LAVT-RIS/scripts/baseline_refzom_angle.sh

#!/bin/bash
#SBATCH --job-name=lvrz1
#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 ?�硧��??������?? 
#SBATCH --output=./logs/rzom_m10_mg12_tmp007_4gpu_bs32_ang.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=4
OUTPUT_DIR=$1
EXP_NAME=$2
MARGIN=$3
TEMP=$4
MODE=$5
MLW=$6
PORT=7852


# TRAIN
# addzero is set to none for default. we don't include zero target cases in MRaCL
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun \
    --nproc_per_node=$GPUS --master_port=$PORT train_refzom_angle.py \
    --model lavt_one \
    --dataset ref-zom \
    --splitBy final \
    --split test \
    --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


# rzom_m10_mg12_tmp007_4gpu_bs32_ang original
# # sbatch ./scripts/baseline_refzom_angle.sh ./models/rzom_m10_mg12_tmp007_4gpu_bs32_ang rzom_m10_mg12_tmp007_4gpu_bs32_ang 12 0.07 hardpos_only 0.10