Official code for "InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization" (ICLR 2020, spotlight)

Thanks for your work first!

I am testing on unsupervised graph classification. I saw in the paper experiments on 6 well known dataset: MUTAG, PTC, REDDIT-BINARY, REDDIT-5K, IMDB-BINARY and IMDB-MULTI were conducted. Among these datasets some have node/edge labels (like MUTAG), while some just contain necessarily three files A.txt, graph_indicator.txt and graph_labels.txt (like IMDB-BINARY, IMDB-MULTI). For dataset with node labels, I see the code initialized node feature h(0) as a one-hot vector from the node label. While when I was testing with dataset without node labels, I met the "float division by zero" error when initializing the Linear() in the first gc_layer in the Encoder class of 'gin.py'.

I tried to debug based on current code and found that this is because the input 'num_feature' is 0 when there are no node labels provided. I am wondering how do you initialize the node features on the dataset without node labels, for example using the node degree instead? I also tried to find additional parameters and different handling this exceptions, while I didn't find it. I am wondering could you please possible give some suggestions or see if there is any modification that should be made when testing on dataset without labels, based on the current code? Thanks!

Hi, we notice that there is a gap between the performance on the QM9 dataset reported and that we reproduced using your code. e.g. for HOMO(2), the reported RMSE is 0.0060, while with your code it's 0.1560.

Hi @fanyun-sun,

Thanks for your nice work and released code first!

When I ran the code on the Reddit-BINARY dataset, I got the problem as follows:

Could you pls help me with the bug? Thanks in advance!

Thanks for sharing you work, it's interesting and solid. But I want to know what is the exact from of mutual infomation in equation(8)'s third term. Thanks!

I'm curious about the utility of PriorDiscriminator and GlobalDiscriminator, it seems that in your default parameter setting the code set prior=False and does not use PriorDiscriminator to cacluate part of the loss

Hi @fanyun-sun,

Thanks for your released code! I have two minor questions about the reported results as follows:

1. When I ran the code, the saved log file reported the performance of 20 epochs. I wonder how to obtain the reported results in the paper (e.g. 89.10 for MUTAG dataset).

Here are the printed results for MUTAG dataset: "linearsvc": [0.7950584795321638, 0.8650584795321636, 0.8351169590643275, 0.868216374269006, 0.8592397660818714, 0.8723391812865497, 0.8939766081871345, 0.8783625730994153, 0.8770175438596493, 0.855877192982456, 0.8981286549707603, 0.8586549707602339, 0.8731578947368421, 0.8789473684210527, 0.8550584795321639, 0.8631286549707602, 0.8656432748538011, 0.8717251461988305, 0.8503508771929825, 0.8814327485380117, 0.8831286549707602],

2. Could you pls explain the meaning of 10-fold cross-validation accuracy for me? Maybe you can point out the specific implementation in the code.

Thanks a lot!

When I look at :

 error += (model(data) * std - data.y * std).abs().sum().item()  # MAE

I am curious what the .sum is doing here? Aren't these just the errors for each entry in the batch, why do you need to sum them?

Hi,

I find that as the training goes (beyond 20 epochs), the loss will gradually become negative. May I ask if this is harmful to downstream tasks?

Thank you!