Skip inconsistent models
Opened this issue · 2 comments
Is your feature request related to a problem? Please describe.
Hi,
Playing this CNN models with different kernel size and alignment it sounds like a bit tricky to be able to launch the tuner.
Often it results in a wrong model.
It sounds like it's a question which was already asked, but now, do you have any solution for that?
Describe the solution you'd like
I feel like it could be fixed user side, in the callback which returns the model.
This callback could return an exception keras_tuner knows (for example kt.InvalidModel), to let them skip silently this combination of hyperparameters.
As the combination on it's own have no meaning for keras tuner, I don't see other good solution. And it probably could be used for different use cases (such as the amount of parameters, or what even).
I can try to play with that idea if you like.
What do you think?
Sounds like we can already raise a FailedTrialError() from the user model_builder, for an early skip of a model.
Ill try that next time i have the hands on.
It is not working.
But it sounds like such "skip" behavior would be easy to implement the same way.