numdifftools package¶
Subpackages¶
- numdifftools.tests package
- Submodules
- numdifftools.tests.conftest module
- numdifftools.tests.test_extrapolation module
- numdifftools.tests.test_hessian module
- numdifftools.tests.test_limits module
- numdifftools.tests.test_multicomplex module
- numdifftools.tests.test_nd_algopy module
- numdifftools.tests.test_numdifftools module
- numdifftools.tests.test_numdifftools_docstrings module
- Module contents
Submodules¶
numdifftools.core module¶
numerical differentiation functions:
Derivative, Gradient, Jacobian, and Hessian
Author: Per A. Brodtkorb
Created: 01.08.2008 Copyright: (c) pab 2008 Licence: New BSD
Based on matlab functions derivest.m gradest.m hessdiag.m, hessian.m and jacobianest.m version 1.0 released 12/27/2006 by John D’Errico (e-mail: woodchips@rochester.rr.com)
Also based on the python functions approx_fprime, approx_fprime_cs, approx_hess_cs, approx_hess1, approx_hess2 and approx_hess3 in the statsmodels.tools.numdiff module released in 2014 written by Josef Perktold.
-
numdifftools.core.
dea3
(v0, v1, v2, symmetric=False)[source]¶ Extrapolate a slowly convergent sequence
Parameters: v0, v1, v2 : array-like
3 values of a convergent sequence to extrapolate
Returns: result : array-like
extrapolated value
abserr : array-like
absolute error estimate
-
class
numdifftools.core.
Derivative
(f, step=None, method='central', order=2, n=1, full_output=False)[source]¶ Bases:
numdifftools.core._Derivative
Calculate n-th derivative with finite difference approximation
Parameters: f : function
function of one array f(x, *args, **kwds)
step : float, array-like or StepGenerator object, optional
Defines the spacing used in the approximation. Default is MinStepGenerator(base_step=step, step_ratio=None) if step or method in in [‘complex’, ‘multicomplex’], otherwise MaxStepGenerator(step_ratio=None, num_extrap=14) The results are extrapolated if the StepGenerator generate more than 3 steps.
method : {‘central’, ‘complex’, ‘multicomplex’, ‘forward’, ‘backward’}
defines the method used in the approximation
order : int, optional
defines the order of the error term in the Taylor approximation used. For ‘central’ and ‘complex’ methods, it must be an even number.
n : int, optional
Order of the derivative.
full_output : bool, optional
If full_output is False, only the derivative is returned. If full_output is True, then (der, r) is returned der is the derivative, and r is a Results object.
Returns: der : ndarray
array of derivatives
Notes
Complex methods are usually the most accurate provided the function to differentiate is analytic. The complex-step methods also requires fewer steps than the other methods and can work very close to the support of a function. The complex-step derivative has truncation error O(steps**2) for n=1 and O(steps**4) for n larger, so truncation error can be eliminated by choosing steps to be very small. Especially the first order complex-step derivative avoids the problem of round-off error with small steps because there is no subtraction. However, this method fails if f(x) does not support complex numbers or involves non-analytic functions such as e.g.: abs, max, min. Central difference methods are almost as accurate and has no restriction on type of function. For this reason the ‘central’ method is the default method, but sometimes one can only allow evaluation in forward or backward direction.
For all methods one should be careful in decreasing the step size too much due to round-off errors.
Higher order approximation methods will generally be more accurate, but may also suffer more from numerical problems. First order methods is usually not recommended.
-
n
¶
-
-
class
numdifftools.core.
Jacobian
(f, step=None, method='central', order=2, full_output=False)[source]¶ Bases:
numdifftools.core.Gradient
Calculate Jacobian with finite difference approximation
Parameters: f : function
function of one array f(x, *args, **kwds)
step : float, array-like or StepGenerator object, optional
Defines the spacing used in the approximation. Default is MinStepGenerator(base_step=step, step_ratio=None) if step or method in in [‘complex’, ‘multicomplex’], otherwise MaxStepGenerator(step_ratio=None, num_extrap=14) The results are extrapolated if the StepGenerator generate more than 3 steps.
method : {‘central’, ‘complex’, ‘multicomplex’, ‘forward’, ‘backward’}
defines the method used in the approximation
order : int, optional
defines the order of the error term in the Taylor approximation used. For ‘central’ and ‘complex’ methods, it must be an even number.
full_output : bool, optional
If full_output is False, only the derivative is returned. If full_output is True, then (der, r) is returned der is the derivative, and r is a Results object.
