Skip to content

Latest commit

 

History

History
40 lines (36 loc) · 1.7 KB

README.md

File metadata and controls

40 lines (36 loc) · 1.7 KB
Raw

A Multigrid Method for Efficiently Training Video Models

Chao-Yuan Wu, Ross Girshick, Kaiming He, Christoph Feichtenhofer, Philipp Krähenbühl
In CVPR, 2020. [Paper]


Getting started

To enable multigrid training, add MULTIGRID.LONG_CYCLE True and/or MULTIGRID.SHORT_CYCLE True when training your model. (Default multigrid training uses both long and short cycles; See paper for details.) For example,

python tools/run_net.py \
  --cfg configs/Charades/SLOWFAST_16x8_R50.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
  MULTIGRID.LONG_CYCLE True \
  MULTIGRID.SHORT_CYCLE True \

This should train multiple times faster than training without multigrid training. Note that multigrid training might induce higher IO overhead. Systems with faster IO (e.g., with efficient local disk) might enjoy more speedup. Please see MODEL_ZOO.md for more examples of multigrid training.

Citing Multigrid Training

If you use multigrid training or the models from MODEL_ZOO in your research, please use the following BibTeX entry.

@inproceedings{multigrid2020,
  Author    = {Chao-Yuan Wu and Ross Girshick and Kaiming He and Christoph Feichtenhofer
               and Philipp Kr\"{a}henb\"{u}hl},
  Title     = {{A Multigrid Method for Efficiently Training Video Models}},
  Booktitle = {{CVPR}},
  Year      = {2020}}