    'depth': [1.0, 60.0, 1.0],
}

voxel_size = [0.1, 0.1, 0.2]

numC_Trans = 80

multi_adj_frame_id_cfg = (1, 3, 1)

model = dict(
    type='BEVDet4D',
    align_after_view_transfromation=False,
    num_adj=len(range(*multi_adj_frame_id_cfg)),
    img_backbone=dict(
        pretrained='torchvision://resnet50',
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(2, 3),
        frozen_stages=-1,
        norm_cfg=dict(type='BN', requires_grad=True),
        norm_eval=False,
        with_cp=True,
        style='pytorch'),
    img_neck=dict(
        type='CustomFPN',
        in_channels=[1024, 2048],
        out_channels=512,
        num_outs=1,
        start_level=0,
        out_ids=[0]),
    img_view_transformer=dict(
        type='LSSViewTransformer',
        grid_config=grid_config,
        input_size=data_config['input_size'],
        in_channels=512,
        out_channels=numC_Trans,
        downsample=16),
    img_bev_encoder_backbone=dict(
        type='CustomResNet',
        numC_input=numC_Trans * (len(range(*multi_adj_frame_id_cfg))+1),
        num_channels=[numC_Trans * 2, numC_Trans * 4, numC_Trans * 8]),
    img_bev_encoder_neck=dict(
        type='FPN_LSS',
        in_channels=numC_Trans * 8 + numC_Trans * 2,
        out_channels=256),
    pre_process=dict(
        type='CustomResNet',
        numC_input=numC_Trans,
        num_layer=[2,],
        num_channels=[numC_Trans,],
        stride=[1,],
        backbone_output_ids=[0,]),
    pts_bbox_head=dict(
        type='CenterHead',
        in_channels=256,
        tasks=[
            dict(num_class=1, class_names=['car']),
            dict(num_class=2, class_names=['truck', 'construction_vehicle']),
            dict(num_class=2, class_names=['bus', 'trailer']),
            dict(num_class=1, class_names=['barrier']),
            dict(num_class=2, class_names=['motorcycle', 'bicycle']),
            dict(num_class=2, class_names=['pedestrian', 'traffic_cone']),
        ],
        common_heads=dict(
            reg=(2, 2), height=(1, 2), dim=(3, 2), rot=(2, 2), vel=(2, 2)),
        share_conv_channel=64,
        bbox_coder=dict(
            type='CenterPointBBoxCoder',
            pc_range=point_cloud_range[:2],
            post_center_range=[-61.2, -61.2, -10.0, 61.2, 61.2, 10.0],
            max_num=500,
            score_threshold=0.1,
            out_size_factor=8,
            voxel_size=voxel_size[:2],
            code_size=9),
        separate_head=dict(
            type='SeparateHead', init_bias=-2.19, final_kernel=3),
        loss_cls=dict(type='GaussianFocalLoss', reduction='mean'),
        loss_bbox=dict(type='L1Loss', reduction='mean', loss_weight=0.25),
        norm_bbox=True),
    # model training and testing settings
    train_cfg=dict(
        pts=dict(
            point_cloud_range=point_cloud_range,
            grid_size=[1024, 1024, 40],
            voxel_size=voxel_size,
            out_size_factor=8,
            dense_reg=1,
            gaussian_overlap=0.1,
            max_objs=500,
            min_radius=2,
            code_weights=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0])),
    test_cfg=dict(
        pts=dict(
            pc_range=point_cloud_range[:2],
            post_center_limit_range=[-61.2, -61.2, -10.0, 61.2, 61.2, 10.0],
            max_per_img=500,
            max_pool_nms=False,
            min_radius=[4, 12, 10, 1, 0.85, 0.175],
            score_threshold=0.1,
            out_size_factor=8,
            voxel_size=voxel_size[:2],
            pre_max_size=1000,
            post_max_size=83,

            # Scale-NMS
            nms_type=[
                'rotate', 'rotate', 'rotate', 'circle', 'rotate', 'rotate'
            ],
            nms_thr=[0.2, 0.2, 0.2, 0.2, 0.2, 0.5],
            nms_rescale_factor=[
                1.0, [0.7, 0.7], [0.4, 0.55], 1.1, [1.0, 1.0], [4.5, 9.0]
            ])))

