NumPy tutorial for ML@LPC

This repository hosts the material of a tutorial to NumPy in the context of a working group at Laboratoire de Physique de Clermont related to machine learning and its application in (high energy) physics. This tutorial is split into three parts, trying go from first principles to current (known) limitations for High Energy Physics (HEP). A PDF document containing the notes of this tutorial is also available in this repository. All notebooks can be ran in a browser using both mybinder (issue of dead kernel though) and google Colaboratory (requires chrome and a google account).

Contact: [email protected]

1. Introduction to NumPy

This first part, detailed in this notebook, motivates the use of NumPy inside python and, also, why it is interesting to use python! The basic NumPy object, numpy.array, is introduced emphasizing differences with usual python lists. Then the three key concepts on which NumPy is based are presented (broadcasting, indexing/slicing, vectorization) and simple examples are given to illustrate each of them (in particular, how to get ride of loops). Finally, a quick tour of two important tools interfaced with NumPy is given, namely pandas (database-like objet), matplotlib (plotting package) and a very tiny glimps of scipy (scientific python). A notebook with exercises is proposed to quickly practice these concepts using a electric pulse analysis.

2. Typical use cases in high energy physics

This second part, detailed in this notebook, presents few typical standard computations in HEP based on multi-dimensional data using NumPy. The playground here is a sample of 1 million random "observations", each made of ten 3D vectors (r0, ..., r9) with ri=(x,y,z). One can think of ten positions in space or ten RGB colors. First, the concept presented in the introduction are used with this ndarray, and then more complex computations are discussed such as. A notebook with exercices to practice these concepts is also available. Typical computations include:

• combinatorics over the collection of 10 ri's
• distances between all (ri, rj) pairs
• distribution of |ri|^2 for only few selected r among the 10 ri ( with a criteria to be defined)
• compute an arbitrary function of ri over just a few selected point among the 10.

3. Limitations and possibles workarounds

This third part, detailed in this notebook, presents a realistic study of collider data using the concept presented in the two first sections. In particular, the notion of jagged array is introduced and the associated limitation in numpy is mitigated using few functions based on numpy for HEP. Few concrete cases are treated (such as computing invariant masses, finding the closest jet of the less isolated lepton) in order to illustrate how to use numpy for realistic event-by-event computation without explicit for loops.

---

Instructions to run notebooks into colab

1. Upload files (data or code): open the left panel (small gray arrow), go to tab Files and go to Upload
2. Import local module: the normal import my_module will not work and you need to import the file as follow (after uploading the my_module.py file):

from importlib.machinery import SourceFileLoader
module = SourceFileLoader('my_module', 'my_module.py').load_module()

After this, the command help(module) or module.function will work in the notebook.