feeding decoded output into training network

[rwth-i6]

feeding decoded output into training network

[albertz]

Yes, we can. :)
I'm doing that in multiple ways already. What I would recommend now is to use the extra-net mechanism. See the documentation here. See also the very simple example here. A much more complex example is here.

Basically, when you prefix a layer with the name "extra.search:", it (and all its sublayers) will have the search flag enabled, and thus you can use ChoiceLayer or DecideLayer and they operate like in search. Note, for a RecLayer, if you just want to use its decoded output, and do not need to backpropagate through the search, then you should explicitly disable the back-propagation mechanism for the RecLayer, via "back_prop": False (only for the layer where you do the search, and don't need to backpropagate through), as this makes it much faster and more memory efficient.

[ashu5644]

Ok, thanks. I will look into the document.

[albertz]

Closing this now. Please comment further if sth is unclear.