default.yaml

task:
  name: classification
  output_dir_name: classification
  batch_size: 512
  max_epochs: 80
  max_steps: null
  target_metric: val_top5_err
  target_metric_mode: min
  # davit_tiny takes too long to train for us; performance was fine with custom stem
  # make sure to comment out the other model when switching to davit
  # model:
  #   name: davit_tiny.msft_in1k
  #   stem: custom_conv  # {default, wrn_conv, custom_conv}
  # wide res net
  label_smoothing: 0.0
  model:
   name: wide_resnet50_2
   kernel_size: 3
   stride: 1
   padding: 1
   maxpool: True
  train_transforms:
    random_crop:
      use: True
      size: 64
      padding: 4
      padding_mode: reflect
    horizontal_flip:
      use: True
      p: 0.5
    trivial_augment:
      use: True
engine:
  devices: 4
  sbatch_args:
    time: 05:00:00
optimizer:
  name: adamw_baseline
  weight_decay: 1.0
evaluation:
  plot:
    metric: test_top1_acc
    test_metric_mode: max
    format: "2.2"
    limits: [55, 72]