search

YKQ98 / Matformer

Official code for Periodic Graph Transformers for Crystal Material Property Prediction (NeurIPS 2022)
MIT License
78 stars 8 forks source link

reproduction of MP e_form #12

Closed 292916808 closed 1 year ago

292916808 commented 1 year ago

Hi, thanks for providing the nice codebase! I am currently trying to reproduce the results on MP e_form. However, I can only get a list of ckpts like the one below and the test MAE is around 0.91. The other tasks on MP didn't have a similar issue and the numbers look correct. image image

Do you have any thoughts of this? Thank you!

YKQ98 commented 1 year ago

What learning rate are you using?

292916808 commented 1 year ago

it's 0.01 as default

YKQ98 commented 1 year ago

It is the training failure of some random seed, the validation never gets better after 190 epoch. I suggest you take a different random seed, or decrease learning rate to 0.008 to stabilize it.

292916808 commented 1 year ago

Oh I meant 0.001 as the default value (sorry about the typo). So it would be 0.0008 after decreasing right?

YKQ98 commented 1 year ago

What about using the code directly, with the default learning rate scheduler? I set everything up in the code and it should be fine to use the default settings.

292916808 commented 1 year ago

Yeah, but I was indeed using the default setting... I will try tuning lr a bit and see if I can reproduce it.