Correcting trained model fails.

, , ,
1

I manually annotated a set of data, and then trained a model using

prodigy train model --ner dataset,eval:eval-dataset -m en_core_web_trf --gpu-id 0

I then try to use the trained model with

 prodigy ner.correct new-dataset model/model-best/ ~/data/data.jsonl --exclude dataset

Which gives:

ValueError: Cannot deserialize model: mismatched structure

This seems to be the same problem as NER training produces invalid config.cfg? - #4 by bob_ln. Is there a workaround that would allow me to correct this trained model? I tried removing tagger, parser, attribute_ruler from pipeline in config.cfg, which allows the model to load, but results in loads of warnings.

2

Hi! Sorry about that. The problem here are the frozen components in your pipeline so after upgrading spacy-transformers you should be able to just use the same workaround @SofieVL provided in the thread and re-source the components from the pipeline you want to keep:

3

That worked, thanks :).

4

Hi, I'm doing basically the same thing but can't work out how to alter the config so that it works.

I can follow the instructions in a .py file and the model loads - and looks like it's doing well - but not with the ner.correct

Would some kind soul put me out of my misery please?

M

Thanks!

5

OK, I have this working now by disabling some components in the [nlp] section, then loading replacements in the [components] section.

Is this the intended approach?

Though I get a warning about 'automatically converting a transformer component from v1 to v1.1. Doesn't say which one though...

Thanks!

M