train_video.sh

#!/usr/bin/env sh

ADAPTER_CHANNEL=$1
GPUS=${GPUS:-8}
PORT=$((12000 + $RANDOM % 20000))
MASTER_ADDR=${MASTER_ADDR:-127.0.0.1}



DATASET=K400
CLUSTER=True \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch \
       --master_addr=$MASTER_ADDR \
       --nproc_per_node=$GPUS \
       --master_port=$PORT \
       --use_env \
       main_video.py \
            --batch_size 16 \
            --cls_token \
            --finetune VIT_BASE_IN21K \
            --dist_eval \
            --output_dir "./video_output/IN21K/0.5/${DATASET}" \
            --drop_path 0.0 \
            --blr 1e-3 \
            --epochs 12 \
            --warmup_epochs 2 \
            --weight_decay 0.01 \
            --dataset "${DATASET}" \
            --ffn_adapt \
            --auto_remove \
            --token_target_ratio 0.5


DATASET=SSV2
CLUSTER=True \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch \
       --master_addr=$MASTER_ADDR \
       --nproc_per_node=$GPUS \
       --master_port=$PORT \
       --use_env \
       main_video.py \
            --batch_size 16 \
            --cls_token \
            --finetune VIT_BASE_IN21K \
            --dist_eval \
            --output_dir "./video_output/IN21K/0.5/${DATASET}" \
            --drop_path 0.0 \
            --blr 1e-3 \
            --epochs 50 \
            --warmup_epochs 2 \
            --eval_freq 5 \
            --weight_decay 0.01 \
            --dataset "${DATASET}" \
            --ffn_adapt \
            --auto_remove \
            --token_target_ratio 0.5