ipyopt

Ipopt solves general nonlinear programming problems of the form

min f(x)

under the constraints

g_l <= g(x) <= g_u,
x_l <= x <= x_u,

where x is n dimensional and g(x) is m dimensional.

Goal

Provide as much performance as possible. This is also reflected in the fact that the shipped scipy.optimize.minimize ipopt method deviates in some concerns from the usual methods in scipy. If you are interested in a more scipy like interface, have a look at cyipopt.

Installation

Note the pypi repository currently only provides 🐧 linux wheels.

pip install [--user] ipyopt

This will install a precompiled binary version from pypi. Please note, that the precompiled binary is linked against the unoptimized reference implementation of blas/lapack. If you want to take advantage of optimized versions of blas/lapack, compile from source:

pip install --no-binary ipyopt ipyopt

In this case, you also need Ipopt and Numpy. On a debian based system:

sudo apt-get install python3-numpy coinor-ipopt

If coinor-ipopt does not link correctly, you might have to compile ipopt yourself. See the section Build below or .ci/Dockerfile on how this can be done.

Usage

You can use ipyopt like this:

import ipyopt
# define your call back functions
nlp = ipyopt.Problem(...)
nlp.solve(...)

For an example, see examples/hs071.py.

Note that the ipyopt.Problem.solve(.) mutates some its arguments, including the initial guess for the variables, and the multipliers.

For maximal performance, there is also support for PyCapsules / scipy.LowLevelCallable. By using this approach, there will be no C++ <-> Python interactions during Ipopt optimization. Here is an example test/c_capsules/ (C code) and test/test_ipyopt.py (Python code using the PyCapsules provided by the C code).

For more details and hints, see the docs.

Build

ipyopt depends on the following packages:

1. A compiler and a linker, e.g. gcc, ld
2. Ipopt
3. Numpy
4. Python.h (part of the python source code, you can download it from Python.org)

To build from source, first, get the latest source code using:

git clone https://gitlab.com/ipyopt-devs/ipyopt.git

Check whether a file ipopt.pc was distributed with your Ipopt installation. If this is the case and ipopt.pc is in the search path of pkg-config (on unix systems: /usr/lib/pkgconfig, /usr/share/pkgconfig, /usr/local/lib/pkgconfig, /usr/local/share/pkgconfig), nothing has to be modified.

In this case run

python setup.py build
sudo python setup.py install

If pkg-config is not available for your system, you will need to pass appropriate information to setup.py by setting the environment variable CFLAGS. Example:

CFLAGS="-I/usr/include/coin/ -l/usr/lib64 -lipopt -lmumps_common -ldmumps -lzmumps -lsmumps -lcmumps -llapack -lblas -lblas -lblas -lm  -ldl' ./setup.py build
sudo python setup.py install

If you have an ipopt.pc which is not in the pkg-config search path, specify the path via the PKG_CONFIG_PATH environment variable (see below). If you cannot find an ipopt.pc in your ipopt installation, there is an example pc file pkgconfig/ipopt.pc. Copy it to a location (best of all directly in a subfolder named pkgconfig of your Ipopt installation) and edit it to reflect the library and include paths of the dependencies.

Then do

PKG_CONFIG_PATH=<dir containing ipopt.pc> python setup.py build
sudo python setup.py install

Testing

Unit tests:

python -m unittest

Run examples:

Use the following command under the examples directory.

python hs071.py

The file examples/hs071.py contains a toy optimization problem. If everything is OK, ipyopt will invoke Ipopt to solve it for you. This python file is self-documented and can be used as a template for writing your own optimization problems.

Hessian Estimation: since Hessian estimation is usually tedious, Ipopt can solve problems without Hessian estimation. ipyopt also supports this feature. The file examples/hs071.py demonstrates the idea. If you provide the ipyopt.Problem constructor with an eval_h callback function, Ipopt will delegate the Hessian matrix calculation to your function (otherwise Ipopt will approximate Hessian for you).

Contributing

1. Fork it.
2. Create a branch (git checkout -b new_branch)
3. Commit your changes (git commit -am "your awesome message")
4. Push to the branch (git push origin new_branch)
5. Create a merge request

Credits

• Modifications on logger made by OpenMDAO at NASA Glenn Research Center, 2010 and 2011
• Added "eval_intermediate_callback" by OpenMDAO at NASA Glenn Research Center, 2010 and 2011
• Modifications on the SAFE_FREE macro made by Guillaume Jacquenot, 2012
• Changed logger from code contributed by alanfalloon
• Originally developed by Eric Xu when he was a PhD student at Washington University and issued under the BSD license. Original repository: xuy/pyipopt.