det_map/config/agent/map_agent.yaml

_target_: det_map.map.map_agent.MapAgent
_convert_: 'all'


is_train: &is_train
  is_train: True

point_cloud_range: &point_cloud_range
  pc_range: [ -15.0, -30.0, -10.0, 15.0, 30.0, 10.0 ]

lidar_filter_ranges: &lidar_filter_ranges
  x_range: (-15.0, 15.0)
  y_range: (-30.0, 30.0)
  z_range: (-10.0, 10.0)

voxel_size: &voxel_size
  voxel_size: [0.075, 0.075, 20.0]

img_voxel_size: &img_voxel_size
  voxel_size: [0.3, 0.3, 20.0]


dbound: &dbound
  dbound: [1.0, 35.0, 0.5]

grid_config: &grid_config
  grid_config:
    x: (-15.0, 15.0, 0.6)
    y: (-30.0, 30.0, 0.6)
    z: (-10.0, 10.0, 20.0)
    depth: (1.0, 60.0, 0.5)

img_norm_cfg : &img_norm_cfg
  img_norm_cfg:
    mean: [123.675, 116.28, 103.53]
    std: [58.395, 57.12, 57.375]
    to_rgb: True

map_classes: &map_classes
  map_classes: ['divider', 'ped_crossing','boundary', 'centerline']

#fixed_ptsnum_per_gt_line: &fixed_ptsnum_per_gt_line
#  fixed_ptsnum_per_gt_line: 20

#fixed_ptsnum_per_pred_line: &fixed_ptsnum_per_pred_line
#  fixed_ptsnum_per_pred_line: 20

eval_use_same_gt_sample_num_flag: &eval_use_same_gt_sample_num_flag
  eval_use_same_gt_sample_num_flag: True



#_pos_dim_: &_pos_dim_
#  _pos_dim_: 128

#_ffn_dim_: &_ffn_dim_
#  _ffn_dim_: 512

#_num_levels_: &_num_levels_
#  _num_levels_: 1

#bev_h_: &bev_h_
#  bev_h_: 100

#bev_w_: &bev_w_
#  bev_w_: 200

#queue_length: &queue_length
#  queue_length: 1

aux_seg : &aux_seg_cfg
  aux_seg:
    use_aux_seg: False
    bev_seg: False
    pv_seg: False
    seg_classes: 1
    feat_down_sample: 32
    pv_thickness: 1

#z_cfg : &z_cfg
#
#  pred_z_flag: True
#  gt_z_flag: True
  
model:
  _target_: det_map.map.map_model.MapModel
  _convert_: 'all'
  use_grid_mask: True
  video_test_mode: False
  pretrained:
    img: ckpts/resnet50-19c8e357.pth

  img_backbone:
    type: ResNet
    depth: 50
    num_stages: 4
    out_indices: [3]
    frozen_stages: 1
    norm_cfg: 
      type: BN
      requires_grad: False
    norm_eval: True
    style: pytorch
  img_neck:
    type: FPN
    in_channels: [2048]
    out_channels: 256
    start_level: 0
    add_extra_convs: on_output
    num_outs: 1
    relu_before_extra_convs: True
  pts_bbox_head:
    type: MapTRv2Head
    <<: *point_cloud_range
    bev_h: 100
    bev_w: 50
    num_query: 900
    num_vec_one2one: 20
    num_vec_one2many: 300
    k_one2many: 6
    num_pts_per_vec: 20
    num_pts_per_gt_vec: 20
    dir_interval: 1
    query_embed_type: 'instance_pts'
    transform_method: 'minmax'
    gt_shift_pts_pattern: 'v2'
    num_classes: 2
    in_channels: 256
    sync_cls_avg_factor: True
    with_box_refine: True
    as_two_stage: False
    code_size: 2
    code_weights: None
    <<: *aux_seg_cfg
#    z_cfg: *z_cfg
    transformer: 
      type: MapTRPerceptionTransformer
      bev_h: 100
      bev_w: 50
#      fuser:
#        type: 'ConvFuser'
#        in_channels: [256, 256]
#        out_channels: 256
      num_cams: 2
#      z_cfg: *z_cfg
      rotate_prev_bev: False
      use_shift: True
      use_can_bus: False
      embed_dims: 256
      encoder:
        type: 'SpatialDecoder'
        num_layers: 1
        <<: *point_cloud_range
        grid_config:
          x: [-15.0, 15.0, 0.6]
          y: [-30.0, 30.0, 0.6]
          z: [ -10.0, 10.0, 20.0 ]
        data_config:
          input_size: [256, 704]
        transformerlayers:
          type: 'SpatialDecoderLayer'
          attn_cfgs:
            - type: 'SpatialCrossAttention'
              <<: *point_cloud_range
              num_cams: 2
              dropout: 0.0
              embed_dims: 256
              deformable_attention:
                type: 'MSDeformableAttention'
                embed_dims: 256
                num_points: 8
                num_levels: 1
          ffn_cfgs:
            type: 'FFN'
            embed_dims: 256
            feedforward_channels: 1024
            ffn_drop: 0.0
            act_cfg:
              type: 'ReLU'
              inplace: True
          feedforward_channels: 1024
          ffn_dropout: 0.0
          operation_order: ['cross_attn', 'norm' ,'ffn', 'norm']
      decoder:
        type: MapTRDecoder
        num_layers: 6
        return_intermediate: True
        transformerlayers:
          type: DecoupledDetrTransformerDecoderLayer
          num_vec: 20
          num_pts_per_vec: 20
          attn_cfgs:
            - type: MultiheadAttention
              embed_dims: 256
              num_heads: 8
              dropout: 0.1
            - type: MultiheadAttention
              embed_dims: 256
              num_heads: 8
              dropout: 0.1
            - type: CustomMSDeformableAttention
              embed_dims: 256
              num_levels: 1
          feedforward_channels: 512
          ffn_dropout: 0.1
          operation_order: ['self_attn', 'norm', 'self_attn', 'norm', 'cross_attn', 'norm', 'ffn', 'norm']

