https://www.tensorflow.org/addons/api_docs/python/tfa/activations
In order to conform with the current API standard, all activations must:
- Be a
tf.function
unless it is a straightforward call to a custom op or likely to be retraced. - Register as a keras global object so it can be serialized properly:
@tf.keras.utils.register_keras_serializable(package='Addons')
- Simple unittests that demonstrate the layer is behaving as expected.
- To run your
tf.functions
in eager mode and graph mode in the tests, you can use the@pytest.mark.usefixtures("maybe_run_functions_eagerly")
decorator. This will run the tests twice, once normally, and once withtf.config.run_functions_eagerly(True)
. - Add activation name to activations_test.py to test serialization.
- Update the CODEOWNERS file