InputInvariance for CNN

[h-moghaddam]

I'm trying to evaluate different explainability methods using the InputInvariance metric. I have faced some issues:
1- If there is a sequential container in pytorch implementation the first layer needs to be specified more precisely. something like : module = modules[1][0] at here.
2- I can't understand why torch.unique() has been used here and here.
3- I expect the model's output to remain the same when you pass the shifted input from the modified model.

Minimum acceptance criteria

• The output of the original input passed through the original network is the same as the shifted input passed through the modified network
  @annahedstroem