iassd_kitti.yaml

batch_size: 8  #on 4 gpus, total bs = 32
epochs: 80

train_dataset:
  type: KittiPCDataset
  dataset_root: datasets/KITTI
  class_names: [ "Car", "Pedestrian", "Cyclist"]
  use_road_plane: True
  transforms:
    - type: LoadPointCloud
      dim: 4
      use_dim: 4
    - type: RemoveCameraInvisiblePointsKITTIV2
    - type: SamplingDatabase
      min_num_points_in_box_per_class:
        Car: 5
        Pedestrian: 5
        Cyclist: 5
      max_num_samples_per_class:
        Car: 20
        Pedestrian: 15
        Cyclist: 15
      ignored_difficulty: [ -1 ]
      database_anno_path: datasets/KITTI/kitti_train_gt_database/anno_info_train.pkl
      database_root: datasets/KITTI
      class_names: [ "Car", "Pedestrian", "Cyclist" ]
    - type: RandomVerticalFlip
    - type: GlobalRotate
      min_rot: -0.78539816
      max_rot: 0.78539816
    - type: GlobalScale
      min_scale: 0.95
      max_scale: 1.05
    - type: FilterBBoxOutsideRange
      point_cloud_range: &point_cloud_range [ 0, -40, -3, 70.4, 40, 1 ]
    - type: FilterPointOutsideRange
      point_cloud_range: *point_cloud_range
    - type: SamplePoint
      num_points: 16384
    - type: ShufflePoint
    - type: ConvertBoxFormat
  mode: train

val_dataset:
  type: KittiPCDataset
  dataset_root: datasets/KITTI
  class_names: [ "Car", "Pedestrian", "Cyclist"]
  transforms:
    - type: LoadPointCloud
      dim: 4
      use_dim: 4
    - type: RemoveCameraInvisiblePointsKITTIV2
    - type: FilterPointOutsideRange
      point_cloud_range: *point_cloud_range
    - type: SamplePoint
      num_points: 16384
  mode: val

model:
  type: IASSD

  backbone:
    type: IASSD_Backbone
    npoint_list: [4096, 1024, 512, 256, null, 256]
    sample_method_list: &sample_method_list ["D-FPS", "D-FPS", "ctr_aware", "ctr_aware", null, null]
    radius_list: [[0.2,0.8], [0.8,1.6], [1.6,4.8], [], [], [4.8, 6.4]]
    nsample_list: [[16,32], [16,32], [16,32], [], [], [16, 32]]
    mlps: [[[16,16,32],  [32,32,64]],
           [[64,64,128],  [64,96,128]],
           [[128,128,256],  [128,256,256]],
           [],
           [128],
           [[256,256,512], [256,512,1024]]]
    layer_types: ["SA_Layer", "SA_Layer", "SA_Layer", "SA_Layer", "Vote_Layer", "SA_Layer"]
    dilated_group: [False, False, False, False, False, False]
    aggregation_mlps: [[64], [128], [256], [256], [], [512]]
    confidence_mlps: [[], [128], [256], [], [], []]
    layer_input: [0, 1, 2, 3, 4, 3]
    ctr_index: [-1, -1, -1, -1, -1, 5]
    max_translate_range: [3., 3., 2.]
    input_channel: 4
    num_classes: 3

  head:
    type: IASSD_Head
    input_channel: 512
    cls_fc: [256, 256]
    reg_fc: [256, 256]
    num_classes: 3

    target_config:
      gt_extra_width: [0.2, 0.2, 0.2]
      extra_width: [1.0, 1.0, 1.0]
      box_coder_config: {
        'angle_bin_num': 12,
        'use_mean_size': True,
        'mean_size': [
            [3.9, 1.6, 1.56],
            [0.8, 0.6, 1.73],
            [1.76, 0.6, 1.73]
        ]
      }

    loss_config:
      loss_cls: WeightedClassificationLoss
      loss_reg: WeightedSmoothL1Loss
      loss_ins: WeightedClassificationLoss
      sample_method_list: *sample_method_list
      corner_loss_regularization: True
      centerness_regularization: True
      centerness_regularization_sa: True
      loss_weight: {
          'ins_aware_weight': [0, 1.0, 1.0],
          'vote_weight': 1.0,
          'point_cls_weight': 1.0,
          'point_box_weight': 1.0,
          'corner_weight': 1.0,
          'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
          'dir_weight': 0.2
      }

  post_process_cfg:
    score_thresh: 0.1
    nms_config:
      nms_thresh: 0.01
      nms_pre_maxsize: 4096
      nms_post_maxsize: 500

optimizer:
  type: AdamWOnecycle
  clip_grad_by_norm: 10.0
  learning_rate: 0.01
  beta1: 0.9
  beta2: 0.99
  weight_decay: 0.01

lr_scheduler:
  type: OneCycle
  total_step: 9280  #change to your correspondent total iter
  lr_max: 0.01
  moms: [0.95, 0.85]
  div_factor: 10
  pct_start: 0.4