TorchKM¶

TorchKM is a PyTorch-based library for kernel machines with fast training and model selection. It is designed for users who want the statistical behavior of classical kernel methods while taking advantage of GPU-friendly linear algebra.

The main idea behind TorchKM is simple: in many kernel-machine workflows, the bottleneck is not fitting one model, but fitting many models across cross-validation folds and tuning parameters. TorchKM integrates training and tuning so that repeated matrix computations can be reused.

What TorchKM provides¶

TorchKM currently focuses on binary kernel classifiers, kernel quantile regression, and related model-selection routines. It provides:

kernel support vector classification;
kernel distance-weighted discrimination;
kernel logistic regression;
kernel quantile regression;
pathwise model selection over a grid of regularization values;
exact cross-validation reuse for kernel machines;
GPU acceleration through PyTorch/CUDA, with CPU fallback;
Nyström approximation for larger data sets;
a scikit-learn-style estimator interface for common workflows.

Why use TorchKM?¶

TorchKM is useful when you want to tune nonlinear kernel classifiers without repeatedly refitting a separate model for every fold and every regularization value. It is especially helpful when the full training-and-tuning pipeline is the expensive part of the analysis.

Quick links¶

Citation¶

If you use TorchKM in academic work, please cite the software paper listed in the repository README.