Train with LibAUC Trainer

Author: Siqi Guo, Gang Li, Tianbao Yang

Introduction

This tutorial shows a two-stage workflow using the built-in LibAUC Trainer:

  1. Pretrain a ResNet18 model with cross-entropy style binary training.

  2. Load that checkpoint and optimize AUROC with AUCMLoss + PESG.

The Trainer CLI entry point is:

python -m libauc.trainer.run_trainer --config_file <path_to_yaml>

Step 1: CE pretraining config

Create a YAML file named ce_config.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_ce_cifar10
  SEED: 2026

  epochs: 100
  batch_size: 128
  eval_batch_size: 256
  sampling_rate: 0.5
  num_workers: 0
  decay_epochs: [0.5, 0.75]

  loss: BCELoss
  optimizer: Adam
  optimizer_kwargs:
    lr: 1.0e-3
    weight_decay: 1.0e-4

  output_path: ./output
  resume_from_checkpoint: false
  save_checkpoint_every: 5

Run CE pretraining

python -m libauc.trainer.run_trainer --config_file ce_config.yaml

After training, the checkpoint is expected at:

./output/resnet18_ce_cifar10/epoch_100.pt

Step 2: AUCMLoss optimization config

Create a second YAML file named aucmloss_config.yaml:

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

model:
  name: resnet18
  pretrained: true
  pretrained_path: "./output/resnet18_ce_cifar10/epoch_100.pt"
  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

Run AUROC optimization

python -m libauc.trainer.run_trainer --config_file aucmloss_config.yaml

Expected outputs

  • CE stage checkpoints: ./output/resnet18_ce_cifar10/

  • AUCMLoss stage checkpoints: ./output/resnet18_AUCMLoss_cifar10/

  • Validation/test AUROC logs shown during Trainer evaluation callbacks.

Tips

  • If you train on CPU, set smaller batch_size values.

  • You can override selected training args from CLI, e.g.:

python -m libauc.trainer.run_trainer --config_file aucmloss_config.yaml --epochs 20 --batch_size 64