    positional_encoding:
      type: LearnedPositionalEncoding
      num_feats: 128
      row_num_embed: 100
      col_num_embed: 50
    loss_cls:
      type: FocalLoss
      use_sigmoid: True
      gamma: 2.0
      alpha: 0.25
      loss_weight: 2.0
    loss_bbox:
      type: L1Loss
      loss_weight: 0.0
    loss_iou:
      type: GIoULoss
      loss_weight: 0.0
    loss_pts:
      type: PtsL1Loss
      loss_weight: 5.0
    loss_dir:
      type: PtsDirCosLoss 
      loss_weight: 0.005
    loss_seg:
      type: SimpleLoss
      pos_weight: 4.0
      loss_weight: 1.0
    loss_pv_seg:
      type: SimpleLoss 
      pos_weight: 1.0
      loss_weight: 2.0
#  train_cfg:
#    pts:
#      grid_size: [512, 512, 1]
#      <<: *voxel_size
#      point_cloud_range: [ -15.0, -30.0, -10.0, 15.0, 30.0, 10.0 ]
#      out_size_factor: 4
#      assigner:
#        type: MapTRAssigner
#        cls_cost:
#          type: FocalLossCost
#          weight: 2.0
#        reg_cost:
#          type: BBoxL1Cost
#          weight: 0.0
#          box_format: 'xywh'
#        iou_cost:
#          type: IoUCost
#          iou_mode: 'giou'
#          weight: 0.0
#        pts_cost:
#          type: OrderedPtsL1Cost
#          weight: 5
#        pc_range: [ -15.0, -30.0, -10.0, 15.0, 30.0, 10.0 ]

pipelines:
  lidar_filter:
    _target_: det_map.data.pipelines.filter_lidar.LiDARFilter
    _convert_: 'all'
    close_radius: 1.0
    <<: *lidar_filter_ranges

  #  only include in training
  point_shuffle:
    _target_: det_map.data.pipelines.point_shuffle.PointShuffle
    <<: *is_train

  lidar_aug:
    _target_: det_map.data.pipelines.lidar_aug.LiDARAug
    bda_aug_conf:
      rot_lim: (-22.5 * 2, 22.5 * 2)
      scale_lim: (0.9, 1.1)
      flip_dx_ratio: 0.5
      flip_dy_ratio: 0.5
      tran_lim: (0.5, 0.5, 0.5)
    <<: *lidar_filter_ranges
    #    if no aug for map, set this is_train to False
    <<: *is_train

  depth:
    _target_: det_map.data.pipelines.prepare_depth.LiDAR2Depth
    <<: *grid_config

  img:
    _target_: det_map.data.pipelines.prepare_img.PrepareImageInputs
    _convert_: 'all'
    opencv_pp: True
    #    Flag should be False in Eval!!!!
    <<: *is_train
    data_config:
      input_size: (256, 704)
      src_size: (900, 1600)
      # Augmentation
      resize: (-0.06, 0.44)
      rot: (-5.4, 5.4)
      flip: True
      crop_h: (0.0, 0.0)
      random_crop_height: True
      vflip: True
      resize_test: 0.04
      pmd:
        brightness_delta: 32
        contrast_lower: 0.5
        contrast_upper: 1.5
        saturation_lower: 0.5
        saturation_upper: 1.5
        hue_delta: 18
        rate: 0.5

#<<: *is_train
checkpoint_path: null
hidden_layer_dim: 512
lr: 1e-4