| #SBATCH --job-name=lavt_mlw005-res | |
| #SBATCH --nodes=1 | |
| #SBATCH --gres=gpu:4 | |
| #SBATCH --time=0-12:00:00 # d-hh:mm:ss, job time limit | |
| #SBATCH --mem=75000 # cpu memory size | |
| #SBATCH --cpus-per-task=6 | |
| #SBATCH --output=./logs/gref_hp05_m10_tmp005_resume.log | |
| source ${HOME}/.bashrc | |
| source ${HOME}/miniconda3/bin/activate base | |
| conda activate cris | |
| cd /home/s1/chaeyunkim/LAVT-RIS | |
| # gref_hp05_m10_tmp005_resume | |
| # --metric_loss_weight 0.05 \ | |
| # sbatch ./scripts/baseline_test_lr_resume_mlw005.sh ./models/gref_hp05_m10_tmp005 gref_hp05_m10_tmp005 10 0.05 hardpos_only_refined /home/s1/chaeyunkim/LAVT-RIS/models/gref_hp05_m10_tmp005/model_best_gref_hp05_m10_tmp005.pth | |
| #--metric_loss_weight 0.05 \ | |
| # gref_hp05_m10_tmp005_orig_resume | |
| # sbatch ./scripts/baseline_test_lr_resume_mlw005.sh ./models/gref_hp05_m10_tmp005_orig gref_hp05_m10_tmp005_orig 10 0.05 hardpos_only /home/s1/chaeyunkim/LAVT-RIS/models/gref_hp05_m10_tmp005_orig/model_best_gref_hp05_m10_tmp005_orig.pth | |
| export NCCL_P2P_DISABLE=1 | |
| export NVIDIA_TF32_OVERRIDE=0 | |
| # # core args | |
| # BATCH_SIZE=64 | |
| GPUS=4 | |
| OUTPUT_DIR=$1 | |
| EXP_NAME=$2 | |
| MARGIN=$3 | |
| TEMP=$4 | |
| MODE=$5 | |
| RESUME=$6 | |
| # # Create the directory if it does not exist | |
| # if [[ ! -d "${OPT_DIR}/${EXP_NAME}" ]]; then | |
| # echo "Directory ${OPT_DIR}/${EXP_NAME} does not exist. Creating it..." | |
| # mkdir -p "${OPT_DIR}/${EXP_NAME}" | |
| # fi | |
| # TRAIN | |
| # hardpos_only, hardpos_only_rev | |
| python_args="--model lavt_one \ | |
| --dataset refcocog \ | |
| --splitBy umd \ | |
| --output-dir ${OUTPUT_DIR} \ | |
| --model_id ${EXP_NAME} \ | |
| --batch-size 4 \ | |
| --lr 0.00005 \ | |
| --wd 1e-2 \ | |
| --swin_type base \ | |
| --pretrained_swin_weights ./pretrained_weights/swin_base_patch4_window12_384_22k.pth \ | |
| --epochs 40 \ | |
| --img_size 480 \ | |
| --metric_learning \ | |
| --margin_value ${MARGIN} \ | |
| --temperature ${TEMP} \ | |
| --resume ${RESUME} \ | |
| --metric_mode ${MODE} \ | |
| --metric_loss_weight 0.05 \ | |
| --exclude_multiobj " | |
| python -m torch.distributed.launch --nproc_per_node=$GPUS train.py $python_args | |
| # python test.py --model lavt_one --swin_type base --dataset refcocog --splitBy umd --split test --resume ./models/gref_umd/lavt_test_lr/gref_testlr_4gpu.pth --workers 4 --ddp_trained_weights --window12 --img_size 480 | |