jeppe742/autodiff

A simple framework for doing automatical differentiation

AutoDiff - A framework for doing automatical differentiation

Basic Usage

The framework works around the AdFloat object. To evaluate a function for a value, you need to create this value as an AdFloat

from autodiff import AdFloat
x = AdFloat(27)

simple operations like + - * / is overloaded, so you can just do

f = lambda x: x*2 
value = f(x)

value is now an AdFloat, which means that you can access both the derivative, and the evaluated value by

print(value.dx)
> 2
print(value.x)
> 52

For more complex functions, you have to import the implemented versions.

from autodiff import AdFloat, sin
x = AdFloat(3.14159265)
f = lambda x: sin(x)
value = f(x)
print(value.dx)
> -1

Chaining derivatives

The power of automatic differentiation comes from the fact, that one can evaluate derivatives of arbitrary functions without nummeric approximations.
To work with functions with multiple terms, simply write them as normal

from autodiff import AdFloat, sin, exp
x = AdFloat(1)
f = lambda x: sin(exp(x*4)) + (x+2)**2
print(f(x).dx)
> -74.94977972639266 

Installation

Windows

AutoDiff is available as a precompiled wheel for windows, and can be installed using

pip install autodiff

Other platforms

For other platforms it can be build from source by downloading the repository, and running
python setup.py install

This requires that you have installed a c compiler

autodiff's Issues

typo

https://github.com/jeppe742/AutoDiff/blob/master/autodiff/AdFloat_py.py#L36

mistake b.dx*b.x Tests need improvement:)
IDK if you (or somebody) use this library, but I borrowed some code from here for quick&dirty simple autodiff implementation (autograd has less obvious code) and mentioned this bug.
Sorry, I'm lazy to do PR.