Beta test: spaCy Curated Transformer pipelines

[danieldk]

We have recently announced the first the tech preview of Curated Transformers. Curated Transformers is a transformer library for PyTorch that is developed by Explosion. It provides a curated set of state-of-the-art models that are composed of reusable building blocks.

Since Curated Transformers is a pure PyTorch library, we have also released a library with Thinc model wrappers and a spaCy pipeline component called spacy-curated-transformers. With this library, you can use Curated Transformers in your own pipelines. We have prepared an example project that shows how you can train a pipeline with Curated Transformers.

Today, we are releasing experimental spaCy transformer pipelines that use Curated Transformers. These pipelines are used when you download a transformer model using the spaCy 3.7.0 development release:

$ pip install spacy==3.7.0.dev
$ spacy download de_dep_news_trf

If you are using our transformer pipelines with spaCy 3.6 or earlier versions, it would be very helpful if you could test these new transformer pipelines and report significant regressions in speed or accuracy or for that matter, any other issues that you might encounter.

Available Curated Transformer pipelines

Model name	Language
ca_core_news_trf	Catalan
da_core_news_trf	Danish
de_dep_news_trf	German
en_core_web_trf	English
es_dep_news_trf	Spanish
fr_dep_news_trf	French
ja_core_news_trf	Japanese
sl_core_news_trf	Slovenian
uk_core_news_trf	Ukrainian
zh_core_web_trf	Chinese

[adrianeboyd]

The default pipelines with spacy download have just been updated to support replace_listeners (pipeline version with spacy validate should show 3.7.1).