sweep_neurips_db.yml

program: train.py
method: grid
metric:
  goal: minimize
  name: eval_loss

parameters:
  model:
    values: [resnet18, resnet18_ciga_ssl, resnet50, vit_small_r26_s32_224]
  pretrained:
    # The train.py script will skip the not pretrained resnet18_ciga_ssl. It doesn't
    # make sense to test non-pretrained resnet18_ciga_ssl because that is the same
    # as resnet18 from scratch.
    values: [true, false]
  data-dir:
    values:
      - data/task1_tumor_notumor/images/
      - data/task2_tils/images-tcga-tils/pancancer/
      - data/task3_msi_crc_ffpe/images/
      - data/task4_msi_crc_frozen/images/
      - data/task5_msi_stad_ffpe/images/
      - data/task6_precancer_vs_benign_polyps/images-split/
  seed:
    value: 42
  lr:
    value: 0.0001
  freeze-encoder:
    value: false
  epochs:
    value: 500
  sched:
    value: cosine
  opt:
    value: adamw
  opt-eps:
    value: .001
  warmup-lr:
    value: 1e-6
  weight-decay:
    value: 0.01
  drop:
    value: 0.3
  drop-path:
    value: 0.2
  model-ema:
    value: true
  model-ema-decay:
    value: 0.9999
  aa:
    value: rand-m9-mstd0.5
  remode:
    value: pixel
  reprob:
    value: 0.2
  amp:
    value: true
  train-split:
    value: train
  val-split:
    value: val
  workers:
    value: 4
  num-classes:
    value: 2
  batch-size:
    value: 84
  decay-epochs:
    value: 2.4
  decay-rate:
    value: .97
  log-wandb:
    value: true
  experiment:
    value: neuripsdb
  output:
    value: sweep-output/

command:
  - ${env}
  - ${interpreter}
  - ${program}
  - ${args_no_boolean_flags}