RIS-DMMI/train_repro.sh

#!/bin/bash
#SBATCH --job-name=dmmibs12-gg
#SBATCH --partition=a6000
#SBATCH --gres=gpu:1
#SBATCH --time=12-00:00:00  # d-hh:mm:ss, job time limit
#SBATCH --mem=32000 # cpu memory size 
#SBATCH --cpus-per-task=6 
#SBATCH --output=./trainlog/dmmi_gref_google_bs12_repro.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

# dmmi_swin_hardpos_only

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


# TRAIN
CUDA_VISIBLE_DEVICES=0 torchrun \
    --nproc_per_node=$GPUS --master_port=2947 train.py \
    --model dmmi_swin \
    --dataset refcocog \
    --split val \
    --splitBy google \
    --output_dir ${OUTPUT_DIR} \
    --model_id ${EXP_NAME} \
    --batch-size 12 \
    --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 \
    --resume /data2/projects/chaeyun/RIS-DMMI/experiments/dmmi_gref_google_bs12_repro/model_best_dmmi_gref_google_bs12_repro.pth \
    --epochs 40 \
    --img_size 480


# sbatch train_repro.sh ./experiments/dmmi_refzom_bs12_repro dmmi_refzom_bs12_repro

# sbatch train_repro.sh ./experiments/dmmi_gref_google_bs12_repro dmmi_gref_google_bs12_repro

# /data2/projects/chaeyun/RIS-DMMI/experiments/dmmi_gref_google_bs12_repro/model_best_dmmi_gref_google_bs12_repro.pth
# /data2/projects/chaeyun/RIS-DMMI/experiments/dmmi_refzom_bs12_repro/model_best_dmmi_refzom_bs12_repro.pth