Captum: n_perturb_samples on captum.metrics.sensitivity_max

[gkakogeorgiou]

The n_perturb_samples variable is also used in Lime definition. This means that you can not use n_perturb_samples to **kwargs on metrics.sensitivity_max, in case you want to use Sensitivity algorithm on Lime for specific n_perturb_samples. I propose to change the name of n_perturb_samples on sensitivity_max algorithm.

[NarineK]

Thank you for bringing this up, @gkakogeorgiou !
n_perturb_samples is used in both in sensitivity_max and infidelity metrics.
For Lime and KernelSHAP we can probably rename that variable to n_surrogate_samples ?
What do you think, @vivekmig ?

[vivekmig]

Yeah, this is a good point! We should avoid argument name overlaps with metrics. Could we possibly just use n_samples in Lime and KernelSHAP, which would match other attribution methods?

[NarineK]

n_samples sounds good. One thing only - n_samples is also used in noise tunnel. It is less likely that someone will use Lime and KernelSHAP w/noise tunnel but just in case ?