Implementing feedback equalization in a tensor flow

Question

I want to implement Direct Feedback Alignemnt in Tensorflow.

Reference document: https://arxiv.org/pdf/1609.01596v5.pdf , Nøkland (2016)

I implemented a simple network that does DFA in pure Python, explicitly using backprop, I just switched the BP return pass with one of the DFAs.

where B1 and B2 are two random matrices with corresponding sizes.

Training in tensor motion obviously makes backprop the default. Where and what do I need to change in Tensorflow code to get this learning behavior?