LAVT-RIS/scripts/baseline_test_lr_angle2.sh

#!/bin/bash
#SBATCH --job-name=angle2
#SBATCH --partition=a6000
#SBATCH --gres=gpu:4
#SBATCH --time=13-11:30:00 # d-hh:mm:ss ??��??, ��???? job?? max time limit ????�� 
#SBATCH --mem=60000 # cpu memory size 
#SBATCH --cpus-per-task=8 # cpu ����?����?? 
#SBATCH --output=./logs/gref_m10_mg10_tmp007_4gpu_bs32_ang.log

ml purge
ml load cuda/11.3
eval "$(conda shell.bash hook)"
conda activate cris 

cd /data2/projects/chaeyun/LAVT-RIS/

# gref_m10_mg10_tmp007_4gpu_bs32_ang
# # sbatch ./scripts/baseline_test_lr_angle2.sh ./models/gref_m10_mg10_tmp007_4gpu_bs32_ang gref_m10_mg10_tmp007_4gpu_bs32_ang 10 0.07 hardpos_only 0.10 

# todo after 241208
# gref_m10_mg10_tmp010_4gpu_bs32_ang
# mlw 0.10 margin 8 tmp 0.10 original
# # bash ./scripts/baseline_test_lr_angle.sh ./models/gref_m10_mg10_tmp010_4gpu_bs32_ang gref_m10_mg10_tmp010_4gpu_bs32_ang 10 0.10 hardpos_only 0.10 > ./logs/gref_m10_mg10_tmp010_4gpu_bs32_ang.log 2>&1


# # bash ./scripts/baseline_test_lr.sh ./models/gref_m10_mg08_tmp010_4gpu_bs32_ang gref_m10_mg08_tmp010_4gpu_bs32_ang 8 0.10 hardpos_only 0.10 > ./logs/gref_m10_mg08_tmp010_4gpu_bs32_ang.log 2>&1


# gref_m10_mg12_tmp007_4gpu_bs32_ang
# gref_m10_mg12_tmp010_4gpu_bs32_ang
# gref_m10_mg10_tmp007_4gpu_bs32_ang
# gref_m10_mg10_tmp010_4gpu_bs32_ang
# gref_m10_mg15_tmp010_4gpu_bs32_ang



export NCCL_P2P_DISABLE=1
export NVIDIA_TF32_OVERRIDE=0

GPUS=4
OUTPUT_DIR=$1
EXP_NAME=$2
MARGIN=$3
TEMP=$4
MODE=$5
MLW=$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 8 \
--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} \
--metric_mode ${MODE} \
--metric_loss_weight ${MLW} \
--exclude_multiobj "

python -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=6872 train_angle.py $python_args

# sbatch ./scripts/baseline_test_lr.sh ./models/gref_hp10_m10_tmp003 gref_hp10_m10_tmp003 10 0.03 hardpos_only_refined
# sbatch ./scripts/baseline_test_lr.sh ./models/gref_hp10_m10_tmp010 gref_hp10_m10_tmp010 10 0.10 hardpos_only_refined
# sbatch ./scripts/baseline_test_lr.sh ./models/gref_hp10_m15_tmp005 gref_hp10_m15_tmp005 15 0.05 hardpos_only_refined
# sbatch ./scripts/baseline_test_lr.sh ./models/gref_hp10_m20_tmp005 gref_hp10_m20_tmp005 20 0.05 hardpos_only_refined


# 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