floret-vecs as drop-in replacement in _sm models

[concepttt-ma]

Just a brief question: is it possible to use one of the _sm pretrained models (for taggin, parsing, ner,..) and have the model yield floret vectors (which I have already trained and created a spacy model for)?
I mean, the model should then access the floret vecs when the .vector (property) is called on the tokens.

Of course, I could use an extension to get/set the floret-vector if needed (with an "additional" spacy model with floret vecs only); or are the user_hooks a more apt tool for this? Anyway, I was just wondering if there is an easy way to create such a hybrid model.

Any help is appreciated
have a good day

[adrianeboyd]

Sure, as long as the model doesn't currently have vectors you should be able to do:

floret_nlp = spacy.load("floret_model")
nlp = spacy.load("en_core_web_sm")
nlp.vocab.vectors = floret_nlp.vocab.vectors

[concepttt-ma]

Hi Adriane
Thanks for the lightening fast answer!
ok, cool that works - it would still imply that I start both models alongside, right? (guess, that this replacement has to be done at runtime and is not configurable on the level of the model (let's say a customized version of the _sm model)).

[adrianeboyd]

You can save any pipeline with to_disk and load from a directory or package with spacy package as you wish.

[concepttt-ma]

Perfect!