Does fairseq transform align can returns the alignment result?

[itaybar]

I am using fairseq transform align and Im trying to get the alignment results with no success. this is possible somehow?

[minhthuc2502]

I think I don't have much knowledge about this. @vince62s Do you have any idea?

[vince62s]

it is supported I think: https://github.com/OpenNMT/CTranslate2/blob/master/docs/guides/fairseq.md?plain=1#L1-L9 but never used

[itaybar]

I run the model successfully but cant get the align output (an extra output in this architecture that represent the alignments)

[ymoslem]

My understanding is that this also applies to models trained with the guided alignment feature in OpenNMT-tf. According to issue https://github.com/OpenNMT/CTranslate2/issues/767, the feature is not explicitly available; however…

Alignments can be easily inferred from the attention vectors

For a model trained with Fairseq Transformer Align, I am not sure if this would work, but I assume it depends on the model itself.

[itaybar]

Its working, Thanks