Returns: jacob : array
Jacobian
See also
Notes
Complex methods are usually the most accurate provided the function to differentiate is analytic. The complex-step methods also requires fewer steps than the other methods and can work very close to the support of a function. The complex-step derivative has truncation error O(steps**2) for n=1 and O(steps**4) for n larger, so truncation error can be eliminated by choosing steps to be very small. Especially the first order complex-step derivative avoids the problem of round-off error with small steps because there is no subtraction. However, this method fails if f(x) does not support complex numbers or involves non-analytic functions such as e.g.: abs, max, min. Central difference methods are almost as accurate and has no restriction on type of function. For this reason the ‘central’ method is the default method, but sometimes one can only allow evaluation in forward or backward direction.
For all methods one should be careful in decreasing the step size too much due to round-off errors.
Higher order approximation methods will generally be more accurate, but may also suffer more from numerical problems. First order methods is usually not recommended.
If f returns a 1d array, it returns a Jacobian. If a 2d array is returned by f (e.g., with a value for each observation), it returns a 3d array with the Jacobian of each observation with shape xk x nobs x xk. I.e., the Jacobian of the first observation would be [:, 0, :]
-
class
numdifftools.core.
Gradient
(f, step=None, method='central', order=2, full_output=False)[source]¶ Bases:
numdifftools.core.Derivative
Calculate Gradient with finite difference approximation
Parameters: f : function
function of one array f(x, *args, **kwds)
step : float, array-like or StepGenerator object, optional
Defines the spacing used in the approximation. Default is MinStepGenerator(base_step=step, step_ratio=None) if step or method in in [‘complex’, ‘multicomplex’], otherwise MaxStepGenerator(step_ratio=None, num_extrap=14) The results are extrapolated if the StepGenerator generate more than 3 steps.
method : {‘central’, ‘complex’, ‘multicomplex’, ‘forward’, ‘backward’}
defines the method used in the approximation
order : int, optional
defines the order of the error term in the Taylor approximation used. For ‘central’ and ‘complex’ methods, it must be an even number.
full_output : bool, optional
If full_output is False, only the derivative is returned. If full_output is True, then (der, r) is returned der is the derivative, and r is a Results object.
Returns: grad : array
gradient
See also
Notes
Complex methods are usually the most accurate provided the function to differentiate is analytic. The complex-step methods also requires fewer steps than the other methods and can work very close to the support of a function. The complex-step derivative has truncation error O(steps**2) for n=1 and O(steps**4) for n larger, so truncation error can be eliminated by choosing steps to be very small. Especially the first order complex-step derivative avoids the problem of round-off error with small steps because there is no subtraction. However, this method fails if f(x) does not support complex numbers or involves non-analytic functions such as e.g.: abs, max, min. Central difference methods are almost as accurate and has no restriction on type of function. For this reason the ‘central’ method is the default method, but sometimes one can only allow evaluation in forward or backward direction.
For all methods one should be careful in decreasing the step size too much due to round-off errors.
Higher order approximation methods will generally be more accurate, but may also suffer more from numerical problems. First order methods is usually not recommended.
-
class
numdifftools.core.
Hessian
(f, step=None, method='central', full_output=False)[source]¶ Bases:
numdifftools.core._Derivative
Calculate Hessian with finite difference approximation
Parameters: f : function
function of one array f(x, *args, **kwds)
step : float, array-like or StepGenerator object, optional
Defines the spacing used in the approximation. Default is MinStepGenerator(base_step=step, step_ratio=None) if step or method in in [‘complex’, ‘multicomplex’], otherwise MaxStepGenerator(step_ratio=None, num_extrap=14) The results are extrapolated if the StepGenerator generate more than 3 steps.
method : {‘central’, ‘complex’, ‘multicomplex’, ‘forward’, ‘backward’}
defines the method used in the approximation
full_output : bool, optional
If full_output is False, only the derivative is returned. If full_output is True, then (der, r) is returned der is the derivative, and r is a Results object.
