Open wanghonghui1998 opened 1 month ago
Hi, thanks for this issue :+1:. You are right, we carried on experiments with 16 trajectories instead of 8 for DINo. I will try to upload the dataset somewhere, I don't know which upload option is the most practical (I am not sure hugging face supports .h5
format). I will let you know today when it is done and add a link to the dataset.
I uploaded it on gdrive here: https://drive.google.com/drive/folders/1RlH4E7-dnKOEA6Zur2EdK-HuEaW_PC8w?usp=sharing Can you tell me if that works for you ? Thanks for asking for the dataset !
Hi,
Thank you for sharing the GDrive link. I apologize for the delayed response.
I've successfully replicated the results for π=100%. However, I'm encountering difficulties with π=20% and π=5%. The performance is slightly worse than what's reported in the paper, especially for π=5% where the model is overfitting when encoding vorticity. I've used the configurations from Appendix Table 5.
Do you have any suggestions for mitigating the overfitting issue or enhancing performance for lower π values?
Thank you for your continued assistance!
Hi there,
Thanks for sharing the code!
I have successfully reproduced the results on the Navier-Stokes dataset. However, I encountered an issue with the Shallow-Water dataset shared in the repository (https://github.com/mkirchmeyer/DINo). The repository contains only 8 trajectories for training, while your paper mentions that there should be 16. I would like to request access to the complete dataset to accurately reproduce the results.
Additionally, I noticed that two implicit neural representations (INRs) are used to learn the height and vorticity for the Shallow-Water dataset, according to your repository (https://github.com/LouisSerrano/coral). Could you also provide the complete configuration files necessary to reproduce these results?
Thank you very much for your time and assistance. I look forward to your response.
Best regards