meettyj / nosmog Goto Github PK

Thanks for sharing this interesting work !

Could you share your settings about lamb on different datasets ?

Meet error while running following command: python train_student.py --exp_setting ind --teacher "SAGE" --student "MLP" --dataset "ogbn-products" --num_exp 10 --max_epoch 200 --patience 30 --device 0 --device 1

File "/data/zhou/NOSMOG-master/train_student.py", line 468, in run
position_feature_obs[idx_cur_node_id] = torch.tensor(temp_position_feature, dtype=torch.float32)
IndexError: index 2396430 is out of bounds for dimension 0 with size 2006411

It looks like the code try to access unobserved test_ind idx in a tensor where only observed idx are contained.

We noticed that NOSMOG showed superior performance and inference speed. However, during the training process, it relies on node label information. Can NOSMOG be used for unsupervised or self-supervised graph representation learning and then node clustering? How to improve it?

Dear authors,

What is the difference between this "standard setting" and the transductive setting in table 2? Would it be the standard setting used in [1], where the idx of training/valid/test set is fixed among different seeds?

Furthermore, the release code also contains transductive setting, is it corresponded to the transductive setting in table 2?

[1] Kipf, T. N., & Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907.

Thank you for sharing this work!

Could you please explain what is the difference between the results of NOSMOG (Tab.1) and the NOSMOG/transductive results in Tab.2, if any??

(somewhat similar to:
#2
I saw the issue is closed, but was yet unsure what is the answer).

Thank you!
Niv.