optipy 0.1.0

Numerical optimization/minimization

optipy

optipy contains a generic optimization/minimization method. Its creation was motivated by the absence of an implementation of Newton's method with a custom Hessian solver in SciPy (see this bug).

The mandatory Rosenbrock example:

a = 1.0
b = 100.0

def fun(x):
    return (a-x[0])**2 + b*(x[1] - x[0]**2)**2

def jac(x):
    return numpy.array([
        -2*(a-x[0]) - 4*b*(x[1] - x[0]**2) * x[0],
        2*b*(x[1] - x[0]**2)
        ])

def hess_inv(x, grad):
    hess = numpy.array([
        [2 + 8*b*x[0]**2 - 4*b*(x[1] - x[0]**2), -4*b*x[0]],
        [-4*b*x[0], 2*b]
        ])
    return -numpy.linalg.solve(hess, grad)

sol = optipy.minimize(
    fun=fun,
    x0=[-1.0, 3.5],
    jac=jac,
    get_search_direction=hess_inv,
    atol=1.0e-5
    )

This is basically the exact Newton method. When setting get_search_direction to lambda x, grad: -grad, one gets the steepest descent method. For larger computations, one will typically replace this with a tailored solver, e.g., a preconditioned Krylov solver.

The return type is largely compatible with SciPy's generic return type, OptmizeResult.

Installation

optipy is available from the Python Package Index, so simply do

pip install -U optipy

to install or upgrade. Use sudo -H to install as root or the --user option of pip to install in $HOME.

Testing

To run the optipy unit tests, check out this repository and type

pytest

Distribution

To create a new release

1. bump the __version__ number,

2. publish to PyPi and tag on GitHub:

   $ make publish
   

License

optipy is published under the MIT license.