spaCy 3.0 support, is Rasa 2.0 ready?

[koaning]

spaCy 3.0 will be out, probably this month. It makes sense to check compatibility. The release notes on pypi suggests there's already dev versions.

[koaning]

So it seems like spaCy will have transformer support for a lot of Non-English languages. The thing though is that the vectors are retrieved differently. This is the API that is currently in spaCy 3.0 nightly;

token_emb, sent_emb = nlp("i duck to hide")._.trf_data.tensors
token_emb.shape, sent_emb.shape

For transformer based models you can't fetch the tensors directly via vector anymore but via ._.trf_data.

Also note that the token_emb do not match the shape of the tokens that spaCy finds. More likely these tokens correspond to the bytepair-something tokens that the internal transformer is using.