Add init fill-config-transformer CLI command

[shadeMe]

Description

This command reads the Hugging Face model name and revision from the initialize.components.transformer.encoder_loader config section, fetches its config from the HF Model Hub and fills in the entry point parameters for the same.

Types of change

New feature

Checklist

• I confirm that I have the right to submit this contribution under the project's MIT license.
• I ran the tests, and all new and existing tests passed.
• My changes don't require a change to the documentation, or if they do, I've added all required information.

[danieldk]

LGTM, feel free to merge if @adrianeboyd is also ok with it.

[adrianeboyd]

I could run the CLI command for the prefilled config snippets, but there's still no way for users to get to an initial config:

nlp.add_pipe("curated_transformer")

or

spacy init config -p curated_transformer

both fail with:

Config validation error
curated_transformer.model -> vocab_size	field required
{'@architectures': 'spacy-curated-transformers.XlmrTransformer.v1', 'piece_encoder': {'@architectures': 'spacy-curated-transformers.XlmrSentencepieceEncoder.v1'}, 'with_spans': {'@architectures': 'spacy-curated-transformers.WithStridedSpans.v1'}}

[shadeMe]

I could run the CLI command for the prefilled config snippets, but there's still no way for users to get to an initial config:

Fixed in #20.

[adrianeboyd]

Can you tell the user based on the downloaded model which kind of architecture they probably need? (Like, I was even confused by the naming of something that looks like bert but wasn't.)

[shadeMe]

I've expanded the error message to display either the expected HF model type for the given curated transformers architecture (in case the type of the HF model is not supported) or the correct curated transformers architecture (if the HF model type is supported).

[svlandeg]

(still reviewing - first few comments)

[svlandeg]

I don't like the way that fill doesn't mean quite the same thing as in fill-config, but the alternatives are probably worse.

I was actually thinking the same earlier. What about update instead of fill? 🤔

[shadeMe]

I was actually thinking the same earlier. What about update instead of fill? 🤔

I'd still keep fill as the grand majority of the fields being added are actually missing in the config, i.e., there are no keys for them (as opposed to updating the values of existing keys).

[svlandeg]

Looks great!