petarv- / dgi Goto Github PK

Hi Petar,
I have read your idea abot Deep Graph Infomax.
While right now something really puzzles me: if I want to add features for my edges and I still want to do unsupervised or self-supervised learning. Can I just simply change the encoder and remain all the other function unchanged? Such as corruption function or readout function?

Hello, the paper said: "an explicit (stochastic) corruption function (\tilde{X}, \tilde{A}) = C(X, A)."
However, in the code, I only find the corruption on node attributes:
idx = np.random.permutation(nb_nodes)
shuf_fts = features[:, idx, :]

I can not find the corruption on the graph structure, why? Does this not affect the final result?

thanks

yes this is a brief cora experiment,
but the data is so different from original data set cora.
could anybody explain what they mean?
thanks a lot!

I tried to run the released execute.py on Pubmed. However, it seems that it takes 19.25 GB during back propagation.

Is this the correct behavior? Is there any solution to bypass this problem and replicate the paper reported number?

I saw the DGI implementation which only contains Cora dataset on your github

I was wondering whether you could kindly share your DGI implementation which contains Reddit and PPI dataset!

Thank you very much！

great paper!
I have a question about the objective function, is it same as cross_enpty?
looking forward to your reply!

Hi Petar,
Can you provide the time cost on Reddit and PPI datasets?
Now I am doing some work to scale up Graph Auto Encoder, and I wonder how efficient is this method?

Thanks

HI, great work ! But I have a question regarding how to repeat the experiments.
While getting the mean accuracy, you used the fixed representation and only repeated the logistic regression part. Shall we repeat the whole pretext training + downstream task altogether ?
Or is there any reference for the reason of doing so ?
Many thanks

Hi,

Thank you for making the code available. I would like to ask a question about a remark you made in your paper about the PPI dataset. On page 8, paragraph "Inductive learning on multiple graphs", you noted that 40% of the node has all-zero feature vector. However, the feature vectors loaded using GraphSAGE (http://snap.stanford.edu/graphsage/) is dense. Did you use a different set of feature vectors or a different PPI dataset?

Thank you for your time! If I misunderstood something, please kindly point out my mistake.

Edit: Sorry I made a mistake when I check the feature vectors. It was indeed 42% all zeros.

What is the kind of your dataset(cora).
How can i make my graph like this.
thank you very much.

Great work! I really enjoy reading the paper.

However, will the codes to replicate the reported performance on reddit be released? If so, is there a planned schedule?

Thank you!

Hi,

Can you please share your code for training and evaluating PPI dataset as well?

Thanks

Hi, I was wondering that the learned representations tend to conserve their unique information or common information. Maximizing Mutual information between patch and summary vector is to find more information between them, but the discriminator wants to distinguish the samples. So, I am confused.

Hi, I think there's an error in AvgReadout with a mask. The mask summation should be performed along the first dimension only.

It is
return torch.sum(seq * msk, 1) / torch.sum(msk)
but should be
return torch.sum(seq * msk, 1) / torch.sum(msk, 1) # Note the dimension here.

It seems that the sigmoid function is missing in layers/discriminator.py[line30].
Explained in your paper, the logistic sigmoid nonlinearity is used to convert scores into probabilities of (h_i, s) being a positive example.

Hello, how do you make the data set under the data folder?