boxer2d_R_101_5x_lsj.yaml

includes:
- config/base_boxer2d_detection.yaml

dataset_config:
  detection:
    use_mask: true
    processors:
      image_train_processor:
        type: compose
        params:
          preprocessors:
          - type: random_horizontal_flip
            params:
              prob: 0.5
          - type: resize_scale
            params:
              min_scale: 0.4
              max_scale: 2.0
              image_size: 1024
          - type: random_size_crop_v2
            params:
              image_size: 1024
          - type: to_tensor
            params: {}
          - type: normalize
            params:
              mean:
              - 0.485
              - 0.456
              - 0.406
              std:
              - 0.229
              - 0.224
              - 0.225

model_config:
  boxer2d:
    use_mask: true
    loss:
      params:
        dice_loss_coef: 5
        mask_loss_coef: 5
        use_mask: ${model_config.boxer2d.use_mask}
    backbone:
      type: resnet101

scheduler:
  type: multi_step
  params:
    use_warmup: true
    mode: iter
    lr_ratio: 0.1
    lr_steps:
    - 327778
    - 355092
    warmup_iterations: 10
    warmup_factor: 1.0

optimizer:
  type: adamw
  params:
    lr: 0.0002
    lr_backbone: 2.0e-05
    weight_decay: 0.05

training:
  max_update: 368750
  max_norm: 0.01
  iou_type:
  - bbox
  - segm