slurm_train.sh

#!/usr/bin/env bash

set -x
export PYTHONPATH=`pwd`:$PYTHONPATH

PARTITION=$1
JOB_NAME=$2
CONFIG=$3
WORK_DIR=$4
GPUS=${GPUS:-8}
GPUS_PER_NODE=${GPUS_PER_NODE:-8}
SRUN_ARGS=${SRUN_ARGS:-""}
PY_ARGS=${PY_ARGS:-"--validate"}


srun -p ${PARTITION} \
    --job-name=${JOB_NAME} \
    --gres=gpu:${GPUS_PER_NODE} \
    --ntasks=${GPUS} \
    --ntasks-per-node=${GPUS_PER_NODE} \
    --kill-on-bad-exit=1 \
    ${SRUN_ARGS} \
    python -u tools/train.py ${CONFIG} --work_dir=${WORK_DIR} --launcher="slurm" ${PY_ARGS}