Training with Webdataset under distributed setting

This repo contains code to handle large dataset and perform efficient IO operations to speed up the training. This code is written for BP4D and BP4D+ dataset but can be extended to any dataset.

• ./trainer/save_pickles.py Contains code to save all the filepaths per subject and task. Can be grouped however you want.
• ./trainer/extract_chips.py Contains code to perform face detection for each group and save it into a .tar format. It also saves the meta information such as filename, class, subject_name and task.
• ./trainer/extract_openface.py Extracts OpenFace features by loading the tar files. Uses my external library TorchOpenFace. See repo README and wiki for build details.
• ./trainer/train_wds Trains a model using webdataset library for IO loading under distributed setting. Uses SLURM based training and Torch DDP.