This library offers comprehensive support for widely used WHAR (Wearable Human Activity Recognition) datasets, including:

• automated downloading from original sources and data extraction
• parsing into a unified, standardized data format
• configurable pre-processing (e.g., resampling, windowing) and post-processing (e.g., normalization)
• dataset splitting for common evaluation protocols such as LOSO and K-Fold cross-validation
• built-in caching and multi-processing for improved performance
• seamless integration with PyTorch and TensorFlow

The library currently includes out-of-the-box support for 33 datasets (listed below). Additional WHAR datasets can be easily integrated by defining a custom configuration with an associated parser and registering it with the framework.

This library does not host any datasets. To use a dataset, please visit its original website and make sure you understand and agree to the dataset’s terms and conditions.

pip install "git+https://github.com/teco-kit/whar-datasets.git"

from whar_datasets import (
    Loader,
    LOSOSplitter,
    PostProcessingPipeline,
    PreProcessingPipeline,
    TorchAdapter,
    WHARDatasetID,
    get_dataset_cfg,
)

# create cfg for WISDM dataset
cfg = get_dataset_cfg(WHARDatasetID.WISDM)

# create and run pre-processing pipeline
pre_pipeline = PreProcessingPipeline(cfg)
activity_df, session_df, window_df = pre_pipeline.run()

# create LOSO splits
splitter = LOSOSplitter(cfg)
splits = splitter.get_splits(session_df, window_df)
split = splits[0]

# create and run post-processing pipeline for the specific split
post_pipeline = PostProcessingPipeline(cfg, pre_pipeline, window_df, split.train_indices)
samples = post_pipeline.run()

# create dataloaders for the specific split
loader = Loader(session_df, window_df, post_pipeline.samples_dir, samples)
adapter = TorchAdapter(cfg, loader, split)
dataloaders = adapter.get_dataloaders(batch_size=64)

If you use the WHAR Datasets library in your research, please cite our paper:

@inproceedings{burzer2025whar,
  title={WHAR Datasets: An Open Source Library for Wearable Human Activity Recognition},
  author={Burzer, Maximilian and King, Tobias and Riedel, Till and Beigl, Michael and R{\"o}ddiger, Tobias},
  booktitle={Companion of the 2025 ACM International Joint Conference on Pervasive and Ubiquitous Computing},
  pages={1315--1322},
  year={2025}
}