Returns: hess : ndarray
array of partial second derivatives, Hessian
See also
Notes
Complex methods are usually the most accurate provided the function to differentiate is analytic. The complex-step methods also requires fewer steps than the other methods and can work very close to the support of a function. The complex-step derivative has truncation error O(steps**2) for n=1 and O(steps**4) for n larger, so truncation error can be eliminated by choosing steps to be very small. Especially the first order complex-step derivative avoids the problem of round-off error with small steps because there is no subtraction. However, this method fails if f(x) does not support complex numbers or involves non-analytic functions such as e.g.: abs, max, min. Central difference methods are almost as accurate and has no restriction on type of function. For this reason the ‘central’ method is the default method, but sometimes one can only allow evaluation in forward or backward direction.
For all methods one should be careful in decreasing the step size too much due to round-off errors.
Computes the Hessian according to method as: ‘forward’ (1), ‘central’ (2) and ‘complex’ (3):
(1)
(2)
(3)
where
is a vector with element
is one and the rest are zero and
is a scalar spacing
.
-
class
numdifftools.core.
Hessdiag
(f, step=None, method='central', order=2, full_output=False)[source]¶ Bases:
numdifftools.core.Derivative
Calculate Hessian diagonal with finite difference approximation
Parameters: f : function
function of one array f(x, *args, **kwds)
step : float, array-like or StepGenerator object, optional
Defines the spacing used in the approximation. Default is MinStepGenerator(base_step=step, step_ratio=None) if step or method in in [‘complex’, ‘multicomplex’], otherwise MaxStepGenerator(step_ratio=None, num_extrap=14) The results are extrapolated if the StepGenerator generate more than 3 steps.
method : {‘central’, ‘complex’, ‘multicomplex’, ‘forward’, ‘backward’}
defines the method used in the approximationorder : int, optional defines the order of the error term in the Taylor approximation used. For ‘central’ and ‘complex’ methods, it must be an even number.
full_output : bool, optional
If full_output is False, only the derivative is returned. If full_output is True, then (der, r) is returned der is the derivative, and r is a Results object.
Returns: hessdiag : array
hessian diagonal
See also
Notes
Complex methods are usually the most accurate provided the function to differentiate is analytic. The complex-step methods also requires fewer steps than the other methods and can work very close to the support of a function. The complex-step derivative has truncation error O(steps**2) for n=1 and O(steps**4) for n larger, so truncation error can be eliminated by choosing steps to be very small. Especially the first order complex-step derivative avoids the problem of round-off error with small steps because there is no subtraction. However, this method fails if f(x) does not support complex numbers or involves non-analytic functions such as e.g.: abs, max, min. Central difference methods are almost as accurate and has no restriction on type of function. For this reason the ‘central’ method is the default method, but sometimes one can only allow evaluation in forward or backward direction.
For all methods one should be careful in decreasing the step size too much due to round-off errors.
Higher order approximation methods will generally be more accurate, but may also suffer more from numerical problems. First order methods is usually not recommended.
-
class
numdifftools.core.
MinStepGenerator
(base_step=None, step_ratio=2, num_steps=None, offset=0, scale=None, num_extrap=0, use_exact_steps=True, check_num_steps=True)[source]¶ Bases:
object
Generates a sequence of steps
where steps = base_step * step_ratio ** (np.arange(num_steps) + offset)
Parameters: base_step : float, array-like, optional
Defines the base step, if None, then base_step is set to EPS**(1/scale)*max(log(1+|x|), 1) where x is supplied at runtime through the __call__ method.
step_ratio : real scalar, optional, default 2
Ratio between sequential steps generated. Note: Ratio > 1 If None then step_ratio is 2 for n=1 otherwise step_ratio is 1.6
num_steps : scalar integer, optional, default n + order - 1 + num_extrap
defines number of steps generated. It should be larger than n + order - 1
offset : real scalar, optional, default 0
offset to the base step
scale : real scalar, optional
scale used in base step. If not None it will override the default computed with the default_scale function.
