Esempio n. 1
0
    def test_custom(self):

        np.random.seed(123)
        X, y = datasets.make_regression(n_samples=15, n_features=3)

        breast_cancer = datasets.load_breast_cancer()
        Z = breast_cancer.data
        t = breast_cancer.target

        regr = linear_model.BayesianRidge()
        regr2 = gaussian_process.GaussianProcessClassifier()

        fit_obj = ns.CustomRegressor(
            obj=regr,
            n_hidden_features=10,
            direct_link=False,
            bias=False,
            nodes_sim="sobol",
            activation_name="relu",
            n_clusters=0,
        )

        fit_obj2 = ns.CustomRegressor(
            obj=regr,
            n_hidden_features=9,
            direct_link=False,
            bias=True,
            nodes_sim="halton",
            activation_name="sigmoid",
            n_clusters=2,
        )

        fit_obj3 = ns.CustomRegressor(
            obj=regr,
            n_hidden_features=8,
            direct_link=True,
            bias=False,
            nodes_sim="uniform",
            activation_name="tanh",
            n_clusters=3,
        )

        fit_obj4 = ns.CustomRegressor(
            obj=regr,
            n_hidden_features=7,
            direct_link=True,
            bias=True,
            nodes_sim="hammersley",
            activation_name="elu",
            n_clusters=2,
        )

        fit_obj5 = ns.CustomClassifier(
            obj=regr2,
            n_hidden_features=3,
            direct_link=True,
            bias=True,
            nodes_sim="hammersley",
            activation_name="elu",
            n_clusters=4,
        )

        fit_obj6 = ns.CustomClassifier(
            obj=regr2,
            n_hidden_features=3,
            direct_link=True,
            bias=True,
            nodes_sim="hammersley",
            activation_name="elu",
            n_clusters=4,
            col_sample=0.8,
        )

        fit_obj7 = ns.CustomClassifier(
            obj=regr2,
            n_hidden_features=3,
            direct_link=True,
            bias=True,
            nodes_sim="hammersley",
            activation_name="elu",
            n_clusters=2,
            col_sample=0.8,
            row_sample=0.9,
        )

        fit_obj8 = ns.CustomRegressor(
            obj=regr,
            n_hidden_features=10,
            direct_link=False,
            bias=False,
            nodes_sim="sobol",
            activation_name="relu",
            n_clusters=0,
            col_sample=0.9,
            row_sample=1.0,
        )

        fit_obj9 = ns.CustomRegressor(
            obj=regr,
            n_hidden_features=10,
            direct_link=False,
            bias=False,
            nodes_sim="sobol",
            activation_name="relu",
            n_clusters=2,
            cluster_encode=False,
            col_sample=0.9,
            row_sample=1.0,
        )

        index_train = range(10)
        index_test = range(10, 15)
        X_train = X[index_train, :]
        y_train = y[index_train]
        X_test = X[index_test, :]
        y_test = y[index_test]

        fit_obj.fit(X_train, y_train)
        err = fit_obj.predict(X_test) - y_test
        rmse = np.sqrt(np.mean(err**2))

        fit_obj2.fit(X_train, y_train)
        err2 = fit_obj2.predict(X_test) - y_test
        rmse2 = np.sqrt(np.mean(err2**2))

        fit_obj3.fit(X_train, y_train)
        err3 = fit_obj3.predict(X_test) - y_test
        rmse3 = np.sqrt(np.mean(err3**2))

        fit_obj4.fit(X_train, y_train)
        err4 = fit_obj4.predict(X_test) - y_test
        rmse4 = np.sqrt(np.mean(err4**2))

        err4_1 = fit_obj4.predict(X_test[0, :]) - y_test[0]
        rmse4_1 = np.sqrt(np.mean(err4_1**2))

        fit_obj8.fit(X_train, y_train)
        err8 = fit_obj8.predict(X_test) - y_test
        rmse8 = np.sqrt(np.mean(err8**2))

        fit_obj5.fit(Z[0:100, :], t[0:100])
        pred5 = fit_obj5.predict(Z[106, :])

        fit_obj6.fit(Z[0:100, :], t[0:100])
        pred6 = fit_obj6.predict(Z[106, :])
        pred6_proba1 = fit_obj6.predict_proba(Z[100, :])
        pred6_proba2 = fit_obj6.predict_proba(Z[100:120, :])

