Template scripts to run sbatch jobs

jobarray_template.slurm

#!/bin/bash
#SBATCH --job-name=<some_name>
#SBATCH --account=<some_account>    # optional, count job hours from this account
#SBATCH --ntasks=1                  # number of tasks (a single process here)
#SBATCH --gres=gpu:1                # number of GPUs (a single GPU here)
#SBATCH --cpus-per-task=4           # number of cores, mostly for data loader workers
#SBATCH --hint=nomultithread        # optional, restrict to physical cores and not logical ones
#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
#SBATCH --output=logs_%A_%a.out     # output file, %A is the id of the array job and %a is the number of the job in the array
#SBATCH --error=logs_%A_%a.out      # error file (same as output is fine)
#SBATCH --array=0-99                # execute as a job array (end is counted in start-end)

# Read a list of hyper-parameter values (stored as an array called lrs/bs in this case)
IFS=$'\n' read -d '' -r -a lrs < learning_rates.lst
IFS=$'\n' read -d '' -r -a bs < batch_sizes.lst

args=()
for lr in "${lrs[@]}"
do
  for b in "${bs[@]}"
  do
    args+=("--lr ${lr} --batch-size ${b}")
  done
done

# Load conda environment
conda activate myenv

# Echo launched commands from here
set -x

# Launch the job
python train.py ${args[${SLURM_ARRAY_TASK_ID}]}

single_job_template.slurm

#!/bin/bash
#SBATCH --job-name=<some_name>
#SBATCH --account=<some_account>    # optional, count job hours from this account
#SBATCH --ntasks=1                  # number of tasks (a single process here)
#SBATCH --gres=gpu:1                # number of GPUs (a single GPU here)
#SBATCH --cpus-per-task=4           # number of cores, mostly for data loader workers
#SBATCH --hint=nomultithread        # optional, restrict to physical cores and not logical ones
#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
#SBATCH --output=logs.out           # output file
#SBATCH --error=logs.out            # error file (same as output is fine)

# Load conda environment
conda activate myenv

# Echo launched commands from here
set -x

# Launch the job
python train.py --lr 0.01 --batch-size 32