ASDA/scripts/train_rccp_sbert.sh

#!/bin/bash
#SBATCH --job-name=asda-rccp-pj5
#SBATCH --partition=a6000
#SBATCH --gres=gpu:1
#SBATCH --time=13-11:30:00
#SBATCH --mem=28000
#SBATCH --cpus-per-task=3
#SBATCH --output=./exp_rccp_projection/pj_rccp_m12_tmp007_refcoarse_nofltroiou_bs32.log

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

cd /data2/projects/chaeyun/ASDA

export NCCL_P2P_DISABLE=1
export NVIDIA_TF32_OVERRIDE=1
export NCCL_IB_TIMEOUT=100
export NCCL_IB_RETRY_CNT=15
export MASTER_PORT=8128

BS=32
SAVENAME=pj_rccp_m12_tmp007_refcoarse_nofltroiou_bs32
MARGIN=12
TEMP=0.07
MODE=hardpos_only_refined
FILTER_THRES=0.99
FUSE_MODE=refined_coarse

# Running options
# pj_rccp_m10_tmp007_coarse_nofltroiou_bs28

# pending 
# pj_rccp_m10_tmp007_coarse_fthr070_oiou_bs28


# TRAIN
export CUDA_VISIBLE_DEVICES=0
python_args="--dataset refcoco+ \
--splitBy unc \
--ngpu 1 --batch_size ${BS} \
--savename ${SAVENAME} --time 17 \
--metric_learning --use_projections \
--margin_value ${MARGIN} \
--filter_thres ${FILTER_THRES} \
--temperature ${TEMP} \
--metric_mode ${MODE} \
--fuse_mode ${FUSE_MODE} "

python train_rcc_sbert_oiou.py $python_args