addons

addons copied to clipboard

The current implementation of MultiOptimizer puts grads and vars inside the optimizer_specs variable. Since the grads and vars are not serializable, any model that uses MultiOptimizer class cannot be saved or checkpointed. This PR will only extract the necessary optimizer_specs components that are useful for re-instantiating the MultiOptimizer instance inside get_config() method, leading to a serializable MultiOptimizer implementation.

https://github.com/tensorflow/addons/pull/2719