Wormulon

This package exists to support machine learning projects that run on clusters. It makes interacting with Slurm less of a nightmare through the cunning use of submitit and wandb.

Installation

git clone [email protected]:mweiss17/wormulon.git cd wormulon pip install -e .

Usage

Use Wormulon after you've developed your models and made a vague pass at debugging your scripts -- once you are ready to start burning up the cluster. Instead of writing a crappy bash script, or dropping in some home-rolled project-specific scripts, just run: submit scripts/<your-train-script>.py experiments/<exp-1> templates/<template-1>

If you want to do a hyper-parameter search, then run:

salvo -dry --- scripts/<your-train-script>.py ~/scratch/experiments/exp1-{job_idx} --inherit templates/exp1 --config.use_wandb True --config._wandb_group group-1 --config._wandb_run_name exp1-{job_idx} --config.seed {job_idx} --- generators/random_search.py -d lr~log_uniform[0.01,0.001] -n 10