pyADiff: A simple, pure python algorithmic differentiation package¶
pyADiff is a (yet) very basic algorithmic differentiation package, which implements forward and adjoint/reverse mode differentiation. If you are looking for a fully-featured and faster library, have a look at google/jax, autograd or dco/c++ (or many more), but if you are interested in a package where you are able to quickly “look under the hood”, you may be right here.
Contents¶
Basic Usage¶
Suppose we want to compute the gradient of the function \(f(x_0, x_1) = 2 x_0 x_1^2\). This is a rather trivial task, because by simple calculus the gradient is:
Nevertheless we use this example illustrate the use of pyADiff.
import pyADiff as ad
# define the function f
def f(x):
return 2.*x[0]*x[1]**2.
# call the gradient function of pyADiff
df = ad.gradient(f)
x = [0.5, 2.0]
# Call the function f and the gradient function df
y = f(x)
dy = df(x)
print("f({}) = {}".format(x, y)) # prints f([0.5, 2.0]) = 4.0
print("f'({}) = {}".format(x, dy)) # prints f'([0.5, 2.0]) = [8. 4.]
Which corresponds to the evaluation of the analytic gradient.
Motivation¶
My motivation to start this project arose from curiosity while listening to the lecture Computational Differentiation by Uwe Naumann at RWTH Aachen University. So basically I tried to understand the concepts from the lecture by implementing them by myself. In the end I was (positively) surprised with the outcome and decided to bundle it in a python package. Additionaly this gave me the chance to learn about python packaging, distributing, documentation, …