def test_lifecycle(self,
loss,
subnetwork_generator,
max_iteration_steps,
want_loss,
want_metrics=None,
want_metrics_names=None,
metrics=None,
logits_dimension=1,
ensemblers=None,
ensemble_strategies=None,
evaluator=None,
adanet_loss_decay=0.9,
dataset=None,
epochs=None,
steps_per_epoch=None):
keras_model = model.Model(subnetwork_generator=subnetwork_generator,
max_iteration_steps=max_iteration_steps,
logits_dimension=logits_dimension,
ensemblers=ensemblers,
ensemble_strategies=ensemble_strategies,
evaluator=evaluator,
adanet_loss_decay=adanet_loss_decay,
filepath=self.test_subdirectory)
keras_model.compile(loss=loss, metrics=metrics)
if want_metrics_names is None:
want_metrics_names = ["loss"]
# Make sure we have access to metrics_names immediately after compilation.
self.assertEqual(want_metrics_names, keras_model.metrics_names)
if dataset is None:
dataset = lambda: tf.data.Dataset.from_tensors( # pylint: disable=g-long-lambda
({
"x": XOR_FEATURES
}, XOR_LABELS)).repeat()
keras_model.fit(dataset,
epochs=epochs,
steps_per_epoch=steps_per_epoch)
eval_results = keras_model.evaluate(dataset, steps=3)
self.assertAlmostEqual(want_loss, eval_results[0], places=3)
if metrics:
self.assertAllClose(want_metrics, eval_results[1:], 1e-3, 1e-3)
# TODO: Predict not currently working for BinaryClassHead and
# MultiClassHead.
if loss == "mse":
prediction_data = lambda: tf.data.Dataset.from_tensors(
({ # pylint: disable=g-long-lambda
"x": XOR_FEATURES
}))
predictions = keras_model.predict(prediction_data)
self.assertLen(predictions, 4)
def test_loss_exceptions(self):
"""Check that ValueError is raised when from_logits=False for loss."""
keras_model = model.Model(subnetwork_generator=SimpleGenerator(
[_DNNBuilder("dnn")]),
max_iteration_steps=1)
loss = tf.keras.losses.BinaryCrossentropy(from_logits=False)
with self.assertRaises(ValueError):
keras_model.compile(loss=loss)
def test_compile_exceptions(self):
keras_model = model.Model(subnetwork_generator=SimpleGenerator(
[_DNNBuilder("dnn")]),
max_iteration_steps=1)
train_data = tf.data.Dataset.from_tensors(([[1., 1.]], [[1.]]))
predict_data = tf.data.Dataset.from_tensors(([[1., 1.]]))
with self.assertRaises(RuntimeError):
keras_model.fit(train_data)
with self.assertRaises(RuntimeError):
keras_model.evaluate(train_data)
with self.assertRaises(RuntimeError):
keras_model.predict(predict_data)
def test_lifecycle(self,
subnetwork_generator,
max_iteration_steps,
want_loss,
ensemblers=None,
ensemble_strategies=None,
evaluator=None,
adanet_loss_decay=0.9,
dataset=None,
epochs=None,
steps_per_epoch=None):
keras_model = model.Model(subnetwork_generator=subnetwork_generator,
max_iteration_steps=max_iteration_steps,
ensemblers=ensemblers,
ensemble_strategies=ensemble_strategies,
evaluator=evaluator,
adanet_loss_decay=adanet_loss_decay,
filepath=self.test_subdirectory)
keras_model.compile(loss="mse")
# Make sure we have access to metrics_names immediately after compilation.
self.assertEqual(["loss"], keras_model.metrics_names)
if dataset is None:
dataset = tf.data.Dataset.from_tensors(
(XOR_FEATURES, XOR_LABELS)).repeat()
keras_model.fit(dataset,
epochs=epochs,
steps_per_epoch=steps_per_epoch)
eval_results = keras_model.evaluate(dataset, steps=3)
self.assertAlmostEqual(want_loss, eval_results["loss"], places=3)
prediction_data = tf.data.Dataset.from_tensors((XOR_FEATURES))
predictions = keras_model.predict(prediction_data)
self.assertLen(predictions, 4)