This is needed so that dense layers can be specifically named so that the dense layer variable scope names line up with the pre-trained models.

https://github.com/tensorflow/tensorflow/blob/590d6eef7e91a6a7392c8ffffb7b58f2e0c8bc6b/tensorflow/python/layers/core.py#L185

I don't see the mechanism to do this in the TF.Net keras layer so for now I'll just make my own little dense layer. This is just a heads up that w/o fixing this it may be hard to load some pre-trained models.