autoptim: automatic differentiation + optimization

Do you have a new machine learning model that you want to optimize, and do not want to bother computing the gradients? Autoptim is for you.

Warning:

As of version 0.3, pytorch has been replaced with autograd for automatic differentiation. It makes the interfacing with numpy even simpler.

Short presentation

Autoptim is a small Python package that blends autograd's automatic differentiation in scipy.optimize.minimize.

The gradients are computed under the hood using automatic differentiation; the user only provides the objective function:

import numpy as np
from autoptim import minimize


def rosenbrock(x):
    return (1 - x[0]) ** 2 + 100 * (x[1] - x[0] ** 2) ** 2


x0 = np.zeros(2)

x_min, _ = minimize(rosenbrock, x0)
print(x_min)

>>> [0.99999913 0.99999825]

It comes with the following features:

• Natural interfacing with Numpy: The objective function is written in standard Numpy. The input/ output of autoptim.minimize are Numpy arrays.

• Smart input processing: scipy.optimize.minimize is only meant to deal with one-dimensional arrays as input. In autoptim, variables can be multi-dimensional arrays or lists of arrays.

• Preconditioning: Preconditioning is a simple way to accelerate minimization, by doing a change of variables. autoptim makes preconditioning straightforward.

Disclaimer

This package is meant to be as easy to use as possible. As so, some compromises on the speed of minimization are made.

Installation

To install, use pip:

pip install autoptim

Dependencies

• numpy>=1.12
• scipy>=0.18.0
• autograd >= 1.2

Examples

Several examples can be found in autoptim/tutorials