Wrapping a textcat in a custom component

[NixBiks]

I have a custom component that runs text categorization on each sentence in a Doc. This was build in a kinda hacky way in v2. Now I want to update it to work in v3 with the new config system.

Initially I was thinking something like this

from spacy.pipeline.textcat_multilabel import MultiLabel_TextCategorizer


@Language.factory("my_component", default_config={"textcat": None})
def create_my_component(nlp, name, textcat: MultiLabel_TextCategorizer):
    return MyComponent(textcat)


class MyComponent:
    def __init__(self, textcat):
        this.textcat = textcat

and with a config like this

[components.my_component]
factory = "my_component"

[components.my_component.textcat]
factory = "textcat_multilabel"
scorer = {"@scorers":"spacy.textcat_multilabel_scorer.v1"}
threshold = 0.5

[components.my_component.textcat.model]
@architectures = "spacy.TextCatBOW.v2"
exclusive_classes = false
ngram_size = 1
no_output_layer = false
nO = null

I have my component registered fine but I'm getting the following error when I'm using spacy.load

Config validation error

my_component -> textcat   instance of MultiLabel_TextCategorizer expected

{
  'nlp': <spacy.lang.en.Englishobjectat0x117057f10>,
  'name': 'my_component',
  'textcat': {
    'factory': 'textcat_multilabel',
    'scorer': {
      '@scorers': 'spacy.textcat_multilabel_scorer.v1'
    },
    'threshold': 0.5,
    'model': {
      '@architectures': 'spacy.TextCatBOW.v2',
      'exclusive_classes': False,
      'ngram_size': 1,
      'no_output_layer': False,
      'nO': None
    }
  },
  '@factories': 'my_component'
}

I didn't expect it to work since I still need to point to where my textcat model can actually be found but yeah...

[darrkj]

Is the issue in having the signatures not aligned. The factory passes(nlp, name, textcat: MultiLabel_TextCategorizer) to create_my_component but in the init method you only supply self, textcat so where it expects to get a nlp it gets the textcat component. That json shows the three things the factory should supply but the error is stating that init is expecting an instance of MultiLabel_TextCategorizer.

[NixBiks]

Ah yes that was actually just a typo in this issue. I do have textcat in my actual code. I've updated my question accordingly

[NixBiks]

Workaround

My current workaround is to have only this part in my config

[components.my_component]
factory = "my_component"

Then I load the component like this

import spacy

from pathlib import Path
from spacy.pipeline.textcat_multilabel import MultiLabel_TextCategorizer


@Language.factory("my_component", default_config={"textcat": None})
def create_my_component(nlp, name, textcat: MultiLabel_TextCategorizer):
    textcat = spacy.load(
        Path(__file__).parent / "model", vocab=nlp.vocab
    ).get_pipe("textcat_multilabel")
    return MyComponent(textcat)


class MyComponent:
    def __init__(self, textcat):
        this.textcat = textcat

Note this requires an entire pipeline in ./model folder with it's own config.cfg. So I don't see this as optimal really

[polm]

Sorry for the delayed reply to this.

Have you looked at the spancat? It's still not as well documented as our other components, but it's a pretty natural fit for labelling each sentence in a document.

About your component / issue here more generally, it might be easier to subclass the textcat and modify some of the methods on it to get the behavior your want rather than putting it inside another component. Things like serialization etc. should also just work.

Another approach is of course to split your docs up, using one doc for each sentence, and then using a normal pipeline with textcat and reassembling the docs later.