MilesCranmer / easy_distributed_hyperopt

Do distributed hyperparameter optimization using only a shared folder between processes.

Easy distributed hyperopt

The code in allows you to run distributed hyperopt with only a shared folder: all trials will be stored and shared between nodes using that folder.

No head node, no distributed frameworks. Just a shared folder.

To use,

1. Create an empty folder to store all the trial scores. The folder should be accessible from all your nodes.
2. Fill in example.py with the foldername, your desired optimization function and hyperparameters to optimize over.
3. Then, run example.py on all your nodes.
4. Kill the script when you want to stop.
5. Set the foldername in print_best_model.py and then run it.