This package is a simpler version of the ABXpy package, for smaller task with less constraints.

Input and output format are easier to use.

$ python setup.py build && python setup.py install

To do an ABX evaluation:

import numpy as np
from abx_numpy import abx

# generate 120 random features of dimension 4 splitted in 3 classes
features = np.random.randint(0, 10, (120, 4))
classes = np.array(np.random.randint(0, 4, (120,)), dtype='S1')

# compute the ABX discrimination score
distance = lambda x, y: np.linalg.norm(x-y)
abx(classes, features, distance)

The main function (abx) takes as input:

• classes: list, [n_items], the class to which each items belong (any type). E.g. ['class 1', 'class 1', 'class 2'].
• features: array_like, [n_items, n_dim], the features for each items (in the same order as classes, any type). E.g. [[2, 3], [1, 3], [5, 4]]
• distance_function: callable, distance function to use for the calculation. E.g. euclidian distance. There are example of distances in the 'distances' module.

See '2D_normal_example.py' in 'examples/' for the code. A dataset is randomly generated: datapoints from 3 classes are sampled from 3 different gaussians. We evaluate the ABX discriminability between those 3 classes.

Average abx score: 0.82

This score correspond to the proportion of successful ABX trials (chance level is at 0.5).

Discrimination matrix:

The results confirm what we can see: Class 1 is confused with Class 2 but not with Class 3. Class 2 is confused with both. And Class 3 is confused with neither.

Note that the ordering of the array is the following: the score shown is the discriminability of the class in the column when compared to the class in the row. For instance, for the 1st row and 2nd column (0.76), X and A are from class 1 and B is from class 2.

Complete documentation can be found here.