search

hongzimao / gcn_pytorch

9 stars 3 forks source link

Some questions about Decim #1

Open zhangy0822 opened 3 years ago

zhangy0822 commented 3 years ago

Hi，

We have a project involving graph neural network and reinforcement learning, so we are very interested in Decia. Because we are not familiar with the version of Tensorflow1.x, it is difficult for us to understand your code. I was very lucky to see your pyTorch version of the code, which I think is very helpful.

I see that you are using GCN. Meanwhile, Pytorch-Geometric also provides some interfaces for GCN. What's different about messaging? Is your code the same as GCNCov in Pytorch-Geometric?

I look forward to your reply. Thank you

hongzimao commented 3 years ago

Compared to other implementation, there might be some subtle difference in how we parametrize the message function (equation 1 in Decima paper). When we developed Decima, pytorch geometric library or TensorFlow 2.0+ wasn't there (otherwise we would have used libraries with dynamic graphs). I'm not very familiar with GCNCov, but the if the parametrized message passing function is general enough, GCNCov should contain a family of functions that resembles our message passing scheme. At the end of the day, how you parametrize the message function should depend on your problem structure. Hope this helps. Thanks!

zhangy0822 commented 3 years ago

Ok, thank you very much for your reply. I recently learned this code (gcn-pytorch). Is this part mainly to implement equation 1 in Decima paper?

According to the Decima paper, does the message passing go from leaf to root? The DAG Summary is then computed based on all the nodes in a job. The GCNCov layer in PyTorch Geometric seems to calculate only one massage passing from neighbor nodes, not from leaf to root. I don't know if that's a major part of the message passing scheme that you have designed.

I have another problem that has been puzzling me. What is the significance of 'num_steps' in gcn-pytorch? Is it the same as' max_depth 'in the TensorFlow code? Does it represent the maximum number of message passing from leave to root? How does he control it in training?

I look forward to your reply. Thank you

hongzimao commented 3 years ago

Thanks for the question here too. Replied in https://github.com/hongzimao/decima-sim/issues/37.