def test_l1_neural_network_regressor_with_proximal_bundle():
X, y = load_boston(return_X_y=True)
X_scaled = StandardScaler().fit_transform(X)
X_train, X_test, y_train, y_test = train_test_split(X_scaled,
y,
train_size=0.75,
random_state=123456)
net = NeuralNetworkRegressor(
(FullyConnected(13, 13, relu), FullyConnected(13, 1, linear)),
loss=mean_absolute_error,
optimizer=ProximalBundle,
max_iter=150)
net.fit(X_train, y_train)
assert net.score(X_test, y_test) >= 0.83
def test_neural_network_regressor_with_stochastic_optimizer():
X, y = load_boston(return_X_y=True)
X_scaled = StandardScaler().fit_transform(X)
X_train, X_test, y_train, y_test = train_test_split(X_scaled,
y,
train_size=0.75,
random_state=1)
net = NeuralNetworkRegressor(
(FullyConnected(13, 13, sigmoid), FullyConnected(
13, 13, sigmoid), FullyConnected(13, 1, linear)),
loss=mean_squared_error,
optimizer=Adam,
learning_rate=0.02)
net.fit(X_train, y_train)
assert net.score(X_test, y_test) >= 0.84
def test_l2_neural_network_regressor_with_stochastic_optimizer():
X, y = load_boston(return_X_y=True)
X_scaled = StandardScaler().fit_transform(X)
X_train, X_test, y_train, y_test = train_test_split(X_scaled,
y,
train_size=0.75,
random_state=123456)
net = NeuralNetworkRegressor(
(FullyConnected(13, 13, sigmoid), FullyConnected(13, 1, linear)),
loss=mean_squared_error,
optimizer=StochasticGradientDescent,
learning_rate=0.01,
momentum_type='nesterov',
momentum=0.9)
net.fit(X_train, y_train)
assert net.score(X_test, y_test) >= 0.83
def test_perceptron_regressor_with_line_search_optimizer():
# aka linear regression
X, y = load_boston(return_X_y=True)
net = NeuralNetworkRegressor(
(FullyConnected(13, 1, linear, fit_intercept=False), ),
loss=mean_squared_error,
optimizer=Newton).fit(X, y)
assert np.allclose(net.coefs_[0], net.loss.x_star())
def test_perceptron_ridge_regressor_with_line_search_optimizer():
# aka ridge regression
X, y = load_boston(return_X_y=True)
lmbda = 0.1
net = NeuralNetworkRegressor((FullyConnected(
13, 1, linear, coef_reg=L2(lmbda), fit_intercept=False), ),
loss=mean_squared_error,
optimizer=Newton).fit(X, y)
assert np.allclose(
net.coefs_[0].ravel(),
np.linalg.inv(X.T.dot(X) + np.identity(net.loss.ndim) * lmbda).dot(
X.T).dot(y))