name: "BDA_Deeplabv3plus_ResNet101_CCE"
model_type: "MaskedTower"
model: "MaskedDeepLabV3PlusResNet"
task: "BDA"
training:
  training_parameters:
    gradient_clip_val: 10.0
    alpha: 0.25
    batch_size: 1
    gamma: 2
    grad_accumulation: 8
    l1_reg: 1.0e-06
    l2_reg: 1.0e-05
    log_images_every_n_steps: 500
    mask_x: 1024
    mask_y: 1024
    max_epochs: 1000
    optimizer_parameters:
      name: "AdamW"
      lr: 0.0005
      weight_decay: 0.00001
    output_class_weights_strategy: "ipw"
    normalize_weights: False
    output_class_weights:
      background: 1.0e-06
      no damage: 1.76
      major damage: 9.32
      minor damage: 3.68
      destroyed: 33.4
      un-classified: 40.06
    tile_x: 1024
    tile_y: 1024
    loss_parameters:
      loss: [["cross entropy", 1.0]]
      gamma: 2
      alpha: 0.25
      bins: 10
      momentum: 0.9
      smooth: 1.0e-06
validation:
  validation_parameters:
    validation_monitor: "val_macro_f1"
    validation_monitor_mode: "max"
    validation_checkpoint_save_top_k: 3
    validation_reduce_lr_on_plateau_cooldown: 3
    validation_reduce_lr_on_plateau_patience: 3
    batch_size: 2
input:
  normalized_inputs: true
  mask_input: false
model_parameters:
  n_cls: 6
  encoder_parameters:
    depth: 101
    num_stages: 4
    out_indices: [0, 1, 2, 3]
    dilations: [1, 1, 2, 4]
    strides: [1, 2, 1, 1]
    norm_eval: False
    contract_dilation: True
  decoder_parameters:
    in_channels: 2048
    in_index: 3
    channels: 512
    dilations: [1, 12, 24, 36]
    c1_in_channels: 256
    c1_channels: 48
    dropout_ratio: 0.1
    align_corners: False
  auxiliary_head_parameters:
    in_channels: 1024
    in_index: 2
    channels: 256
    num_convs: 1
    concat_input: False
    dropout_ratio: 0.1
    align_corners: False