How can we customize pipeline?

🤗Transformers
1

How pipeline handles custom pre-processing or post-processing. For example, different doc_stride, max_seq_length in Question Answering?

2

I dont think that is possible. But lets hear from someone experienced.

3

Hi @s4sarath

it’s possible to pass different doc_stride, max_seq_length etc directly to the __call__ method, i.e when calling the pipeline, have a look at the docstring of __call__ of QA pipeline here

4

Agree. But I am asking more of like a custom preprocessing. Let’s say I want to add token type I’d 2 to question and 0s to context .

5

In that case, you could subclass the pipeline to add the custom behavior.

6

That’s what I was looking for .