The code requires so much memory

[chandlerbang]

The code requires so much memory

[ChandlerBang]

Hi shaolyy,

Q1: I used GPU V100 (32GB) for Pubmed dataset since ProGNN can only deal with dense matrix. For other datasets, 12GB GPU Memory should be enough.

Q2: The parameter settings of GCN are lr=0.01, weight_decay=5e-4, epoch=200, dropout=0.5 (same as that in the GCN paper). For the performance inconsistency, I just checked my original implementation and found the difference is caused by the different data splits. In my experiments of Pubmed dataset, I mistakenly used get_train_test_val(..., stratify=onehot_labels) while it is get_train_test_val(..., stratify=labels)in other datasets.

Now I've updated the code in train.pyto generate the same data splits as in my experiments. Now you can update this repository and match the performance.

Thanks.

[shaoyf9]

Thank you very much for your reply, I will update the code.