Changelog and Road map


Our Changelog can be found on the KeOps Github repository.

To-do list

  • Put reference paper on Arxiv.

  • Fully document the inner C++ API and algorithms.

  • Provide R bindings.

  • Add support for tensor (and not just vector) variables.

  • Allow users to backprop through the .min() or .max() reduction of a LazyTensor.

  • Add support for the advanced indexing of LazyTensors. Users should be able to extract sub-matrices as LazyTensors or genuine NumPy arrays / PyTorch tensors to perform e.g. Nyström approximation without having to implement twice the same kernel formula.

  • Add support for the block construction of LazyTensors, using a BlockLazyTensor([[A, B], [C, D]]) syntax.

  • Write new tensor() and array() methods for LazyTensors, allowing users to cast their symbolic operators as explicit matrices whenever possible.

  • Add support for the Fast and Furious Method and other Multipole or Nyström-like approximations. By the start of 2020, we hope to provide a simple K.tol = 1e-3 syntax for LazyTensors.