In-context Freeze-Thaw Bayesian Optimization for Hyperparameter Optimization

This repository contains the official code for our ICML 2024 paper. This main branch provides the Freeze-Thaw PFN surrogate (FT-PFN) surrogate model as a drop-in surrogate for multi-fidelity Bayesian Optimization loops. Along with the synthetic prior generation and training code. To reproduce experiments from the above paper version, please refer to the branch icml-2024.

To use the ifBO algorithm in practice, please refer to NePS, a package for hyperparameter optimization that maintains the latest, improved ifBO version (TBA, TODO).

Setup

conda create -n ifBO-env python=3.10 setuptools
conda activate ifBO-env
pip install -e .

Surrogate versions

Version	Identifier	Notes
0.0.1	ICML '24 submission	FT-PFN from ifBO, trained on LCNet curves, DPL power law, broke scaling law

Surrogate usage API

# To initialize the surrogate and load pretrained weights
from ifbo.surrogate import FTPFN

model = FTPFN()

NOTE: This creates a .model/ directory in the current working directory for the surrogate model. To have control over this specify a target_path: Path when initializing.

To cite:

If using our surrogate, code, experiment setup, kindly cite using:

@inproceedings{
  rakotoarison-icml24,
  title={In-Context Freeze-Thaw Bayesian Optimization for Hyperparameter Optimization},
  author={H. Rakotoarison and S. Adriaensen and N. Mallik and S. Garibov and E. Bergman and F. Hutter},
  booktitle={Forty-first International Conference on Machine Learning},
  year={2024},
  url={https://openreview.net/forum?id=VyoY3Wh9Wd}
}