frankwolfe-py

This package is a python wrapper for FrankWolfe.jl.

Installation

First, download the package frankwolfe-py.

Move to the root of the package and execute on a terminal the following command:

pip install .

Now you can use the FrankWolfe.jl functions with

from frankwolfepy import frankwolfe

and the wrapper of the objective function with

from frankwolfepy import wrapper

Usage

Each objective function must be wrapped if the algorithm is used with verbose=true.

In each file, import wrapper and wrap the objective function using:

f = wrapper.wrap_objective_function(f).

A simple example :

from frankwolfepy import frankwolfe
from frankwolfepy import wrapper
import numpy as np

def f(x): #objective function
    return np.linalg.norm(x)**2

f = wrapper.wrap_objective_function(f) #wrap the objective function

def grad(storage,x): #gradient computation
    for i in range(len(x)):
        storage[i] = x[i]

# Create the Linear Minimization Oracle
lmo_prob = frankwolfe.ProbabilitySimplexOracle(1)

# Compute first point
x0 = frankwolfe.compute_extreme_point(lmo_prob,np.zeros(5))

#Solve the optimisation problem using vanilla Frank-Wolfe algorithm
frankwolfe.frank_wolfe(
    f,
    grad,
    lmo_prob,
    x0,
    max_iteration=1000,
    line_search=frankwolfe.Agnostic(),
    verbose=True,
)

For documentation on FrankWolfe.jl: https://zib-iol.github.io/FrankWolfe.jl/dev/

Wrapping arbitrary Julia evaluations

The most flexible way to wrap Julia expressions is through seval which returns the result of evaluating a string as Julia code. This lets you for instance use type parameters like Vector{Int} or FrankWolfe.LpNormLMO{1}:

import juliacall
lmo = juliacall.Main.seval("FrankWolfe.LpNormLMO{1}(0.95)")