Can't re-implement DeepInteraction-base on nuScenes val set

[funnyjojo]

When I use the pre-trained model r50_fpn_voxel_0075.pth to train DeepInteraction-base on the nuScenes dataset, I can't reproduce the results in the paper, especially the truck category. I don't know where the problem is. I was wondering if there is additional data preprocessing
Here is the results.

	mAP	NDS	Car	Truck	C.V.	Bus	Trailer	Barrier	M.T.	Bicycle	Ped.	T.C.
paper	69.9	72.7	88.5	64.4	30.1	79.2	44.6	76.4	79.0	67.8	88.9	80.0
my result	67.62	70.97	88.1	49.7	29.1	77.8	43.8	75.7	77.3	66.4	88.9	79.4

Here is the environment I used:

GPU: V100 * 2
python3.7 + pytorch1.9.1 + CUDA11.1
TorchVision: 0.10.1+cu111
OpenCV: 4.5.5
MMCV: 1.3.18
MMCV Compiler: GCC 7.5
MMDetection: 2.14.0
MMSegmentation: 0.14.1
MMDetection3D: 0.17.1+f110797

The paper uses 8 * V100 and samples_per_gpu=2, and I use 2 * V100 and samples_per_gpu=1, so I modified lr from 0.0001 to 0.0000125. Here is my config file:

plugin=True
plugin_dir='projects/mmdet3d_plugin/'

