def test_errors_loss_output_activation():
"""Make sure cross-entropy loss with activations not equal to
`tensorflow.nn.sigmoid` or `tensorflow.nn.softmax` fails."""
# This data will not actually be fit.
# I am just using it to call the `fit` method.
X = np.ones((1000, 4))
# There are two code paths for this test. One for all features
# using the default loss and one for a mix of losses.
# All features use the default loss.
ae = Autoencoder(loss='cross-entropy', output_activation=tf.exp)
with pytest.raises(ValueError) as e:
ae.fit(X)
assert "'cross-entropy' loss!" in str(e), (
"Wrong error raised for testing 'cross-entropy' loss with "
"output activation that is not allowed for all features!")
# Not all features use the default loss.
ae = Autoencoder(loss='cross-entropy',
output_activation=tf.exp,
sigmoid_indices=[0])
with pytest.raises(ValueError) as e:
ae.fit(X)
assert "'cross-entropy' loss!" in str(e), (
"Wrong error raised for testing 'cross-entropy' loss with "
"output activation that is not allowed for a subset of features!")
def test_refitting():
"""Make sure that refitting resets internals."""
X = iris.data # Use the iris features.
X = MinMaxScaler().fit_transform(X)
# Use digitize to make a binary features.
for i in range(X.shape[1]):
bins = [0.0, np.median(X[:, i]), 1.1]
X[:, i] = np.digitize(X[:, i], bins) - 1.0
ae = Autoencoder(hidden_units=(1,),
n_epochs=1000,
random_state=4556,
learning_rate=1e-2,
keep_prob=1.0,
loss='cross-entropy',
output_activation=tf.nn.sigmoid)
ae.fit(X)
assert ae.input_layer_size_ == 4, ("Input layer is the wrong size for "
"the Autoencoder!")
X_small = X[:, 0:-1]
assert X_small.shape != X.shape, "Test data for refitting does not work!"
ae.fit(X_small)
assert ae.input_layer_size_ == 3, ("Input layer is the wrong size for "
"the Autoencoder!")
def test_errors_overlapping_sigmoid_softmax_indixes():
"""Make overlapping sigmoid and softmax indices raises an error."""
# This data will not actually be fit.
# I am just using it to call the `fit` method.
X = np.ones((1000, 4))
ae = Autoencoder(loss='blah',
sigmoid_indices=[0],
softmax_indices=[[0, 2]])
with pytest.raises(ValueError) as e:
ae.fit(X)
assert "Sigmoid indices and softmax indices" in str(e), (
"Wrong error raised for overlapping sigmoid and softmax indices")
def test_monitor_ae():
"""Test the monitor keyword."""
# Use the iris features.
X = iris.data
X = MinMaxScaler().fit_transform(X)
ae = Autoencoder(hidden_units=(3, 2,),
n_epochs=7500,
random_state=4556,
learning_rate=DEFAULT_LEARNING_RATE,
keep_prob=1.0,
hidden_activation=tf.nn.sigmoid,
encoding_activation=tf.nn.sigmoid,
output_activation=tf.nn.sigmoid)
def _monitor(epoch, est, stats):
assert epoch <= 1000, "The autoencoder has been running too long!"
if stats['loss'] < 0.2:
assert epoch > 10, "The autoencoder returned too soon!"
return True
else:
return False
ae.fit(X, monitor=_monitor)
def test_errors_unallowed_loss():
"""Make sure unallowed losses cause an error."""
# This data will not actually be fit.
# I am just using it to call the `fit` method.
X = np.ones((1000, 4))
# There are two code paths for this test. One for all features
# using the default loss and one for a mix of losses.
# All features use the default loss.
ae = Autoencoder(loss='blah')
with pytest.raises(ValueError) as e:
ae.fit(X)
assert "Loss 'blah'" in str(e), (
"Wrong error raised for testing unallowed losses!")
# Not all features use the default loss.
ae = Autoencoder(loss='blah', sigmoid_indices=[0])
with pytest.raises(ValueError) as e:
ae.fit(X)
assert "Loss 'blah'" in str(e), (
"Wrong error raised for testing unallowed losses!")