FSharp data science conference '23 notes
Our approach to establishing what the FsLab ecosystem is missing in terms of a 'goto' package for performant numerical algebra, was done according to the G(oal), R(eality), O(ptions), W(ay forward) flow. The main outcomes are:
- Tensor is the central concept, meaning a multi-dimensional array of type T, not constrainted to 2 dimensions.
- Operations/Algorithms on the Tensor objects will be dispatched to third party packages via "Providers", e.g. to TorchSharp. Relevant algorithms are svd, eig, linear solves,.. .
- Finding a suitable abstraction for a Provider is the core challenge of making the library work; a playground will be setup to iterate on ideas.
- The aim is to make a FsTensor package where the Tensor data structure will have it's own module for indexing, reshaping, broadcasting,... A second part/module within the package will be for Algorithms and will contain the definition of Provider(s).
Below are photos taken from the notes made during our discussions.
