Skip to content

Latest commit

 

History

History
88 lines (59 loc) · 3.39 KB

README.md

File metadata and controls

88 lines (59 loc) · 3.39 KB
Raw

This is a Python script for printing performance profiles as defined by E. D. Dolan and J. J. Moré.

It is based on perfprof.m from the MATLAB Guide by D. J. Higham and N. J. Higham.

Performance profiles

Performance profiles are a mechanism to visualise the performance of multiple algorithms on multiple test problems. Given m problems and n algorithms, we're interested in the relative performance of all algorithms across the entire problem set.

Let t = t(i, j) > 0 be a measure of the performance of solver j on problem i, where lower means "better". Common choices for t are:

  • execution/CPU time;
  • number of iterations (assuming the algorithms are iterative);
  • the number of evaluations of some reference function (e.g., the ordering predicate in a sorting algorithm, or the objective function in an optimisation algorithm).

Given this data, a typical performance profile may look like this:

Example

Each algorithm has one line plot, where a point (x, y) means that, for x of the problem set, the algorithm in question was within a factor of y of the respective best algorithm. For example:

  • the point ≈(1, 0.6) means that Alg2 was the fastest algorithm on around 60% of the problem set;
  • the point ≈(1.5, 0.4) means that Alg3 was within a factor of 1.5 of the respective best algorithm for 40% of problems;
    • note that the "best" algorithm may be different for each problem;
  • the point ≈(1.5, 0.95) means that Alg1 was within a factor of 1.5 of the respective best algorithm for 95% of problems;
  • etc.

Generally speaking, an algorithm is considered efficient (relative to the others) when its performance profile comes close to the top left corner (1, 1).

It is possible for algorithms to fail on certain problems. This can be achieved by simply setting the performance measure t(i, j) to +inf or NaN.

Usage examples

import matplotlib.pyplot as plt
import perfprof

palette = ['-r', ':b', '--c', '-.g', '-y']

perfprof.perfprof(data, palette)
plt.show()

Marking y-intercepts

Markers can be inserted using the standard matplotlib pattern.

import matplotlib.pyplot as plt
import perfprof

palette = ['o-r', 'o:b', 'o--c', 'o-.g', 'o-y']

perfprof.perfprof(data, palette, markersize=4, markevery=[0])
plt.show()

Displaying legends

Legends can be displayed using matplotlib.pyplot.legend.

import matplotlib.pyplot as plt
import perfprof

palette = ['o-r', 'o:b', 'o--c', 'o-.g', 'o-y']
legend = ['Algorithm 1', 'Algorithm 2']

perfprof.perfprof(data, palette, markersize=4, markevery=[0])
plt.legend(legend)
plt.show()

Why another implementation?

Multiple implementations of performance profiles already exist in the public domain.

The design of perfprof was driven by a few key desires:

  • Simplicity: provide a clearly scoped, easy to use implementation that integrates with matplotlib;
  • Flexibility: unlock the full power of matplotlib for plot styling, legends, subplots etc.;
  • Robustness: the implementation must work in all edge cases including inf, NaN, etc.;
  • Usability: full Python3 compatibility and sensible defaults where possible.