-
Notifications
You must be signed in to change notification settings - Fork 116
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
pretrained models #6
Comments
Hi, I didn't think of it but it is a good idea! However I don't have the checkpoints at the moment. |
Hi @rampasek , Have you tried to finetune the pretrained checkpoints on the OGBG-MolPCBA and OGBG-MolHIV datasets like Graphormer? Best, |
Hi @yanzhangnlp, I briefly tired a basic fine-tuning and didn't get to a competitive performance in the initial experiment. For Graphormer, the authors used FLAG [1] and considerable hyper-parameter search (see Appendix B.2 of the original Graphormer paper). Perhaps that would be needed here too. Best, [1] Kong et al. Flag: Adversarial data augmentation for graph neural networks. arXiv:2010.09891, 2020 |
Hi @rampasek Thank you for your reply! I am also trying to reproduce the results of GPS on the ogbg-code2 dataset. I used the suggested hparams but the results are horrible. May I ask if you have any suggestions to resolve this issue? I attach the training log below for your reference. Best, train: {'epoch': 0, 'time_epoch': 1247.29074, 'eta': 36171.43156, 'eta_hours': 10.04762, 'loss': 8.51734692, 'lr': 0.0, 'params': 12454066, 'time_iter': 0.09783, 'precision': 0.0004024746553718845, 'recall': 0.0008608895973593895, 'f1': 0.0005234065686151103}
|
Hi @yanzhangnlp, I'll look into it in the next few days and try to rerun Here are the logs from 10 runs on which the paper results are based: https://github.com/rampasek/GraphGPS/blob/main/final-results.zip For Sorry for the problems, I'll try to get back to you in the next few days. |
Hi @rampasek , Yes, I use the default config file. Best, |
Hi @yanzhangnlp, Here is W&B with 10 repeated runs (random seeds from 0 to 9) for the default So far it seems to be replicating the curve I posted before (currently the jobs are still running as I’m writing this). I run the code that is on GitHub, no modification at all, I made a clean copy and re-downloaded the dataset. I use Python 3.9, PyTorch 1.10, PyG 2.0.4. I also ran Please let me know if you figure out what may be the issue, it seems to work ok on my end. |
Hi @msadegh97, please see PR #11 for running inference using a pre-trained GPS model. With that all questions in this issue are closed. |
Hi @rampasek ,
Do you have any plan to release your checkpoints for the PCQM4M dataset?
The text was updated successfully, but these errors were encountered: