scannet_model.yaml

CLASS_NAMES: [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18]
BATCH_SIZE: 8
DATA_CONFIG:
    _BASE_CONFIG_: /tools/cfgs/scannet_configs/scannet.yaml


MODEL:
    NAME: CenterPoint

    VFE:
        NAME: MeanVFE

    BACKBONE_3D:
        NAME: VoxelBackBone8x

    MAP_TO_BEV:
        NAME: HeightCompression
        NUM_BEV_FEATURES: 640 # for Height Compression, it doesnot create more channels like PointPillarNet, so this para should be consistent with output of Backbone_3d, which is 128

    BACKBONE_2D:
        NAME: BaseBEVBackbone

        LAYER_NUMS: [5]
        LAYER_STRIDES: [1]
        NUM_FILTERS: [128]
        UPSAMPLE_STRIDES: [2]
        NUM_UPSAMPLE_FILTERS: [256]

    DENSE_HEAD:
        NAME: CenterHead_single
        CLASS_AGNOSTIC: False

        USE_DIRECTION_CLASSIFIER: False

        TARGET_ASSIGNER_CONFIG:  # ToDo
            MAX_OBJS: 64
            VOXEL_SIZE: [0.025, 0.025, 0.05]
            OUT_SIZE_FACTOR: 4
            GAUSSIAN_OVERLAP: 0.01  # relevant to gaussian radius see: centerpoint_head_single.py/gaussian_radius---default 0.1
            MIN_RADIUS: 1   #have to be an interger  --default1
            BOX_CODER: ResidualCoder

        LOSS_CONFIG:
            LOSS_WEIGHTS: {
                'cls_weight': 1.0,
                'loc_weight': 2.0,
                'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
            }

    POST_PROCESSING:
        RECALL_THRESH_LIST: [0.3, 0.5, 0.7]
        SCORE_THRESH: 0.05 # original 0.1
        OUTPUT_RAW_SCORE: False

        EVAL_METRIC: kitti

        NMS_CONFIG:
            MULTI_CLASSES_NMS: False
            NMS_TYPE: nms_gpu
            NMS_THRESH: 0.25
            NMS_PRE_MAXSIZE: 4096
            NMS_POST_MAXSIZE: 500


OPTIMIZATION:  # all default except bs_per_gpu
    BATCH_SIZE_PER_GPU: 8
    NUM_EPOCHS: 200

    OPTIMIZER: adam_onecycle
    LR: 0.003
    WEIGHT_DECAY: 0.01
    MOMENTUM: 0.9

    MOMS: [0.95, 0.85]
    PCT_START: 0.4
    DIV_FACTOR: 10
    DECAY_STEP_LIST: [35, 45]
    LR_DECAY: 0.1
    LR_CLIP: 0.0000001

    LR_WARMUP: False
    WARMUP_EPOCH: 1

    GRAD_NORM_CLIP: 10