Failing to restore RectifiedAdam optimizer from checkpoint

[olesalscheider]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 18.04
• TensorFlow version and how it was installed (source or binary): source, nightly
• TensorFlow-Addons version and how it was installed (source or binary): source, nightly
• Python version: 3.6.7
• Is GPU used? (yes/no): yes

Describe the bug

When restoring the RectifiedAdam optimizer from a checkpoint, restoring fails with the following error message:
ValueError: Tensor conversion requested dtype float32 for Tensor with dtype int32: <tf.Tensor: shape=(), dtype=int32, numpy=0>

The reason is that RectifiedAdam stores total_steps as an integer hyper-parameter:
https://github.com/tensorflow/addons/blob/master/tensorflow_addons/optimizers/rectified_adam.py#L134

OptimizerV2 however creates variables by calling add_weight without a dtype in _create_hypers. In other words, it assumes that all hyper-parameters are floating point parameters.

Storing the total_steps parameter as float fixes the problem for me.

[olesalscheider]

olesalscheider@cceaa69 seems to fix it for me, but I am not sure if this is the intended way.

[bhack]

Can you open a PR?