Add support for function_to_apply in classification

[coderrg]

For classification models, users should have some way of specifying the function to be applied to the logits rather than softmax always being used since the logits do not always correspond to mutually exclusive classes.

Take, for example, a model's logits representing whether a piece of text conveys negative, neutral, or positive sentiment in a text classification task. In this case, the softmax activation function is apt as it ensures that the probabilities in the final output vector sum to 1. However, if the logits denote whether the text is toxic, obscene, insulting, etc. (as seen in Detoxify models), the sigmoid activation function would be more appropriate since the values in the final output vector do not necessarily have to sum to 1.

HF pipelines used softmax as the mandatory activation function over multi-class model logits until huggingface/transformers#8328, when they added support for the function_to_apply parameter which can be specified by the user.

I am proposing support for a function_to_apply parameter here as well. The options are the same as implemented in HF: softmax, sigmoid, and none (none applies no activation function and just returns the raw logits). softmax is the default since it preserves backward compatibility and seems to be the most common case.

[coderrg]

Thank you for your feedback, @jonatanklosko! I have implemented your suggested changes.

[jonatanklosko]

Awesome, thanks :)