Do you have the implementation for your Structured version of GRasP that was mentioned in the paper?
johnrachwan123 opened this issue · 1 comments
johnrachwan123 commented
Do you have the implementation for your Structured version of GRasP that was mentioned in the paper?
StijnVerdenius commented
Hi,
As mentioned in the appendix, structured GraSP didn't work very well. Unfortunately, I don't have the code for that anymore because of that reason.
However, you can reproduce it, by doing the same as SNAP, yet instead using the second-order GraSP saliency scores for the gates instead of the first order ones used in SNAP. Both SNAP and GraSP can be found in the /models/criterions
subfolder