def test_partial_fit_classification():
# Test partial_fit on classification.
# `partial_fit` should yield the same results as 'fit' for binary and
# multi-class classification.
for X, y in classification_datasets:
X = X
y = y
mlp = MLPClassifier(solver='sgd',
max_iter=100,
random_state=1,
tol=0,
alpha=1e-5,
learning_rate_init=0.2)
with ignore_warnings(category=ConvergenceWarning):
mlp.fit(X, y)
pred1 = mlp.predict(X)
mlp = MLPClassifier(solver='sgd',
random_state=1,
alpha=1e-5,
learning_rate_init=0.2)
for i in range(100):
mlp.partial_fit(X, y, classes=np.unique(y))
pred2 = mlp.predict(X)
assert_array_equal(pred1, pred2)
assert mlp.score(X, y) > 0.95
def plot_on_dataset(X, y, ax, name):
# for each dataset, plot learning for each learning strategy
print("\nlearning on dataset %s" % name)
ax.set_title(name)
X = MinMaxScaler().fit_transform(X)
mlps = []
if name == "digits":
# digits is larger but converges fairly quickly
max_iter = 15
else:
max_iter = 400
for label, param in zip(labels, params):
print("training: %s" % label)
mlp = MLPClassifier(random_state=0, max_iter=max_iter, **param)
# some parameter combinations will not converge as can be seen on the
# plots so they are ignored here
with warnings.catch_warnings():
warnings.filterwarnings("ignore",
category=ConvergenceWarning,
module="mrex")
mlp.fit(X, y)
mlps.append(mlp)
print("Training set score: %f" % mlp.score(X, y))
print("Training set loss: %f" % mlp.loss_)
for mlp, label, args in zip(mlps, labels, plot_args):
ax.plot(mlp.loss_curve_, label=label, **args)
def test_multilabel_classification():
# Test that multi-label classification works as expected.
# test fit method
X, y = make_multilabel_classification(n_samples=50,
random_state=0,
return_indicator=True)
mlp = MLPClassifier(solver='lbfgs',
hidden_layer_sizes=50,
alpha=1e-5,
max_iter=150,
random_state=0,
activation='logistic',
learning_rate_init=0.2)
mlp.fit(X, y)
assert mlp.score(X, y) > 0.97
# test partial fit method
mlp = MLPClassifier(solver='sgd',
hidden_layer_sizes=50,
max_iter=150,
random_state=0,
activation='logistic',
alpha=1e-5,
learning_rate_init=0.2)
for i in range(100):
mlp.partial_fit(X, y, classes=[0, 1, 2, 3, 4])
assert mlp.score(X, y) > 0.9
# Make sure early stopping still work now that spliting is stratified by
# default (it is disabled for multilabel classification)
mlp = MLPClassifier(early_stopping=True)
mlp.fit(X, y).predict(X)
def test_tolerance():
# Test tolerance.
# It should force the solver to exit the loop when it converges.
X = [[3, 2], [1, 6]]
y = [1, 0]
clf = MLPClassifier(tol=0.5, max_iter=3000, solver='sgd')
clf.fit(X, y)
assert clf.max_iter > clf.n_iter_
def test_adaptive_learning_rate():
X = [[3, 2], [1, 6]]
y = [1, 0]
clf = MLPClassifier(tol=0.5,
max_iter=3000,
solver='sgd',
learning_rate='adaptive')
clf.fit(X, y)
assert clf.max_iter > clf.n_iter_
assert 1e-6 > clf._optimizer.learning_rate
def test_early_stopping_stratified():
# Make sure data splitting for early stopping is stratified
X = [[1, 2], [2, 3], [3, 4], [4, 5]]
y = [0, 0, 0, 1]
mlp = MLPClassifier(early_stopping=True)
with pytest.raises(
ValueError,
match='The least populated class in y has only 1 member'):
mlp.fit(X, y)
def test_sparse_matrices():
