ValueError: Cannot get dimension 'nO' for model 'sparse_linear': value unset

[NixBiks]

How to reproduce the behaviour

I have a pipeline with three components that can't work together for some reason. Currently I can't make a minimal reproducible example right away - I'd need some pointers on how to locate the issue.

The following pipeline

pipeline = ["textcat_multilabel", "sentencizer", "financial_entities"]

throws this error (when loading the model)

ValueError: Cannot get dimension 'nO' for model 'sparse_linear': value unset

Diving deeper into the error message, reveals that specifically it breaks when building financial_entities that wraps an extended EntityRuler. I.e. when running ruler.add_patterns.

The very odd thing is that the pipeline works just fine in these constellations

pipeline = ["textcat_multilabel", "financial_entities"]
pipeline = ["sentencizer", "financial_entities"]
pipeline = ["textcat_multilabel", "sentencizer"]

Any ideas how to debug this?

Your Environment

• spaCy version: 3.2.1
• Platform: macOS-12.1-arm64-arm-64bit
• Python version: 3.8.12

[NixBiks]

Okay so I finally got a little wiser on the issue.

My textcat_multilabel model was trained with a modified tokenizer using

before_init = {"@callbacks":"customize_tokenizer"}

in config.cfg.

In __init__ in my financial_entities I have ruler.add_patterns. If I instead move that to a method, e.g. add_patterns on financial_entities and then call that after the pipeline has been loaded. So

nlp = spacy.load("./pipeline")
nlp.get_pipe("financial_entities").add_patterns()

All in all I'm confused here but it seems to me that I have one tokenizer when __init__ all my components. And then the modified tokenizer after all __init__ have been called. Does that sound right?

[adrianeboyd]

Because entity ruler phrase patterns run the pipeline (up to the current position in the pipeline), you want to load patterns in initialize rather than __init__ so that the previous components have already been initialized. The entity ruler itself already supports this directly, but you can add whatever additional kwargs to your method that you need for the custom component: https://spacy.io/api/entityruler#initialize

[NixBiks]

Okay then I don't understand how to do it. I need to make sure that my initialize method will be called and thereby adds the patterns when I use spacy.load. I assume I won't have to retrain. Currently this is how it looks

    def initialize(self, get_examples: Callable[[], Iterable[Example]], *, nlp: Optional[Language] = None):
        print("Initializing financial entities")
        self.first_ruler.initialize(get_examples, nlp=nlp, patterns=self.first_rule_patterns)
        self.second_ruler.initialize(get_examples, nlp=nlp, patterns=self.second_rule_patterns)

If I understand what you are saying then I need to run either spacy train or spacy assemble (the latter is unknown to me). I don't want to retrain since I have ~35k documents. If I run `spacy assemble then I get the following error

ValueError: [E143] Labels for component 'textcat_multilabel' not initialized

I did train it and I also have a meta.json with the results so I don't understand that error.

Another approach could be to create a callback which adds the patterns but it seems wrong - I'd assume the component itself should have this logic as with the initialize approach.

[NixBiks]

I just realized the solution. Put it in from_disk instead.

    def from_disk(
        self, path, *, exclude
    ):
        self.first_ruler.add_patterns(self.first_rule_patterns)
        self.second_ruler.add_patterns(self.second_rule_patterns)

[adrianeboyd]

You'd want initialize to take the patterns as kwargs following nlp so that they are only loaded from some local path once, and not every time the pipeline is loaded. You don't want to load from any local paths as settings in __init__ or in settings in the [components] block. Then you would need to make sure that your custom component can serialize and reload the entity ruler patterns as part of to/from_disk/bytes instead of relying on local paths.

After you have modified the custom component, you can put together the final pipeline with spacy assemble, which would use source to include already-trained components like a textcat and would initialize the remaining components.

If you can restructure the pipeline to use top-level entity ruler components instead, this might be simpler to implement overall since they already handle the serialization for you. If your component doesn't have any other state/settings it needs to be aware of, then you could use Language.component instead of Language.factory.

[NixBiks]

I can't use the original entity ruler unfortunately since I extend it, customize it and wrap a few of them. All my patterns are written in python so the easiest solution is the solution above where I only implement from_disk, not to_disk nor initialize. No need to serialize, but when deserializing the pipeline I add patterns (already loaded in python memory) via from_disk. Of course I could also serialize all the patterns to JSON and deserialize from JSON but there is no need really since I write them in python. My custom components with the data hardcoded I guess.

Thanks for your help and please correct me if I'm wrong in my conclusions here.

[adrianeboyd]

Without more concrete details about the custom component implementation: I am not really sure.

If you're generating all the patterns in __init__ without referencing any external data, then you could potentially initialize a phrase matcher from docs created with nlp.make_doc rather than nlp, but this isn't an option for the entity ruler, which only expects string phrase patterns.