Input Invariance misses model input bias shift

[dkrako]

Description

Currently the InputInvariance metric just adds a bias to x_batch. According to the original paper, you should then add an offset in the opposite direction to the model input bias. Attribution methods are input invariant, if they don't change under this setup.

The current implementation misses the second half, so it is very unrealistic that the metric actually returns True for any setup.

Here's the reference: https://arxiv.org/abs/1711.00867

Minimum acceptance criteria

• Add offset to model input bias.
• Add at least one test case parametrization where InputInvariance returns True