Error parsing config overrides - `nlp.tokenizer` sections does not exist - spacy

Question

I encountered a weird problem in Spacy and I am not sure whether I am doing something wrong or it is a genuine bug.
I use Spacy project and create a default config file via:
python -m spacy init config spacy.cfg
Then I try to load an NLP object using this config:
import spacy
config = spacy.util.load_config('./spacy.cfg')
nlp = spacy.load("en_core_web_sm", config=config)
When do this I get the following error:
ConfigValidationError:
Error parsing config overrides
nlp -> tokenizer -> #tokenizers not a section value that can be overwritten
By looking at what is going on inside through PDB I noticed that section nlp.tokenizer is not created. Instead the config stores the following ugly item within the NLP section:
'tokenizer': '{"#tokenizers":"spacy.Tokenizer.v1"}'
which does not seem to look allright.
I am using Spacy v3.0.3 on Ubuntu 20.04.2 LTS.

Answer 1

There's nothing wrong with that tokenizer setting. en_core_web_sm already contains its own config including a tokenizer, which is what you can't override.
You want to load your config starting from a blank pipeline instead of a pretrained pipeline:
nlp = spacy.blank("en", config=config)
Be aware that the language en needs to match here with the spacy init config language setting or it won't be able to load the config.