-
class
numdifftools.core.
MaxStepGenerator
(step_max=2.0, step_ratio=2.0, num_steps=15, step_nom=None, offset=0, num_extrap=0, use_exact_steps=False, check_num_steps=True)[source]¶ Bases:
numdifftools.core.MinStepGenerator
Generates a sequence of steps
- where
- steps = base_step * step_ratio ** (-np.arange(num_steps) + offset) base_step = step_max * step_nom
Parameters: max_step : float, array-like, optional default 2
Defines the maximum step
step_ratio : real scalar, optional, default 2
Ratio between sequential steps generated. Note: Ratio > 1
num_steps : scalar integer, optional, default n + order - 1 + num_extrap
defines number of steps generated. It should be larger than n + order - 1
step_nom : default maximum(log1p(abs(x)), 1)
Nominal step.
offset : real scalar, optional, default 0
offset to the base step: max_step * nom_step
-
class
numdifftools.core.
Richardson
(step_ratio=2.0, step=1, order=1, num_terms=2)[source]¶ Bases:
object
Extrapolates as sequence with Richardsons method
Notes
Suppose you have series expansion that goes like this
L = f(h) + a0 * h^p_0 + a1 * h^p_1+ a2 * h^p_2 + ...
where p_i = order + step * i and f(h) -> L as h -> 0, but f(0) != L.
If we evaluate the right hand side for different stepsizes h we can fit a polynomial to that sequence of approximations. This is exactly what this class does.
numdifftools.extrapolation module¶
Created on 28. aug. 2015
@author: pab
-
class
numdifftools.extrapolation.
Dea
(limexp=3)[source]¶ Bases:
object
LIMEXP is the maximum number of elements the epsilon table data can contain. The epsilon table is stored in the first (LIMEXP+2) entries of EPSTAB.
-
class
numdifftools.extrapolation.
Richardson
(step_ratio=2.0, step=1, order=1, num_terms=2)[source]¶ Bases:
object
Extrapolates as sequence with Richardsons method
Notes
Suppose you have series expansion that goes like this
L = f(h) + a0 * h^p_0 + a1 * h^p_1+ a2 * h^p_2 + ...
where p_i = order + step * i and f(h) -> L as h -> 0, but f(0) != L.
If we evaluate the right hand side for different stepsizes h we can fit a polynomial to that sequence of approximations. This is exactly what this class does.
-
numdifftools.extrapolation.
convolve
(sequence, rule, **kwds)[source]¶ Wrapper around scipy.ndimage.convolve1d that allows complex input.
numdifftools.info module¶
Introduction to Numdifftools¶
Numdifftools is a suite of tools written in Python to solve automatic numerical differentiation problems in one or more variables. Finite differences are used in an adaptive manner, coupled with a Richardson extrapolation methodology to provide a maximally accurate result. The user can configure many options like; changing the order of the method or the extrapolation, even allowing the user to specify whether complex, multicomplex, central, forward or backward differences are used. The methods provided are:
- Derivative:
- Computates the derivative of order 1 through 10 on any scalar function.
- Gradient:
- Computes the gradient vector of a scalar function of one or more variables.
- Jacobian:
- Computes the Jacobian matrix of a vector valued function of one or more variables.
- Hessian:
- Computes the Hessian matrix of all 2nd partial derivatives of a scalar function of one or more variables.
- Hessdiag:
- Computes only the diagonal elements of the Hessian matrix
All of these methods also produce error estimates on the result.
Numdifftools also provide an easy to use interface to derivatives calculated with AlgoPy. Algopy stands for Algorithmic Differentiation in Python. The purpose of AlgoPy is the evaluation of higher-order derivatives in the forward and reverse mode of Algorithmic Differentiation (AD) of functions that are implemented as Python programs.
