FAQ

Does TorchKM require a GPU?

No. The high-level estimators choose CUDA when it is available and otherwise use CPU. You can also pass device="cpu" explicitly for small examples, tests, and debugging.

What labels should I use for classification?

The low-level classification solvers use labels in {-1, +1}. The high-level estimators accept any two distinct labels, map them internally, and return predictions in the original label space.

How is the tuning parameter selected?

Pass candidate values with Cs or let TorchKM create a log-spaced grid from C_max, C_min, and nC. The estimator uses cross-validation with cv folds and stores the selected value as best_C_.

When should I use low_rank=True?

Use the Nyström approximation when the full kernel matrix is too large or too slow for your workflow. The current high-level low-rank classifier path is for raw-feature RBF-kernel workflows.

Which estimators should I start with?

Start with the scikit-learn-style estimators in torchkm.estimators:

from torchkm.estimators import (
    TorchKMSVC,
    TorchKMDWD,
    TorchKMLogit,
    TorchKMKQR,
)

Use TorchKMSVC, TorchKMDWD, and TorchKMLogit for binary classification, and TorchKMKQR for kernel quantile regression.

Will benchmark times match the paper exactly?

No. Wall-clock times depend on hardware, PyTorch and CUDA versions, system load, and benchmark setup. Benchmark documentation should be read as a protocol, not a promise of identical timings.