The implementation of our VideoMAE supports multi-node distributed training. We provide the off-the-shelf scripts in the scripts folder.
- For example, to pre-train VideoMAE ViT-Base on Something-Something V2 with 64 GPUs (8 nodes x 8 GPUs), you can run
OUTPUT_DIR='YOUR_PATH/ssv2_videomae_pretrain_base_patch16_224_frame_16x2_tube_mask_ratio_0.9_e800'
DATA_PATH='YOUR_PATH/list_ssv2/train.csv'
OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=8 \
--master_port 12320 --nnodes=8 \
--node_rank=0 --master_addr=$ip_node_0 \
run_mae_pretraining.py \
--data_path ${DATA_PATH} \
--mask_type tube \
--mask_ratio 0.9 \
--model pretrain_videomae_base_patch16_224 \
--decoder_depth 4 \
--batch_size 32 \
--num_frames 16 \
--sampling_rate 2 \
--opt adamw \
--opt_betas 0.9 0.95 \
--warmup_epochs 40 \
--save_ckpt_freq 20 \
--epochs 801 \
--log_dir ${OUTPUT_DIR} \
--output_dir ${OUTPUT_DIR}
on the first node. On other nodes, run the same command with --node_rank 1
, ..., --node_rank 7
respectively. --master_addr
is set as the ip of the node 0.
-
For example, to pre-train VideoMAE ViT-Base on Kinetics400 with 64 GPUs (8 nodes x 8 GPUs), you can run
OUTPUT_DIR='YOUR_PATH/k400_videomae_pretrain_base_patch16_224_frame_16x4_tube_mask_ratio_0.9_e800' DATA_PATH='YOUR_PATH/list_kinetics-400/train.csv' OMP_NUM_THREADS=1 python3 -m torch.distributed.launch --nproc_per_node=8 \ --master_port 12320 --nnodes=8 \ --node_rank=0 --master_addr=$your_node_0_ip \ run_mae_pretraining.py \ --data_path ${DATA_PATH} \ --mask_type tube \ --mask_ratio 0.9 \ --model pretrain_videomae_base_patch16_224 \ --decoder_depth 4 \ --batch_size 32 \ --num_frames 16 \ --sampling_rate 4 \ --opt adamw \ --opt_betas 0.9 0.95 \ --warmup_epochs 40 \ --save_ckpt_freq 20 \ --epochs 801 \ --log_dir ${OUTPUT_DIR} \ --output_dir ${OUTPUT_DIR}
on the first node. On other nodes, run the same command with
--node_rank 1
, ...,--node_rank 7
respectively.--master_addr
is set as the ip of the node 0.
- Here the batch size is 32 (
batch_size
per gpu) * 8 (nodes
) * 8 (gpus per node) = 2048. lr
here is the base learning rate and is set to1.5e-4
as default. Theactual lr
is computed by the linear scaling rule:actual lr
=lr
* total batch size / 256.- [Fixed]
We have observed accidental interrupt in the last epoch when conduct the experiment on V100 GPUs (torch 1.6.0). This interrupt is caused by the scheduler of learning rate. We naively set--epochs 801
to walk away from issue :)
To help the community to reproduce our results on slurm cluster, we also provide the the off-the-shelf script.
For example, to pre-train VideoMAE ViT-Base on Kinetics400 with 64 GPUs (8 nodes x 8 GPUs), you can run
export MASTER_PORT=$((12000 + $RANDOM % 20000))
export OMP_NUM_THREADS=1
OUTPUT_DIR='YOUR_PATH/k400_videomae_pretrain_base_patch16_224_frame_16x4_tube_mask_ratio_0.9_e800'
DATA_PATH='YOUR_PATH/list_kinetics-400/train.csv'
JOB_NAME=$1
PARTITION=${PARTITION:-"video"}
# 8 for 1 node, 16 for 2 node, etc.
GPUS=${GPUS:-64}
GPUS_PER_NODE=${GPUS_PER_NODE:-8}
CPUS_PER_TASK=${CPUS_PER_TASK:-8}
SRUN_ARGS=${SRUN_ARGS:-""}
PY_ARGS=${@:2}
# batch_size can be adjusted according to the graphics card
srun -p $PARTITION \
--job-name=${JOB_NAME} \
--gres=gpu:${GPUS_PER_NODE} \
--ntasks=${GPUS} \
--ntasks-per-node=${GPUS_PER_NODE} \
--cpus-per-task=${CPUS_PER_TASK} \
--kill-on-bad-exit=1 \
${SRUN_ARGS} \
python -u run_mae_pretraining.py \
--data_path ${DATA_PATH} \
--mask_type tube \
--mask_ratio 0.9 \
--model pretrain_videomae_base_patch16_224 \
--decoder_depth 4 \
--batch_size 32 \
--num_frames 16 \
--sampling_rate 4 \
--opt adamw \
--opt_betas 0.9 0.95 \
--warmup_epochs 40 \
--save_ckpt_freq 20 \
--epochs 801 \
--log_dir ${OUTPUT_DIR} \
--output_dir ${OUTPUT_DIR} \
${PY_ARGS}