Eric Hunsberger

I'm also encountering this problem. Part of the problem appears to be that `QuantizeWrapper` doesn't account for `self.trainable == False` in `trainable_weights` and `non_trainable_weights` (c.f. `tf.keras.layers.Layer`, which does account for...

I've started work on a patch here: https://github.com/hunse/model-optimization/pull/1. It works for ensuring that the layer's trainable parameters (e.g. kernel, bias) do not get trained when `trainable=False` on the layer. @metinsuloglu,...

OK, I've added a test to https://github.com/hunse/model-optimization/pull/1. Let me know if/when I can make a PR here (since the CONTRIBUTING.md document says not to make a PR until an issue...

Thanks @js1010. I was able to get something working for my own code base, by modifying the `clone_model_with_weights` function that's used in `quantize_apply` to use Keras's `SharedObjectSavingScope` and `SharedObjectLoadingScope`. I...

This was caused by 4caba86b9c4f12d2be17394ee3eff1af8d4e0e31, because we changed a number of `except Exception` statements to more specific exceptions, but obviously missed catching the exception that causes this problem. We might...

I don't see any problem with this.

Modifying the error population is slightly different, because if your error population is actually in neurons then you're limited in terms of the resolution on the learning rate. While having...

One thing to keep in mind is that a connection can have multiple learning rules, and we should really allow separate rates for each rule. So for example with Option...

> My hope is that we can do something like "you can set the learning rate to a constant via the constructor or as an attribute (as we currently have...

You could also pose this as a NNLS problem by making the activities for the inhibitory neurons negative, and having all positive weights. Then you could use a solver like...