pretrain

Pre-training VideoMAEv2

VideoMAEv2 has additional decoder masking compared to VideoMAE. Our codebase supports multi-node slurm training and multi-node distributed training. We provide the off-the-shelf slurm training scripts in the pre-train scripts folder. Below we give an example of the pre-training script.

Slurm Train

To pre-train VideoMAEv2 ViT-giant on UnlabeledHybrid with 64 A100-80G (8 nodes x 8 GPUs), you can use the following script file script/pretrain/vit_g_hybrid_pt.sh.

#!/usr/bin/env bash
set -x  # print the commands

export MASTER_PORT=$((12000 + $RANDOM % 20000))  # Randomly set master_port to avoid port conflicts
export OMP_NUM_THREADS=1  # Control the number of threads

OUTPUT_DIR='YOUR_PATH/work_dir/vit_g_hybrid_pt_1200e'  # Your output folder for deepspeed config file, logs and checkpoints
DATA_PATH='YOUR_PATH/data/hybrid_train.csv'  # The data list file path.
# pretrain data list file follows the following format
# for the video data line: video_path, 0, -1
# for the rawframe data line: frame_folder_path, start_index, total_frames

JOB_NAME=$1  # the job name of the slurm task
PARTITION=${PARTITION:-"video"}  # Name of the partition
# 8 for 1 node, 16 for 2 node, etc.
GPUS=${GPUS:-64}  # Number of GPUs
GPUS_PER_NODE=${GPUS_PER_NODE:-8}  # Number of GPUs in each node
CPUS_PER_TASK=${CPUS_PER_TASK:-14}  # Number of CPU cores allocated, number of tasks equal to the number of GPUs used
SRUN_ARGS=${SRUN_ARGS:-""}  # Other slurm task args
PY_ARGS=${@:2}  # Other training args

# Please refer to `run_mae_pretraining.py` for the meaning of the following hyperreferences
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 \
        --async \
        ${SRUN_ARGS} \
        python -u run_mae_pretraining.py \
        --data_path ${DATA_PATH} \
        --mask_type tube \
        --mask_ratio 0.9 \
        --decoder_mask_type run_cell \
        --decoder_mask_ratio 0.5 \
        --model pretrain_videomae_giant_patch14_224 \
        --decoder_depth 4 \
        --batch_size 32 \
        --with_checkpoint \
        --num_frames 16 \
        --sampling_rate 4 \
        --num_sample 4 \
        --num_workers 10 \
        --opt adamw \
        --lr 6e-4 \
        --clip_grad 0.02 \
        --opt_betas 0.9 0.95 \
        --warmup_epochs 30 \
        --save_ckpt_freq 5 \
        --epochs 300 \
        --log_dir ${OUTPUT_DIR} \
        --output_dir ${OUTPUT_DIR} \
        ${PY_ARGS}

Start training by running

bash script/pretrain/vit_g_hybrid_pt.sh hybrid_pretrain

, where 'hybrid_pretrain' is the job name.

Dist Train

The above slurm training script can be modified to distributed training script as follows:

#!/usr/bin/env bash
set -x  # print the commands

export MASTER_PORT=${MASTER_PORT:-12320}  # You should set the same master_port in all the nodes

OUTPUT_DIR='YOUR_PATH/work_dir/vit_g_hybrid_pt_1200e'  # Your output folder for deepspeed config file, logs and checkpoints
DATA_PATH='YOUR_PATH/data/hybrid_train.csv'  # The data list file path.
# pretrain data list file follows the following format
# for the video data line: video_path, 0, -1, 0
# for the rawframe data line: frame_folder_path, start_index, total_frames, 0

N_NODES=${N_NODES:-8}  # Number of nodes
GPUS_PER_NODE=${GPUS_PER_NODE:-8}  # Number of GPUs in each node
SRUN_ARGS=${SRUN_ARGS:-""}  # Other slurm task args
PY_ARGS=${@:3}  # Other training args

# Please refer to `run_mae_pretraining.py` for the meaning of the following hyperreferences
OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=${GPUS_PER_NODE} \
        --master_port ${MASTER_PORT} --nnodes=${N_NODES} --node_rank=$1 --master_addr=$2 \
        run_mae_pretraining.py \
        --data_path ${DATA_PATH} \
        --mask_type tube \
        --mask_ratio 0.9 \
        --decoder_mask_type run_cell \
        --decoder_mask_ratio 0.5 \
        --model pretrain_videomae_giant_patch14_224 \
        --decoder_depth 4 \
        --batch_size 32 \
        --with_checkpoint \
        --num_frames 16 \
        --sampling_rate 4 \
        --num_sample 4 \
        --num_workers 10 \
        --opt adamw \
        --lr 6e-4 \
        --clip_grad 0.02 \
        --opt_betas 0.9 0.95 \
        --warmup_epochs 30 \
        --save_ckpt_freq 5 \
        --epochs 300 \
        --log_dir ${OUTPUT_DIR} \
        --output_dir ${OUTPUT_DIR} \
        ${PY_ARGS}

Start training by run

NODE_RANK=0  # 0 for the first node 0, 1 for the second node, and so on.
# MASTER_ADDR should be set as the ip of current node

bash dist_train_vit_g_hybrid_pt.sh $NODE_RANK $MASTER_ADDR

at each node.