Skip to content

An automatic differentiation library for Python+NumPy.

License

Notifications You must be signed in to change notification settings

PTNobel/AutoDiff

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

47 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

auto_diff

An automatic differentiation library for Python+NumPy

How To Use

There are five public elements of the API:

  • AutoDiff is a context manager and must be entered with a with statement. The __enter__ method returns a new version of x that must be used to instead of the x passed as a parameter to the AutoDiff constructor.

  • value, jacobian, get_value_and_jacobian, these functions, which must be called in an AutoDiff context, extract the value, Jacobian, or both from a dependent variable.

  • get_value_and_jacobians, if multiple vectors are passed in as arguments to AutoDiff, this method returns a tuple of Jacobians wrt to the different variables.

If you are using get_value_and_jacobian, x must be a 2D column vector, and the value you must be parsing for the derivative must also be a 2D column vector. In most other cases, how to convert to a Jacobian Matrix is non-obvious. If you wish to deal with those cases see the paragraph after the example.

auto_diff also supports using sparse matrices instead of ndarrays to store the Jacobians. Simple use the SparseAutoDiff context manager instead of AutoDiff. Also if you use SparseAutoDiff, you need to verify that your code and none of non-NumPy dependencies use the np.ndarray constructor for a floating point vector. If using SparseAutoDiff, get_value_and_jacobian, jacobian, and get_value_and_jacobians return scipy.sparse.lil_matrixes instead of ndarrays.

Example

import auto_diff
import numpy as np

# Define a function f
# f can have other arguments, if they are constant wrt x
# Define the input vector, x

with auto_diff.AutoDiff(x) as x:
    f_eval = f(x, u)
    y, Jf = auto_diff.get_value_and_jacobian(f_eval)

# y is the value of f(x, u) and Jf is the Jacobian of f with respect to x.

If you need both the Jacobian wrt to x and u,

with auto_diff.AutoDiff(x, u) as (x, u):
    f_eval = f(x, u)
    y, (Jfx, Jfu) = auto_diff.get_value_and_jacobians(f_eval)

# y is the value of f(x, u), Jfx is the Jacobian of f with respect to x, and
# Jfu is the Jacobian of f with respect to u.

Finally, if f and x are very high-dimensional, then we can use SparseAutoDiff to save memory.

with auto_diff.SparseAutoDiff(x, u) as (x, u):
    f_eval = f(x, u)
    y, (Jfx, Jfu) = auto_diff.get_value_and_jacobians(f_eval)

# y is the value of f(x, u), Jfx is the Jacobian of f with respect to x, and
# Jfu is the Jacobian of f with respect to u.
# Jfx and Jfu are instances of scipy.sparse.lil_matrix.

We can also differentiate functions from arbitrarily shaped numpy arrays to arbitrarily shaped outputs. Let y = f(x), where x is a numpy array of shape x.shape, and y is is the output of the function we wish to differentiate, f.

We can then access a numpy array of shape (*y.shape, *x.shape), by accessing y.der. This represents the gradients of each component of y with respect to x. To find the gradient of the norm of a vector x, for example one can do

import auto_diff
import numpy as np
x = np.array([[np.pi], [3.0], [17.0]])

with auto_diff.AutoDiff(x) as x:
    print(np.linalg.norm(x).der)

Restrictions

  • You must import numpy and use that object, rather then do something like from numpy import ..., where ... is either * or just function names.

Crashes, Bug Reports, and Feedback: Email parthnobel@berkeley.edu

There are missing features right now. I'm working on them, feel free to email me if you want something prioritized.

How It Works

Parth Nobel. 2020. Auto_diff: an automatic differentiation package for Python. In Proceedings of the 2020 Spring Simulation Conference (SpringSim '20). Society for Computer Simulation International, San Diego, CA, USA, Article 10, 1–12. https://dl.acm.org/doi/10.5555/3408207.3408219

Prerequisite

A version of NumPy >= 1.17 may be required. Bugs on older versions have always raised errors, so there should be nothing to worry about.

Author: Parth Nobel (Github: /PTNobel, parthnobel@berkeley.edu) Version: 0.3