LAVT-RIS/scripts/baseline_repro_refzom.sh

#!/bin/bash
#SBATCH --job-name=lavt-rz-trial
#SBATCH --partition=a6000
#SBATCH --gres=gpu:2
#SBATCH --time=13-11:30:00 # d-hh:mm:ss ??��?, ��??? job?? max time limit ???�� 
#SBATCH --mem=28000 # cpu memory size 
#SBATCH --cpus-per-task=4 # cpu �Ʃ���? 
#SBATCH --output=./logs/lavt_refzom_repro_bs32.log

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

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

export NCCL_P2P_DISABLE=1
export NVIDIA_TF32_OVERRIDE=0

GPUS=2
OUTPUT_DIR=$1
EXP_NAME=$2
PORT=7852

# TRAIN
# hardpos_only, hardpos_only_rev
CUDA_VISIBLE_DEVICES=0,1 torchrun \
    --nproc_per_node=$GPUS --master_port=$PORT train_refzom.py \
    --model lavt_one \
    --dataset ref-zom \
    --splitBy final \
    --split test \
    --output-dir ${OUTPUT_DIR} \
    --model_id ${EXP_NAME} \
    --batch-size 16 \
    --lr 0.00005 \
    --wd 1e-2 \
    --swin_type base \
    --pretrained_swin_weights ./pretrained_weights/swin_base_patch4_window12_384_22k.pth \
    --resume /data2/projects/chaeyun/LAVT-RIS/models/refzom_lavt_bs32_repro/model_best_refzom_lavt_bs32_repro.pth \
    --epochs 40 \
    --img_size 480 

# lavt-rzom
# sbatch ./scripts/baseline_test_lr2.sh ./models/gref_m05_tmp010_4gpu_bs32_orig gref_m05_tmp010_4gpu_bs32_orig 10 0.10 hardpos_only 0.05


# gref_m10_tmp005_4gpu_bs32
# margin 10 tmp 0.05 refined
# sbatch ./scripts/baseline_test_lr2.sh ./models/gref_m10_tmp005_4gpu_bs32 gref_m10_tmp005_4gpu_bs32 10 0.05 hardpos_only_refined 

# sbatch ./scripts/baseline_repro_refzom.sh ./models/refzom_lavt_bs32_repro refzom_lavt_bs32_repro

# sbatch ./scripts/baseline_repro_refzom.sh ./models/refzom_lavt_bs16_repro refzom_lavt_bs16_repro