machine learning - Model selection with dropout training neural network -

i've been studying neural networks bit , learned dropout training algorithm. there excellent papers out there understand how works, including ones authors.

so built neural network dropout training (it easy) i'm bit confused how perform model selection. understand, looks dropout method used when training final model obtained through model selection.

as test part, papers talk using complete network halved weights, not mention how use in training/validation part (at least ones read).

i thinking using network without dropout model selection part. makes me find net performs n neurons. then, final training (the 1 use train network test part) use 2n neurons dropout probability p=0.5. assures me have n neurons active on average, using network @ right capacity of time.

is correct approach?

by way, i'm aware of fact dropout might not best choice small datasets. project i'm working on has academic purposes, it's not needed use best model data, long stick machine learning practices.

using different model model selection , different final training never approach. supposed use exact same procedure (with dropout, same probability etc.) during model selection. in fact probability hyperparameter need fit during model's selection. in fact there no difference between model selection , without dorpout, why noone writing in papers - noting changed.