def test_sequential_model_saving_without_input_shape(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with self.cached_session():
model = keras.models.Sequential()
model.add(keras.layers.Dense(2))
model.add(keras.layers.RepeatVector(3))
model.add(keras.layers.TimeDistributed(keras.layers.Dense(3)))
model.compile(
loss=keras.losses.MSE,
optimizer='rmsprop',
metrics=[
keras.metrics.categorical_accuracy,
keras.metrics.CategoricalAccuracy(name='cat_acc')
],
weighted_metrics=[
keras.metrics.categorical_accuracy,
keras.metrics.CategoricalAccuracy(name='cat_acc2')
],
sample_weight_mode='temporal')
x = np.random.random((1, 3))
y = np.random.random((1, 3, 3))
model.train_on_batch(x, y)
out = model.predict(x)
model.save(saved_model_dir, save_format=save_format)
new_model = keras.models.load_model(saved_model_dir)
self._assert_same_weights_and_metrics(model, new_model)
out2 = new_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
def _assert_same_weights_and_metrics(self, model, loaded_model):
"""Checks that the loaded weights and metrics are the same as the original.
Args:
model: original model
loaded_model: loaded model
"""
self.assertAllClose(model.weights, loaded_model.weights)
if loaded_model.optimizer:
if test_utils.get_save_format() == 'tf':
# TODO(b/153110928): Keras TF format doesn't restore optimizer weights
# currently.
return
self.assertAllClose(model.optimizer.weights,
loaded_model.optimizer.weights)
# In V1/Graph mode, the model isn't built, so the metrics are not loaded
# immediately (requires model to be called on some data before building
# metrics).
check_metrics = tf.__internal__.tf2.enabled() and tf.executing_eagerly()
if check_metrics:
self.assertAllEqual([m.name for m in model.metrics],
[m.name for m in loaded_model.metrics])
def test_functional_model_with_getitem_op_layer(self):
inp = keras.Input(shape=(8))
out = inp[:]
model = keras.Model(
inputs=[inp],
outputs=out)
batch_size = 7
x = tf.stack([
tf.range(8) for _ in range(batch_size)])
args = [x]
expected = x[:]
self.assertAllEqual(model(args), expected)
self.assertAllEqual(model.predict(args, batch_size=batch_size), expected)
# Make sure it can be successfully saved and loaded.
save_format = test_utils.get_save_format()
saved_model_dir = self._save_model_dir()
keras.models.save_model(model, saved_model_dir, save_format=save_format)
loaded_model = keras.models.load_model(saved_model_dir)
self.assertAllEqual(loaded_model(args), expected)
self.assertAllEqual(loaded_model.predict(args, batch_size=batch_size),
expected)
def test_basic_saving_and_loading(self, model_fn):
save_format = test_utils.get_save_format()
custom_objects = self.get_custom_objects()
if "subclassed_in_functional" in model_fn.__name__:
subclass_custom_objects = {
"MySubclassModel": model_architectures.MySubclassModel,
}
custom_objects.update(subclass_custom_objects)
elif "subclassed" in model_fn.__name__ and save_format == "h5":
self.skipTest(
"Saving the model to HDF5 format requires the model to be "
"a Functional model or a Sequential model."
)
saved_model_dir = self._save_model_dir()
model_data = model_fn()
model = model_data.model
x_test, y_test = self.get_test_data(
model_data.input_shape, model_data.target_shape
)
model.compile("rmsprop", "mse")
model.train_on_batch(x_test, y_test)
# Save model.
out1 = model.predict(x_test)
keras.models.save_model(model, saved_model_dir, save_format=save_format)
# Load model.
loaded_model = keras.models.load_model(
saved_model_dir, custom_objects=custom_objects
)
out2 = loaded_model.predict(x_test)
self.assertAllClose(out1, out2, atol=1e-05)
def test_weight_loading(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with self.cached_session():
a = keras.layers.Input(shape=(2, ))
x = keras.layers.Dense(3)(a)
b = keras.layers.Dense(1)(x)
model = keras.models.Model(a, b)
x = np.random.random((3, 2))
ref_y = model.predict(x)
weights = model.get_weights()
model.set_weights(weights)
y = model.predict(x)
self.assertAllClose(ref_y, y)
with self.assertRaises(ValueError):
model.set_weights(weights[1:])
with self.assertRaises(ValueError):
model.set_weights(weights[::-1])
model.save_weights(saved_model_dir, save_format=save_format)
model.load_weights(saved_model_dir)
y = model.predict(x)
self.assertAllClose(ref_y, y)
def test_sequential_model_saving_2(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with tf.Graph().as_default(), self.cached_session():
# test with custom optimizer, loss
class CustomOp(optimizer_v1.RMSprop):
pass
def custom_loss(y_true, y_pred):
return keras.losses.mse(y_true, y_pred)
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3,)))
model.add(keras.layers.Dense(3))
model.compile(loss=custom_loss, optimizer=CustomOp(), metrics=['acc'])
x = np.random.random((1, 3))
y = np.random.random((1, 3))
model.train_on_batch(x, y)
out = model.predict(x)
keras.models.save_model(model, saved_model_dir, save_format=save_format)
new_model = keras.models.load_model(
saved_model_dir,
custom_objects={'CustomOp': CustomOp,
'custom_loss': custom_loss})
self._assert_same_weights_and_metrics(model, new_model)
out2 = new_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
def test_nested_model_weight_loading(self):
save_format = test_utils.get_save_format()
saved_model_dir = self._save_model_dir()
batch_size = 5
shape = (None, None, 3)
with self.cached_session():
def gen_model():
def seq_model():
model = keras.models.Sequential([
keras.layers.Conv2D(3, 1, input_shape=shape),
keras.layers.BatchNormalization()])
return model
x = inner_inputs = keras.layers.Input((None, None, 3))
x = seq_model()(x)
x = seq_model()(x)
inner_model = keras.models.Model(inner_inputs, x)
inputs = keras.layers.Input(shape)
return keras.models.Model(inputs, inner_model(inputs))
model = gen_model()
x = np.random.random((batch_size, 1, 1, 3))
ref_y = model.predict(x)
model.save_weights(saved_model_dir, save_format=save_format)
model = gen_model()
model.load_weights(saved_model_dir)
y = model.predict(x)
self.assertAllClose(y, ref_y)
def test_warning_when_saving_invalid_custom_mask_layer(self):
class MyMasking(keras.layers.Layer):
def call(self, inputs):
return inputs
def compute_mask(self, inputs, mask=None):
mask = tf.not_equal(inputs, 0)
return mask
class MyLayer(keras.layers.Layer):
def call(self, inputs, mask=None):
return tf.identity(inputs)
samples = np.random.random((2, 2))
model = keras.Sequential([MyMasking(), MyLayer()])
model.predict(samples)
with warnings.catch_warnings(record=True) as w:
model.save(self._save_model_dir(), test_utils.get_save_format())
self.assertIn(generic_utils.CustomMaskWarning,
{warning.category for warning in w})
# Test that setting up a custom mask correctly does not issue a warning.
class MyCorrectMasking(keras.layers.Layer):
def call(self, inputs):
return inputs
def compute_mask(self, inputs, mask=None):
mask = tf.not_equal(inputs, 0)
return mask
# This get_config doesn't actually do anything because our mask is
# static and doesn't need any external information to work. We do need a
# dummy get_config method to prevent the warning from appearing, however.
def get_config(self, *args, **kwargs):
return {}
model = keras.Sequential([MyCorrectMasking(), MyLayer()])
model.predict(samples)
with warnings.catch_warnings(record=True) as w:
model.save(self._save_model_dir(), test_utils.get_save_format())
self.assertNotIn(generic_utils.CustomMaskWarning,
{warning.category for warning in w})
def test_saving_without_compilation(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3,)))
model.add(keras.layers.Dense(3))
model.compile(loss='mse', optimizer='sgd', metrics=['acc'])
keras.models.save_model(model, saved_model_dir, save_format=save_format)
model = keras.models.load_model(saved_model_dir)
def test_saving_right_after_compilation(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with self.cached_session():
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3,)))
model.add(keras.layers.Dense(3))
model.compile(loss='mse', optimizer='sgd', metrics=['acc'])
if not tf.compat.v1.executing_eagerly_outside_functions():
model._make_train_function()
keras.models.save_model(model, saved_model_dir, save_format=save_format)
model = keras.models.load_model(saved_model_dir)
def test_saving_with_tf_optimizer(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3,)))
model.add(keras.layers.Dense(3))
model.compile(loss='mse',
optimizer=tf.compat.v1.train.AdadeltaOptimizer(0.1),
metrics=['acc'])
keras.models.save_model(model, saved_model_dir, save_format=save_format)
model = keras.models.load_model(saved_model_dir)
def test_save_and_load(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
save_kwargs = test_utils.get_save_kwargs()
if ((save_format == 'h5' or not save_kwargs.get('save_traces', True)) and
test_utils.get_model_type() == 'subclass'):
# HDF5 format currently does not allow saving subclassed models.
# When saving with `save_traces=False`, the subclassed model must have a
# get_config/from_config, which the autogenerated model does not have.
return
with self.cached_session():
model = test_utils.get_model_from_layers(
[keras.layers.Dense(2),
keras.layers.RepeatVector(3),
keras.layers.TimeDistributed(keras.layers.Dense(3))],
input_shape=(3,))
model.compile(
loss=keras.losses.MSE,
optimizer=keras.optimizers.optimizer_v2.rmsprop.RMSprop(lr=0.0001),
metrics=[
keras.metrics.categorical_accuracy,
keras.metrics.CategoricalCrossentropy(
name='cce', label_smoothing=tf.constant(0.2)),
],
weighted_metrics=[
keras.metrics.categorical_crossentropy,
keras.metrics.CategoricalCrossentropy(
name='cce', label_smoothing=tf.constant(0.2)),
],
sample_weight_mode='temporal')
x = np.random.random((1, 3))
y = np.random.random((1, 3, 3))
model.train_on_batch(x, y)
out = model.predict(x)
keras.models.save_model(
model, saved_model_dir, save_format=save_format,
**save_kwargs)
loaded_model = keras.models.load_model(saved_model_dir)
self._assert_same_weights_and_metrics(model, loaded_model)
out2 = loaded_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
eval_out = model.evaluate(x, y)
eval_out2 = loaded_model.evaluate(x, y)
self.assertArrayNear(eval_out, eval_out2, 0.001)
def test_primitive_attrs_contain_no_extraneous_strings(self):
if h5py is None:
self.skipTest('h5py required to run this test')
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
model = keras.models.Sequential()
model.add(keras.layers.Dense(1, input_shape=[2]))
model.save(saved_model_dir, save_format=save_format)
if save_format in ['tf', 'tensorflow']:
return
h5file = h5py.File(saved_model_dir, 'r')
self.assertRegex(h5file.attrs['keras_version'], r'^[\d]+\.[\d]+\.[\S]+$')
def test_saving_constant_initializer_with_numpy(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
model = keras.models.Sequential()
model.add(
keras.layers.Dense(
2,
input_shape=(3,),
kernel_initializer=keras.initializers.Constant(np.ones((3, 2)))))
model.add(keras.layers.Dense(3))
model.compile(loss='mse', optimizer='sgd', metrics=['acc'])
keras.models.save_model(model, saved_model_dir, save_format=save_format)
model = keras.models.load_model(saved_model_dir)
def test_model_saving_to_pre_created_h5py_file(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with tf.Graph().as_default(), self.cached_session():
inputs = keras.Input(shape=(3,))
x = keras.layers.Dense(2)(inputs)
outputs = keras.layers.Dense(3)(x)
model = keras.Model(inputs, outputs)
model.compile(
loss=keras.losses.MSE,
optimizer=optimizer_v1.Adam(),
metrics=[
keras.metrics.categorical_accuracy,
keras.metrics.CategoricalAccuracy()
])
x = np.random.random((1, 3))
y = np.random.random((1, 3))
model.train_on_batch(x, y)
out = model.predict(x)
keras.models.save_model(model, saved_model_dir, save_format=save_format)
loaded_model = keras.models.load_model(saved_model_dir)
out1 = loaded_model.predict(x)
self.assertAllClose(out, out1, atol=1e-05)
if save_format in ['tf', 'tensorflow']:
return
# Test h5 format specifically
fd, fname = tempfile.mkstemp('.h5')
with h5py.File(fname, mode='r+') as h5file:
keras.models.save_model(model, h5file)
loaded_model = keras.models.load_model(h5file)
out2 = loaded_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
# Test non-default options in h5
with h5py.File(
'_', driver='core', mode='w', backing_store=False) as h5file:
keras.models.save_model(model, h5file)
loaded_model = keras.models.load_model(h5file)
out2 = loaded_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
# Cleanup
os.close(fd)
os.remove(fname)
def test_save_uncompiled_model_with_optimizer(self):
with self.cached_session() as session:
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
model = keras.models.Sequential([keras.layers.Dense(1, input_shape=(3,))])
# Set the model's optimizer but don't compile. This can happen if the
# model is trained with a custom training loop.
model.optimizer = keras.optimizers.optimizer_v2.rmsprop.RMSprop(lr=0.0001)
if not tf.executing_eagerly():
session.run([v.initializer for v in model.variables])
model.save(saved_model_dir, save_format=save_format)
if save_format in ['tf', 'tensorflow']:
loaded = keras.models.load_model(saved_model_dir)
self.assertIsInstance(
loaded.optimizer,
keras.optimizers.optimizer_v2.optimizer_v2.OptimizerV2)
def test_shared_objects_wrapper(self):
"""Tests that shared layers wrapped with `Wrapper` restore correctly."""
input_ = keras.Input(shape=(1,))
unwrapped = keras.layers.Layer(name='unwrapped')
wrapped = keras.layers.Wrapper(unwrapped, name='wrapped')
model = keras.Model(inputs=input_,
outputs=[unwrapped(input_), wrapped(input_)])
# Test recreating directly from config
config = model.get_config()
loaded = keras.Model.from_config(config)
self.assertIs(loaded.layers[1], loaded.layers[2].layer)
# Test saving and loading to disk
save_format = test_utils.get_save_format()
saved_model_dir = self._save_model_dir()
keras.models.save_model(model, saved_model_dir, save_format=save_format)
loaded = keras.models.load_model(saved_model_dir)
self.assertIs(loaded.layers[1], loaded.layers[2].layer)
def test_multi_output_metrics_name_stay_same(self, fit):
"""Tests that metric names don't change with each save/load cycle.
e.g. "head_0_accuracy" should not become "head_0_head_0_accuracy" after
saving and loading a model.
Arguments:
fit: Whether the model should be fit before saving.
"""
# This doesn't work at all, so we can't check whether metric names are
# correct.
if not tf.executing_eagerly() and not fit:
self.skipTest('b/181767784')
input_ = keras.Input((4,))
model = keras.Model(
input_,
[keras.layers.Softmax(name='head_0')(keras.layers.Dense(3)(input_)),
keras.layers.Softmax(name='head_1')(keras.layers.Dense(5)(input_))])
metric = keras.metrics.BinaryAccuracy()
model.compile(optimizer='rmsprop',
loss='mse',
metrics={'head_0': [metric, 'accuracy']})
x = np.random.rand(2, 4)
y = {'head_0': np.random.randint(2, size=(2, 3)),
'head_1': np.random.randint(2, size=(2, 5))}
# Make sure metrix prefixing works the same regardless of whether the user
# has fit the model before saving.
if fit:
model.fit(x, y, verbose=0)
# Save and reload.
save_format = test_utils.get_save_format()
saved_model_dir = self._save_model_dir()
keras.models.save_model(model, saved_model_dir, save_format=save_format)
loaded = keras.models.load_model(saved_model_dir)
