Skip to content
This repository has been archived by the owner on Apr 13, 2021. It is now read-only.

henryiii/hepvector

Repository files navigation

This package has been completely redesigned and released as vector, use pip install vector to obtain. The array-at-a-time computations from HEPVector are present, along with Numba support, Awkward support, and much more!


HEPVector

image Documentation Status Updates

Numpy based vectors for general purpose calculation and physics. Designed for the SciKit-HEP project, but can be used stand-alone with no dependencies except Numpy (extra performance can be obtained if Numba is present).

Features

The key design feature is array support. While non-array is also possible (backward compatible), arrays are many times faster and utilize all of Numpy's speed. An example of an array:

v = Vector3D(1, [2,3], 4)

would make two vectors, [1,2,4] and [1,3,4] (standard Numpy casting rules). In fact, the vector classes are simply subclasses of NDArrays, so anything that works with arrays works with Vectors, and Vectors can be cast to/from Numpy without a copy.

  • Optional speedups with Numba for a few parts (more eventually)
  • Phase space generator, similar speed (within a factor of 2 or so) to compiled ROOT, but fully in Python! Even works on an iPad. Also makes a good example of what HEPvector code looks like.
  • Lorentz, 3D, and 2D vectors, and all common operations in a single vector base class. More metrics and dimensions can be added by users.
  • Free software: BSD license
  • Documentation: https://hepvector.readthedocs.io

Performance

To take advantage of the underlying numpy machinery, you should use arrays of vectors when possible. For example, if you use the included generation of phase space:

  • 1 at a time: 5.18 ± 0.06 ms each
  • 1,000,000 at a time: 0.80 ± 0.04 µs each (803 ms total)

You can see it is 6,000 times faster if you generate 1,000,000 events using arrays instead of placing event-per-event code in loops.

Usage

The most important thing to remember is that this is just a lightly modified subclass of np.ndarray (just like np.array and np.matrix). The first dimension is the vector components; all Vector3D shapes will start with 3, for example. If you index, start your index with a : to capture all dimensions; if you don't, you'll get a normal np.array out. You can use .view to convert back and forth (no memory is copied).

You can create a vector several ways:

v = LorentzVector(x,y,z,t) # can be values or arrays
v = arr.view(LorentzVector) # Convert an existing array
v = LorentzVector.from_vector(arr) # Can also copy
v = LorentzVector.from_pandas(df) # DataFrames, too
v = Vector3D.from_spherical_coords(r, theta, phi)
v = Vector3D.from_cylindrical_coords(rho, phi, z)

You can use a variety of methods:

dot_product = a.dot(b)
cross_product = a.cross(b)
boost_vec = a.boost()

and others, see the docs. HEPvector is just a simple wrapper on top of a numpy array, where the first dimension represents the components of the vector. Any operation that does not apply to vectors (such as the transpose .T) uses the underlying numpy functionality and may return a numpy array instead of a Vector.

Credits

HEPvector was created by Henry Schreiner.

This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.