Implemented `RobertaForSequenceClassification`

[Seun-Ajayi]

Proposed changes

Implemented RobertaForSequenceClassification in the roberta.py module. Added RobertaClassificationHead along the line and added SequenceClassifierOutput class to the modelling_outputs.py script since it is the output for the sequence classifier.

Types of changes

What types of changes does your code introduce? Put an x in the boxes that apply

• [x] New Model Architecture
• [ ] Bugfix (non-breaking change which fixes an issue)
• [ ] New feature (non-breaking change which adds functionality)
• [ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
• [ ] Documentation Update (if none of the other choices apply)

For multi_label_classification problem type, BCEWithLogitsLoss is used but this is not present in mlx-core module

            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

so I had to use binary_cross_entropy

            elif self.config.problem_type == "multi_label_classification":
                loss_fct = nn.losses.binary_cross_entropy()
                loss = loss_fct(logits, labels)

How else could I have handled this?

[ToluClassics]

Can you please add a test case?

[Seun-Ajayi]

Implemented RobertaForTokenClassification and RobertaForQuestionAnswering too.

CrossEntropyLoss method takes ignore_index as an argument but this parameter is not included in mlx's version of it, so I just did without the parameter. This is in the RobertaForQuestionAnswering class

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)

My workaround

            loss_fct = nn.losses.cross_entropy()

[Seun-Ajayi]

Can you please add a test case?

I would

[Seun-Ajayi]

We are good now

[ToluClassics]

We are good now

Thank you!