HPOMax

Easy-to-configure hyper-parameters (HPs) optimization for Deep Learning simulations on top of Scikit-Optimize.

This library allows to parse easily the parameters for your model. Then, you can configure the HPs that will be optimized, and run the skopt.gp_minimize function. The template folder allows to define new HPO processes.

This library was used for HPO applied to physics-informed neural networks in:

@misc{https://doi.org/10.48550/arxiv.2205.06704,
  doi = {10.48550/ARXIV.2205.06704},
  url = {https://arxiv.org/abs/2205.06704},
  author = {Escapil-Inchauspé, Paul and Ruz, Gonzalo A.},
  title = {Hyper-parameter tuning of physics-informed neural networks: Application to Helmholtz problems},
  publisher = {arXiv},
  year = {2022},
}

How it works?

First, go to (or copy) the template folder:

cd template

• In config.yaml, define all the parameters for your model. This will allow you to run different models with one line of code as the parameters will be parsed directly from the terminal.
• Define your model in model.py. The model is with two functions:
  • create_model: define and compile your model
  • train_model: train the model and return its accuracy as an output
• Set the HPO in model_gp.