Giters

microsoft / AI2BMD

AI-powered ab initio biomolecular dynamics simulation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Results Reproduction

ShihaoShao-GH opened this issue · comments

commented

Many thanks to your awesome Geoformer which gave the community another solid, and insightful work. But unfortunately I encountered some results reproduction issue with it.
I downloaded your energy_U0.ckpt and testing on it only gave me 0.0054 that is different from the number 0.0043 in your paper. I strictly followed the guide in this repo and the only thing that seems obvious different to me is that I used torch=1.10.0+cuda11.3. But I didn't feel like it would be the case leading to the issue. Can you give me some hints about this? Thanks in advance!

commented

Hi Shihao,

Sorry for the delay. We rerun the inference process and could reproduce our results:
image

Maybe the cuda/torch version influences the final inference results.
Also u could retrain the Geoformer on QM9-energy U0 following the instructions in README.

Best,
Yusong and Shaoning

commented

Hi Shihao,

Sorry for the delay. We rerun the inference process and could reproduce our results: image

Maybe the cuda/torch version influences the final inference results. Also u could retrain the Geoformer on QM9-energy U0 following the instructions in README.

Best, Yusong and Shaoning

Thanks for the response! I will go create an exact same environment to see if everything works well.

commented

It works. Thanks for your help, Yusong and Shaoning!