yaysummeriscoming

For backprop to work, training needs to be done in float precision

Yes, but the forward pass also needs to be done in float precision for backprop to work. For binary inference, you'll need to use another software package.

Unfortunately I haven't kept up to date with the latest developments in binary networks - it's been a few years since I made this. That being said, I've heard that...

Hi, are you using standalone keras with tensorflow 1?

This example is purely for image loading. DALI itself has added support for many other datatypes, such as audio. I suggest you check out the official docs, there's some examples...

The activations are binarised at the input to the next layer. as for the final layer: my understanding is that whilst the convolution multiplications are binary (xnor operation), the summing...

FYI theres another really good paper that has come out that discusses this a bit further and also has some big improvement on the original BinaryNet implemented here. It's a...

Thanks :) You're correct about the passthroughSign function, that got me too. Remember that only the weight * activation multiplies are binary, summation is done in higher precision. Regarding HardTanh...

+1

+1 on tensorrt support. Benchmarks against Triton would also be great!