Reworking KerasTensor?

[AdityaKane2001]

@haifeng-jin @fchollet

This issue is both a bug report and a feature request. https://github.com/tensorflow/tensorflow/issues/48531 shows an issue of KerasTensor. Of course, the docstring elaborates it even further.

IMO, given the position of Keras, KerasTensor should be completely interoperable with tf.Tensor, ideally being able to use interchangeably. However, given the current design decision to declare a separate class KerasTensor and then overload all methods to it, I am assuming that simply inheriting KerasTensor from the public tf.Tensor is not a solution. Hence, my assumption is that a refactor or revamp of KerasTensor is a plausible, albeit difficult solution.

Personally, I am excited to work on this, but I would need some form of technical help from the team to understand the codebase and related implications of the change.