search

N3PDF / evolutionary_keras

An evolutionary algorithm implementation for Keras
GNU General Public License v3.0
10 stars 7 forks source link

update NGA #20

Closed RoyStegeman closed 4 years ago

RoyStegeman commented 4 years ago

Should we just return an error if the network does not have only weight-bias structures? As in that case the NGA functionality is not well suited.

scarlehoff commented 4 years ago

Yes, and exception like "ValueError" might be suitable. Like raise ValueError("The chosen optimizer blabla is not suitable because of this and that")

scarlehoff commented 4 years ago

This would be a nice feature to have but it needs to take into account the issue discussed in #21 as we cannot rely on the input layer being just Dense.

RoyStegeman commented 4 years ago

In that case I think the only way to do this is to check whether the shapes of the trainable parameters are such that they could correspond to such a structure.

But is it really worth adding these kind of features if there are no plans to use them?

scarlehoff commented 4 years ago

Well, it is an enhancement that would be nice. Wouldn't say it's a required feature by any means, no.

Mainly because it would be not trivial to do a check general enough that can capture any possible situation...