No weight update due to sign function

[jonkoi]

Hi,

I don't think weights get updated because of the tf.sign function cutting off backprop gradients:

binarized_filters = tf.sign(tiled_filters + expanded_stddev, name="binarized_filters")

i believe this line is needed:

with tf.get_default_graph().gradient_override_map({"Sign": "Identity"}):

[layog]

Sorry, for the late reply.

Yes, shift_parameters and beta_values are currently not trainable due to tf.sign function. It's mentioned on the bottom of the README.md.

I didn't get time to fix that one but I guess your solution will work, will try and publish the results.

[martinkersner]

@jonkoi You are right, the implementation is not correct. The first two ABC layers behave only as constant feature extractors and all the learning happens in the last two float fully connected layers.

[RealDataLLC]

Even with the weight update, it seems to not be learning on any dataset beside MNIST. Did you have success @jonkoi?