bevformer_tiny_r50_fpn_nuscenes.yml

batch_size: 1
epochs: 24
train_dataset:
  type: NuscenesMVDataset
  dataset_root: ./datasets/nuscenes
  ann_file: ./datasets/nuscenes/bevformer_nuscenes_annotation_train.pkl
  queue_length: 3
  use_valid_flag: True
  mode: train
  class_names: [
            'car', 'truck', 'construction_vehicle', 'bus', 'trailer',
            'barrier', 'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
        ]
  transforms:
    - type: LoadMultiViewImageFromFiles
      to_float32: True
    - type: PhotoMetricDistortionMultiViewImage
    - type: LoadAnnotations3D
      with_bbox_3d: True
      with_label_3d: True
    - type: SampleRangeFilter
      point_cloud_range: [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
    - type: SampleNameFilter
      classes: [
            'car', 'truck', 'construction_vehicle', 'bus', 'trailer',
            'barrier', 'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
        ]
    - type: NormalizeMultiviewImage
      mean: [123.675, 116.28, 103.53]
      std: [58.395, 57.12, 57.375]
      to_rgb: True
    - type: RandomScaleImageMultiViewImage
      scales: [0.5]
    - type: PadMultiViewImage
      size_divisor: 32
    - type: SampleFilerByKey
      keys: ['gt_bboxes_3d', 'gt_labels_3d', 'img']

val_dataset:
  type: NuscenesMVDataset
  dataset_root: ./datasets/nuscenes
  ann_file: ./datasets/nuscenes/bevformer_nuscenes_annotation_val.pkl
  queue_length: 3
  mode: val
  class_names: ['car', 'truck', 'construction_vehicle', 'bus', 'trailer',
                    'barrier', 'motorcycle', 'bicycle', 'pedestrian',
                    'traffic_cone']
  transforms:
    - type: LoadMultiViewImageFromFiles
      to_float32: True
    - type: NormalizeMultiviewImage
      mean: [123.675, 116.28, 103.53]
      std: [58.395, 57.12, 57.375]
      to_rgb: True
    - type: RandomScaleImageMultiViewImage
      scales: [0.5]
    - type: PadMultiViewImage
      size_divisor: 32
    - type: SampleFilerByKey
      keys: ['img']

optimizer:
  type: AdamW
  weight_decay: 0.01
  grad_clip:
    type: ClipGradByGlobalNorm
    clip_norm: 35

lr_scheduler:
  type: LinearWarmup
  learning_rate:
    type: CosineAnnealingDecayByEpoch
    learning_rate: 0.0002
    T_max: 24
    eta_min: 0.0000002
  warmup_steps: 500
  start_lr: 0.00006666666
  end_lr: 0.0002

model:
  type: BEVFormer
  use_grid_mask: True
  video_test_mode: True
  backbone:
    type: $paddledet.ResNet
    depth: 50
    lr_mult_list: [0.1, 0.1, 0.1, 0.1]
    return_idx: [3]
  neck:
    type: $paddledet.FPN
    in_channels: [2048]
    out_channel: 256
    spatial_scales: [0.03125]
    has_extra_convs: True
    extra_stage: 0
    use_c5: False
  pts_bbox_head:
    type: 'BEVFormerHead'
    bev_h: 50
    bev_w: 50
    num_classes: 10
    in_channels: 256
    num_query: 900
    sync_cls_avg_factor: True
    with_box_refine: True
    as_two_stage: False
    positional_encoding:
      type: 'LearnedPositionalEncoding'
      num_feats: 128
      row_num_embed: 50
      col_num_embed: 50
    transformer:
      type: 'PerceptionTransformer'
      rotate_prev_bev: True
      use_shift: True
      use_can_bus: True
      embed_dims: 256
      encoder:
        type: 'BEVFormerEncoder'
        num_layers: 3
        point_cloud_range: [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
        num_points_in_pillar: 4
        return_intermediate: False
        transformerlayers:
          type_name: 'BEVFormerLayer'
          attn_cfgs:  [
            {
              type_name: 'TemporalSelfAttention',
              embed_dims: 256,
              num_levels: 1
            },
            {
              type_name: 'SpatialCrossAttention',
              point_cloud_range: [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0],
              deformable_attention: {
                type_name: 'MSDeformableAttention3D',
                embed_dims: 256,
                num_points: 8,
                num_levels: 1
              },
              embed_dims: 256
            }
          ]
          feedforward_channels: 512
          ffn_dropout: 0.1
          operation_order: ['self_attn', 'norm', 'cross_attn', 'norm', 'ffn', 'norm']
      decoder:
        type: 'DetectionTransformerDecoder'
        num_layers: 6
        return_intermediate: True
        transformerlayers:
          type_name: 'DetrTransformerDecoderLayer'
          attn_cfgs: [
            {
              type_name: 'MultiheadAttention',
              embed_dims: 256,
              num_heads: 8,
              dropout: 0.1
            },
            {
              type_name: 'CustomMSDeformableAttention',
              embed_dims: 256,
              num_levels: 1
            },
          ]
          feedforward_channels: 512
          ffn_dropout: 0.1
          operation_order: ['self_attn', 'norm', 'cross_attn', 'norm', 'ffn', 'norm']
    bbox_coder:
      type: 'NMSFreeCoder'
      post_center_range: [-61.2, -61.2, -10.0, 61.2, 61.2, 10.0]
      point_cloud_range: [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
      max_num: 300
      voxel_size: [0.2, 0.2, 8]
      num_classes: 10
    loss_cls:
      type: 'WeightedFocalLoss'
      use_sigmoid: True
      gamma: 2.0
      alpha: 0.25
      loss_weight: 2.0
    loss_bbox:
      type: 'L1Loss'
      loss_weight: 0.25
    loss_iou:
      type: 'GIoULoss'
      loss_weight: 0.0
    assigner:
      type: 'HungarianAssigner3D'
      cls_cost:
        type: 'FocalLossCost'
        weight: 2.0
      reg_cost:
        type: 'BBox3DL1Cost'
        weight: 0.25
      iou_cost:
        type: 'IoUCost'
        weight: 0.0 # Fake cost. This is just to make it compatible with DETR head.
    sampler:
      type: 'PseudoSampler'