wds_example

This project hosts a sample code that demonstrates how to use WebDataset on video data. The code includes the implementation of the Facebook SlowFast network and the dataloader of Kinetics-Sounds.

Environment setup

You can create a conda environment by

conda create -y -n wds python=3.7.6

and activate it with

conda activate wds

Install the required packages as follows:

# Install via conda-forge for ffmpeg
conda install -y -c conda-forge av
bash install.sh

Kinetics-Sounds

• Create a directory named datasets.
• Download and decompress Kinetics-Sounds (download link) into datasets/KineticsSounds

Runing Experiments

Details of experiment configuration are in config.py. To run an experiment

python run.py --cfg_file SLOWFAST_8x8_R50.yaml

You can manually evaluate a checkpoint

python run.py --cfg_file SLOWFAST_8x8_R50.yaml --test TRAIN.ENABLE False TEST.ENABLE True TEST.CHECKPOINT_FILE_PATH [path-to-checkpoint]

Note

• As of September 4, 2020, using MultiDataset and ResizedDataset simultaneously gives rise to an AttributeError (issue). Thus, please DO NOT USE multiprocessing for data loading (that is, DATA_LOADER.NUM_WORKERS = 0).
• The above issue is solved: when using multiprocessing for data loading, the number of workers cannot exceed the number of shard .tar files.