def testConfigurableConv2DAlternateOpSuffixes(self):
# ConfigurableConv2D accepts both 'Conv2D' and 'ConfigurableConv2D' as op
# suffixes in the parameterization to be compatible with structures learned
# using keras.layers.Conv2D or ConfigurableConv2D.
valid_parameterization_1 = {
'conv1/Conv2D': 1,
}
out1 = ops.ConfigurableConv2D(
parameterization=valid_parameterization_1,
filters=10,
kernel_size=3,
name='conv1')(self.inputs)
self.assertEqual(out1.shape.as_list()[-1], 1)
valid_parameterization_2 = {
'conv2/ConfigurableConv2D': 2,
}
out2 = ops.ConfigurableConv2D(
parameterization=valid_parameterization_2,
filters=10,
kernel_size=3,
name='conv2')(self.inputs)
self.assertEqual(out2.shape.as_list()[-1], 2)
# Only one op suffix variant should exist in the parameterization.
bad_parameterization = {
'conv3/Conv2D': 1,
'conv3/ConfigurableConv2D': 2
}
with self.assertRaises(KeyError):
_ = ops.ConfigurableConv2D(parameterization=bad_parameterization,
filters=10,
kernel_size=3,
name='conv3')(self.inputs)
def testParameterizeNamesWithSlashes(self):
conv1 = ops.ConfigurableConv2D(parameterization={
'name/with/slashes/ConfigurableConv2D':
1,
},
filters=10,
kernel_size=3,
name='name/with/slashes')
out1 = conv1(self.inputs)
self.assertEqual(out1.shape.as_list()[-1], 1)
conv2 = ops.ConfigurableConv2D(
parameterization={'name//with///multislash/ConfigurableConv2D': 2},
filters=10,
kernel_size=3,
name='name//with///multislash')
out2 = conv2(self.inputs)
self.assertEqual(out2.shape.as_list()[-1], 2)
# When Keras calls tf.variable_scope with N trailing slashes,
# tf.variable_scope will create a scope with N-1 trailing slashes.
conv3 = ops.ConfigurableConv2D(
parameterization={'name/ends/with/slashes//ConfigurableConv2D': 3},
filters=10,
kernel_size=3,
name='name/ends/with/slashes///')
out3 = conv3(self.inputs)
self.assertEqual(3, out3.shape.as_list()[-1])
def testParameterizeDuplicateNames(self):
parameterization = {
'conv1/ConfigurableConv2D': 1,
'conv1_1/ConfigurableConv2D': 2,
'configurable_conv2d/ConfigurableConv2D': 3,
'configurable_conv2d_1/ConfigurableConv2D': 4,
}
conv1 = ops.ConfigurableConv2D(parameterization=parameterization,
filters=10,
kernel_size=3,
name='conv1')
conv1_1 = ops.ConfigurableConv2D(parameterization=parameterization,
filters=10,
kernel_size=3,
name='conv1')
conv_default_name = ops.ConfigurableConv2D(
parameterization=parameterization, filters=10, kernel_size=3)
conv_default_name_1 = ops.ConfigurableConv2D(
parameterization=parameterization, filters=10, kernel_size=3)
out = conv1(self.inputs)
self.assertEqual(1, out.shape.as_list()[-1])
out = conv1_1(self.inputs)
self.assertEqual(2, out.shape.as_list()[-1])
out = conv_default_name(self.inputs)
self.assertEqual(3, out.shape.as_list()[-1])
out = conv_default_name_1(self.inputs)
self.assertEqual(4, out.shape.as_list()[-1])
def testConfigurableConv2DFunctionality(self):
out = ops.ConfigurableConv2D(filters=5, kernel_size=2)(self.inputs)
expected = keras_layers.Conv2D(filters=5, kernel_size=2)(self.inputs)
self.assertAllEqual(out.shape, expected.shape)
self.assertIn(
'configurable_conv2d/ConfigurableConv2D',
[op.name for op in tf.get_default_graph().get_operations()])
def testConfigurableConv2DParameterization(self):
# with default name
conv1 = ops.ConfigurableConv2D(
parameterization={'configurable_conv2d/ConfigurableConv2D': 1},
filters=10,
kernel_size=3)
out1 = conv1(self.inputs)
self.assertEqual(1, out1.shape.as_list()[-1])
# with custom name
conv2 = ops.ConfigurableConv2D(
parameterization={'conv2/ConfigurableConv2D': 2},
filters=10,
kernel_size=3,
name='conv2')
out2 = conv2(self.inputs)
self.assertEqual(2, out2.shape.as_list()[-1])
def testStrictness(self):
parameterization = {
'unused_conv/Conv2D': 2,
}
conv_not_strict = ops.ConfigurableConv2D(
parameterization=parameterization,
is_strict=False,
filters=10,
kernel_size=3)
conv_strict = ops.ConfigurableConv2D(parameterization=parameterization,
is_strict=True,
filters=10,
kernel_size=3)
# extra ops in the parameterization are ok
out = conv_not_strict(self.inputs)
self.assertEqual(10, out.shape.as_list()[-1])
# when strict=True, all ops in the parameterization must be used
with self.assertRaises(KeyError):
out = conv_strict(self.inputs)