Batch-Processing for Noise Tunnel

[yvesrychener]

🚀 Feature

Batch-Processing for Noise Tunnel

Motivation

Methods like IntegratedGradients, which perform multiple steps per input have an internal_batch_size argument. However, Noise Tunnel lacks this as far as I can see. I think it would be quite useful to have a similar argument for Noise Tunnel, since underlying methods like Saliency cannot do batch-processing. This means that OOM issues arise with NoiseGrad-Saliency (and others).

Pitch

Add an argument batch_size_nt, controlling the noise-tunnel internal batch size.

Alternatives

Add a internal_batch_size argument to all attribution methods, but this would be a lot of work.

Additional context

None

[NarineK]

Hi @yvesrychener, that's a good feature. We thought about it but haven't implemented it yet. Feel free to submit a PR if you are interested. It would be good also to create test cases with IG, Saliency, etc and make sure that all tests pass.

[NarineK]

#555 got merged! Closing this issue.
Thank you bringing it up @yvesrychener !