CUKETA

CUKETA, pronounced /sʊketa/, is not only a Czech word for zucchini, but also CUDA Accelerated Minimally Random Convolutional Kernel Transform.

CUKETA is a CUDA implementation of the MiniRocket algorithm [1], or rather it's Julia re-implementation from the TimeSeriesClassification.jl [2] package.

This project came to life as a semestral project from the NI-GPU | GPU Architectures and Programming class at FIT CTU in Prague.

For more detail see the report paper in report/main.pdf, which quickly introduces the MiniRocket algorithm and goes into more details regarding the CUDA implementation and benchmarks.

Benchmark

The lower the shape, the better. :)

Usage

See usage in main.cu in the main() function. Mostly it boils down to:

auto [dilations, num_features_per_dilation, biases] = fit(X, num_examples, input_length);
auto transform_res = transform(X, num_examples, input_length, dilations, num_features_per_dilation, biases);

Compile code

To compile the code and benchmarks, you should have GCC, CMake, Julia, and CUDA SDK installed on your machine.
The versions should not really matter as long as it's "recent enough", i.e. GCC 12+, CMake 3.28+, Julia 1.9+ and CUDA 12+.
For Python you'll also need packages sktime, numpy and numba, for Julia you'll need TimeSeriesClassification.jl.

Steps to compile and run the CUDA code:

1. Have fun!
2. Use CMake to compile the code
3. Run the compiled binary
4. Have fun!

Run benchmarks

See code in each file for more details.

Python:

python mrmini.py

Julia:

julia mrmini.jl

CUDA: See main.cu for more details. Don't forget to set correct input sizes.

# Compiled binaries will be in this folder
mkdir compiled
# Compile the code for each block size specified in the `batch_compile.py` file
python batch_compile.py
# Run the compiled binaries
python batch_run_all.sh

Citations

[1] Dempster, A., Schmidt, D. F., and Webb, G. I., “MINIROCKET: A Very Fast (Almost) Deterministic Transform for Time Series Classification”, arXiv e-prints, 2020. doi:10.48550/arXiv.2012.08791.

[2] https://github.com/antoninkriz/TimeSeriesClassification.jl