captain-pool/hydra-example

Using Hydra with Weights and Biases

Hydra is a widely used configuaration and experiment management tool developed by Facebook Research. In this repository, we make a minimal implementation of integrating Hydra with Weights and Biases

In this repository, we demonstrate how hydra can be used to configure a ML project. Here we reimplement the notebook of the Group Normalization Report and demonstrate how Hydra can be used to configure the codebase.

Usage

NOTE: This repo relies on a submodule normalizations To properly use the repository clone the repository recursively.

$ git clone --recursive https://github.com/captain-pool/hydra-example

To pull the submodule after cloning:

$ git submodule update --init

Runing Experiments

For single run (with default architecture):

$ python3 main.py dataset=mnist

For running Hydra-Multirun over different dataset(dataset/) and architectures (experiments/)

$ python3 main.py dataset=cifar10,mnist experiments=group,batch,instance

NOTE: Running the codebase creates a project called: hydra-example on your W&B account by default. You can configure where to send the runs by editing the wandb/defaults.yaml configuration file.