baldassarreFe / rpl-workshop

Workshop on GPU and slurm usage at RPL (KTH)

GPU and SLURM Workshop

KTH - Royal Institute of Technology

Initial setup

Clone repository, create conda environment, and install package in editable mode:

cd ~
git clone https://github.com/baldassarreFe/rpl-workshop
cd rpl-workshop

conda env create -n workshop --file conda.yaml
conda activate workshop
pip install --editable .

Plain training

python -m workshop.train \
    --runpath "path/to/runs/folder" \
    --datapath "path/to/data/folder" \
    --batch_size 64 \
    --learning_rate .001 \
    --weight_decay .00001 \
    --number_epochs 3 \
    --number_workers 2 \
    --device 'cuda'

Slurm commands

Single job:

sbatch slurm/single_job.sbatch

Grid search using environment variables:

./slurm/grid_search.sh

Grid search using job arrays:

./slurm/grid_search.sh