vorpy 0.9.0

VictOR dods' PYthon package (mostly math-related)

Vorpy

A Python package for (usually math-related) code that I've written that finds use over and over.

Vorpy is short for "VictOR dods' PYthon package". Plus if you Google "vorpy", you come up with some pretty hilarious shit, so it seems like a good name choice.

Contents

• vorpy : The root module for the vorpy package.
• vorpy.apply_along_axes : Multi-dimensional generalization of numpy.apply_along_axis.
• vorpy.pickle : A module which facilitates disk-based caching of data via pickle files. Useful when using data that is time-consuming to generate but does not often change.
• vorpy.symbolic : Symbolic calculus module. This module eases use of sympy, facilitating the use of vector/tensor calculus (via numpy.ndarrays of symbols), and which allows compilation of symbolic functions into Python code, caching the generated code if desired. That process is known as 'lambdification' in sympy. Take a look at the cached_lambdified function within this source.
• vorpy.symplectic_integration : Module implementing symplectic integrators; available is a family of separable Hamiltonian integrators and a nonseparable Hamiltonian integrator.
• vorpy.tensor : Module providing functions for working with tensor values, notably the contract function, which does the same thing as numpy.einsum but allows dtype=object.

How to Install

Vorpy can be installed via pip using the following command:

pip install vorpy

Or, it can be installed directly from this github.com repository using the following command:

pip install --upgrade git+https://github.com/vdods/vorpy.git

Installation can also be made from a local copy of the vorpy package repo:

pip install --upgrade path/to/vorpy

where path/to/vorpy is the root directory of the project; the path containing setup.py (as well as this README.md file). Apropos: see pip install --editable <path/url>

To uninstall, use the following obvious command:

pip uninstall vorpy

A tarball of a particular tagged version can be obtained from github using a particular link. The latest version, with git tag v0.4.3, can be downloaded from:

https://github.com/vdods/vorpy/archive/v0.4.3.tar.gz

Extracting this will create and populate the vorpy-0.4.3 directory.

Running Tests

The suite of unit tests can be run via the command:

python setup.py test

Release Notes

• 0.9.0 : Added integrate_tangent_flow, added stuff to vorpy.tensor and vorpy.symplectic, added code for Bernstein polynomials and Bezier curves, some experimental work on strongly typed manifolds concepts, and a bunch of other miscellaneous work and improvements.
• 0.8.0 : Added the vorpy.linalg module.
• 0.7.0 : Modification to the symbolic polynomial generating functions.
• 0.6.0 : Added the vorpy.manifold and vorpy.symplectic modules, which implement some basic calculus in those categories.
• 0.5.1 : Fixed a bug where the Python 2 reduce was moved to the functools module, and so caused a NameError in Python 3.
• 0.5.0 : Added vorpy.tensor.
• 0.4.3 : More updates to be able to submit to pypi. Made tests generate files to the test_artifacts directory.
• 0.4.2 : Updates to be able to submit to pypi -- the Python Package Index. Also a minor exception printing improvement.
• 0.4.1 : Changed vorpy.pickle to use the dill module (which can pickle lambda expressions) instead of the builtin pickle module.
• 0.4.0 : Added vorpy.pickle.
• 0.3.1 : Added a means to salvage results from symplectic integration if an exception is raised during integration.
• 0.3.0 : Added vorpy.symplectic_integration.nonseparable_hamiltonian.
• 0.2.0 : Added vorpy.symplectic_integration.separable_hamiltonian.
• 0.1.0 : Added vorpy.apply_along_axes.
• 0.0.0 : Initial release. Added vorpy.symbolic.

To-Do List

• Maybe use boost's python numpy stuff to write C++ using numpy constructs (ndarray) -- https://jitcode.readthedocs.io/en/v1.4.0/

• Look into using JiTCODE to make fast numerics -- https://jitcode.readthedocs.io/en/v1.4.0/

• Use L^inf norm for local truncation error estimate, instead of L^2 norm. This should be faster to compute.

• Need to add a shape check in vorpy.symbolic.lambdified to verify that the argument has the expected shape. Ran into a hard-to-find bug that was the result of passing in the wrong shape and it happily not complaining.

• Make a version of vorpy.tensor.contract which allows for general contractions, instead of fixed letter-based contractions.

