Christopher

Hey @dkrako thank you for the PR! I actually had thought of this feature before, so on that end I think it's definitely worth to have! ## Implementation I originally...

Hey @dkrako sorry for the delay. > Regarding your draft prototype, I don't really understand why you are summing up the canonizers. I also don't really understand the purpose of...

> The correct setup of `ctx` should then also be tested, right? Yes, that would be good. I would keep it general and maybe verify using a `DummyComposite` implementation that...

Hey @dkrako > Sorry for the long delay, my schedule was just too tight during the last weeks (but hopefully you will get a new citation soon as I have...

Hey Shivam, check out out the documentation, there is a section on [how-to extract the attribution scores per layer](https://zennit.readthedocs.io/en/latest/how-to/get-intermediate-relevance.html).

Hey Shivam, turns out it is currently a little more involved to do in Zennit. In [this paper](https://arxiv.org/pdf/2109.04236.pdf), they directly modified `zennit/core.py` to also compute the modified gradients wrt. the...

Hey @MaxH1996 I noticed I forgot to add the `grad_sink` parameter for some of the rules that are directly based on `Hook`. (I am thinking doing something else and not...

Hey @MikiFER thanks a lot for writing the issue! This is intended behavior due to how the gradient-modification is implemented. I will explain a little bit why this is the...

Hey @MikiFER , > 1. About these 2 new Identity nodes that "sandwich" ... This is precisely what happens! In the modified backward pass, the gradient will act as the...

Hey @MikiFER sorry for the even more belated response, I was exceptionally busy with my PhD thesis and just was not able to find any time for Zennit. > Could...