# Data
dataset_type = 'NuScenesDataset'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')

bda_aug_conf = dict(
    rot_lim=(-22.5, 22.5),
    scale_lim=(0.95, 1.05),
    flip_dx_ratio=0.5,
    flip_dy_ratio=0.5)

train_pipeline = [
    dict(
        type='PrepareImageInputs',
        is_train=True,
        data_config=data_config,
        sequential=True),
    dict(
        type='LoadAnnotationsBEVDepth',
        bda_aug_conf=bda_aug_conf,
        classes=class_names),
    dict(type='ObjectRangeFilter', point_cloud_range=point_cloud_range),
    dict(type='ObjectNameFilter', classes=class_names),
    dict(type='DefaultFormatBundle3D', class_names=class_names),
    dict(
        type='Collect3D', keys=['img_inputs', 'gt_bboxes_3d', 'gt_labels_3d'])
]

test_pipeline = [
    dict(type='PrepareImageInputs', data_config=data_config, sequential=True),
    dict(
        type='LoadAnnotationsBEVDepth',
        bda_aug_conf=bda_aug_conf,
        classes=class_names,
        is_train=False),
    dict(
        type='LoadPointsFromFile',
        coord_type='LIDAR',
        load_dim=5,
        use_dim=5,
        file_client_args=file_client_args),
    dict(
        type='MultiScaleFlipAug3D',
        img_scale=(1333, 800),
        pts_scale_ratio=1,
        flip=False,
        transforms=[
            dict(
                type='DefaultFormatBundle3D',
                class_names=class_names,
                with_label=False),
            dict(type='Collect3D', keys=['points', 'img_inputs'])
        ])
]

input_modality = dict(
    use_lidar=False,
    use_camera=True,
    use_radar=False,
    use_map=False,
    use_external=False)

share_data_config = dict(
    type=dataset_type,
    classes=class_names,
    modality=input_modality,
    img_info_prototype='bevdet4d',
    multi_adj_frame_id_cfg=multi_adj_frame_id_cfg,
)

test_data_config = dict(
    pipeline=test_pipeline,
    ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')

data = dict(
    samples_per_gpu=16,
    workers_per_gpu=0,
    train=dict(

        data_root=data_root,
        ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
        pipeline=train_pipeline,
        classes=class_names,
        test_mode=False,
        use_valid_flag=True,
        # we use box_type_3d='LiDAR' in kitti and nuscenes dataset
        # and box_type_3d='Depth' in sunrgbd and scannet dataset.
        box_type_3d='LiDAR'),
    val=test_data_config,
    test=test_data_config)

for key in ['train', 'val', 'test']:
    data[key].update(share_data_config)

# Optimizer
optimizer = dict(type='AdamW', lr=2e-4, weight_decay=1e-7)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=200,
    warmup_ratio=0.001,
    step=[20,])
runner = dict(type='EpochBasedRunner', max_epochs=20)

custom_hooks = [
    dict(
        type='MEGVIIEMAHook',
        init_updates=10560,
        priority='NORMAL',
    ),

]

[HuangJunJie2017]

@thom966 I think the config you used is reasonable except for the missing of SequentialControlHook SequentialControlHook offers stable convergence for bevdet4d Recommended config would be like this:

# Copyright (c) Phigent Robotics. All rights reserved.

_base_ = ['../_base_/datasets/nus-3d.py', '../_base_/default_runtime.py']
# Global
# If point cloud range is changed, the models should also change their point
# cloud range accordingly
point_cloud_range = [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
# For nuScenes we usually do 10-class detection
class_names = [
    'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
    'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]

data_config = {
    'cams': [
        'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
        'CAM_BACK', 'CAM_BACK_RIGHT'
    ],
    'Ncams':
    6,
    'input_size': (256, 704),
    'src_size': (900, 1600),

