Nonnegative weights

[Rasmuskh]

Hi,
What is a good way to initialize weights to be nonnegative in flax?

My current approach is to rewrite the glorot_uniform and variance_scaling functions, to do what I want. That works but also adds a lot of boilerplate code to my project.

I hope there is a nicer/cleaner way to do this.

[cgarciae]

@Rasmuskh Creating custom initializers is fine, if they are taking too much space consider creating an initializers.py module and then import them from there.

[Rasmuskh]

Thanks,
Yes, moving all the custom initializers to a separate module would make things cleaner.
In my case since I was using custom layers anyways, I ended up just enforcing nonnegativity inside the layer definition instead.