Subclassing tf.Tensor for both Graph and Eager Execution

Question

I would like to subclass the tf.Tensor class. The idea is that the objects of subclasses should behave like Tensors (i.e. I can use them to make any kind of tf operations) but they should also possess others attributes which provide them specific behaviors inside my framework.

Up to now, I worked in Graph mode and I simply do something like this:

class EnrichedTensor(tf.Tensor):

    def __init__(self, tensor, other_stuff):
        super(EnrichedTensor, self).__init__((
                              op=tensor.op,
                              value_index=tensor.value_index,
                              dtype=tensor.dtype)
        self.other_stuff = other_stuff

Now, I would like to do the same to work only in eager mode, but I really don't know (and I didn't find anything) about EagerTensor instantiation. Obviously, the op attribute does not make sense anymore.

I tried to work on the creation of the object through the __new__ method but I found problems in subclassing my EnrichedTensor and following the creation path.

So, I was wondering if is there any way to do this cleanly and in a "sound" way.