point_cloud_range = [-54.0, -54.0, -5.0, 54.0, 54.0, 3.0]
class_names = [
    'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
    'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
voxel_size = [0.075, 0.075, 0.2]
out_size_factor = 8
pillar_size = [voxel_size[0]*out_size_factor, voxel_size[1]*out_size_factor, point_cloud_range[5]-point_cloud_range[2]]
evaluation = dict(interval=1)
dataset_type = 'NuScenesDataset'
data_root = 'data/nuscenes/'
input_modality = dict(
    use_lidar=True,
    use_camera=True,
    use_radar=False,
    use_map=False,
    use_external=False)
img_scale = (800, 448)
num_views = 6
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
    dict(
        type='LoadPointsFromFile',
        coord_type='LIDAR',
        load_dim=5,
        use_dim=[0, 1, 2, 3, 4],
    ),
    dict(
        type='LoadPointsFromMultiSweeps',
        sweeps_num=10,
        use_dim=[0, 1, 2, 3, 4],
    ),
    dict(type='LoadAnnotations3D', with_bbox_3d=True, with_label_3d=True),
    dict(type='LoadMultiViewImageFromFiles', to_float32=True),
    dict(
        type='GlobalRotScaleTrans',
        rot_range=[-0.3925 * 2, 0.3925 * 2],
        scale_ratio_range=[0.9, 1.1],
        translation_std=[0.5, 0.5, 0.5]),
    dict(
        type='RandomFlip3D',
        sync_2d=True,
        flip_ratio_bev_horizontal=0.5,
        flip_ratio_bev_vertical=0.5),
    dict(type='PointsRangeFilter', point_cloud_range=point_cloud_range),
    dict(type='ObjectRangeFilter', point_cloud_range=point_cloud_range),
    dict(type='ObjectNameFilter', classes=class_names),
    dict(type='PointShuffle'),
    dict(type='ScaleImageMultiViewImage', scales=img_scale),
    dict(type='NormalizeMultiviewImage', **img_norm_cfg),
    dict(type='PadMultiViewImage', size_divisor=32),
    dict(type='DefaultFormatBundle3D', class_names=class_names),
    dict(type='Collect3D', keys=['points', 'img', 'gt_bboxes_3d', 'gt_labels_3d'])
]
test_pipeline = [
    dict(
        type='LoadPointsFromFile',
        coord_type='LIDAR',
        load_dim=5,
        use_dim=[0, 1, 2, 3, 4],
    ),
    dict(
        type='LoadPointsFromMultiSweeps',
        sweeps_num=10,
        use_dim=[0, 1, 2, 3, 4],
    ),
    dict(type='LoadMultiViewImageFromFiles', to_float32=True),
    dict(
        type='MultiScaleFlipAug3D',
        img_scale=img_scale,
        pts_scale_ratio=1,
        flip=False,
        transforms=[
            dict(
                type='GlobalRotScaleTrans',
                rot_range=[0, 0],
                scale_ratio_range=[1.0, 1.0],
                translation_std=[0, 0, 0]),
            dict(type='RandomFlip3D'),
            dict(type='ScaleImageMultiViewImage', scales=img_scale),
            dict(type='NormalizeMultiviewImage', **img_norm_cfg),
            dict(type='PadMultiViewImage', size_divisor=32),
            dict(
                type='DefaultFormatBundle3D',
                class_names=class_names,
                with_label=False),
            dict(type='Collect3D', keys=['points', 'img'])
        ])
]
data = dict(
    samples_per_gpu=1,
    workers_per_gpu=4,
    train=dict(
        type='CBGSDataset',
        dataset=dict(
            type=dataset_type,
            data_root=data_root,
            ann_file=data_root + 'nuscenes_infos_train.pkl',
            load_interval=1,
            pipeline=train_pipeline,
            classes=class_names,
            modality=input_modality,
            test_mode=False,
            box_type_3d='LiDAR')),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file=data_root + 'nuscenes_infos_val.pkl',
        load_interval=1,
        pipeline=test_pipeline,
        classes=class_names,
        modality=input_modality,
        test_mode=True,
        box_type_3d='LiDAR'),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file=data_root + 'nuscenes_infos_val.pkl',
        load_interval=1,
        pipeline=test_pipeline,
        classes=class_names,
        modality=input_modality,
        test_mode=True,
        box_type_3d='LiDAR'))
model = dict(
    type='DeepInteraction',
    freeze_img=True,
    freeze_pts=True,
    img_backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        norm_cfg=dict(type='BN', requires_grad=True),
        norm_eval=True,
        style='pytorch'),
    img_neck=dict(
        type='FPN',
        in_channels=[256, 512, 1024, 2048],
        out_channels=256,
        num_outs=5),
    pts_voxel_layer=dict(
        max_num_points=10,
        voxel_size=voxel_size,
        max_voxels=(120000, 160000),
        point_cloud_range=point_cloud_range),
    pts_pillar_layer=dict(
        max_num_points=20,
        voxel_size=pillar_size,
        max_voxels=(30000, 60000),
        point_cloud_range=point_cloud_range),
    pts_voxel_encoder=dict(
        type='HardSimpleVFE',
        num_features=5,
    ),
    pts_middle_encoder=dict(
        type='SparseEncoder',
        in_channels=5,
        sparse_shape=[41, 1440, 1440],
        output_channels=128,
        order=('conv', 'norm', 'act'),
        encoder_channels=((16, 16, 32), (32, 32, 64), (64, 64, 128), (128, 128)),
        encoder_paddings=((0, 0, 1), (0, 0, 1), (0, 0, [0, 1, 1]), (0, 0)),
        block_type='basicblock'),
    pts_backbone=dict(
        type='SECOND',
        in_channels=256,
        out_channels=[128, 256],
        layer_nums=[5, 5],
        layer_strides=[1, 2],
        norm_cfg=dict(type='BN', eps=0.001, momentum=0.01),
        conv_cfg=dict(type='Conv2d', bias=False)),
    pts_neck=dict(
        type='SECONDFPN',
        in_channels=[128, 256],
        out_channels=[256, 256],
        upsample_strides=[1, 2],
        norm_cfg=dict(type='BN', eps=0.001, momentum=0.01),
        upsample_cfg=dict(type='deconv', bias=False),
        use_conv_for_no_stride=True),
    imgpts_neck=dict(
        type='DeepInteractionEncoder',
        num_layers=2,
        in_channels_img=256,
        in_channels_pts=sum([256, 256]),
        hidden_channel=128,
        bn_momentum=0.1,
        bias='auto'
    ),
    pts_bbox_head=dict(
        type='DeepInteractionDecoder',
        num_views=num_views,
        out_size_factor_img=4,
        num_proposals=200,
        auxiliary=True,
        hidden_channel=128,
        num_classes=len(class_names),
        num_mmpi=4,
        num_heads=8,
        learnable_query_pos=False,
        initialize_by_heatmap=True,
        nms_kernel_size=3,
        ffn_channel=256,
        dropout=0.1,
        bn_momentum=0.1,
        activation='relu',
        common_heads=dict(center=(2, 2), height=(1, 2), dim=(3, 2), rot=(2, 2), vel=(2, 2)),
        bbox_coder=dict(
            type='TransFusionBBoxCoder',
            pc_range=point_cloud_range[:2],
            voxel_size=voxel_size[:2],
            out_size_factor=out_size_factor,
            post_center_range=[-61.2, -61.2, -10.0, 61.2, 61.2, 10.0],
            score_threshold=0.0,
            code_size=10,
        ),
        loss_cls=dict(type='FocalLoss', use_sigmoid=True, gamma=2, alpha=0.25, reduction='mean', loss_weight=1.0),
        loss_bbox=dict(type='L1Loss', reduction='mean', loss_weight=0.25),
        loss_heatmap=dict(type='GaussianFocalLoss', reduction='mean', loss_weight=1.0),
    ),
    train_cfg=dict(
        pts=dict(
            dataset='nuScenes',
            assigner=dict(
                type='HungarianAssigner3D',
                iou_calculator=dict(type='BboxOverlaps3D', coordinate='lidar'),
                cls_cost=dict(type='FocalLossCost', gamma=2, alpha=0.25, weight=0.15),
                reg_cost=dict(type='BBoxBEVL1Cost', weight=0.25),
                iou_cost=dict(type='IoU3DCost', weight=0.25)
            ),
            pos_weight=-1,
            gaussian_overlap=0.1,
            min_radius=2,
            grid_size=[1440, 1440, 40],  # [x_len, y_len, 1]
            voxel_size=voxel_size,
            out_size_factor=out_size_factor,
            code_weights=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.2, 0.2],
            point_cloud_range=point_cloud_range)),
    test_cfg=dict(
        pts=dict(
            dataset='nuScenes',
            grid_size=[1440, 1440, 40],
            out_size_factor=out_size_factor,
            pc_range=point_cloud_range[0:2],
            voxel_size=voxel_size[:2],
            nms_type=None,
        )))
optimizer = dict(type='AdamW', lr=0.0000125, weight_decay=0.01)  
optimizer_config = dict(grad_clip=dict(max_norm=0.1, norm_type=2))
lr_config = dict(
    policy='cyclic',
    target_ratio=(10, 0.0001),
    cyclic_times=1,
    step_ratio_up=0.4)
momentum_config = dict(
    policy='cyclic',
    target_ratio=(0.8947368421052632, 1),
    cyclic_times=1,
    step_ratio_up=0.4)
total_epochs = 6
checkpoint_config = dict(interval=1)
log_config = dict(
    interval=50,
    hooks=[dict(type='TextLoggerHook'),
           dict(type='TensorboardLoggerHook')])
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = None
load_from = 'pretrained/r50_fpn_voxel_0075.pth'
resume_from = None
workflow = [('train', 1)]
gpu_ids = range(0, 2)
find_unused_parameters = True

[Alexander0Yang]

You should use the appropriate batch size to reproduce the result.
In my experience, batch size >= 16 is necessary

[funnyjojo]

Thank you. I'll give it a try.