train:
  seed: 42
  epochs: 100
  batch_size: 8
  accum_steps: 1
  amp: true
  num_workers: 8
  device: cuda

dataset:
  name: lung_ultrasound_seg
  root: data/lung_ultrasound
  task_name: b
  image_size: [256, 256]
  in_channels: 3
  num_classes: 1
  train_split: train
  val_split: val
  test_split: test
  mask_suffix: .png
  image_suffix: .png

model:
  name: swin_unet
  encoder_name: swinv2_base_patch4_window12_192_22k
  in_channels: 3
  out_channels: 1
  img_size: 256
  drop_rate: 0.0
  drop_path_rate: 0.2

pretrain:
  enabled: true
  source: imagenet22k
  checkpoint: weights/swinv2_base_patch4_window12_192_22k.pth
  strict: false

loss:
  task_name: b
  task_mode: binary

metrics:
  task_mode: binary
  metrics:
    - name: dice
    - name: iou

optimizer:
  name: adamw
  lr: 5.0e-5
  weight_decay: 0.05
  betas: [0.9, 0.999]

scheduler:
  name: cosine
  warmup:
    name: linear
    params:
      start_factor: 0.1
      total_iters: 10
  params:
    T_max: 100
    eta_min: 1.0e-6

augmentation:
  train:
    random_flip: true
    random_rotate_90: true
    random_resized_crop: false
    random_brightness_contrast: true
    random_gaussian_noise: true
  val:
    center_crop: false

validation:
  enabled: true
  interval: 1
  metrics: [dice, iou]
  save_best: true
  monitor: dice
  mode: max

checkpoint:
  dir: outputs/segmentation/train_seg_b
  save_last: true
  save_best_only: false
  keep_top_k: 3

logging:
  log_interval: 20
  use_tensorboard: true
  tensorboard_dir: outputs/tensorboard/train_seg_b