If I want to use the SequenceClassifier pipeline for something like reranking, I am (sort of) able to do so using the exposed forward_t method. The problem is that I will need to first encode the inputs using the model's tokenizer. I can get a ref to the tokenizer using get_tokenizer, but if I want to pass in tokenizer params (IE max_len and device) to tokenizer.tokenize, I cannot get them from the SequenceClassificationModel, because they are private fields and there are not any get methods like there are for the tokenizer itself.
Alternatively, you could add a method to wrap calls to SequenceClassificationModel.tokenizer.tokenize and pass these parameter in from the model instance.
If I want to use the
SequenceClassifierpipeline for something like reranking, I am (sort of) able to do so using the exposedforward_tmethod. The problem is that I will need to first encode the inputs using the model's tokenizer. I can get a ref to the tokenizer usingget_tokenizer, but if I want to pass in tokenizer params (IEmax_lenanddevice) totokenizer.tokenize, I cannot get them from theSequenceClassificationModel, because they are private fields and there are not any get methods like there are for the tokenizer itself.Alternatively, you could add a method to wrap calls to
SequenceClassificationModel.tokenizer.tokenizeand pass these parameter in from the model instance.