Exposing registered components via entry points.

[coltonflowers1]

I would like to include a custom-registered function within a config.cfg without having to have users pass in the code for these registered functions to the training CLI via the --code argument. Entry points have allowed me to do this with component factories, and according to the official documentation here, I can also do this with custom languages, displacy colors, and custom lookups, but there does not seem to be anything for plain registered functions. In fact, the docs here only reference the --code solution in order to make sure your custom registered function are imported prior to training.

[adrianeboyd]

There is not a separate registry for components because components are just a wrapper for factory registry entries.

However the entry points behavior is pretty confusing because:

• importing code related to an entry point can run additional code through decorators
• there is a hidden internal entry points registry for factories just to make it easy to import+run code with decorators, and it's the decorators that are important not the entry point names

So with this code + entry point, you'll end up with a factory for the name my_component:

from spacy import Language


@Language.component("my_component")
def my_method(doc):
    return doc

spacy_factories =
    _some_unused_name = my_package.components:my_method

The component would also have been registered in the same way if you'd added this as an entry point in misc or some other category, but it's less confusing if you're not adding extraneous entries to other registries.

This is pretty confusing (I was initially not 100% sure without testing this out), we should add some better examples to the docs...

[coltonflowers1]

My apologies, I edited the second link in my post as it was pointing to the wrong section of the docs. If I understand you correctly, you are saying that the spacy_factories end point can be used to store functions I define using the misc registry? I'm not actually trying to register a pipeline component but a function I use within the initialization of a custom component because I want more control over that function in my configuration. I assumed spacy_factories was only for pipeline components because of the following in the docs: "[The spacy factories end point is a] group of entry points for pipeline component factories, keyed by component name. Can be used to expose custom components defined by another package."

[adrianeboyd]

I thought you mean a component with Language.component vs Language.factory. The registries in this link are the main ones with some custom internal behavior related to each registry: https://spacy.io/usage/saving-loading#entry-points

Most of the other registries are just to make the names easier to filter/remember, and the specific registry and entry point names don't actually matter as long as it returns a function with the right signature: https://spacy.io/api/top-level#registry

You could have:

scorer = {"@batchers": "my_corpus_reader.v1"}

I mean, I don't recommend doing this, but just to demonstrate.

Using misc is fine with:

spacy_misc =
    my_component.v1 = my_package.components:my_component

If you define the entry point in your package setup, then you don't need the decorator in the code.

As an example with a number of different entry points:

https://github.com/explosion/spacy-legacy/blob/eecb096ed01ead79a7f8d0fc64631becf37d4458/setup.cfg#L35-L57

[coltonflowers1]

I see so I am ok to refer to a function registered in any registry using entry points and the entry point name doesn't actually matter. I think that clears up my confusion. Thanks!

[adrianeboyd]

Just to clarify, the entry point name does matter for everything except spacy_factories. The registered function will show up with that name in the registry.