Documentation is at: http://numdifftools.readthedocs.org/
Code and issue tracker is at https://github.com/pbrod/numdifftools.
Latest stable release is at http://pypi.python.org/pypi/Numdifftools.
To test if the toolbox is working paste the following in an interactive python session:
import numdifftools as nd
nd.test(coverage=True, doctests=True)
Getting Started¶
Compute 1’st and 2’nd derivative of exp(x), at x == 1:
>>> import numpy as np
>>> import numdifftools as nd
>>> fd = nd.Derivative(np.exp) # 1'st derivative
>>> fdd = nd.Derivative(np.exp, n=2) # 2'nd derivative
>>> np.allclose(fd(1), 2.7182818284590424)
True
>>> np.allclose(fdd(1), 2.7182818284590424)
True
Nonlinear least squares:
>>> xdata = np.reshape(np.arange(0,1,0.1),(-1,1))
>>> ydata = 1+2*np.exp(0.75*xdata)
>>> fun = lambda c: (c[0]+c[1]*np.exp(c[2]*xdata) - ydata)**2
>>> Jfun = nd.Jacobian(fun)
>>> np.allclose(np.abs(Jfun([1,2,0.75])), 0) # should be numerically zero
True
Compute gradient of sum(x**2):
>>> fun = lambda x: np.sum(x**2)
>>> dfun = nd.Gradient(fun)
>>> dfun([1,2,3])
array([ 2., 4., 6.])
Compute the same with the easy to use interface to AlgoPy:
>>> import numdifftools.nd_algopy as nda
>>> import numpy as np
>>> fd = nda.Derivative(np.exp) # 1'st derivative
>>> fdd = nda.Derivative(np.exp, n=2) # 2'nd derivative
>>> np.allclose(fd(1), 2.7182818284590424)
True
>>> np.allclose(fdd(1), 2.7182818284590424)
True
Nonlinear least squares:
>>> xdata = np.reshape(np.arange(0,1,0.1),(-1,1))
>>> ydata = 1+2*np.exp(0.75*xdata)
>>> fun = lambda c: (c[0]+c[1]*np.exp(c[2]*xdata) - ydata)**2
>>> Jfun = nda.Jacobian(fun, method='reverse')
>>> np.allclose(np.abs(Jfun([1,2,0.75])), 0) # should be numerically zero
True
Compute gradient of sum(x**2):
>>> fun = lambda x: np.sum(x**2)
>>> dfun = nda.Gradient(fun)
>>> dfun([1,2,3])
array([ 2., 4., 6.])
numdifftools.limits module¶
Created on 27. aug. 2015
@author: pab Author: John D’Errico e-mail: woodchips@rochester.rr.com Release: 1.0 Release date: 5/23/2008
-
class
numdifftools.limits.
Limit
(f, step=None, method='above', order=4, full_output=False)[source]¶ Bases:
object
Compute limit of a function at a given point
Parameters: f : callable
function of one array f(z, *args, **kwds) to compute the limit for. The function, f, is assumed to return a result of the same shape and size as its input, z.
step: float, complex, array-like or StepGenerator object, optional
Defines the spacing used in the approximation. Default is MinStepGenerator(base_step=step, step_ratio=4)
method : {‘above’, ‘below’}
defines if the limit is taken from above or below
order: positive scalar integer, optional.
defines the order of approximation used to find the specified limit. The order must be member of [1 2 3 4 5 6 7 8]. 4 is a good compromise.
Returns: limit_fz: array like
estimated limit of f(z) as z –> z0
info:
Only given if full_output is True and contains the following: error estimate: ndarray
95 uncertainty estimate around the limit, such that abs(limit_fz - f(z0)*(z-z0)) < error_estimate
- final_step: ndarray
final step used in approximation
-
class
numdifftools.limits.
