attention output in marian-server

[Jörg Tiedemann]

Hi,

Is it possible to obtain hard and soft alignments also via the marian-server output?

I'd like to highlight some connections in an on-line demo and this would be very useful for that purpose.

Jörg

[Roman Grundkiewicz]

Hi,

Yes, it's possible. Here is an example from our regression tests: https://github.com/marian-nmt/marian-regression-tests/blob/master/tests/server/test_ende_align.sh

Roman

[Jörg Tiedemann]

OK - great. Related question to alignment with transformers: Do I need to specify the layer when generating the attention output? I saw that this is a parameter in training for the guided alignment training but I didn't see the same parameter for the decoder. Is that stored in the model somehow?

[Roman Grundkiewicz]

Yes, the value of --transformer-guided-alignment-layer is stored in the model.npz file.

Roman