finetune

Open wangzeyu135798 opened this issue 4 years ago • 2 comments

Hi: When I use pre-trained model finetune downstream graph_classification dataset RDT-B and RDT-M, I meet one problem. In deep learning finetune process, RDT-B and RDT-M, the train accuracy is very high, sometimes near to 1 while test accuracy is very low, it should be overfitting problem. Have you met this before and how did you deal with it if it does? Thanks.

Jun 28 '21 00:06 wangzeyu135798

Hi @wangzeyu135798 ,

Thanks for your interest in this repo. This is not expected behavior of this repository and we haven't met this. Could please check you package versions and post your python, torch, dgl versions, and reproducible training outputs here? Thanks!

Best,

Jun 28 '21 02:06 qibinc

python 3.7.7 dgl-cu101 0.4.1 In finetune process, I run 100 epochs. When testing rdt-b dataset, the 100th epoch train_accuracy is 0.991 while the test_accuracy is 0.835

Jun 28 '21 03:06 wangzeyu135798