Benchmarking `sklearn.semi_supervised` `n_iter_` as a function of model and data characteristics

bench_semi_supervised_n_iter.py

import numpy as np
from sklearn import datasets
from sklearn.semi_supervised import LabelPropagation, LabelSpreading
###for n_samples in [20, 200, 2000, 20000]:
###    X, y = datasets.make_classification(n_samples=n_samples, n_classes=3, n_informative=3)
for (X, y) in [datasets.load_iris(return_X_y=True)]:
    for model in [LabelPropagation(max_iter=1000),
                  #LabelSpreading(alpha=0.01),
                  #LabelSpreading(alpha=0.1),
                  #LabelSpreading(alpha=0.3)
                  ]:
        for p in [.1, .15, .2, .25, .5, .75]:
            n_iter = []
            for i in range(20):
                random_unlabeled_points = np.random.rand(len(y)) < p
                labels = np.copy(y)
                labels[random_unlabeled_points] = -1
                model.fit(X, labels)
                n_iter.append(model.n_iter_)
            print(len(y), p, model.alpha, np.mean(n_iter), np.std(n_iter), sep='\t')