This project provides four AutoML hyperparameter optimization pipelines using the Optuna framework to identify the best-performing configuration for both image and two-dimensional datasets. Sepcifically, it includes two scripts for image data, one for Bayesian optimization and random search, and another using successive halving, as well as two parallel scripts for two-dimensional data.
After cloning the repository, install the required dependencies using the requirements.txt file.
All scripts can be run with the following command-line arguments:
- name: (required) the name of the study
- sampler: (required, except for successive halving scripts) what type of search strategy, choose between Bayesian optimization or random search
- resume: (optional) Add this flag if you want to resume an existing study instead of starting a new one
- dataset: (required only for two-dimensional scripts) Specifies the type of dataset to use, with the following options; swissroll, spiral, or moons
When running the script, it will create a new folder called studies inside this folder, it will create a subfolder based on the data type, and within that, another subfolder using the study name provided as an argument. This nested directory will store all study-related outputs, including the study checkpoint for later resumption and the generated images at each trial.
Note: The ddpm folder is included so that each trial can load the necessary model files for evaluation.
python3 HPO.py --name butterflies-rs --sampler TPE
python3 HPO-2D.py --name swissroll-rs --sampler TPE --dataset swissroll