def testKernelRegularizerWithReuse(self):
regularizer = lambda x: tf.reduce_sum(x) * 1e-3
inputs = tf.random.uniform((5, 3), seed=1)
_ = core_layers.dense(inputs,
2,
name="my_dense",
kernel_regularizer=regularizer)
self.assertEqual(
len(
tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.REGULARIZATION_LOSSES)),
1,
)
_ = core_layers.dense(
inputs,
2,
name="my_dense",
kernel_regularizer=regularizer,
reuse=True,
)
self.assertEqual(
len(
tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.REGULARIZATION_LOSSES)),
1,
)
def testFunctionalDenseTwiceReuse(self):
with self.cached_session():
inputs = tf.random.uniform((5, 3), seed=1)
core_layers.dense(inputs, 2, name='my_dense')
vars1 = tf.compat.v1.trainable_variables()
core_layers.dense(inputs, 2, name='my_dense', reuse=True)
vars2 = tf.compat.v1.trainable_variables()
self.assertEqual(vars1, vars2)
def testFunctionalDenseTwice(self): inputs = tf.random.uniform((5, 3), seed=1) core_layers.dense(inputs, 2) vars1 = _get_variable_dict_from_varstore().values() core_layers.dense(inputs, 2) vars2 = _get_variable_dict_from_varstore().values() self.assertEqual(len(vars1), 2) self.assertEqual(len(vars2), 4)
def forward_pass(self, inputs, training=None):
out = core_layers.dense(inputs, self.units, name="dense_one",
kernel_initializer=tf.compat.v1.ones_initializer(),
kernel_regularizer="l2")
with tf.compat.v1.variable_scope("nested_scope"):
out = core_layers.dense(
out, self.units, name="dense_two",
kernel_initializer=tf.compat.v1.ones_initializer(),
kernel_regularizer="l2")
return out
def testFunctionalDenseWithCustomGetter(self):
called = [0]
def custom_getter(getter, *args, **kwargs):
called[0] += 1
return getter(*args, **kwargs)
with tf.compat.v1.variable_scope('test', custom_getter=custom_getter):
inputs = tf.random.uniform((5, 3), seed=1)
core_layers.dense(inputs, 2)
self.assertEqual(called[0], 2)
def testFunctionalDenseInitializerFromScope(self):
with tf.compat.v1.variable_scope(
'scope',
initializer=tf.compat.v1.ones_initializer()), self.cached_session():
inputs = tf.random.uniform((5, 3), seed=1)
core_layers.dense(inputs, 2)
self.evaluate(tf.compat.v1.global_variables_initializer())
weights = _get_variable_dict_from_varstore()
self.assertEqual(len(weights), 2)
# Check that the matrix weights got initialized to ones (from scope).
self.assertAllClose(weights['scope/dense/kernel'].read_value(),
np.ones((3, 2)))
# Check that the bias still got initialized to zeros.
self.assertAllClose(weights['scope/dense/bias'].read_value(), np.zeros(
(2)))
def testFunctionalDense(self):
with self.cached_session():
inputs = tf.random.uniform((5, 3), seed=1)
outputs = core_layers.dense(
inputs, 2, activation=tf.nn.relu, name='my_dense')
self.assertEqual(
len(tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES)), 2)
self.assertEqual(outputs.op.name, 'my_dense/Relu')
def testFunctionalDenseInScope(self):
with self.cached_session():
with tf.compat.v1.variable_scope('test'):
inputs = tf.random.uniform((5, 3), seed=1)
core_layers.dense(inputs, 2, name='my_dense')
var_dict = _get_variable_dict_from_varstore()
var_key = 'test/my_dense/kernel'
self.assertEqual(var_dict[var_key].name, '%s:0' % var_key)
with tf.compat.v1.variable_scope('test1') as scope:
inputs = tf.random.uniform((5, 3), seed=1)
core_layers.dense(inputs, 2, name=scope)
var_dict = _get_variable_dict_from_varstore()
var_key = 'test1/kernel'
self.assertEqual(var_dict[var_key].name, '%s:0' % var_key)
with tf.compat.v1.variable_scope('test2'):
inputs = tf.random.uniform((5, 3), seed=1)
core_layers.dense(inputs, 2)
var_dict = _get_variable_dict_from_varstore()
var_key = 'test2/dense/kernel'
self.assertEqual(var_dict[var_key].name, '%s:0' % var_key)
def forward_pass(self, inputs, training=None):
if training:
out = core_layers.dense(inputs, self.units, name="dense_training")
else:
out = core_layers.dense(inputs, self.units, name="dense_no_training")
return out
