Refactor gradient and jacobians as multi-dimensional arrays #20
Labels
enhancement
New feature or request
medium
This is expected to be medium.
students 🎯
A label indicating that this issue is an expected feature for the student software project.
This is an issue reserved for the TU Delft Student Software Project '23 - LaplaceRedux
I think this is a fairly challenging problem, so do expect to spend some time on this. It may involve some breaking changes to the core architecture of the package, but on the bright side, it should help with some outstanding issues.
Currently gradients and Jacobians are mapped to two-dimensional arrays, since so far I had not considered training LA in batches (see related #19). This currently leads to problems for the implementation of GGN for the multi-class cases (see related #17). Refactoring things properly should help. In this context it may be worth using
Tullio.jl
for multi-dimensional array computations.The text was updated successfully, but these errors were encountered: