RIS-DMMI/train_ace_bash_tlr.sh

#!/bin/bash
#SBATCH --job-name=dbs12-tlr
#SBATCH --partition=a6000
#SBATCH --gres=gpu:2
#SBATCH --time=12-00:00:00  # d-hh:mm:ss, job time limit
#SBATCH --mem=60000 # cpu memory size 
#SBATCH --cpus-per-task=6

# Job configuration
LOG_NAME=$6  
LOG_FILE="./trainlog/${LOG_NAME}.log"
GPUS=2
OUTPUT_DIR=$1
EXP_NAME=$2
MARGIN=$3
TEMP=$4
MODE=$5
MASTER_PORT=$7

# Environment setup
module purge
module load cuda/11.8
eval "$(conda shell.bash hook)"
conda activate risall

cd /data2/projects/chaeyun/RIS-DMMI

export NCCL_P2P_DISABLE=1
export NVIDIA_TF32_OVERRIDE=0

# Run the training script
torchrun \
    --nproc_per_node=$GPUS \
    --master_port=$MASTER_PORT \
    train_rev_textlr.py \
    --model dmmi_swin_hardpos_only \
    --dataset refcocog \
    --splitBy umd \
    --output_dir ${OUTPUT_DIR} \
    --model_id ${EXP_NAME} \
    --batch-size 6 \
    --lr 0.00005 \
    --wd 1e-2 \
    --window12 \
    --swin_type base \
    --pretrained_backbone /data2/projects/chaeyun/LAVT-RIS/pretrained_weights/swin_base_patch4_window12_384_22k.pth \
    --epochs 40 \
    --img_size 480 \
    --metric_learning \
    --margin_value ${MARGIN} \
    --temperature ${TEMP} \
    --metric_mode ${MODE} \
    --exclude_multiobj \
    2>&1 | tee $LOG_FILE

# sbatch train_ace_bash_tlr.sh ./experiments/dmmi_grefu_ace_/gref_m10_tmp007_lrh_bs12 gref_m10_tmp007_lrh_bs12 10 0.05 hardpos_only dmmi_ACE_gref_m10_tmp007_lrh_bs12 8236