RIS-DMMI/train_ace_bs4.sh

#!/bin/bash
#SBATCH --job-name=dbs6-ace3
#SBATCH --partition=a6000
#SBATCH --gres=gpu:1
#SBATCH --time=12-00:00:00  # d-hh:mm:ss, job time limit
#SBATCH --mem=28000 # cpu memory size 
#SBATCH --cpus-per-task=4
#SBATCH --output=./trainlog/dmmi_ACE_gref_m10_tmp005_bs6.log

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

conda activate risall

cd /data2/projects/chaeyun/RIS-DMMI

export NCCL_P2P_DISABLE=1
export NVIDIA_TF32_OVERRIDE=0

GPUS=1
OUTPUT_DIR=$1
EXP_NAME=$2
MARGIN=$3
TEMP=$4
MODE=$5
MASTER_PORT=5728


# TRAIN
# hardpos_only, hardpos_only_rev
python_args="--model dmmi_swin_hardpos_only \
--dataset refcocog \
--splitBy umd \
--output_dir ${OUTPUT_DIR} \
--model_id ${EXP_NAME} \
--batch-size 6 \
--lr 0.00005 \
--wd 1e-2 \
--window12 \
--swin_type base \
--pretrained_backbone /data2/projects/chaeyun/LAVT-RIS/pretrained_weights/swin_base_patch4_window12_384_22k.pth \
--epochs 40 \
--img_size 480 \
--metric_learning \
--margin_value ${MARGIN} \
--temperature ${TEMP} \
--metric_mode ${MODE} \
--exclude_multiobj "

CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node=$GPUS --master_port=$MASTER_PORT train_rev.py $python_args

# python -m torch.distributed.launch --nproc_per_node=$GPUS train_rev.py $python_args
# CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=$GPUS train_rev.py $python_args


# sbatch train_ace_bs4.sh ./experiments/dmmi_grefu_ace/gref_m10_tmp007_bs6 gref_m10_tmp007_bs6 10 0.07 hardpos_only 

# sbatch train_ace_bs4.sh ./experiments/dmmi_grefu_ace/gref_m12_tmp007_bs6 gref_m12_tmp007_bs6 12 0.07 hardpos_only 

# sbatch train_ace_bs4.sh ./experiments/dmmi_grefu_ace/gref_m10_tmp005_bs6 gref_m10_tmp005_bs6 10 0.05 hardpos_only