NER transformers example pipeline

[AdrianKrebs]

I've used the ner_demo (https://github.com/explosion/projects/tree/v3/pipelines/ner_demo) as the basis for my NER pipeline and it worked well so far.
To further increase the accuracy, I want to switch to fine-tuning a transformer model. I read through the documentation but couldn't find any bootstrapped project configuration (similar to ner_demo).
Because I'm struggling to migrate the project.yml and the generated config.cfg to en_core_web_trf, I was wondering if there are any project.yml samples for transformer models out there?

[adrianeboyd]

If you just run spacy project run all, you can add -G to the create-config command to generate a config with transformer+ner. You'll want to set the gpu_id at the top before training for reasonable training speeds (although I think this toy example will still train relatively quickly on CPU if you just want to try it once; we wouldn't recommend training on CPU for non-toy dataset sizes).

    script:
      - "python -m spacy init config --lang ${vars.lang} --pipeline ner configs/config.cfg -G --force"

(The optional train-with-vectors command that's not part of all will fail because the generated config doesn't have a tok2vec in it to enable vectors for.)