as_scipy_functional

odl.as_scipy_functional(func, return_gradient=False)

Wrap op as a function operating on linear arrays.

This is intended to be used with the scipy solvers.

Parameters

funcFunctional.: A functional that should be wrapped
return_gradientbool, optional: True if the gradient of the functional should also be returned, False otherwise.

Returns

functioncallable: The wrapped functional.
gradientcallable, optional: The wrapped gradient. Only returned if return_gradient is true.

Examples

Wrap functional and solve simple problem (here toy problem min_x ||x||^2):

>>> func = odl.functionals.L2NormSquared(odl.rn(3))
>>> scipy_func = odl.as_scipy_functional(func)
>>> from scipy.optimize import minimize
>>> result = minimize(scipy_func, x0=[0, 1, 0])
>>> np.allclose(result.x, [0, 0, 0])
True

The gradient (jacobian) can also be provided:

>>> func = odl.functionals.L2NormSquared(odl.rn(3))
>>> scipy_func, scipy_grad = odl.as_scipy_functional(func, True)
>>> from scipy.optimize import minimize
>>> result = minimize(scipy_func, x0=[0, 1, 0], jac=scipy_grad)
>>> np.allclose(result.x, [0, 0, 0])
True

Notes

If the data representation of op’s domain is of type NumpyTensorSpace, this incurs no significant overhead. If the space type is CudaFn or some other nonlocal type, the overhead is significant.