-
Notifications
You must be signed in to change notification settings - Fork 7
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Introduce Tests for the Multi-Transform #6
Comments
What kind of tests would you be interested in? In terms of the performance of fits (RMSEs, MD stability ect.) or that the the transforms work as expected (more along the lines of the test you already wrote). Here are some dimer curves, where we can clearly see that the ACE has different cutoffs depending on the species. I can supply some RMSE plots as-well, but these will only really have meaning if I also optimise setting a species independent cutoff. |
Unit tests should specify what the implementation is meant to achieve. They are just intended to check that the implementation is consistent with what you want it to do. E.g., my test of the symmetric basis checks that the gradient is indeed the gradient of the value, and it checks that it is invariant under actions of the symmetry group. All this for relatively small configurations and small basis sets just for testing. That's the kind of test I'm envisioning. |
For a transform, I test that the inverse and derivatives are correct, and that saving and loading works as expected. But that's about it. Maybe there are other things that should be tested. But what else should be tested for multi-transforms? Suppose that somebody made a PR to change the multi-transform implementation. What do you want to check before you accept the PR? |
I've jotted down some ideas. What we want to test:
How we will test:
site_energies(IP, train_data[1].at)
Create potential with random coefficients: c = rand(length(basis))
IP = JuLIP.MLIPs.combine(basis, c) |
this sounds all very sensible. thanks for the summary and sorry for the slow reply. Give it a try in a PR? |
these new transforms are not currently tested. I'm not sure how, and would be grateful if others can take this on.
@ LarsSchaaf @davkovacs
The text was updated successfully, but these errors were encountered: