def testKerasStyleAddWeight(self):
keras_layer = keras_base_layer.Layer(name="keras_layer")
with backend.name_scope("foo"):
keras_variable = keras_layer.add_weight(
"my_var", [2, 2], initializer=tf.compat.v1.zeros_initializer()
)
self.assertEqual(keras_variable.name, "foo/my_var:0")
with backend.name_scope("baz"):
old_style_layer = base_tf_layers.Layer(name="my_layer")
# Test basic variable creation.
variable = old_style_layer.add_weight(
"my_var", [2, 2], initializer=tf.compat.v1.zeros_initializer()
)
self.assertEqual(variable.name, "my_layer/my_var:0")
with base_tf_layers.keras_style_scope():
layer = base_tf_layers.Layer(name="my_layer")
# Assert that the layer was not instrumented as a Keras layer
self.assertFalse(layer._instrumented_keras_api)
# Test basic variable creation.
with backend.name_scope("bar"):
variable = layer.add_weight(
"my_var", [2, 2], initializer=tf.compat.v1.zeros_initializer()
)
self.assertEqual(variable.name, "bar/my_var:0")
def testLayerProperties(self):
layer = base_layers.Layer(name='my_layer')
self.assertEqual(layer.variables, [])
self.assertEqual(layer.trainable_variables, [])
self.assertEqual(layer.non_trainable_variables, [])
if not tf.executing_eagerly():
# updates, losses only supported in GRAPH mode
self.assertEqual(layer.updates, [])
self.assertEqual(layer.losses, [])
self.assertEqual(layer.built, False)
layer = base_layers.Layer(name='my_layer', trainable=False)
self.assertEqual(layer.trainable, False)
# Assert that the layer was not instrumented as a Keras layer
self.assertFalse(layer._instrumented_keras_api)
def testActivityRegularizer(self):
with tf.Graph().as_default():
regularizer = tf.reduce_sum
layer = base_tf_layers.Layer(activity_regularizer=regularizer)
x = tf.compat.v1.placeholder('int32')
layer(x)
self.assertEqual(len(layer.get_losses_for(x)), 1)
def testReusePartitionedVariablesAndRegularizers(self):
with tf.Graph().as_default():
regularizer = lambda x: tf.reduce_sum(x) * 1e-3
partitioner = tf.compat.v1.fixed_size_partitioner(3)
for reuse in [False, True]:
with tf.compat.v1.variable_scope(
tf.compat.v1.get_variable_scope(),
partitioner=partitioner,
reuse=reuse,
):
layer = base_tf_layers.Layer(name="my_layer")
_ = layer.add_weight(
"reg_part_var",
[4, 4],
initializer=tf.compat.v1.zeros_initializer(),
regularizer=regularizer,
)
self.assertEqual(
len(
tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.REGULARIZATION_LOSSES
)
),
3,
)
def testAddWeight(self):
layer = base_layers.Layer(name='my_layer')
# Test basic variable creation.
variable = layer.add_variable(
'my_var', [2, 2], initializer=tf.compat.v1.zeros_initializer())
self.assertEqual(variable.name, 'my_layer/my_var:0')
self.assertEqual(layer.variables, [variable])
self.assertEqual(layer.trainable_variables, [variable])
self.assertEqual(layer.non_trainable_variables, [])
if not tf.executing_eagerly():
self.assertEqual(
layer.variables,
tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES))
# Test non-trainable variable creation.
# layer.add_variable should work even outside `build` and `call`.
variable_2 = layer.add_variable(
'non_trainable_var', [2, 2],
initializer=tf.compat.v1.zeros_initializer(),
trainable=False)
self.assertEqual(layer.variables, [variable, variable_2])
self.assertEqual(layer.trainable_variables, [variable])
self.assertEqual(layer.non_trainable_variables, [variable_2])
if not tf.executing_eagerly():
self.assertEqual(
len(
tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES)), 1)
regularizer = lambda x: tf.reduce_sum(x) * 1e-3
_ = layer.add_variable('reg_var', [2, 2],
initializer=tf.compat.v1.zeros_initializer(),
regularizer=regularizer)
self.assertEqual(len(layer.losses), 1)
added_variable = [False]
# Test that sync `ON_READ` variables are defaulted to be non-trainable.
variable_3 = layer.add_variable(
'sync_on_read_var', [2, 2],
initializer=tf.compat.v1.zeros_initializer(),
synchronization=tf.VariableSynchronization.ON_READ,
aggregation=tf.compat.v1.VariableAggregation.SUM)
self.assertEqual(layer.non_trainable_variables,
[variable_2, variable_3])
@tf.function
def function_adds_weight():
if not added_variable[0]:
layer.add_variable(
'reg_var_from_function', [2, 2],
initializer=tf.compat.v1.zeros_initializer(),
regularizer=regularizer)
added_variable[0] = True
function_adds_weight()
self.assertEqual(len(layer.losses), 2)
def testLayerProperties(self):
layer = base_tf_layers.Layer(name='my_layer')
self.assertEqual(layer.variables, [])
self.assertEqual(layer.trainable_variables, [])
self.assertEqual(layer.non_trainable_variables, [])
if not tf.executing_eagerly():
# updates, losses only supported in GRAPH mode
self.assertEqual(layer.updates, [])
self.assertEqual(layer.losses, [])
self.assertEqual(layer.built, False)
layer = base_tf_layers.Layer(name='my_layer', trainable=False)
self.assertEqual(layer.trainable, False)
# Assert that the layer was not instrumented as a Keras layer
self.assertFalse(layer._instrumented_keras_api)
# Assert this was instrumented as a legacy layer
self.assertTrue(
keras_base_layer.keras_api_gauge.get_cell('legacy_layer').value())
keras_base_layer.keras_api_gauge.get_cell('legacy_layer').set(False)
def testInvalidTrainableSynchronizationCombination(self):
layer = base_tf_layers.Layer(name='my_layer')
with self.assertRaisesRegex(
ValueError, 'Synchronization value can be set to '
'VariableSynchronization.ON_READ only for non-trainable variables. '
'You have specified trainable=True and '
'synchronization=VariableSynchronization.ON_READ.'):
_ = layer.add_weight(
'v', [2, 2],
initializer=tf.compat.v1.zeros_initializer(),
synchronization=tf.VariableSynchronization.ON_READ,
trainable=True)
def testKerasStyleAddWeight(self):
keras_layer = keras_base_layer.Layer(name='keras_layer')
with backend.name_scope('foo'):
keras_variable = keras_layer.add_variable(
'my_var', [2, 2], initializer=tf.compat.v1.zeros_initializer())
self.assertEqual(keras_variable.name, 'foo/my_var:0')
with backend.name_scope('baz'):
old_style_layer = base_layers.Layer(name='my_layer')
# Test basic variable creation.
variable = old_style_layer.add_variable(
'my_var', [2, 2], initializer=tf.compat.v1.zeros_initializer())
self.assertEqual(variable.name, 'my_layer/my_var:0')
with base_layers.keras_style_scope():
layer = base_layers.Layer(name='my_layer')
# Assert that the layer was not instrumented as a Keras layer
self.assertFalse(layer._instrumented_keras_api)
# Test basic variable creation.
with backend.name_scope('bar'):
variable = layer.add_variable(
'my_var', [2, 2], initializer=tf.compat.v1.zeros_initializer())
self.assertEqual(variable.name, 'bar/my_var:0')
def testInvalidTrainableSynchronizationCombination(self):
layer = base_tf_layers.Layer(name="my_layer")
with self.assertRaisesRegex(
ValueError,
"Synchronization value can be set to "
"VariableSynchronization.ON_READ only for non-trainable variables. "
"You have specified trainable=True and "
"synchronization=VariableSynchronization.ON_READ.",
):
_ = layer.add_weight(
"v",
[2, 2],
initializer=tf.compat.v1.zeros_initializer(),
synchronization=tf.VariableSynchronization.ON_READ,
trainable=True,
)
def testInt64Layer(self):
layer = base_tf_layers.Layer(name='my_layer', dtype='int64')
layer.add_weight('my_var', [2, 2])
self.assertEqual(layer.name, 'my_layer')
def testInt64Layer(self):
layer = base_tf_layers.Layer(name="my_layer", dtype="int64")
layer.add_weight("my_var", [2, 2])
self.assertEqual(layer.name, "my_layer")
