AUROC — AUCMLoss + PESG

The canonical setup for binary imbalanced classification. AUCMLoss with PESG directly maximizes AUROC using a minimax formulation.

Config

recipes/config_auc.yaml

dataset:
  name: cifar10
  eval_splits: [val, test]
  kwargs:
    imratio: 0.1

model:
  name: resnet18
  pretrained: false
  num_classes: 1
  in_channels: 3

metrics:
  - AUROC

training:
  project_name: libauc
  experiment_name: resnet18_AUCMLoss_cifar10
  SEED: 2026
  epochs: 100
  batch_size: 128
  eval_batch_size: 256
  sampling_rate: 0.2
  num_workers: 0
  decay_epochs: [0.5, 0.75]
  loss: AUCMLoss
  optimizer: PESG
  output_path: ./output
  resume_from_checkpoint: false
  save_checkpoint_every: 5

automax:
  deterministic: true
  n_trials: 5
  SEED: 42
  name: resnet18_AUCMLoss_cifar10
  output_directory: ./automax_output
  overwrite: true

Run

python -m src.auto_trainer \
  --config_file recipes/config_auc.yaml

Hyperparameter search space

AutoMAX tunes the following parameters for AUCMLoss + PESG:

Parameter	Source	Type	Notes
lr	optimizer	log-uniform range	Learning rate
epoch_decay	optimizer	range	LR decay factor applied at decay_epochs
weight_decay	optimizer	log-uniform range	L2 regularization
momentum	optimizer	categorical	Momentum coefficient
margin	loss	categorical	Surrogate loss margin