tsxtract 0.1.0

Hardware-accelerated time series feature extraction using JAX

tsxtract

Hardware-accelerated time series feature extraction using JAX.

NOTE: tsxtract is still under development. Please report any bugs by creating an issue.

Please star the repository if you like tsxtract.

You can find the documentation here

Why tsxtract?

• Fast: All extraction operations are vectorized
• Hardware-accelerated: Run on CPU, GPU or TPU
• Easy-to-use: One function is all you need

Example usage

from tsxtract.extraction import extract_features

# Dataset is a 3d-numpy or jax array with following dimensions:
# (samples, channels, length)
features = extract_features(dataset)

print(type(features)) # dict
print(features["mean"].shape) # jax.Array of size (samples, channels)

Check out main.py for a complete usage example.

Installation

Using uv

Step 1: Install uv: pip install uv

Step 2: Clone this repository: git clone https://github.com/ctseidler/tsxtract.git .

Step 3: Create a new virtual environment: uv venv --python 3.12

Step 4: Activate the virtual environment: source .venv/bin/activate

Step 5: Install the package as editable from source: uv pip install -e .

Step 6: Test your setup by executing the main.py script: uv run main.py

Overview of Extracted Features

• Maximum
• Mean
• Minimum

Contributing

Contributions are welcome! Please open an issue, if you have any feature request. You can also implement it by forking the repository and creating a pull-request upon completion. Please make sure that your feature is covered by unittests (see test/). Current test coverage is 100%.

Development setup:

• Install the package locally as mentioned above.
• Run uv sync to install dev dependencies.
• Run the unit tests prior to a commit (pre-commit): uv run coverage run -m pytest
• Check the coverage report to identify missing test coverage: uv run coverage report -m

Roadmap

Version 0.2:

• Test CPU and GPU support
• Add example notebook for CPU and GPU extraction

Version 0.3:

• Add additional features
• Add features with customizable parameters
• Add configuration options

Version 0.4:

• Add support to custom features
• Allow configuration as dict
• Allow configuration as json

Version 0.5:

• Add frequency-based features
• Make package compatible for Python 3.10 and 3.11
• Allow easy IO, e.g., by integrating polars

Version 1.0:

• Performance benchmark
• Add project logo

Authors

Christian T. Seidler

See also

• tsfresh: Time Series Feature extraction based on scalable hypothesis test
• TSFEL: Time Series Feature Extraction Library
• pycatch22: CAnonical Time-series CHaracteristics in Python
• seglearn: An sklearn extension for machine learning time series or sequences

License

MIT