neurotools.stats.minimize module

Helper functions for minimization and optimization

exception neurotools.stats.minimize.FailureError[source]

Bases: RuntimeError

Re-named RuntimeError to distinguish error conditions arising from failure of optimization routines.

neurotools.stats.minimize.minimize_retry(objective, initial, jac=None, hess=None, verbose=False, printerrors=True, failthrough=True, tol=1e-05, simplex_only=False, show_progress=False, dontuse={}, maxfeval=None, maxgeval=None, **kwargs)[source]

Call scipy.optimize.minimize, retrying a few times in case one solver doesn’t work.

This addresses unresolved bugs that can cause exceptions in some of the gradient-based solvers in Scipy. If we happen upon these bugs, we can continue optimization using slower but more robused methods.

Ultimately, this routine falls-back to the gradient-free Nelder-Mead simplex algorithm, although it will try to use faster routines if the hessian and gradient are providede.

Parameters:

  • objective (objective, passed to scipy.minimize)

  • initial (initial parameter guess)

  • jac ((optional) jacobian, passed to scipy.minimize)

  • hess ((optional) Hessian, passed to scipy.minimize)

  • verbose (print extra information)

  • failthrough (return best params found so far, even if minimization fails)

  • tol (convergence tolerance, passed to scipy.minimize)

  • simple_only (Force it to use only the Nelder-Mead simplex optimizer)

  • show_progress (Print status updates during minimization)

  • dontuse (Set of methods not to try)

  • maxfeval (Maximum number of objective function evaluations to allow)

  • maxgeval (Maximum number of gradient function evaluations to allow)