def test_network_creation(self, input_size, model_id, endpoint_filter_scale): """Test creation of RevNet family models.""" tf.keras.backend.set_image_data_format('channels_last') network = revnet.RevNet(model_id=model_id) inputs = tf.keras.Input(shape=(input_size, input_size, 3), batch_size=1) endpoints = network(inputs) network.summary() self.assertAllEqual([ 1, input_size / 2**2, input_size / 2**2, 128 * endpoint_filter_scale ], endpoints['2'].shape.as_list()) self.assertAllEqual([ 1, input_size / 2**3, input_size / 2**3, 256 * endpoint_filter_scale ], endpoints['3'].shape.as_list()) self.assertAllEqual([ 1, input_size / 2**4, input_size / 2**4, 512 * endpoint_filter_scale ], endpoints['4'].shape.as_list()) self.assertAllEqual([ 1, input_size / 2**5, input_size / 2**5, 832 * endpoint_filter_scale ], endpoints['5'].shape.as_list())
def test_input_specs(self, input_dim): """Test different input feature dimensions.""" tf.keras.backend.set_image_data_format('channels_last') input_specs = tf.keras.layers.InputSpec(shape=[None, None, None, input_dim]) network = revnet.RevNet(model_id=56, input_specs=input_specs) inputs = tf.keras.Input(shape=(128, 128, input_dim), batch_size=1) _ = network(inputs)
def test_serialize_deserialize(self): # Create a network object that sets all of its config options. kwargs = dict( model_id=56, activation='relu', use_sync_bn=False, norm_momentum=0.99, norm_epsilon=0.001, kernel_initializer='VarianceScaling', kernel_regularizer=None, ) network = revnet.RevNet(**kwargs) expected_config = dict(kwargs) self.assertEqual(network.get_config(), expected_config) # Create another network object from the first object's config. new_network = revnet.RevNet.from_config(network.get_config()) # Validate that the config can be forced to JSON. _ = new_network.to_json() # If the serialization was successful, the new config should match the old. self.assertAllEqual(network.get_config(), new_network.get_config())