Training NER component using pre-trained tok2vec

[mikelgda]

I want to train a NER component from scratch and I suppose it would be faster to re-use the tok2vec component from en_core_web_lg. However, I'm having issues defining the configuration file for this.

First I tried using the recommended configuration from the documentation but I think this also trains the tok2veccomponent, which includes the static vectors from tok2vec.

I also tried modifying the configuration to source the tok2vec component from en_core_web_lg and then adding a listener for a NER component created from scratch like this

[nlp]
lang = "en"
pipeline = ["tok2vec","ner"]

[components.ner]
factory = "ner"
incorrect_spans_key = null
moves = null
scorer = {"@scorers":"spacy.ner_scorer.v1"}
update_with_oracle_cut_size = 100

[components.ner.model]
@architectures = "spacy.TransitionBasedParser.v2"
state_type = "ner"
extra_state_tokens = false
hidden_width = 64
maxout_pieces = 2
use_upper = true
nO = null

[components.ner.model.tok2vec]
@architectures = "spacy.Tok2VecListener.v1"
width = ${components.tok2vec.model.encode.width}

[components.tok2vec]
source = "en_core_web_lg"

which is taking the recommended configuration and replacing the tok2vec by the pre-trained component. However, this shows a bug since components.tok2vec.model.encoder.width is not accessible because I am sourcing the tok2vec component.

Last, I also tried using the recommended configuration and adding the tok2vec to the frozen_components and annotating_components lists, but I also get a bug because the tok2vec is not trained although it has the static vectors from en_core_web_lg.

My main question is how to source the tok2vec while training a new NER from scratch and not training the tok2vec.

Thank you!