Avoid using uncessary get_values(MODEL_MAPPING)

[ydshieh]

What does this PR do?

When we switch to torch 2.2 in daily CI, natten causing issue and I temporarily disabled its installation. However, some test files use model_class in get_values(MODEL_MAPPING) which will fail with

RuntimeError: Failed to import transformers.models.dinat.modeling_dinat because of the following error (look up to see its traceback):

even if those models are not requiring natten.

The usage of

get_values(MODEL_MAPPING)`

is slow and unnecessary, we can just use the names (string) rather than import them to get the classes.

This PR changes them to use the class name strings.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ydshieh]

cc @NielsRogge For future vision models, let's avoid using get_values() 🙏

[ydshieh]

Hmm, sorry there is still a tiny issue. Working on it

[ydshieh]

Using stuff like MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES.values() is not reliable, as an element in MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES.values() maybe a list itself. For example, we have

        (
            "perceiver",
            (
                "PerceiverForImageClassificationLearned",
                "PerceiverForImageClassificationFourier",
                "PerceiverForImageClassificationConvProcessing",
            ),
        ),

in MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES.

It's probably better to have something like _get_names_for_values: similar to get_values in src/transformers/models/auto/auto_factory.py but only return names without loading the classes.

I could do that in a follow up PR, or any better suggestion.

[ArthurZucker]

yeah it's faster also to not load the classes and just compare strings

[ArthurZucker]

Love that it is simpler 😉

[ArthurZucker]

yeah it's faster also to not load the classes and just compare strings