MinStepGenerator
(base_step=None, step_ratio=4.0, num_steps=None, offset=0, scale=1.2, use_exact_steps=True)[source]¶ Bases:
object
Generates a sequence of steps
where steps = base_step * step_ratio ** (np.arange(num_steps) + offset)
Parameters: base_step : float, array-like, optional
Defines the base step, if None, then base_step is set to EPS**(1/scale)*max(log(1+|x|), 1) where x is supplied at runtime through the __call__ method.
step_ratio : real scalar, optional, default 4
Ratio between sequential steps generated.
num_steps : scalar integer, optional, default n + order - 1 + num_extrap
defines number of steps generated. It should be larger than n + order - 1
offset : real scalar, optional, default 0
offset to the base step
scale : real scalar, optional
scale used in base step. If not None it will override the default computed with the default_scale function.
numdifftools.multicomplex module¶
Created on 22. apr. 2015
@author: pab
References¶
A METHODOLOGY FOR ROBUST OPTIMIZATION OF LOW-THRUST TRAJECTORIES IN MULTI-BODY ENVIRONMENTS Gregory Lantoine (2010) Phd thesis, Georgia Institute of Technology
USING MULTICOMPLEX VARIABLES FOR AUTOMATIC COMPUTATION OF HIGH-ORDER DERIVATIVES Gregory Lantoine, Ryan P. Russell , and Thierry Dargent ACM Transactions on Mathematical Software, Vol. 38, No. 3, Article 16, April 2012, 21 pages,
M.E. Luna-Elizarraras, M. Shapiro, D.C. Struppa1, A. Vajiac (2012) CUBO A Mathematical Journal Vol. 14, No 2, (61-80). June 2012.
Computation of higher-order derivatives using the multi-complex step method Adriaen Verheyleweghen, (2014) Project report, NTNU
numdifftools.nd_algopy module¶
Numdifftools.nd_algopy¶
This module provide an easy to use interface to derivatives calculated with AlgoPy. Algopy stands for Algorithmic Differentiation in Python.
The purpose of AlgoPy is the evaluation of higher-order derivatives in the forward and reverse mode of Algorithmic Differentiation (AD) of functions that are implemented as Python programs. Particular focus are functions that contain numerical linear algebra functions as they often appear in statistically motivated functions. The intended use of AlgoPy is for easy prototyping at reasonable execution speeds. More precisely, for a typical program a directional derivative takes order 10 times as much time as time as the function evaluation. This is approximately also true for the gradient.
Algoritmic differentiation¶
Algorithmic differentiation (AD) is a set of techniques to numerically evaluate the derivative of a function specified by a computer program. AD exploits the fact that every computer program, no matter how complicated, executes a sequence of elementary arithmetic operations (addition, subtraction, multiplication, division, etc.) and elementary functions (exp, log, sin, cos, etc.). By applying the chain rule repeatedly to these operations, derivatives of arbitrary order can be computed automatically, accurately to working precision, and using at most a small constant factor more arithmetic operations than the original program.
Algorithmic differentiation is not:
Symbolic differentiation, nor Numerical differentiation (the method of finite differences). These classical methods run into problems: symbolic differentiation leads to inefficient code (unless carefully done) and faces the difficulty of converting a computer program into a single expression, while numerical differentiation can introduce round-off errors in the discretization process and cancellation. Both classical methods have problems with calculating higher derivatives, where the complexity and errors increase. Finally, both classical methods are slow at computing the partial derivatives of a function with respect to many inputs, as is needed for gradient-based optimization algorithms. Algoritmic differentiation solves all of these problems.
Reference¶
Sebastian F. Walter and Lutz Lehmann 2013, “Algorithmic differentiation in Python with AlgoPy”, in Journal of Computational Science, vol 4, no 5, pp 334 - 344, http://www.sciencedirect.com/science/article/pii/S1877750311001013
https://en.wikipedia.org/wiki/Automatic_differentiation
https://pythonhosted.org/algopy/index.html
-
class
numdifftools.nd_algopy.
