yanghu819 / graph-mlp Goto Github PK

I was trying to run your code with my dataset, and the number of training data is 18000, while my batch size is 500.
The following line can't work with the mentioned parameters because len(rand_index)<len(idx_train):

File: train.py

rand_indx = torch.tensor(np.random.choice(np.arange(adj_label.shape[0]), batch_size)).type(torch.long).cuda() rand_indx[0:len(idx_train)] = idx_train

Here's the error that I get:
File "train.py", line 92, in get_batch rand_indx[0:len(idx_train)] = idx_train
RuntimeError: The expanded size of the tensor (500) must match the existing size (18000) at non-singleton dimension 0. Target sizes: [500]. Tensor sizes: [18000]

Could you please tell me how I can fix it?

After setting the diagonal of adj_label to zero:

The test accuracy for cora goes down to 0.13 with the same configuration:

which is quite confusing since this modification is in accordance with:

Thank you for your work.

Can you also upload the code that check robustness against corrupted connection in inference which you generate fighre 6 for cora and citeseer dataset to compare Graph-MLP and GCN?

Much appreciated

LINE model use the graph structure to generate the node embedding like pre-train, and the embedding is used for multiple downstream tasks. I think this method is only use original feature as the ramdom initialized embedding and combine the two-step training phase into end-to end training phase.

Hello!I would like to ask what is your specific parameter setting of robustness experiment? I did not reproduce your results with the parameter settings provided in the article. Looking forward to your reply.

Hi, your work is really inspiring and I have one question.

In the paper, you were saying the model would be more robust when facing large-scale graph data and corrupted adjacency information, as it utilizes the adjacency information implicitly, rather like GCN which uses adjacency information directly during the information aggregation phase.

However, I am wondering you still use the adjacency information (even multiply 4 times is possible: 4th power of adj) in calculation Ncontrast Loss, how would this maintain robust performance with massive corrupted adjacency information, given you still need the adjacency information in Ncontrast loss in training?

Is that becuase you only need adjacency information during training rather than both train and test phase? Or some other reason to justify?

I am really confused about that and look forward to your reply.

Thanks a lot