        score6 = fit_obj6.score(Z[100:120, :], t[100:120])

        fit_obj7.fit(Z[0:100, :], t[0:100])
        score7 = fit_obj7.score(Z[100:120, :], t[100:120])

        fit_obj9.fit(X_train, y_train)
        err9 = fit_obj9.predict(X_test) - y_test
        rmse9 = np.sqrt(np.mean(err9**2))

        self.assertTrue(np.allclose(rmse, 64.933610490495667))
        self.assertTrue(np.allclose(rmse2, 12.968755131423396))
        self.assertTrue(np.allclose(rmse3, 26.716371782298673))
        self.assertTrue(np.allclose(rmse4, 3.3480684032382375))
        self.assertTrue(np.allclose(rmse4_1, 3.4339085800638713e-05))
        self.assertTrue(np.allclose(rmse8, 64.93379560651447))
        self.assertTrue(np.allclose(pred5, 1))
        self.assertTrue(np.allclose(pred6, 1))
        self.assertTrue(np.allclose(pred6_proba1[0], 0.50507336437937767))
        self.assertTrue(np.allclose(pred6_proba2[3, 0], 0.4836233209751839))
        self.assertTrue(np.allclose(score6, 0.94999999999999996))
        self.assertTrue(np.allclose(score7, 0.85))
        self.assertTrue(np.allclose(rmse9, 76.12747461579583))
# layer 1 (base layer) ----
print(" \n layer 1 ----- \n")
layer1_regr = RandomForestClassifier(n_estimators=10, random_state=123)

start = time() 

layer1_regr.fit(X_train, y_train)

# Accuracy in layer 1
print(layer1_regr.score(X_test, y_test))


# layer 2 using layer 1 ----
print(" \n layer 2 ----- \n")
layer2_regr = ns.CustomClassifier(obj = layer1_regr, n_hidden_features=5, 
                        direct_link=True, bias=True, 
                        nodes_sim='uniform', activation_name='relu', 
                        n_clusters=2, seed=123)
layer2_regr.fit(X_train, y_train)

# Accuracy in layer 2
print(layer2_regr.score(X_test, y_test))


# layer 3 using layer 2 ----
print(" \n layer 3 ----- \n")
layer3_regr = ns.CustomClassifier(obj = layer2_regr, n_hidden_features=10, 
                        direct_link=True, bias=True, dropout=0.7,
                        nodes_sim='uniform', activation_name='relu', 
                        n_clusters=2, seed=123)
layer3_regr.fit(X_train, y_train)
Esempio n. 3
0
    def test_score(self):

        np.random.seed(123)
        X, y = datasets.make_regression(n_samples=15, n_features=3)
        breast_cancer = datasets.load_breast_cancer()
        Z = breast_cancer.data
        t = breast_cancer.target

        regr = linear_model.BayesianRidge()
        regr2 = gaussian_process.GaussianProcessClassifier()
        regr3 = gaussian_process.GaussianProcessClassifier()

        fit_obj = ns.CustomRegressor(
            obj=regr,
            n_hidden_features=100,
            direct_link=True,
            bias=True,
            nodes_sim="sobol",
            activation_name="relu",
            n_clusters=2,
        )

        fit_obj2 = ns.CustomClassifier(
            obj=regr2,
            n_hidden_features=10,
            direct_link=True,
            dropout=0.1,
            bias=True,
            activation_name="relu",
            n_clusters=3,
        )

        fit_obj3 = ns.CustomRegressor(
            obj=regr,
            n_hidden_features=100,
            dropout=0.6,
            direct_link=True,
            bias=True,
            nodes_sim="sobol",
            activation_name="relu",
            n_clusters=2,
        )

        fit_obj4 = ns.CustomClassifier(
            obj=regr3,
            n_hidden_features=50,
            dropout=0.5,
            direct_link=True,
            bias=True,
            activation_name="tanh",
            n_clusters=2,
        )

        fit_obj.fit(X, y)
        fit_obj2.fit(Z, t)
        fit_obj3.fit(X, y)
        fit_obj4.fit(Z, t)

        self.assertTrue(
            np.allclose(fit_obj.score(X, y), 4846.2057110929481)
            & np.allclose(fit_obj2.score(Z, t), 0.99648506151142358)
            & np.allclose(fit_obj3.score(X, y), 2.1668647954067132e-11)
            & np.allclose(fit_obj4.score(Z, t), 1.0))