search

danielzuegner / nettack

Implementation of the paper "Adversarial Attacks on Neural Networks for Graph Data".
https://www.cs.cit.tum.de/daml/forschung/nettack/
MIT License
218 stars 56 forks source link

Regeneration of results from the paper #4

Closed farezaur closed 5 years ago

farezaur commented 5 years ago

Hi,

Can you guys give any codes/procedures/parameters to regenerate the result on the paper for GCN/CLN/DW?

Thanks

danielzuegner commented 5 years ago

Hi,

right now I don't have the whole pipeline ready for publication as it's pretty specific for my setup. I hope the procedure we followed in generating the results is mostly clear from the descriptions in the paper -- if you have any questions there feel free to reply here or send me an e-mail.

The code for GCN is included in this repo. For CLN (https://github.com/trangptm/Column_networks) and DeepWalk (https://github.com/phanein/deepwalk) we used the code provided by the authors with the standard hyperparameters. Let me know in case anything is unclear there.

Best, Daniel

farezaur commented 5 years ago

Thanks for the reply. So, for the GCN experiments you used the code for the surrogate model that exists in this repo? Not the code in the original GCN repo. Is that correct? Just want to make sure which one should I use.

Thanks in advance Reza

danielzuegner commented 5 years ago

Yes. But make sure to pass with_relu=True so that you don't use the linearized surrogate model.

farezaur commented 5 years ago

Hi, I can't seem to run the CLN from their github. And it seems they did not reply to the issues there as well. Can you tell me how did you run their code? It seems the version of keras, tensorflow, numpy etc has to be same as the authors for running the code.

danielzuegner commented 5 years ago

I have a "fixed" (but pretty brittle) version of the authors' implementation. I have uploaded it here. See "How to run.txt" on which package versions I used to get it to run.