fast-hydra

GPU-optimised PyTorch implementation of HYDRA (HYbrid Dictionary-Rocket Architecture) for fast and accurate time-series classification.

Features

• GPU-accelerated feature extraction for multivariate time series.
• SparseScaler: sparsity-aware feature scaling for both PyTorch and NumPy.
• Scikit-learn compatible API.
• Competitive convolutional kernels with configurable number of dilations to control number of features to extract.

Installation

uv add git+https://github.com/nimanzik/fast-hydra.git

Quick Start

import torch
from fast_hydra import HydraMultivariate, SparseScaler

input_size = 1000  # Length of time series
n_channels = 6   # Number of channels in multivariate time series
batch_size = 32  # Number of samples in a batch

# Extract features from multivariate time series
hydra = HydraMultivariate(
    input_size=input_size,
    n_channels=n_channels,
    n_groups=64,
    n_kernels=8,
    random_state=42,
)
X = torch.randn(batch_size, n_channels, input_size)
features = hydra(X)  # Tensor of shape (batch_size, n_features)

# Scale sparse features
scaler = SparseScaler()
scaled_features = scaler.fit_transform(features)

Requirements

• Numpy >= 2.3.4
• scikit-learn >= 1.7.2
• PyTorch >= 2.9.0

Citation

Dempster, A., Schmidt, D. F., & Webb, G. I. (2023). Hydra: Competing convolutional kernels for fast and accurate time series classification. Data Mining and Knowledge Discovery, 37(5), 1779-1805.

License

This project is licensed under the MIT License. See the LICENSE file for details.