NVIDIA
Add Transolver model
Modulus Pull Request
Description
- Add transolver model from ICML 2024.
Checklist
- [x] I am familiar with the Contributing Guidelines.
- [x] New or existing tests cover these changes.
- [x] The documentation is up to date with these changes.
- [x] The CHANGELOG.md is up to date with these changes.
- [x] An issue is linked to this pull request.
Dependencies
- einops 0.8.0
Hi @luohk19 , thanks for your contribution! Have you verified that your implementation reproduces the paper results? If so, could you please post some results here for reference?
Hi @luohk19 , thanks for your contribution! Have you verified that your implementation reproduces the paper results? If so, could you please post some results here for reference?
I've discovered some mismatches in problem setting between modulus's darcy and transolver's. In modulus, each batch is a new data simulated by algorithm. But in transolver's setting, training set and test set are fixed. So I fixes the process of generation and successfully reproduces the paper results as follows.
And I've also added the reproduction code to the commit.
Please add unit tests for your datapipe and model, change the license headers and add exceptions, and update the readme to descripe the difference between two training recipes, how to validate the model, and include a plot for the results.
Please add unit tests for your datapipe and model, change the license headers and add exceptions, and update the readme to descripe the difference between two training recipes, how to validate the model, and include a plot for the results.
I've added unit tests for new datapipe and the transolver, changed the license headers in transolver, and added more details in readme with a plot and a link to the dataset.