57 lines (53 with data), 2.2 kB

#!/usr/bin/env bash

ARGPARSE_DESCRIPTION="Train recognizer on slurm cluster"
source $(dirname $0)/argparse.bash || exit 1
argparse "$@" <<EOF || exit 1
parser.add_argument('-p', '--partition', default='${PARTITION:-""}', type=str,
                    help='partition (default: env PARTITION)')
parser.add_argument('--gpus', default=${GPUS:-8}, type=int,
                    help='number of gpus to use (default: env GPUS, or 8)')
parser.add_argument('--gpus_per_node', default=${GPUS_PER_NODE:-8}, type=int,
                    help='gpus per node (default: env GPUS_PER_NODE, or 8)')
parser.add_argument('--cpus_per_task', default=${CPUS_PER_TASK:-5}, type=int,
                    help='cpus per task (default: env CPUS_PER_TASK, or 5)')
parser.add_argument('-j', '--job_name', default='', type=str,
                    help='name of job (default: config filename)')
parser.add_argument('--srun_args', default='${SRUN_ARGS:-""}', type=str,
                    help='extra args for srun (default: env SRUN_ARGS)')
parser.add_argument('config', type=str,
                    help='config file')
parser.add_argument('py_args', nargs=argparse.REMAINDER,
                    help='extra args to tools/train.py (can be empty)')
EOF

BARECONFIG="${CONFIG##*/}"
JOB_NAME="${JOB_NAME:-${BARECONFIG%.*}}"

echo partition: "${PARTITION}"
echo jobname: "${JOB_NAME}"
echo config: "${CONFIG}"
echo gpus: "${GPUS}"
echo gpus per node: "${GPUS_PER_NODE}"
echo cpus per task: "${CPUS_PER_TASK}"
echo srun args: "${SRUN_ARGS[@]}"
echo py args: "${PY_ARGS[@]}"
echo command to run:
echo ----
echo srun -p "${PARTITION}" \
    --job-name="${JOB_NAME}" \
    --gres=gpu:${GPUS_PER_NODE} \
    --ntasks=${GPUS} \
    --ntasks-per-node=${GPUS_PER_NODE} \
    --cpus-per-task=${CPUS_PER_TASK} \
    --kill-on-bad-exit=1 \
    ${SRUN_ARGS} \
    python -u tools/train.py ${CONFIG} --launcher="slurm" ${PY_ARGS[@]}
echo ----
echo
srun -p "${PARTITION}" \
    --job-name="${JOB_NAME}" \
    --gres=gpu:${GPUS_PER_NODE} \
    --ntasks=${GPUS} \
    --ntasks-per-node=${GPUS_PER_NODE} \
    --cpus-per-task=${CPUS_PER_TASK} \
    --kill-on-bad-exit=1 \
    ${SRUN_ARGS} \
    python -u tools/train.py ${CONFIG} --launcher="slurm" ${PY_ARGS[@]}