What is MFBML?

Documentation | Installation | GitHub | Tutorials

Summary

mfbml provides provide a general Multi-Fidelity Bayesian Machine Learning framework. The developed MF-BML framework can be used to handle both data scarce and data rich data set scenario depending on the employed algorithm within the framework. The developed MF-BML framework doesn't restrict any algorithm, two configurations are recommended in this repo for handling data scarce and large data set problems respectively.

State of need

mfbml is a package that supports general multi-fidelity Bayesian machine learning. Two practical multi-fidelity Bayesian machine learning algorithms from the paper: 1) Kernel Ridge Regression + Linear Transfer-learning + Gaussian Process Regression (KRR-LR-GPR), implemented based on Numpy; 2) Deep Neural Network + Linear Transfer-learning + Bayesian Neural Network (DNN-LR-BNN), implemented based on Pytorch.

In the particular case of a research environment, `mfbml is designed to easily accommodate further developments, either by improving the already implemented methods or by including new numerical models and techniques.

Authorship and Citation

Author:

• Jiaxiang Yi (J.Yi@tudelft.nl)

Author affiliation:

• Delft University of Technology

arXiv (paper): @misc{yi2024practicalmultifidelitymachinelearning, title={Practical multi-fidelity machine learning: fusion of deterministic and Bayesian models}, author={Jiaxiang Yi and Ji Cheng and Miguel A. Bessa}, year={2024}, eprint={2407.15110}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2407.15110}, }

Get started

Installation

(1). git clone the repo to your local machine

https://github.com/JiaxiangYi96/mfbml.git cd mfbml (2) create a new conda environment with python version 3.10 conda create -n mfbml_env python=3.10 conda activate mfbml_env (3). install dependencies first (a git repo mfpml with branch main and pytorch with cpu installation)

pip install -r requirements.txt pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

(4). go to the local folder where you cloned the repo, and pip install it with editable mode

pip install --verbose --no-build-isolation --editable .

Illustrative examples

1. Kernel Ridge Regression + Linear Transfer-learning + Gaussian Process Regression Notebook

2. Deep Neural Network + Linear Transfer-learning + Bayesian Neural Network Notebook

3. More illustrative examples shown in the paper can be found in the studies folder.

Community Support

If you have any question, please raise an issue on GitHub or contact the developer

License

BSD 3-Clause License, Jiaxiang Yi

All rights reserved.

mfbml is a free and open-source repo published under BSD 3-Clause License.