# Make sure the metrics names from the model before saving match the loaded
# model.
self.assertSequenceEqual(model.metrics_names, loaded.metrics_names)
def test_sequential_model_saving_without_compile(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with self.cached_session():
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3,)))
model.add(keras.layers.RepeatVector(3))
model.add(keras.layers.TimeDistributed(keras.layers.Dense(3)))
x = np.random.random((1, 3))
out = model.predict(x)
# Save the model without any compilation or training.
keras.models.save_model(model, saved_model_dir, save_format=save_format)
new_model = keras.models.load_model(saved_model_dir)
self._assert_same_weights_and_metrics(model, new_model)
out2 = new_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
def test_model_saving_to_new_dir_path(self):
saved_model_dir = os.path.join(self._save_model_dir(), 'newdir',
'saved_model')
save_format = test_utils.get_save_format()
with self.cached_session():
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3,)))
model.add(keras.layers.RepeatVector(3))
model.add(keras.layers.TimeDistributed(keras.layers.Dense(3)))
x = np.random.random((1, 3))
out = model.predict(x)
keras.models.save_model(model, saved_model_dir, save_format=save_format)
new_model = keras.models.load_model(saved_model_dir)
self._assert_same_weights_and_metrics(model, new_model)
out2 = new_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
def test_saving_lambda_numpy_array_arguments(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
if h5py is None:
self.skipTest('h5py required to run this test')
mean = np.random.random((4, 2, 3))
std = np.abs(np.random.random((4, 2, 3))) + 1e-5
inputs = keras.layers.Input(shape=(4, 2, 3))
output = keras.layers.Lambda(lambda image, mu, std: (image - mu) / std,
arguments={'mu': mean, 'std': std})(inputs)
model = keras.models.Model(inputs, output)
model.compile(loss='mse', optimizer='sgd', metrics=['acc'])
keras.models.save_model(model, saved_model_dir, save_format=save_format)
model = keras.models.load_model(saved_model_dir)
self.assertAllClose(mean, model.layers[1].arguments['mu'])
self.assertAllClose(std, model.layers[1].arguments['std'])
def test_saving_model_with_long_weights_names(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with self.cached_session():
x = keras.Input(shape=(2,), name='nested_model_input')
f = x
for i in range(4):
f = keras.layers.Dense(2, name='nested_model_dense_%d' % (i,))(f)
# This layer name will make the `weights_name`
# HDF5 attribute blow out of proportion.
f = keras.layers.Dense(2, name='nested_model_output' + ('x' * (2**14)))(f)
nested_model = keras.Model(inputs=[x], outputs=[f], name='nested_model')
x = keras.Input(shape=(2,), name='outer_model_input')
f = nested_model(x)
f = keras.layers.Dense(2, name='outer_model_output')(f)
model = keras.Model(inputs=[x], outputs=[f])
model.compile(loss='mse', optimizer='adam', metrics=['acc'])
x = np.random.random((1, 2))
y = np.random.random((1, 2))
model.train_on_batch(x, y)
out = model.predict(x)
keras.models.save_model(model, saved_model_dir, save_format=save_format)
model = keras.models.load_model(saved_model_dir)
if save_format in ['h5', 'hdf5', 'keras']:
# Check that the HDF5 files contains chunked array
# of weight names.
with h5py.File(saved_model_dir, 'r') as h5file:
num_weight_arrays = len(
[attr for attr in h5file['model_weights']['nested_model'].attrs
if attr.startswith('weight_names')])
# The chunking of layer names array should have happened.
self.assertGreater(num_weight_arrays, 0)
out2 = model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
def test_load_weights_from_saved_model(self):
save_path = self._save_model_dir()
save_format = test_utils.get_save_format()
if save_format == "h5" and test_utils.get_model_type() == "subclass":
