Testable docstrings

[WindQAQ]

#677 is merged, so we are happy to introduce doctest on the docstring. This feature is awesome that can avoid wrong example usage.

Because the change might be small, please submit one sub-module, for example, activations, per PR. For more details, please refer to https://www.tensorflow.org/community/contribute/docs_ref.

Note that available namespace in our repo is tf for tensorflow, np for numpy and tfa for tensorflow_addons.

• activations: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/activations
• callbacks: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/callback
• image: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/image
• layers: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/layers
• losses: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/losses
• metrics: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/metrics
• optimizers: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/optimizers
• rnn: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/rnn
• seq2seq: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/seq2seq
• text: https://github.com/tensorflow/addons/tree/master/tensorflow_addons/text

It's also really appreciated that one can help improve our doc based on https://www.tensorflow.org/community/contribute/docs_style when scrutinizing docstrings.

[WindQAQ]

Seems that it wasn't integrated in our CI in #677 yet. #2069 should do it.

[ashwin-phadke]

@WindQAQ I am working on optimizers and just wanted to check if the following is the right way to update docstrings

        >>> from tf.keras.optimizers import schedules

        >>> lr_schedule = schedules.TriangularCyclicalLearningRate(
        ... initial_learning_rate=1e-4,
        ... maximal_learning_rate=1e-2,
        ... step_size=2000,
        ... scale_mode="cycle",
        ... name="MyCyclicScheduler")

        >>> model.compile(optimizer=tf.keras.optimizers.SGD(
        ... learning_rate=lr_schedule),
        ... loss='sparse_categorical_crossentropy',
        ... metrics=['accuracy'])

        >>> model.fit(data, labels, epochs=5)

Also do I need to run doctest locally or will it run on the PR directy?