# Test that sparse and dense input matrices output the same results.
X = X_digits_binary[:50]
y = y_digits_binary[:50]
X_sparse = csr_matrix(X)
mlp = MLPClassifier(solver='lbfgs', hidden_layer_sizes=15, random_state=1)
mlp.fit(X, y)
pred1 = mlp.predict(X)
mlp.fit(X_sparse, y)
pred2 = mlp.predict(X_sparse)
assert_almost_equal(pred1, pred2)
pred1 = mlp.predict(X)
pred2 = mlp.predict(X_sparse)
assert_array_equal(pred1, pred2)
def test_early_stopping():
X = X_digits_binary[:100]
y = y_digits_binary[:100]
tol = 0.2
clf = MLPClassifier(tol=tol,
max_iter=3000,
solver='sgd',
early_stopping=True)
clf.fit(X, y)
assert clf.max_iter > clf.n_iter_
valid_scores = clf.validation_scores_
best_valid_score = clf.best_validation_score_
assert max(valid_scores) == best_valid_score
assert best_valid_score + tol > valid_scores[-2]
assert best_valid_score + tol > valid_scores[-1]
def test_lbfgs_classification_maxfun(X, y):
# Test lbfgs parameter max_fun.
# It should independently limit the number of iterations for lbfgs.
max_fun = 10
# classification tests
for activation in ACTIVATION_TYPES:
mlp = MLPClassifier(solver='lbfgs',
hidden_layer_sizes=50,
max_iter=150,
max_fun=max_fun,
shuffle=True,
random_state=1,
activation=activation)
with pytest.warns(ConvergenceWarning):
mlp.fit(X, y)
assert max_fun >= mlp.n_iter_
def test_verbose_sgd():
# Test verbose.
X = [[3, 2], [1, 6]]
y = [1, 0]
clf = MLPClassifier(solver='sgd',
max_iter=2,
verbose=10,
hidden_layer_sizes=2)
old_stdout = sys.stdout
sys.stdout = output = StringIO()
with ignore_warnings(category=ConvergenceWarning):
clf.fit(X, y)
clf.partial_fit(X, y)
sys.stdout = old_stdout
assert 'Iteration' in output.getvalue()
def test_predict_proba_multiclass():
# Test that predict_proba works as expected for multi class.
X = X_digits_multi[:10]
y = y_digits_multi[:10]
clf = MLPClassifier(hidden_layer_sizes=5)
with ignore_warnings(category=ConvergenceWarning):
clf.fit(X, y)
y_proba = clf.predict_proba(X)
y_log_proba = clf.predict_log_proba(X)
(n_samples, n_classes) = y.shape[0], np.unique(y).size
proba_max = y_proba.argmax(axis=1)
proba_log_max = y_log_proba.argmax(axis=1)
assert y_proba.shape == (n_samples, n_classes)
assert_array_equal(proba_max, proba_log_max)
assert_array_equal(y_log_proba, np.log(y_proba))
def test_n_iter_no_change():
# test n_iter_no_change using binary data set
# the classifying fitting process is not prone to loss curve fluctuations
X = X_digits_binary[:100]
y = y_digits_binary[:100]
tol = 0.01
max_iter = 3000
# test multiple n_iter_no_change
for n_iter_no_change in [2, 5, 10, 50, 100]:
clf = MLPClassifier(tol=tol,
max_iter=max_iter,
solver='sgd',
n_iter_no_change=n_iter_no_change)
clf.fit(X, y)
# validate n_iter_no_change
assert clf._no_improvement_count == n_iter_no_change + 1
assert max_iter > clf.n_iter_
def test_alpha():
# Test that larger alpha yields weights closer to zero
X = X_digits_binary[:100]
y = y_digits_binary[:100]
alpha_vectors = []
alpha_values = np.arange(2)
absolute_sum = lambda x: np.sum(np.abs(x))
for alpha in alpha_values:
mlp = MLPClassifier(hidden_layer_sizes=10, alpha=alpha, random_state=1)
with ignore_warnings(category=ConvergenceWarning):
mlp.fit(X, y)
alpha_vectors.append(
np.array(
[absolute_sum(mlp.coefs_[0]),
absolute_sum(mlp.coefs_[1])]))
for i in range(len(alpha_values) - 1):
assert (alpha_vectors[i] > alpha_vectors[i + 1]).all()
def test_lbfgs_classification(X, y):
