search

Shen-Lab / LDM-3DG

[ICLR 2024] "Latent 3D Graph Diffusion" by Yuning You, Ruida Zhou, Jiwoong Park, Haotian Xu, Chao Tian, Zhangyang Wang, Yang Shen
GNU General Public License v3.0
28 stars 6 forks source link

The environmental problem about GPU version #1

Closed BaruaBee closed 3 months ago

BaruaBee commented 3 months ago

Hi Authors,

Congrats on the great work!

As I was reading your team's paper about 3DG, I want to reproduce your excellent work. However, I noticed that you didn't specify the GPU version used to train your model in the paper. Additionally, I observed that the default batch size is 512 and the number of epochs is 10,000. These numbers seem larger than those used in other models I have encountered before. I couldn't find the exact numbers you used to train the unconditional QM9 model in the paper. Therefore, I am wondering: What GPU version did you use to train the unconditional QM9 model, and are the batch size and number of epochs really 512 and 10,000?

Thanks!

yyou1996 commented 3 months ago

Thanks for your interest!

We use the A100 GPU. (40GB). Our latent space here is of fixed dimensionality (i.e. they are just vectors) that bs=512 is affordable. For epoch=10000, we set it as maximum and observe that around thousands the model can converge.

BaruaBee commented 3 months ago

Thanks so much for your reply!