[Sentence Transformers] Complete the update of model_type

[JingyaHuang]

What does this PR do?

• Improve the update of model_type, with the previous update, the model_type remains as the previous one (eg. bert) when displaying the config (which is relatively ok):
  

But if we apply config.to_dict(), the model_type is not changed, so I would suggest to complete the change of model_type thoroughly.

• The clip models in sentence transformers is not fully supported nor tested (the one in the test set actually goes for sentence-transformers-transformer instead of sentence-transformers-clip)

[fxmarty]

Thank you for the notice, would this suggestion work?

[fxmarty]

Suggested change

	model.config = model[0].auto_model.config
	model.config.model_type = "sentence-transformers-transformer"
	model.config.__class__.model_type = "sentence-transformers-transformer"
	model.config = PretrainedConfig.from_dict(model[0].auto_model.config.to_dict())
	model.config.model_type = "sentence-transformers-transformer"

[JingyaHuang]

nope:

[fxmarty]

Suggested change

	model.config = model[0].model.config
	model.config.model_type = "sentence-transformers-clip"
	model.config.__class__.model_type = "sentence-transformers-clip"
	model.config = PretrainedConfig.from_dict(model[0].model.config.to_dict())
	model.config.model_type = "sentence-transformers-clip"

[fxmarty]

Right, the issue comes from https://github.com/huggingface/transformers/blob/881e966aced6f0f208f43d7b7e7e55bc680f8fa5/src/transformers/configuration_utils.py#L901-L902. This sounds like a bug in Transformers to me, self.__class__ is for some reason taking priority over self.

In the meantime we rather fix with:

class SentenceTransformersTransformerPretrainedConfig(PretrainedConfig):
    model_type = "sentence-transformers-transformer"

class SentenceTransformersClipPretrainedConfig(PretrainedConfig):
    model_type = "sentence-transformers-clip"

to avoid modifying class attributes?

[JingyaHuang]

close as this will be tackled in #1647