Add option to pass kernel_* and bias_* regularizers to Coral layer

[gmgeorg]

CORN layer inherits fromDense, so it all usual **kwargs can be passed.

Coral layer inherits from Layer, so kernel_regualizer/bias_regularizer (and other kwargs that would make sense for Coral) dont work as expected (passing them fails init since Layer does not recognized those kwarggs)

[ck37]

Should the Coral layer inherit from Dense instead of Layer?

…

-ck.mobile

On Mar 2, 2022, at 7:11 PM, Georg ***@***.***> wrote:  CORN layer inherits fromDense, so it all usual **kwargs can be passed. Coral layer inherits from Layer, so kernel_regualizer/bias_regularizer (and other kwargs that would make sense for Coral) dont work as expected. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you are subscribed to this thread.

[gmgeorg]

I don't think that will be as trivial, since AFAIU the point of the CORAL layer is that it is a specific (constrained) structure of a fully connected dense layer (which is what CORN solves), which requires overwriting the build method anyway -- so might as well just add the arguments to the inherited Layer class

Looking at Dense() layer implementation, it should be straightforward to pass kernel_regularizer and bias_regularizer as arguments to the CoralOrdinalLayer init, set them as self.kernel_regularizer=regularizers.get(kernel_regularizer) and then pass them to self.add_weight(..., regularizer=self.{kernel|bias}_regularizer) for self.kernel and self.bias respectively.

See https://github.com/keras-team/keras/blob/v2.8.0/keras/layers/core/dense.py#L147

This is mostly to support the same function signagure as a typical Dense layer that has bias/kernel weights.

[ck37]

Belated response but thanks for the explanation on this!