Derivative
(f, n=1, method='forward')[source]¶ Bases:
numdifftools.nd_algopy._Common
Calculate n-th derivative with Algorithmic Differentiation method
Parameters: f : function
function of one array f(x, *args, **kwds)
n : int, optional
Order of the derivative.
method : string, optional {‘forward’, ‘reverse’}
defines method used in the approximation
Returns: der : ndarray
array of derivatives
Notes
Algorithmic differentiation is a set of techniques to numerically evaluate the derivative of a function specified by a computer program. AD exploits the fact that every computer program, no matter how complicated, executes a sequence of elementary arithmetic operations (addition, subtraction, multiplication, division, etc.) and elementary functions (exp, log, sin, cos, etc.). By applying the chain rule repeatedly to these operations, derivatives of arbitrary order can be computed automatically, accurately to working precision, and using at most a small constant factor more arithmetic operations than the original program.
-
class
numdifftools.nd_algopy.
Gradient
(f, method='forward')[source]¶ Bases:
numdifftools.nd_algopy._Common
Calculate Gradient with Algorithmic Differentiation method
Parameters: f : function
function of one array f(x, *args, **kwds)
method : string, optional {‘forward’, ‘reverse’}
defines method used in the approximation
Returns: grad : array
gradient
See also
Notes
Algorithmic differentiation is a set of techniques to numerically evaluate the derivative of a function specified by a computer program. AD exploits the fact that every computer program, no matter how complicated, executes a sequence of elementary arithmetic operations (addition, subtraction, multiplication, division, etc.) and elementary functions (exp, log, sin, cos, etc.). By applying the chain rule repeatedly to these operations, derivatives of arbitrary order can be computed automatically, accurately to working precision, and using at most a small constant factor more arithmetic operations than the original program.
-
class
numdifftools.nd_algopy.
Hessdiag
(f, method='forward')[source]¶ Bases:
numdifftools.nd_algopy.Hessian
Calculate Hessian diagonal with Algorithmic Differentiation method
Parameters: f : function
function of one array f(x, *args, **kwds)
method : string, optional {‘forward’, ‘reverse’}
defines method used in the approximation
Returns: hessdiag : ndarray
Hessian diagonal array of partial second order derivatives.
See also
Notes
Algorithmic differentiation is a set of techniques to numerically evaluate the derivative of a function specified by a computer program. AD exploits the fact that every computer program, no matter how complicated, executes a sequence of elementary arithmetic operations (addition, subtraction, multiplication, division, etc.) and elementary functions (exp, log, sin, cos, etc.). By applying the chain rule repeatedly to these operations, derivatives of arbitrary order can be computed automatically, accurately to working precision, and using at most a small constant factor more arithmetic operations than the original program.
-
class
numdifftools.nd_algopy.
Hessian
(f, method='forward')[source]¶ Bases:
numdifftools.nd_algopy._Common
Calculate Hessian with Algorithmic Differentiation method
Parameters: f : function
function of one array f(x, *args, **kwds)
method : string, optional {‘forward’, ‘reverse’}
defines method used in the approximation
Returns: hess : ndarray
array of partial second derivatives, Hessian
See also
Notes
Algorithmic differentiation is a set of techniques to numerically evaluate the derivative of a function specified by a computer program. AD exploits the fact that every computer program, no matter how complicated, executes a sequence of elementary arithmetic operations (addition, subtraction, multiplication, division, etc.) and elementary functions (exp, log, sin, cos, etc.). By applying the chain rule repeatedly to these operations, derivatives of arbitrary order can be computed automatically, accurately to working precision, and using at most a small constant factor more arithmetic operations than the original program.
-
class
numdifftools.nd_algopy.
Jacobian
(f, method='forward')[source]¶ Bases:
numdifftools.nd_algopy._Common
Calculate Jacobian with Algorithmic Differentiation method
Parameters: f : function
function of one array f(x, *args, **kwds)
method : string, optional {‘forward’, ‘reverse’}
defines method used in the approximation
Returns: jacob : array
Jacobian
See also
Notes
Algorithmic differentiation is a set of techniques to numerically evaluate the derivative of a function specified by a computer program. AD exploits the fact that every computer program, no matter how complicated, executes a sequence of elementary arithmetic operations (addition, subtraction, multiplication, division, etc.) and elementary functions (exp, log, sin, cos, etc.). By applying the chain rule repeatedly to these operations, derivatives of arbitrary order can be computed automatically, accurately to working precision, and using at most a small constant factor more arithmetic operations than the original program.