# TODO(b/173646281): HDF5 format currently does not allow saving
# subclassed models.
return
with self.cached_session():
model = test_utils.get_small_mlp(1, 4, input_dim=3)
data = np.random.random((1, 3))
labels = np.random.random((1, 4))
model.compile(loss="mse", optimizer="rmsprop")
model.fit(data, labels)
model.save(save_path, save_format=save_format)
new_model = test_utils.get_small_mlp(1, 4, input_dim=3)
if test_utils.get_model_type() == "subclass":
# Call on test data to build the model.
new_model.predict(data)
new_model.load_weights(save_path)
self.assertAllClose(model.weights, new_model.weights)
def test_custom_functional_registered(self):
def _get_cls_definition():
class CustomModel(keras.Model):
def c(self):
return 'c'
return CustomModel
cls = _get_cls_definition()
self.assertEqual(cls.__bases__[0], keras.Model)
with self.cached_session() as sess:
input_ = keras.layers.Input(shape=(1,))
output = keras.layers.Dense(1)(input_)
model = cls(input_, output)
# `cls` now inherits from `Functional` class.
self.assertEqual(cls.__bases__[0], functional.Functional)
if not tf.executing_eagerly():
sess.run([v.initializer for v in model.variables])
save_format = test_utils.get_save_format()
saved_model_dir = self._save_model_dir()
keras.models.save_model(model, saved_model_dir, save_format=save_format)
loaded_model = keras.models.load_model(
saved_model_dir, custom_objects={'CustomModel': cls})
self.assertIsInstance(loaded_model, cls)
# Check with "new" `CustomModel` class definition.
new_cls = _get_cls_definition()
# The new `CustomModel` class is *not* derived from `Functional`.
self.assertEqual(new_cls.__bases__[0], keras.Model)
reloaded_model = keras.models.load_model(
saved_model_dir, custom_objects={'CustomModel': new_cls})
self.assertIsInstance(reloaded_model, new_cls)
def test_functional_model_with_custom_loss_and_metric(self):
def _make_model():
inputs = keras.Input(shape=(4,))
x = keras.layers.Dense(8, activation='relu')(inputs)
outputs = keras.layers.Dense(3, activation='softmax')(x)
model = keras.Model(inputs=inputs, outputs=outputs)
custom_loss = keras.layers.Lambda(lambda x: keras.backend.sum(x * x))(x)
model.add_loss(custom_loss)
model.add_metric(custom_loss, aggregation='mean', name='custom_loss')
return model
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with self.cached_session():
model = _make_model()
model.compile(
loss=keras.losses.SparseCategoricalCrossentropy(),
optimizer=optimizers.gradient_descent_v2.SGD(),
metrics=[keras.metrics.SparseCategoricalCrossentropy()])
x = np.random.normal(size=(32, 4))
y = np.random.randint(0, 3, size=32)
model.train_on_batch(x, y)
evaluation_results = model.evaluate(x, y)
# Save and reload model.
model.save(saved_model_dir, save_format=save_format)
del model # Prevent misuse.
loaded_model = keras.models.load_model(saved_model_dir)
loaded_model_eval_results = loaded_model.evaluate(x, y)
# Assert all evaluation results are the same.
self.assertAllClose(evaluation_results, loaded_model_eval_results, 1e-9)
# Check correctness of the loss calculation.
self.assertAllGreater(evaluation_results, 0.)
evaluation_results = dict(
zip(loaded_model.metrics_names, evaluation_results))
self.assertNear(
evaluation_results['sparse_categorical_crossentropy'] +
evaluation_results['custom_loss'], evaluation_results['loss'], 1e-6)
def test_saving_model_with_long_layer_names(self):
saved_model_dir = self._save_model_dir()
save_format = test_utils.get_save_format()
with self.cached_session():
