Skip to content

Latest commit

 

History

History
24 lines (18 loc) · 1.01 KB

GETTING_STARTED.md

File metadata and controls

24 lines (18 loc) · 1.01 KB

Get Started

Running

  • To train and test the "Multi-Modal Multi-Action Video Recognition" framework. You can build variant models via setting M3A_MODE, M3A_MODAL_TYPE, M3A_MODAL_JOINT_TYPE.

    python tools/run_net.py \
  --cfg configs/Mmit/R2PLUS1D_8x2.yaml \
  DATA.PATH_TO_DATA_DIR ${DATA_PATH} \
  DATA.MMIT_VERSION ${MMIT_VERSION} \
  TRAIN.CHECKPOINT_FILE_PATH ${PRETRAIN_PATH} \
  M3A.MODE ${M3A_MODE} \
  M3A.MODAL_TYPE ${M3A_MODAL_TYPE} \
  M3A.MODAL_JOINT_TYPE ${M3A_MODAL_JOINT_TYPE}

    You can also set the variables (DATA_PATH, MMIT_VERSION, PRETRAIN_PATH) in scripts/run_mmit_r2plus1d.sh, and then run the script.

    bash scripts/run_mmit_r2plus1d.sh

  • Download the R(2+1)D-34 IG-65M+Kinetics-400 pretrained model: r2plus1d-34-pretrained-model.pth. This pretrained model is converted from facebookresearch/VMZ.