AutoMAX

AutoMAX is an automated hyperparameter optimization framework for AUC-focused objectives, built on top of LibAUC and SMAC3.

Instead of manually sweeping learning rates, margins, and decay schedules, AutoMAX performs Bayesian search over predefined hyperparameter spaces for each supported loss/optimizer pair, then returns the best configuration for your dataset.

Documentation

Full documentation is available at:

• AutoMAX Documentation

Installation

Clone AutoMAX library

git clone https://github.com/Optimization-AI/AutoMAX.git
cd AutoMAX

Clone the dev branch of libauc 1.5.0

git clone https://github.com/Optimization-AI/LibAUC.git
cd LibAUC
git checkout 1.5.0_dev
cd ..
cp LibAUC/libauc ./libauc

Install other dependencies

conda create -n AutoMAX python=3.10
conda activate AutoMAX
conda install gxx_linux-64 gcc_linux-64 swig
pip install -r requirements.txt

Quick Start

Run AutoTune for AUROC on CIFAR-10 with ResNet-18:

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

Any training field can be overridden from the command line:

python -m src.auto_trainer --config_file ./recipes/config_auc.yaml --epochs 50 --output_path ./runs

Supported Loss / Optimizer Pairs

The supported loss and optimizer pairs can be found at LibAUC document.