Skip to content

stochastic reconfiguration with jvp and vjp

Vicentini Filippo requested to merge github/fork/inailuig/sr into master

Created by: inailuig

Adds the option to do stochastic reconfiguration where the gradients are never stored, but (re-)calculated using jvp and vjp for every matrix-vector product in the iterative solver.

There are several advantages of this approach:

  • works for large numbers of samples and parameters where its impossible to store the gradients
  • non-homogeneous parameters can be treated more easily
  • can be faster for certain kinds of machines

Merge request reports