`batch_size` potentially misleading

[gsarti]

Hi all,

With the 0.4.0 release ferret now supports the usage of batching for methods requiring multiple steps of approximation, such as Lime and Integrated Gradients.

However, the name of the argument, batch_size, can erroneously lead users to believe that attributing multiple examples at once is possible, while it does not seem to be the case. In Captum, the same argument is referred to as internal_batch_size to disambiguate inter-example and intra-example batching.

Do you think it would be more sound to stick to the internal_batch_size nomenclature in this context? This could be especially useful if multi-example batching will be supported in the future.

[g8a9]

Hi! Good catch!

Indeed, we now support batched inference for multi-step methods and plan to support multi-sample attribution in a future release :)

I'm not sure what's the best option here. We are currently passing call arguments to explainers (e.g., for gradient attributions) so as to speak on could already use internal_batch_size.

I like the idea of keeping the lib easy and accessible even to non-experts. In that sense, one could just use batch_size to refer to the multiple steps of approximation, even in the case of multi-example requests. Even though we should explicitly handle the case where both batch_size and internal_batch_size 🤔

Anyway, yours is a good point we'll be addressing in the new task API we are building these days :)