def rank(M, eps=1e-9): U, s, V = svd(M, full_matrices=False) return np.sum(s > eps) bunch = fetch_20newsgroups_vectorized(subset="train") X_train = bunch.data y_train = bunch.target # Reduces dimensionality to make the example faster ch2 = SelectKBest(chi2, k=5000) X_train = ch2.fit_transform(X_train, y_train) bunch = fetch_20newsgroups_vectorized(subset="test") X_test = bunch.data y_test = bunch.target X_test = ch2.transform(X_test) clf = FistaClassifier(C=1.0 / X_train.shape[0], max_iter=200, penalty="trace", multiclass=True) for alpha in (1e-3, 1e-2, 0.1, 0.2, 0.3): print "alpha=", alpha clf.alpha = alpha clf.fit(X_train, y_train) print clf.score(X_test, y_test) print rank(clf.coef_)