sstv2.yaml

voxel_size: [0.15, 0.15, 8]
point_cloud_range: [-54.0, -54.0, -5.0, 54.0, 54.0, 3.0]

model:
  type: BEVFusion_SST
  encoders:
    lidar:
      voxelize:
        point_cloud_range: ${point_cloud_range}
        voxel_size: ${voxel_size}
        max_voxels: [-1, -1]
        max_num_points: -1
        Voxelization: true

      voxel_encoder:
        type: DynamicVFE
        in_channels: 5
        feat_channels: [64, 128]
        with_distance: false
        voxel_size: ${voxel_size}
        with_cluster_center: true
        with_voxel_center: true
        point_cloud_range: ${point_cloud_range}
      
      middle_encoder:
        type: SSTInputLayerV2
        window_shape: [16, 16, 1]
        sparse_shape: [720, 720, 1]
        shuffle_voxels: true
        debug: true
        pos_temperature: 10000
        normalize_pos: false
        mute: true

      backbone:
        type: SSTv2
        d_model: [128, 128, 128, 128, 128, 128, 128, 128]
        nhead: [8, 8, 8, 8, 8, 8, 8, 8]
        num_blocks: 8
        dim_feedforward: [256, 256, 256, 256, 256, 256, 256, 256]
        output_shape: [720, 720]
        num_attached_conv: 3
        conv_in_channel: 128
        conv_out_channel: 128
        debug: true
  decoder:
    backbone:
      in_channels: 128
      out_channels: [64, 128, 256]
      layer_nums: [3, 5, 5]
      layer_strides: [2, 2, 2]
    neck:
      in_channels: [64, 128, 256]
      out_channels: [128, 128, 128]
      upsample_strides: [0.5, 1, 2]

  heads:
    object:
      in_channels: 384
      train_cfg:
        grid_size: [720, 720, 1]
        out_size_factor: 4
      test_cfg:
        grid_size: [720, 720, 1]
        out_size_factor: 4
      bbox_coder:
        out_size_factor: 4

checkpoint_config:
  max_keep_ckpts: 4

data:          
  samples_per_gpu: 4
  workers_per_gpu: 12