# Test lbfgs on classification.
# It should achieve a score higher than 0.95 for the binary and multi-class
# versions of the digits dataset.
X_train = X[:150]
y_train = y[:150]
X_test = X[150:]
expected_shape_dtype = (X_test.shape[0], y_train.dtype.kind)
for activation in ACTIVATION_TYPES:
mlp = MLPClassifier(solver='lbfgs',
hidden_layer_sizes=50,
max_iter=150,
shuffle=True,
random_state=1,
activation=activation)
mlp.fit(X_train, y_train)
y_predict = mlp.predict(X_test)
assert mlp.score(X_train, y_train) > 0.95
assert ((y_predict.shape[0],
y_predict.dtype.kind) == expected_shape_dtype)
def test_learning_rate_warmstart():
# Tests that warm_start reuse past solutions.
X = [[3, 2], [1, 6], [5, 6], [-2, -4]]
y = [1, 1, 1, 0]
for learning_rate in ["invscaling", "constant"]:
mlp = MLPClassifier(solver='sgd',
hidden_layer_sizes=4,
learning_rate=learning_rate,
max_iter=1,
power_t=0.25,
warm_start=True)
with ignore_warnings(category=ConvergenceWarning):
mlp.fit(X, y)
prev_eta = mlp._optimizer.learning_rate
mlp.fit(X, y)
post_eta = mlp._optimizer.learning_rate
if learning_rate == 'constant':
assert prev_eta == post_eta
elif learning_rate == 'invscaling':
assert (mlp.learning_rate_init /
pow(8 + 1, mlp.power_t) == post_eta)
def test_predict_proba_multilabel():
# Test that predict_proba works as expected for multilabel.
# Multilabel should not use softmax which makes probabilities sum to 1
X, Y = make_multilabel_classification(n_samples=50,
random_state=0,
return_indicator=True)
n_samples, n_classes = Y.shape
clf = MLPClassifier(solver='lbfgs', hidden_layer_sizes=30, random_state=0)
clf.fit(X, Y)
y_proba = clf.predict_proba(X)
assert y_proba.shape == (n_samples, n_classes)
assert_array_equal(y_proba > 0.5, Y)
y_log_proba = clf.predict_log_proba(X)
proba_max = y_proba.argmax(axis=1)
proba_log_max = y_log_proba.argmax(axis=1)
assert (y_proba.sum(1) - 1).dot(y_proba.sum(1) - 1) > 1e-10
assert_array_equal(proba_max, proba_log_max)
assert_array_equal(y_log_proba, np.log(y_proba))
def test_predict_proba_binary():
# Test that predict_proba works as expected for binary class.
X = X_digits_binary[:50]
y = y_digits_binary[:50]
clf = MLPClassifier(hidden_layer_sizes=5,
activation='logistic',
random_state=1)
with ignore_warnings(category=ConvergenceWarning):
clf.fit(X, y)
y_proba = clf.predict_proba(X)
y_log_proba = clf.predict_log_proba(X)
(n_samples, n_classes) = y.shape[0], 2
proba_max = y_proba.argmax(axis=1)
proba_log_max = y_log_proba.argmax(axis=1)
assert y_proba.shape == (n_samples, n_classes)
assert_array_equal(proba_max, proba_log_max)
assert_array_equal(y_log_proba, np.log(y_proba))
assert roc_auc_score(y, y_proba[:, 1]) == 1.0
def test_n_iter_no_change_inf():
# test n_iter_no_change using binary data set
# the fitting process should go to max_iter iterations
X = X_digits_binary[:100]
y = y_digits_binary[:100]
# set a ridiculous tolerance
# this should always trigger _update_no_improvement_count()
tol = 1e9
# fit
n_iter_no_change = np.inf
max_iter = 3000
clf = MLPClassifier(tol=tol,
max_iter=max_iter,
solver='sgd',
n_iter_no_change=n_iter_no_change)
clf.fit(X, y)
# validate n_iter_no_change doesn't cause early stopping
assert clf.n_iter_ == max_iter
# validate _update_no_improvement_count() was always triggered
assert clf._no_improvement_count == clf.n_iter_ - 1
def test_warm_start():
X = X_iris
y = y_iris
y_2classes = np.array([0] * 75 + [1] * 75)
y_3classes = np.array([0] * 40 + [1] * 40 + [2] * 70)
y_3classes_alt = np.array([0] * 50 + [1] * 50 + [3] * 50)
y_4classes = np.array([0] * 37 + [1] * 37 + [2] * 38 + [3] * 38)
y_5classes = np.array([0] * 30 + [1] * 30 + [2] * 30 + [3] * 30 + [4] * 30)
# No error raised
clf = MLPClassifier(hidden_layer_sizes=2, solver='lbfgs',
warm_start=True).fit(X, y)
clf.fit(X, y)
clf.fit(X, y_3classes)
for y_i in (y_2classes, y_3classes_alt, y_4classes, y_5classes):
clf = MLPClassifier(hidden_layer_sizes=2,
solver='lbfgs',
warm_start=True).fit(X, y)
message = ('warm_start can only be used where `y` has the same '
'classes as in the previous call to fit.'
