Continuing discussion about enum/classes for model types

[dbirman]

@bruno-f-cruz copying the discussion from #1385 since I'll close that issue soon.

For reference, this is related to https://github.com/AllenNeuralDynamics/aind-data-schema/issues/960. I raised this exact issue here (https://github.com/AllenNeuralDynamics/aind-data-schema/issues/960#issuecomment-2155055024). While this PR solves the specific problem, please consider the larger issue: These enum-like classes can't be easily hashed, and this will always present problems when trying to use Set or Dict, which is a big limitation IMO.

    Yes, same issue. But we're still trapped by the goal/requirement of keeping information about registries attached to their corresponding objects

A few possible solutions come to mind:

    Modify the serialization/schema building behavior to "mock" it as a simple enum (this API is a bit annoying, but we have done it in the past https://github.com/AllenNeuralDynamics/aind-behavior-curriculum/blob/e18289f084ba325df0bfe3c3ada79a5cbf18fcdc/src/aind_behavior_curriculum/curriculum.py#L53). It should be noted that this may not even fix the problem completely depending on what class you deserialize to.
    Attach the information as metadata via annotated (this is a bit clunky...):

from aind_data_schema_models.modalities import Modality
import enum
from typing import Annotated, get_args

class StrModality(enum.Enum):
    BEHAVIOR = "behavior"
    VIDEO = "video"

StrModalityWithMetadata = Annotated[StrModality, Modality]

def materialize_metadata(value: StrModality):
    return get_args(StrModalityWithMetadata)[1].from_abbreviation(value.value)

print(StrModalityWithMetadata.BEHAVIOR)
print(materialize_metadata(StrModality.BEHAVIOR))

a = {StrModalityWithMetadata.BEHAVIOR: 1, StrModalityWithMetadata.VIDEO: 2}
print(a)

    If the information is not necessary during schema building, and only later during analysis, one simple way to go about this is to add a single method (e.g. materialize_enum_reference(modality: Enum.Modality) -> FullyReferencedModality). This method decouples the ontology from the code that generates the fully referenced structure. The fully referenced structure could be fetched from a separate database (or even from the library itself, worst-case scenario).