Add RoBERTaConfig to TransformerConverter

[bhavnicksm]

Hey!

I did a simple hack to add support for RoBERTa models that I have been using for a bit. Please give me feedback if this makes sense or if this is wrong as well.

Thanks! 😄

[michaelfeil]

@bhavnicksm The code between Roberta and Bert is quite different - e.g. compare to https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py and https://github.com/huggingface/transformers/blob/main/src/transformers/models/roberta/modeling_roberta.py - I am sorry to say that your PR likley will need quite some more work to support Roberta.

[mhillebrand]

What if you did this instead?

@register_loader("RobertaConfig")
class RobertaLoader(XLMRobertaLoader):
    @property
    def architecture_name(self):
        return "RobertaForSequenceClassification"

XLMRobertaConfig and RobertaConfig look nearly identical under the hood.