run_vrg.sh

#!/usr/bin/env bash

NODE_ID=$SLURM_PROCID
TOTAL_NODES=1
MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)

GPUS=$1
PORT=${PORT:-29500}

echo 'The TOTAL_NODES is: ' ${TOTAL_NODES:-1}

echo 'The node_rank is: ' ${NODE_ID:-0} 

echo 'The MASTER_ADDR is: ' $MASTER_ADDR

echo 'The GPUS is: ' $GPUS

echo 'The port is: ' $PORT

CURDIR=$(cd $(dirname $0); pwd)
cd $CURDIR
export PYTHONPATH="$PYTHONPATH:$CURDIR"
echo 'The current dir is: ' $CURDIR



WANDB_MODE=offline torchrun \
    --nnodes=${TOTAL_NODES:-1} \
    --node_rank=${NODE_ID:-0} \
    --master_addr=$MASTER_ADDR \
    --nproc_per_node=$GPUS \
    --master_port=$PORT \
    $CURDIR/train.py \
    --data-dir /sfs/weka/scratch/pfq7pm/ai_agent/zhang/imagenet100 \
    --model vit_base_patch16_224 \
    --batch-size 128  \
    --val-split val \
    --decay-epochs 0 \
    --decay-milestones 0 \
    --decay-rate 0 \
    --epochs 50 \
    --lr 3e-4 \
    --opt adamw \
    --num-classes 100 \
    --img-size 224 \
    --log-wandb True \
    --warmup-epochs 0