The activation functions needs to be changed to Htanh() instead of tanh.

[sudo-sh]

p=tf.tanh(wx)
Needs to be replaced by
p=tf.clip_by_value(wx,-1,1)
So, that it becomes the hard tanh function, as described in the paper and validates the straight through estimator.

I.Hubara et.al Binarized Neural Networks: Training Neural Networks with Weights and
Activations Constrained to +1 or -1.

Thanks
-Sudarshan

PS:- Great work done!