    # Augmentation
    'resize': (-0.06, 0.11),
    'rot': (-5.4, 5.4),
    'flip': True,
    'crop_h': (0.0, 0.0),
    'resize_test': 0.00,
}

# Model
grid_config = {
    'x': [-51.2, 51.2, 0.8],
    'y': [-51.2, 51.2, 0.8],
    'z': [-5, 3, 8],
    'depth': [1.0, 60.0, 1.0],
}

voxel_size = [0.1, 0.1, 0.2]

numC_Trans = 80

multi_adj_frame_id_cfg = (1, 3, 1)

model = dict(
    type='BEVDet4D',
    align_after_view_transfromation=False,
    num_adj=len(range(*multi_adj_frame_id_cfg)),
    img_backbone=dict(
        pretrained='torchvision://resnet50',
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(2, 3),
        frozen_stages=-1,
        norm_cfg=dict(type='BN', requires_grad=True),
        norm_eval=False,
        with_cp=True,
        style='pytorch'),
    img_neck=dict(
        type='CustomFPN',
        in_channels=[1024, 2048],
        out_channels=512,
        num_outs=1,
        start_level=0,
        out_ids=[0]),
    img_view_transformer=dict(
        type='LSSViewTransformer',
        grid_config=grid_config,
        input_size=data_config['input_size'],
        in_channels=512,
        out_channels=numC_Trans,
        downsample=16),
    img_bev_encoder_backbone=dict(
        type='CustomResNet',
        numC_input=numC_Trans * (len(range(*multi_adj_frame_id_cfg))+1),
        num_channels=[numC_Trans * 2, numC_Trans * 4, numC_Trans * 8]),
    img_bev_encoder_neck=dict(
        type='FPN_LSS',
        in_channels=numC_Trans * 8 + numC_Trans * 2,
        out_channels=256),
    pre_process=dict(
        type='CustomResNet',
        numC_input=numC_Trans,
        num_layer=[2,],
        num_channels=[numC_Trans,],
        stride=[1,],
        backbone_output_ids=[0,]),
    pts_bbox_head=dict(
        type='CenterHead',
        in_channels=256,
        tasks=[
            dict(num_class=1, class_names=['car']),
            dict(num_class=2, class_names=['truck', 'construction_vehicle']),
            dict(num_class=2, class_names=['bus', 'trailer']),
            dict(num_class=1, class_names=['barrier']),
            dict(num_class=2, class_names=['motorcycle', 'bicycle']),
            dict(num_class=2, class_names=['pedestrian', 'traffic_cone']),
        ],
        common_heads=dict(
            reg=(2, 2), height=(1, 2), dim=(3, 2), rot=(2, 2), vel=(2, 2)),
        share_conv_channel=64,
        bbox_coder=dict(
            type='CenterPointBBoxCoder',
            pc_range=point_cloud_range[:2],
            post_center_range=[-61.2, -61.2, -10.0, 61.2, 61.2, 10.0],
            max_num=500,
            score_threshold=0.1,
            out_size_factor=8,
            voxel_size=voxel_size[:2],
            code_size=9),
        separate_head=dict(
            type='SeparateHead', init_bias=-2.19, final_kernel=3),
        loss_cls=dict(type='GaussianFocalLoss', reduction='mean'),
        loss_bbox=dict(type='L1Loss', reduction='mean', loss_weight=0.25),
        norm_bbox=True),
    # model training and testing settings
    train_cfg=dict(
        pts=dict(
            point_cloud_range=point_cloud_range,
            grid_size=[1024, 1024, 40],
            voxel_size=voxel_size,
            out_size_factor=8,
            dense_reg=1,
            gaussian_overlap=0.1,
            max_objs=500,
            min_radius=2,
            code_weights=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0])),
    test_cfg=dict(
        pts=dict(
            pc_range=point_cloud_range[:2],
            post_center_limit_range=[-61.2, -61.2, -10.0, 61.2, 61.2, 10.0],
            max_per_img=500,
            max_pool_nms=False,
            min_radius=[4, 12, 10, 1, 0.85, 0.175],
            score_threshold=0.1,
            out_size_factor=8,
            voxel_size=voxel_size[:2],
            pre_max_size=1000,
            post_max_size=83,

