BinaryNet implementation in TensorFlow

Question

I recently read a very interesting paper (http://arxiv.org/pdf/1602.02830v3.pdf) suggesting a method for training a CNN with weights and activations constrained to [-1,1]. This is highly beneficial from power/speed perspective.

There are implementations for the method in Torch and Theano publicly available in github: https://github.com/MatthieuCourbariaux/BinaryNet (Theano) https://github.com/itayhubara/BinaryNet (Torch)

I was wondering if the above method can be implemented in TensorFlow ? Has anyone tried implementing this?

Answer 1

Take a look at the TensorFlow GitHub issue #1592. It tracks the progress of the current attempt to add support for binary networks in TensorFlow.