CUROCKET

This repo contains a CUDA-based implementation of the feature-extraction algorithm ROCKET (RandOm Convolutional KErnel Transform) [1] to run ROCKET on GPU. On the test setup, it has up to 11 times higher computational efficiency per watt than ROCKET on CPU. CUROCKET is explained in detail in this paper [2].

Content

• CuRocketUnivariate: univariate version of CUROCKET
• CuRocketMultivariate: multivariate version of CUROCKET
• CuRocketMultivariateMultiGPU: multivariate version of CUROCKET that works on multiple GPUs in parallel. It has more overhead than single GPU, therefore it is only more efficient at a certain dataset size. Currently only mutltiple GPUs of the same type are supported

Setup

You need a system with a CUDA-capable GPU and CUDA installed. The python packages sktime, numba and cupy need to be installed. If you want to use the notebook, install jupyter. Refer to the packages' documentation for a detailed explanation on how to install them. They can be installed via pip. Depending on your CUDA version, you have to install a different version of cupy. In the following example it's for CUDA 12.x.

pip install sktime
pip install numba
pip install cupy-cuda12x
pip install jupyter

Alternatively, you can use the devcontainer.

Usage

The modules are set up in an sklearn-compatible format, so you can use functions fit and transform on an instance of CUROCKET as demonstrated in the following example.

import numpy as np
from CuRocketUnivariate.CuRocket import CuRocketUnivariate

X = np.random.random((100, 1, 1000))
cuda_rocket = CuRocketUnivariate()
cuda_rocket.fit(X)
X_t = cuda_rocket.transform(X)

You can also use this notebook to test out the different versions and compare the calculation time.

Citation

If you use this code, please cite:

@article{stueven_etal_2026,
  author={Stüven, Ole and Moenck, Keno and Schüppstuhl, Thorsten},
  title={CUROCKET: Optimizing ROCKET for GPU},
  year={2026},
}