How to connect previously trained Spacy NER model to new empty Language pipeline?

[justmars]

I've been able to train and load a model using a custom script involving a @Language factory:

# post training
_model = spacy.load("./models/statutes/model-best")
_model.pipe_names
# ['tok2vec', 'ner']
doc = _model('Section 13 of PD No. 1869 provides xxx')
for e in doc.ents:
  print(f"{e.text=} {e.label_=}")
# e.text='PD No. 1869' e.label_='STATUTE'

I'd like to load this model in a separate language object and it seems to work but no entities are detected:

x = spacy.blank("en")
x.add_pipe("ner", name="statute_ner", source=_model)
x.pipe_names
# ['statute_ner']
doc = x('Section 13 of PD No. 1869 provides xxx')
doc.ents
# ()

Note this follows the convention described in #10674.

According to the docs:

Important note for trained components
When reusing components across pipelines, keep in mind that the vocabulary, vectors and model settings must match.

Initially I thought that since tok2vec wasn't in the blank model, it would fail. So I tried loading the en_core_web_sm model which had it but it still failed to detect the entities from the pretrained _model.

I think the vocabulary matches. Wehn running x.to_disk, the vocab found in the _model trained appears in x. So I'm guessing I can rule this out.

There was some discussion about loading files via Language.factories through an __init__.py file found in a dated clip of the quite thorough Dr. WJB Mattingly / Python for Digital Humanities but it seems like this has been eclipsed by spacy v3.

Hope for some guidance on what I'm doing wrong. Thank you!

[justmars]

Ok, upon investigating #10674, @adrianeboyd links to https://github.com/explosion/projects/tree/v3/tutorials/ner_double and I see a curious: drug_nlp.replace_listeners("tok2vec", "ner", ["model.tok2vec"]). Didn't know this was a necessary step so I tried it and it solves my problem. Thank you!

_model = spacy.load("./models/statutes/model-best")
_model.pipe_names
# ['tok2vec', 'ner']
# adapted from solution
_model.replace_listeners("tok2vec", "ner", ["model.tok2vec"]) 
# reuse same code above
x = spacy.blank("en")
x.add_pipe("ner", name="statute_ner", source=_model)
x.pipe_names
# ['statute_ner']
doc = x('Section 13 of PD No. 1869 provides xxx')
for e in doc.ents:
  print(f"{e.text=} {e.label_=}")
# e.text='PD No. 1869' e.label_='STATUTE'

[rmitsch]

Happy you could resolve this, and thanks for posting the solution!