ASDA/scripts/train_gref_sbert.sh

#!/bin/bash
#SBATCH --job-name=asda-fltr
#SBATCH --partition=a6000
#SBATCH --gres=gpu:1
#SBATCH --time=13-11:30:00
#SBATCH --mem=25000
#SBATCH --cpus-per-task=3
#SBATCH --output=./exp/gref_m10_tmp007_fine_nofltr_bs28.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=2721


BS=28
SAVENAME=gref_m10_tmp007_fine_nofltr_bs28
MARGIN=10
TEMP=0.07
MODE=hardpos_only_refined
FILTER_THRES=0.99
FUSE_MODE=fine

## options 
# gref_m10_tmp007_coarse_nofltr_bs28
# gref_m10_tmp007_coarse_fthr065_bs28
# gref_m10_tmp007_coarse_fthr050_bs28
# gref_m10_tmp007_fine_fthr065_bs28
# gref_m10_tmp007_fine_nofltr_bs28

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

python train_gref_sbert.py $python_args

# python train.py --dataset refcoco --splitBy unc --ngpu 1 --batch_size 36 --time 17 --savename refcoco_bs36_repro 
# python train.py --dataset refcocog --splitBy umd --ngpu 1 --batch_size 28 --time 17 --savename gref_umd_bs28_repro
# python train.py --dataset refcocog --splitBy umd --ngpu 1 --batch_size 36 --time 17 --savename gref_umd_bs36_repro
# export CUDA_VISIBLE_DEVICES=0,1; python train.py --dataset refcocog --splitBy umd --ngpu 2 --batch_size 64 --time 17 --savename gref_umd_bs64_repro
# export CUDA_VISIBLE_DEVICES=0,1; python train.py --dataset refcocog --splitBy umd --ngpu 1 --batch_size 64 --time 17 --savename gref_umd_bs64_repro