Large-scale experiment
dgm2 opened this issue · 2 comments
Hi
Could you suggest how to create an example / or how is possible to have a 1) large-scale 2) more challenging experiment in this library?
E.g.
- large number of sequences (tutorial1 is just curve with 20 points ?) making this large scale is just making more points or one can have also more curves?
- any other features that require significant training to achieve performance (e.g. a few hours) (similar to existing benchmarks you mention in your paper)
For example, how to apply any of those to tutorial1 or tutorial2 etc.
Thanks for your work
Hi @dgm2 👋🏼 Thank you for submitting the issue. Ideed the tutorials are thought to be run in just few seconds or maximum one minute, hence for explanation purposes.
If you want a challenging experiment (for PINNs) try to have a look at our examples, maybe the Stokes problem is a good start. For NOs I would suggest you to download the KS dataset which is a very challenging dataset and try with one of our available neural operators (FNO, LNO, AVNO, DeepONet, MIONet) or make your own one. This dataset has been used in literature for complex analysis (see for example Message Passing Neural PDE Solvers).
Finally, if you want to suggest some problems that might be interesting for you and useful for the community we can think to integrate them in a new benchmark section, directly downloadable from the package (or in the examples directory). Let me know how this sounds for you!
I hope this was helpful for you😃. If you like the package leave us a star ⭐️ which helps us grow the community!