Reproducing standard benchmarks using Transolver++

[xjffff]

Hi, and thanks for the great work!
I’m having trouble reproducing the reported Transolver++ results on the six standard benchmarks. My understanding is that Transolver++ (a) reduces in_project_fx and (b) replaces the softmax with Gumbel-Softmax. Using the training hyperparameters from the paper, my runs do not outperform the original Transolver.

1. Are there additional architectural or training changes beyond reducing in_project_fx and using Gumbel-Softmax?
2. Could the reduced parameter count require capacity to be added elsewhere?
3. Has anyone successfully matched the paper’s numbers, and if so, could you share the training scripts and configs?

Thanks in advance!

[yufu-liu]

Hi @xjffff,

Although it has been a while since your post, I just done the same task recently.
The results of my trial are Plasticity: 0.0019, Elasticity: 0.011, Darcy: 0.015, Airfoil: 0.009, Pipe: 0.005, NS2D: 0.18.
(Almost twice larger than the paper's)

Do you have any further improvement?

Thanks!
Liu

[wuhaixu2016]

Hi, thanks for your interest. We have released the checkpoint here: https://drive.google.com/drive/folders/1hN-5hh-i4nZR8chcnBvyjdRTQAd43zW8