numdifftools.run_benchmark module¶
Module contents¶
Introduction to Numdifftools¶
Numdifftools is a suite of tools written in Python to solve automatic numerical differentiation problems in one or more variables. Finite differences are used in an adaptive manner, coupled with a Richardson extrapolation methodology to provide a maximally accurate result. The user can configure many options like; changing the order of the method or the extrapolation, even allowing the user to specify whether complex, multicomplex, central, forward or backward differences are used. The methods provided are:
- Derivative:
- Computates the derivative of order 1 through 10 on any scalar function.
- Gradient:
- Computes the gradient vector of a scalar function of one or more variables.
- Jacobian:
- Computes the Jacobian matrix of a vector valued function of one or more variables.
- Hessian:
- Computes the Hessian matrix of all 2nd partial derivatives of a scalar function of one or more variables.
- Hessdiag:
- Computes only the diagonal elements of the Hessian matrix
All of these methods also produce error estimates on the result.
Numdifftools also provide an easy to use interface to derivatives calculated with AlgoPy. Algopy stands for Algorithmic Differentiation in Python. The purpose of AlgoPy is the evaluation of higher-order derivatives in the forward and reverse mode of Algorithmic Differentiation (AD) of functions that are implemented as Python programs.
Documentation is at: http://numdifftools.readthedocs.org/
Code and issue tracker is at https://github.com/pbrod/numdifftools.
Latest stable release is at http://pypi.python.org/pypi/Numdifftools.
To test if the toolbox is working paste the following in an interactive python session:
import numdifftools as nd
nd.test(coverage=True, doctests=True)
Getting Started¶
Compute 1’st and 2’nd derivative of exp(x), at x == 1:
>>> import numpy as np
>>> import numdifftools as nd
>>> fd = nd.Derivative(np.exp) # 1'st derivative
>>> fdd = nd.Derivative(np.exp, n=2) # 2'nd derivative
>>> np.allclose(fd(1), 2.7182818284590424)
True
>>> np.allclose(fdd(1), 2.7182818284590424)
True
Nonlinear least squares:
>>> xdata = np.reshape(np.arange(0,1,0.1),(-1,1))
>>> ydata = 1+2*np.exp(0.75*xdata)
>>> fun = lambda c: (c[0]+c[1]*np.exp(c[2]*xdata) - ydata)**2
>>> Jfun = nd.Jacobian(fun)
>>> np.allclose(np.abs(Jfun([1,2,0.75])), 0) # should be numerically zero
True
Compute gradient of sum(x**2):
>>> fun = lambda x: np.sum(x**2)
>>> dfun = nd.Gradient(fun)
>>> dfun([1,2,3])
array([ 2., 4., 6.])
Compute the same with the easy to use interface to AlgoPy:
>>> import numdifftools.nd_algopy as nda
>>> import numpy as np
>>> fd = nda.Derivative(np.exp) # 1'st derivative
>>> fdd = nda.Derivative(np.exp, n=2) # 2'nd derivative
>>> np.allclose(fd(1), 2.7182818284590424)
True
>>> np.allclose(fdd(1), 2.7182818284590424)
True
Nonlinear least squares:
>>> xdata = np.reshape(np.arange(0,1,0.1),(-1,1))
>>> ydata = 1+2*np.exp(0.75*xdata)
>>> fun = lambda c: (c[0]+c[1]*np.exp(c[2]*xdata) - ydata)**2
>>> Jfun = nda.Jacobian(fun, method='reverse')
>>> np.allclose(np.abs(Jfun([1,2,0.75])), 0) # should be numerically zero
True
Compute gradient of sum(x**2):
>>> fun = lambda x: np.sum(x**2)
>>> dfun = nda.Gradient(fun)
>>> dfun([1,2,3])
array([ 2., 4., 6.])
See also¶
scipy.misc.derivative