' Previously got [0 1 2], `y` has %s' % np.unique(y_i))
assert_raise_message(ValueError, message, clf.fit, X, y_i)
X, y = fetch_openml('mnist_784', version=1, return_X_y=True)
X = X / 255.
# rescale the data, use the traditional train/test split
X_train, X_test = X[:60000], X[60000:]
y_train, y_test = y[:60000], y[60000:]
mlp = MLPClassifier(hidden_layer_sizes=(50, ),
max_iter=10,
alpha=1e-4,
solver='sgd',
verbose=10,
random_state=1,
learning_rate_init=.1)
mlp.fit(X_train, y_train)
print("Training set score: %f" % mlp.score(X_train, y_train))
print("Test set score: %f" % mlp.score(X_test, y_test))
fig, axes = plt.subplots(4, 4)
# use global min / max to ensure all weights are shown on the same scale
vmin, vmax = mlp.coefs_[0].min(), mlp.coefs_[0].max()
for coef, ax in zip(mlp.coefs_[0].T, axes.ravel()):
ax.matshow(coef.reshape(28, 28),
cmap=plt.cm.gray,
vmin=.5 * vmin,
vmax=.5 * vmax)
ax.set_xticks(())
ax.set_yticks(())
plt.show()
def test_gradient():
# Test gradient.
# This makes sure that the activation functions and their derivatives
# are correct. The numerical and analytical computation of the gradient
# should be close.
for n_labels in [2, 3]:
n_samples = 5
n_features = 10
random_state = np.random.RandomState(seed=42)
X = random_state.rand(n_samples, n_features)
y = 1 + np.mod(np.arange(n_samples) + 1, n_labels)
Y = LabelBinarizer().fit_transform(y)
for activation in ACTIVATION_TYPES:
mlp = MLPClassifier(activation=activation,
hidden_layer_sizes=10,
solver='lbfgs',
alpha=1e-5,
learning_rate_init=0.2,
max_iter=1,
random_state=1)
mlp.fit(X, y)
theta = np.hstack(
[l.ravel() for l in mlp.coefs_ + mlp.intercepts_])
layer_units = ([X.shape[1]] + [mlp.hidden_layer_sizes] +
[mlp.n_outputs_])
activations = []
deltas = []
coef_grads = []
intercept_grads = []
activations.append(X)
for i in range(mlp.n_layers_ - 1):
activations.append(np.empty((X.shape[0], layer_units[i + 1])))
deltas.append(np.empty((X.shape[0], layer_units[i + 1])))
fan_in = layer_units[i]
fan_out = layer_units[i + 1]
coef_grads.append(np.empty((fan_in, fan_out)))
intercept_grads.append(np.empty(fan_out))
# analytically compute the gradients
def loss_grad_fun(t):
return mlp._loss_grad_lbfgs(t, X, Y, activations, deltas,
coef_grads, intercept_grads)
[value, grad] = loss_grad_fun(theta)
numgrad = np.zeros(np.size(theta))
n = np.size(theta, 0)
E = np.eye(n)
epsilon = 1e-5
# numerically compute the gradients
for i in range(n):
dtheta = E[:, i] * epsilon
numgrad[i] = ((loss_grad_fun(theta + dtheta)[0] -
loss_grad_fun(theta - dtheta)[0]) /
(epsilon * 2.0))
assert_almost_equal(numgrad, grad)