• Reorganize module structure to reflect the category theory of the subjects (e.g. vorpy.man for the category of manifolds, vorpy.vec for the category of vector spaces, vorpy.man.sympl for the category of symplectic manifolds, vorpy.man.riem for the category of Riemannian manifolds, etc).

• Add parsing of 'ij,kl->iljk' style string (as in numpy.einsum) to vorpy.tensor.contract, which will just replace the output kwarg.

• Make sure tests doesn't get installed (it looks like it is). OR, put tests under vorpy module (this is probably better).

• Look into using pytorch for faster numerics.

• Consolidate vorpy.symbolic.multiindex_iterator and vorpy.tensor.multiindex_iterator (probably use the one in symbolic because it looks fancier, but it does belong within the vorpy.tensor module).

• Include tests in the dist package, so they can be tested at the install site.

• Make the symbolic module aware of vectorized operations so that fast numpy-implemented ndarray functions can be used instead of structure-forgetting symbolic expressions that are fully written out. For example, the 1st and 2nd total derivatives of a quadratic form are simply matrix expressions which have simple numpy expressions.

• Verify that vorpy.apply_along_axes is automatically parallelized.

• Require numpy.ndarray or tuple be the type of the input array(s) for apply_along_axes, so that extra parens to form a trivial tuple are not necessary in the basic case where there is a single input array.

• In apply_along_axes, allow the base case where input_array_v[0].shape == () and input_axis_v == [].

• Increasing the order of Tao's nonseparable Hamiltonian integrator doesn't actually do what it should, even when using very small timesteps (expected behavior is that the error is orders smaller, but what actually happens is that it's of the same order as say order=2). Examine why this is happening (perhaps published version of Tao paper is updated with correction).

• Create a [symplectic] integrator using the Jacobi-Maupertuis principle.

• Specify an upper bound for integrator on H error, which abs(H - H(qp(0))). When computing the next step, if the H error is above the upper bound, recompute that step with a lower dt. This will involve defining a search strategy for dt that doesn't make it too small.

• Move PendulumNd and KeplerNd from tests dir into vorpy and write tests to symbolically verify all the formulas are correct. The goal would be to create more of these to provide a whole family of dynamical systems that can be used and experimented with.

• Have tests create a test_artifacts subdir which all files that the tests produce are put in, so there's an easy single thing to delete after tests run (or tests can clean this up themselves).

• Switch to pytest.

• For sys.stderr.write-based warnings, create a flag that can be set within the vorpy module to silence them.

• Use hashlib for data caching layer on top of vorpy.pickle.

• Make a human-readable hash, which concatenates words to represent elements of a hash space. This should be convertable to/from a binary string. Examples:

  StupidlyWalkingPerson
  BigThing
  Tree
  

  Maybe also tack on the remaining bits to the end as a decimal or hex number. Examples:

  StupidlyWalkingPersonA07BE5
  Tree10034
  

  See this.

• Use hashing in vorpy.symbolic.cached_lambdified on what gets lambdified so that changes to the function automatically cause the cache to be updated.

• Make a module that provides a nice abstraction for concurrent.futures.ProcessPoolExecutor for parallel processing.

• Allow multiply-nested dirs for cache_dirname in cached_lambdified -- this would correspond to multiply-nested modules for the cached code. Use os.makedirs instead of os.mkdir

References

• Maintainer Release Instructions

• [http://docs.python-guide.org/en/latest/writing/structure/]

• [https://en.wikipedia.org/wiki/Symplectic_integrator]

• [https://en.wikipedia.org/wiki/Energy_drift]

• [https://journals.aps.org/pre/abstract/10.1103/PhysRevE.94.043303] - Molei Tao - Explicit symplectic approximation of nonseparable Hamiltonians: Algorithm and long time performance

• The package index for dill (which is imported by vorpy.pickle) makes the following request:

  If you use dill to do research that leads to publication, we ask that you acknowledge use of dill by citing the following in your publication:
  
      M.M. McKerns, L. Strand, T. Sullivan, A. Fang, M.A.G. Aivazis,
      "Building a framework for predictive science", Proceedings of
      the 10th Python in Science Conference, 2011;
      http://arxiv.org/pdf/1202.1056
  
      Michael McKerns and Michael Aivazis,
      "pathos: a framework for heterogeneous computing", 2010- ;
      http://trac.mystic.cacr.caltech.edu/project/pathos
  
  Please see http://trac.mystic.cacr.caltech.edu/project/pathos or http://arxiv.org/pdf/1202.1056 for further information.