准确率 Accuracy

[chenqi13814529300]

What parameters can be set to achieve the accuracy of the author

[shanjiayao]

Could you please provide more information about the problem?

If you want to retrain our code, the default parameters could achieve the performance in README.

[chenqi13814529300]

It uses PTT on GitHub Yaml default file, using Kitti data set, training car, transformer_ Block uses transformerblock

[chenqi13814529300]

` CLASS_NAMES: Car # ['Car', 'Pedestrian', 'Cyclist', 'Van']

DATA_CONFIG: DATASET: 'KittiTrackingDataset' DATA_PATH: '../data/kitti' DEBUG: False # whether debug mode REF_COOR: 'lidar' # lidar / camera USE_Z_AXIS: True # whether use the z prediction of model LOAD_FROM_DATABASE: True # whether use the memory to save time LIDAR_CROP_OFFSET: 10.0 # the crop offset when training NUM_CANDIDATES_PERFRAME: 4 # generate random offsets per frame SEARCH_INPUT_SIZE: 1024 # the regularized points number of search area TEMPLATE_INPUT_SIZE: 512 # the regularized points number of template area SEARCH_BB_OFFSET: 0.0 SEARCH_BB_SCALE: 1.25 MODEL_BB_OFFSET: 0.0 MODEL_BB_SCALE: 1.25 REFINE_BOX_SIZE: True # whether scale the box by a ratio, adopted from p2b POINT_CLOUD_RANGE: [-1, -1, -1] # -1 means unlimited DATA_SPLIT: { 'train': train, 'test': test } SAMPLED_INTERVAL: 1 # only for training INFO_PATH: { 'train': kitti_infos_train.pkl, 'test': kitti_infos_test.pkl, } FOV_POINTS_ONLY: False POINT_FEATURE_ENCODING: { encoding_type: absolute_coordinates_encoding, used_feature_list: [ 'x', 'y', 'z' ], src_feature_list: [ 'x', 'y', 'z', 'intensity' ], }

MODEL: NAME: PTT # PTT / P2B BACKBONE_3D: NAME: PointNet2BackboneLight DEBUG: False # whether debug mode SA_CONFIG: SAMPLE_METHOD: ['fps', 'sequence', 'sequence'] # fps/sequence/rs ffps not support USE_XYZ: True NORMALIZE_XYZ: True NPOINTS_SEARCH: [512, 256, 128] NPOINTS_TEMPLATE: [256, 128, 64] RADIUS: [0.3, 0.5, 0.7] NSAMPLE: [32, 32, 32] MLPS: [[0, 64, 64, 128], [128, 128, 128, 256], [256, 128, 128, 256]]

SIMILARITY_MODULE:
    NAME: CosineSimAug
    DEBUG: False   # whether debug mode
    MLP:
        CHANNELS: [260, 256, 256, 256]
        BN: True
    CONV:
        CHANNELS: [256, 256, 256]
        BN: True

CENTROID_HEAD:
    NAME: CentroidVotingHead
    DEBUG: False   # whether debug mode
    CLS_USE_SEARCH_XYZ: False
    CLS_FC:
        CHANNELS: [256, 256, 256, 1]
    REG_FC:
        CHANNELS: [259, 256, 256, 259]

    TRANSFORMER_BLOCK:
        ENABLE: True
        NAME: TransformerBlock   # TransformerBlock / MulTransformerBlock
        DIM_INPUT: 256
        DIM_MODEL: 512
        KNN: 16
        N_HEADS: 1
        N_LAYERS: 1

    LOSS_CONFIG:
        CLS_LOSS: BinaryCrossEntropy
        CLS_LOSS_REDUCTION: 'mean'
        CLS_LOSS_POS_WEIGHT: 1.0
        REG_LOSS: smooth-l1
        LOSS_WEIGHTS: {
            'centroids_cls_weight': 0.2,
            'centroids_reg_weight': 1.0
        }

BOX_HEAD:
    NAME: BoxVotingHead
    DEBUG: False   # whether debug mode
    FC: [256, 256, 256, 5]

    SA_CONFIG:
        NPOINTS: 64
        RADIUS: 0.3
        NSAMPLE: 16
        MLPS: [257, 256, 256, 256]
        USE_XYZ: True
        NORMALIZE_XYZ: True
        SAMPLE_METHOD: 'fps'

    TRANSFORMER_BLOCK:
        ENABLE: True
        NAME: TransformerBlock   # TransformerBlock / MulTransformerBlock
        DIM_INPUT: 256
        DIM_MODEL: 512
        KNN: 16
        N_HEADS: 1
        N_LAYERS: 1

    LOSS_CONFIG:
        CLS_LOSS: BinaryCrossEntropy
        CLS_LOSS_REDUCTION: 'none'
        CLS_LOSS_POS_WEIGHT: 2.0
        REG_LOSS: smooth-l1
        LOSS_WEIGHTS: {
            'boxes_cls_weight': 1.5,
            'boxes_reg_weight': 0.2
        }

OPTIMIZATION: DEBUG: False # whether debug mode BATCH_SIZE_PER_GPU: 48 NUM_EPOCHS: 60

optimizer

OPTIMIZER: adam
LR: 0.001
WEIGHT_DECAY: 0
BETAS: [0.5, 0.999]
EPS: 1e-06
# scheduler
SCHEDULER: 'step'
STEP_SIZE: 12
GAMMA: 0.2
# grad clip
GRAD_NORM_CLIP: 10

TRAIN: WITH_EVAL: ENABLE: False START_EPOCH: 3 INTERVAL: 1 TEST: VISUALIZE: False SAVE_PCD: False # TODO check could use or not SHAPE_AGGREGATION: firstandprevious # first/previous/firstandprevious/all REF_BOX: previous_result # previous_result/previous_gt/current_gt

`

[chenqi13814529300]

TransformerBlock / MulTransformerBlock Which one will be better?

[shanjiayao]

As for the performance, the default parameters could achieve the performance in README.

And about the transformer type, we found that only using one layer and one head is better. Hence please use the default TransformerBlock.

[chenqi13814529300]

Thank you. I just tried it yesterday. By default, it can achieve better results