Project Overview
WHAR Datasets is an open-source library for Wearable Human Activity Recognition (WHAR) research. WHAR research often struggles with inconsistent datasets, difficulties in reproducing experiments, and time-consuming manual preprocessing. The library addresses these issues by providing a standardized data format, a configuration-driven design for streamlined workflows, support for nine widely-used WHAR datasets, and integration with PyTorch and TensorFlow. It also uses multiprocessing to accelerate preprocessing by up to 3.8 times and is easily extensible to include new datasets. The library was validated by training TinyHAR and MLP-HAR models, successfully reproducing published results and enabling reliable benchmarking.
Our Goal
The main goal of WHAR Datasets is to enable efficient, reproducible, and comparable research in WHAR. It aims to reduce manual effort and errors while providing a consistent framework for training, evaluating, and benchmarking human activity recognition models.