# This layer name will make the `layers_name` HDF5 attribute blow
# out of proportion. Note that it fits into the internal HDF5
# attribute memory limit on its own but because h5py converts
# the list of layer names into numpy array, which uses the same
# amount of memory for every item, it increases the memory
# requirements substantially.
x = keras.Input(shape=(2,), name='input_' + ('x' * (2**15)))
f = x
for i in range(4):
f = keras.layers.Dense(2, name='dense_%d' % (i,))(f)
model = keras.Model(inputs=[x], outputs=[f])
model.compile(
'adam', loss=keras.losses.MeanSquaredError(), metrics=['acc'])
x = np.random.random((1, 2))
y = np.random.random((1, 2))
model.train_on_batch(x, y)
out = model.predict(x)
keras.models.save_model(model, saved_model_dir, save_format=save_format)
model = keras.models.load_model(saved_model_dir)
if save_format in ['tf', 'tensorflow']:
return
# Check that the HDF5 files contains chunked array
# of layer names.
with h5py.File(saved_model_dir, 'r') as h5file:
num_names_arrays = len([attr for attr in h5file['model_weights'].attrs
if attr.startswith('layer_names')])
# The chunking of layer names array should have happened.
self.assertGreater(num_names_arrays, 0)
out2 = model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
def test_shared_objects(self):
class OuterLayer(keras.layers.Layer):
def __init__(self, inner_layer):
super(OuterLayer, self).__init__()
self.inner_layer = inner_layer
def call(self, inputs):
return self.inner_layer(inputs)
def get_config(self):
return {
'inner_layer': generic_utils.serialize_keras_object(
self.inner_layer)
}
@classmethod
def from_config(cls, config):
return cls(generic_utils.deserialize_keras_object(
config['inner_layer']))
class InnerLayer(keras.layers.Layer):
def __init__(self):
super(InnerLayer, self).__init__()
self.v = self.add_weight(name='v', shape=[], dtype=tf.float32)
def call(self, inputs):
return self.v + inputs
@classmethod
def from_config(cls, config):
return cls()
# Create a model with 2 output layers that share the same inner layer.
inner_layer = InnerLayer()
outer_layer_1 = OuterLayer(inner_layer)
outer_layer_2 = OuterLayer(inner_layer)
input_ = keras.Input(shape=(1,))
model = keras.Model(
inputs=input_, outputs=[outer_layer_1(input_), outer_layer_2(input_)])
# Changes to the shared layer should affect both outputs.
model.layers[1].inner_layer.v.assign(5)
self.assertAllEqual(model(1), [6.0, 6.0])
model.layers[1].inner_layer.v.assign(3)
self.assertAllEqual(model(1), [4.0, 4.0])
# After loading, changes to the shared layer should still affect both
# outputs.
def _do_assertions(loaded):
loaded.layers[1].inner_layer.v.assign(5)
self.assertAllEqual(loaded(1), [6.0, 6.0])
loaded.layers[1].inner_layer.v.assign(3)
self.assertAllEqual(loaded(1), [4.0, 4.0])
loaded.layers[2].inner_layer.v.assign(5)
self.assertAllEqual(loaded(1), [6.0, 6.0])
loaded.layers[2].inner_layer.v.assign(3)
self.assertAllEqual(loaded(1), [4.0, 4.0])
# We'd like to make sure we only attach shared object IDs when strictly
# necessary, so we'll recursively traverse the generated config to count
# whether we have the exact number we expect.
def _get_all_keys_recursive(dict_or_iterable):
if isinstance(dict_or_iterable, dict):
for key in dict_or_iterable.keys():
yield key
for key in _get_all_keys_recursive(dict_or_iterable.values()):
yield key
elif isinstance(dict_or_iterable, str):
return
else:
try:
for item in dict_or_iterable:
for key in _get_all_keys_recursive(item):
yield key
# Not an iterable or dictionary
except TypeError:
return
with generic_utils.CustomObjectScope({
'OuterLayer': OuterLayer, 'InnerLayer': InnerLayer}):
# Test saving and loading to disk
save_format = test_utils.get_save_format()
saved_model_dir = self._save_model_dir()
keras.models.save_model(model, saved_model_dir, save_format=save_format)
loaded = keras.models.load_model(saved_model_dir)
_do_assertions(loaded)
# Test recreating directly from config
config = model.get_config()
key_count = collections.Counter(_get_all_keys_recursive(config))
self.assertEqual(key_count[generic_utils.SHARED_OBJECT_KEY], 2)
loaded = keras.Model.from_config(config)
_do_assertions(loaded)
