hanjun-dai / graph_adversarial_attack Goto Github PK

When I run
git clone [email protected]:Hanjun-Dai/graph_adversarial_attack --recursive

It says permission denied. I need to git clone pytorch_structure2vec separately

The isTerminal(self) function is in located at graph_adversarial_attack/code/graph_attack/rl_common.py.

def isTerminal(self):
if self.n_steps == 2 * self.n_edges:
return True
return False

When initializing, self.n_edges is set 1, so isTerminal(self) would return true when self.n_steps is 2.

Does this mean that after the two hierarchical actions are completed, the MDP ends, and there is only one state transition for a simulation in run_simulation(self) function which is located at graph_adversarial_attack/code/graph_attack/dqn.py?

If this is the case, isn't it inconsistent with the description "Once the agent modifies m edges, the process stops." in section 3.1 of the paper?

Looking forward to your reply, thank you Hanjun.

I am able to successfully run the graph classification, but when I try to attack the graph, I get this error: undefined symbol: __cudaRegisterFatBinaryEnd

This error occurs with each attack method except for the genetic algorithm which works successfully. The error also occurs when running the node_classification.

I am using CUDA 9.1, pytorch 0.3.1, and cffi 1.11.2

Hi, it seems the code only modifies one edge per graph. Is there a simpler way to change the code to multiple edges modification? I have not found a parameter in the code to control the edge budget. Thanks.

Hi, hanjun. Thanks a lot for your great work! I have a question about the hierarchical Q-Learning mentioned in the paper. In equation 11, there are 2M Q functions and the paper claims only two distinct parametrizations are enough. However, when I read the code /node_attack/q_net_node.py, in line 163, there initialize num_steps Q networks which are different from the paper. Does my understanding have any mistakes? Looking forward to your reply!!

Hello! I meet a problem when running the codes:

In custom_func.py, there is a code : "from _ext import my_lib", however, I can't find this pakage named "_ext", could you tell me how can I get it?

During make the codes, I got a mistake, too:

~/RS/graph_adversarial_attack/code/common$ make /bin/nvcc --default-stream per-thread -I/include -I/mkl/include -I/tbb/include -Iinclude -std=c++11 --use_fast_math --compiler-options '-fPIC' -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_50,code=compute_50 -M src/custom_kernel.cu -o _ext/custom_kernel.d -odir _ext make: /bin/nvcc: Command not found make: *** [Makefile:42: _ext/custom_kernel.o] Error 127

I guess it is because there is no pakage named "_ext", so again I wonder how to get this pakage.

I hope that you could reply, thanks!

Hi, the graph embedding used in the paper is structure2vec embedding. Do you know where there is a python 3 implementation of structure2vec without using c++?

I tried to replace the embedding with 2 two layer GCN of hidden dim 64, but I have not get it working. Theoretically I think it should work. Can you please give suggestions on the embedding?

Thanks.

Is this hash function problematic? It seems that this function is equivalent to key = a_t % base.