Deep-learning example for OptunaPruningSweeper

This example is adapted from an official optuna example. It shows how more complicated experiments can be configured with hydra or hydra-zen. Pytorch-Lightning is used to avoid boilerplate for training the neural networks. This example also shows how the OptunaPruningSweeper can be used.

To show a more complicated configuration a ResNet model is implemented adapted from torchvision. However, pre-activation ResNet blocks are used. The FashionMNIST dataset is used.

The hydra specific code/configuration is located in run_hydra.py and config_hydra. The code/configuration specific to hydra_zen can be found in run_hydra_zen.py and config_hydra_zen.

Installation

It is recommended to first create a virtual environment. In this environment install the dependencies with

pip install -r requirements.txt

To run the hyperparameter optimization with pruning install the OptunaPruningSweeper. This plugin has not yeet been added to PyPI. Install it by cloning this repository and execute in your virtual environment

pip install PATH-TO-CLONED-REPOSITORY-OF-HYDRA-OPTUNA-PRUNING-SWEEPER

If you are interested in a template for pytorch-lightning + hydra also take a look at this repository.