            # Scale-NMS
            nms_type=[
                'rotate', 'rotate', 'rotate', 'circle', 'rotate', 'rotate'
            ],
            nms_thr=[0.2, 0.2, 0.2, 0.2, 0.2, 0.5],
            nms_rescale_factor=[
                1.0, [0.7, 0.7], [0.4, 0.55], 1.1, [1.0, 1.0], [4.5, 9.0]
            ])))

# Data
dataset_type = 'NuScenesDataset'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')

bda_aug_conf = dict(
    rot_lim=(-22.5, 22.5),
    scale_lim=(0.95, 1.05),
    flip_dx_ratio=0.5,
    flip_dy_ratio=0.5)

train_pipeline = [
    dict(
        type='PrepareImageInputs',
        is_train=True,
        data_config=data_config,
        sequential=True),
    dict(
        type='LoadAnnotationsBEVDepth',
        bda_aug_conf=bda_aug_conf,
        classes=class_names),
    dict(type='ObjectRangeFilter', point_cloud_range=point_cloud_range),
    dict(type='ObjectNameFilter', classes=class_names),
    dict(type='DefaultFormatBundle3D', class_names=class_names),
    dict(
        type='Collect3D', keys=['img_inputs', 'gt_bboxes_3d', 'gt_labels_3d'])
]

test_pipeline = [
    dict(type='PrepareImageInputs', data_config=data_config, sequential=True),
    dict(
        type='LoadAnnotationsBEVDepth',
        bda_aug_conf=bda_aug_conf,
        classes=class_names,
        is_train=False),
    dict(
        type='LoadPointsFromFile',
        coord_type='LIDAR',
        load_dim=5,
        use_dim=5,
        file_client_args=file_client_args),
    dict(
        type='MultiScaleFlipAug3D',
        img_scale=(1333, 800),
        pts_scale_ratio=1,
        flip=False,
        transforms=[
            dict(
                type='DefaultFormatBundle3D',
                class_names=class_names,
                with_label=False),
            dict(type='Collect3D', keys=['points', 'img_inputs'])
        ])
]

input_modality = dict(
    use_lidar=False,
    use_camera=True,
    use_radar=False,
    use_map=False,
    use_external=False)

share_data_config = dict(
    type=dataset_type,
    classes=class_names,
    modality=input_modality,
    img_info_prototype='bevdet4d',
    multi_adj_frame_id_cfg=multi_adj_frame_id_cfg,
)

test_data_config = dict(
    pipeline=test_pipeline,
    ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')

data = dict(
    samples_per_gpu=8,
    workers_per_gpu=4,
    train=dict(
        data_root=data_root,
        ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
        pipeline=train_pipeline,
        classes=class_names,
        test_mode=False,
        use_valid_flag=True,
        # we use box_type_3d='LiDAR' in kitti and nuscenes dataset
        # and box_type_3d='Depth' in sunrgbd and scannet dataset.
        box_type_3d='LiDAR'),
    val=test_data_config,
    test=test_data_config)

for key in ['train', 'val', 'test']:
    data[key].update(share_data_config)

# Optimizer
optimizer = dict(type='AdamW', lr=2e-4, weight_decay=1e-7)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=200,
    warmup_ratio=0.001,
    step=[20,])
runner = dict(type='EpochBasedRunner', max_epochs=20)

custom_hooks = [
    dict(
        type='MEGVIIEMAHook',
        init_updates=10560,
        priority='NORMAL',
    ),
    dict(
        type='SequentialControlHook',
        temporal_start_epoch=2,
    ),
]

# fp16 = dict(loss_scale='dynamic')

[thom966]

mAP: 0.0984
mATE: 1.0577 mASE: 0.3081 mAOE: 1.0259 mAVE: 0.7268 mAAE: 0.2837 NDS: 0.2173

The score of mAP is still not reproduced. Could you please provide me any potential origin of the code that might caused the problem?

[HuangJunJie2017]

@thom966 maybe your env is problematic, check this by training models with CBGS