model:
  names:
  - sam
  sam:
    checkpoint_name: facebook/sam-vit-base
    image_norm: imagenet
    data_types:
    - semantic_segmentation_img
    train_transforms:
    - random_horizontal_flip
    val_transforms: []
    img_transforms:
    - resize_to_square
    gt_transforms:
    - resize_gt_to_square
    max_img_num_per_col: 1
    frozen_layers:
    - mask_decoder.iou_prediction_head
    - prompt_encoder
    num_mask_tokens: 1
    ignore_label: 255
data:
  image:
    missing_value_strategy: zero
  text:
    normalize_text: false
  categorical:
    minimum_cat_count: 100
    maximum_num_cat: 20
    convert_to_text: true
  numerical:
    convert_to_text: false
    scaler_with_mean: true
    scaler_with_std: true
  document:
    missing_value_strategy: zero
  label:
    numerical_label_preprocessing: standardscaler
  pos_label: null
  column_features_pooling_mode: concat
  mixup:
    turn_on: false
    mixup_alpha: 0.8
    cutmix_alpha: 1.0
    cutmix_minmax: null
    prob: 1.0
    switch_prob: 0.5
    mode: batch
    turn_off_epoch: 5
    label_smoothing: 0.1
  templates:
    turn_on: false
    num_templates: 30
    template_length: 2048
    preset_templates:
    - super_glue
    - rte
    custom_templates: null
optimization:
  optim_type: adamw
  learning_rate: 0.0001
  weight_decay: 0.0001
  lr_choice: single_stage
  lr_decay: 0
  lr_schedule: polynomial_decay
  max_epochs: 30
  max_steps: -1
  warmup_steps: 0.0
  end_lr: 0
  lr_mult: 1
  patience: 10
  val_check_interval: 1.0
  check_val_every_n_epoch: 1
  skip_final_val: false
  gradient_clip_val: 1
  gradient_clip_algorithm: norm
  track_grad_norm: -1
  log_every_n_steps: 10
  top_k: 3
  top_k_average_method: best
  efficient_finetune: lora
  lora:
    module_filter:
    - .*vision_encoder.*attn
    filter:
    - q
    - v
    r: 3
    alpha: 32
  loss_function: structure_loss
  focal_loss:
    alpha: null
    gamma: 2.0
    reduction: mean
  mask2former_loss:
    loss_cross_entropy_weight: 10.0
    loss_mask_weight: 5.0
    loss_dice_weight: 5.0
  extra_trainable_params:
  - .*mask_decoder
env:
  num_gpus: 1
  num_nodes: 1
  batch_size: 4
  per_gpu_batch_size: 1
  eval_batch_size_ratio: 1
  per_gpu_batch_size_evaluation: null
  precision: 16-mixed
  num_workers: 4
  num_workers_evaluation: 2
  accelerator: auto
  fast_dev_run: false
  deterministic: false
  auto_select_gpus: false
  strategy: auto
  deepspeed_allgather_size: 1000000000.0
  deepspeed_allreduce_size: 1000000000.0
  compile:
    turn_on: false
    mode: default
    dynamic: true
    backend: inductor