Acceleration

The pyunlocbox.acceleration module implements acceleration schemes for use with the pyunlocbox.solvers. Pass a given acceleration object as an argument to your chosen solver during its initialization so that the solver can use it.

Interface

The accel base class defines a common interface to all acceleration schemes:

accel.pre(self, functions, x0)	Pre-processing specific to the acceleration scheme.
accel.update_step(self, solver, objective, niter)	Update the step size for the next iteration.
accel.update_sol(self, solver, objective, niter)	Update the solution point for the next iteration.
accel.post(self)	Post-processing specific to the acceleration scheme.

Acceleration schemes

Then, derived classes implement various common acceleration schemes.

dummy()	Dummy acceleration scheme which does nothing.
backtracking([eta])	Backtracking based on a local quadratic approximation of the smooth part of the objective.
fista(**kwargs)	Acceleration scheme for forward-backward solvers.
fista_backtracking([eta])	Acceleration scheme with backtracking for forward-backward solvers.
regularized_nonlinear([k, lambda_, …])	Regularized nonlinear acceleration (RNA) for gradient descent.