CGFormer/scripts/train_rcc_bash.sh

#!/bin/bash

GPUS=4
MASTER_PORT=2948

if [ "$#" -ne 2 ]; then
    echo "Usage: bash train.sh <OUTPUT_DIR> <EXP_NAME>"
    exit 1
fi

OUTPUT_DIR=$1
EXP_NAME=$2

LOG_DIR="./refcoco_filter_exp"
LOG_FILE="${LOG_DIR}/${EXP_NAME}.log"

mkdir -p "${LOG_DIR}"

MARGIN=12
TEMP=0.07
MODE=hardpos_only_sbertsim_refined
MLW=0.1
BATCH_SIZE=48
MIXUP_FQ=False
USE_PROJECTIONS=False

# 

echo "Starting distributed training with ${GPUS} GPUs on port ${MASTER_PORT}..."
echo "Experiment Name: ${EXP_NAME}, Output Dir: ${OUTPUT_DIR}"
echo "Logging to: ${LOG_FILE}"

ml purge
ml load cuda/11.8
eval "$(conda shell.bash hook)"
conda activate risall

cd /data2/projects/chaeyun/CGFormer/

export NCCL_P2P_DISABLE=1
export NVIDIA_TF32_OVERRIDE=1
export NCCL_DEBUG=INFO
export NCCL_TIMEOUT=7200
export NCCL_IB_RETRY_CNT=15

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch \
    --nproc_per_node=${GPUS} \
    --master_port=${MASTER_PORT} \
    train_rcc_sbert.py \
    --config config/config_rcc_ace.yaml \
    --opts TRAIN.batch_size ${BATCH_SIZE} \
           TRAIN.exp_name ${EXP_NAME} \
           TRAIN.output_folder ${OUTPUT_DIR} \
           TRAIN.metric_mode ${MODE} \
           TRAIN.metric_loss_weight ${MLW} \
           TRAIN.margin_value ${MARGIN} \
           TRAIN.temperature ${TEMP} \
           TRAIN.filter_threshold 0.52 \
           TRAIN.mixup_lasttwo ${MIXUP_FQ} \
           TRAIN.use_projections ${USE_PROJECTIONS} \
    > "${LOG_FILE}" 2>&1