netket.optimizer.SR(*args, solver_restart=False, **kwargs)#

Construct the structure holding the parameters for using the Stochastic Reconfiguration/Natural gradient method.

Depending on the arguments, an implementation is chosen. For details on all possible kwargs check the specific SR implementations in the documentation.

You can also construct one of those structures directly.

  • diag_shift – Diagonal shift added to the S matrix

  • method – (cg, gmres) The specific method.

  • iterative – Whether to use an iterative method or not.

  • jacobian – Differentiation mode to precompute gradients can be “holomorphic”, “R2R”, “R2C”, None (if they shouldn’t be precomputed)

  • rescale_shift – Whether to rescale the diagonal offsets in SR according to diagonal entries (only with precomputed gradients)

  • solver_restart (bool) –


The SR parameter structure.