Intro

This is a PyTorch implementation used for video task.

Usage

Install

• Tesla V100 (32G): CUDA 10.1 + PyTorch 1.6.0 + torchvision 0.7.0
• timm 0.4.8
• einops
• easydict

Data Preparation

See DATASET.md.

Training

Start

# video
OMP_NUM_THREADS=1 python3 -m torch.distributed.launch \
    --nproc_per_node=8 --nnodes=8 \
    --node_rank=$1 --master_addr=$2 --master_port=22234 \
    --use_env main_video.py \
    --finetune /path/to/pre_trained/checkpoints \
    --output_dir /path/to/output \
    --batch_size 16 --epochs 90 --blr 0.1 --weight_decay 0.0 --dist_eval \
    --data_path /path/to/SSV2 --data_set SSV2 \
    --ffn_adapt

on each of 8 nodes. --master_addr is set as the ip of the node 0. and --node_rank is 0, 1, ..., 7 for each node.

To obtain the pre-trained checkpoint, see PRETRAIN.md.

Acknowledgement

The project is based on MAE, VideoMAE, timm, and MAM. Thanks for their awesome works.

License

This project is under the MIT license. See LICENSE for details.