def test_respawn(self):
m = 100
k = 5
X = np.random.randn(m, 2)
centroids = np.zeros((k, 2))
centroids = pvml.kmeans_train(X, k, init_centroids=centroids)
Z, _ = pvml.kmeans_inference(X, centroids)
self.assertListEqual(np.unique(Z).tolist(), list(range(k)))
def test_one_per_class(self):
for k in range(1, 12):
with self.subTest(k):
X = np.random.randn(k, 3)
centroids = pvml.kmeans_train(X, k)
Y, _ = pvml.kmeans_inference(X, centroids)
self.assertListEqual(
normalize_labels(Y).tolist(), list(range(k)))
def _linear(self, k, m):
Y = np.arange(m) % k
a = np.linspace(0, 2 * np.pi, m)
X = np.stack([np.cos(a) + 10 * Y, np.sin(a)], 1)
centroids = pvml.kmeans_train(X, k)
Z, _ = pvml.kmeans_inference(X, centroids)
self.assertListEqual(
normalize_labels(Y).tolist(),
normalize_labels(Z).tolist())
def _sort_centroids(self, X, Y):
# K-means labels do not correspond to training labels. A
# categorical classifier is used to reorder the centroids to
# minimize the error.
P, _ = pvml.kmeans_inference(X, self.centroids)
probs, priors = pvml.categorical_naive_bayes_train(P[:, None], Y)
YK = np.arange(self.k)[:, None]
Q, _ = pvml.categorical_naive_bayes_inference(YK, probs, priors)
ii = np.argsort(Q)
self.centroids = self.centroids[ii, :]
def inference(self, X):
ret = pvml.kmeans_inference(X, self.centroids)
return ret