-
Notifications
You must be signed in to change notification settings - Fork 10
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
using inverse of gap_fit error as weights #19
Comments
The command line arguments are documented here. They are not very relevant, since I'm just talking about the julia fitting functions so far, and a fitting "executable" can do whatever it wants to convert from command line arguments to julia function arguments. More importantly (and not documented on that page), the per-configuration is in four separate info fields, |
that should be enough to get us going - thank you! |
I think step 1 is understanding the difference how GAP and ACE weight the training data. Then it should be easy to implement the translation. For ACE look at these:
|
Having looked into it, it seems there are at least two separate tasks.
Does that sound right to start? If so, I'll begin a pull request and we can continue discussing there. |
yep, that sounds right to me. But make sure that the weights are really 1/sig^2 and not 1/sig? @casv2 ??? |
Ah you're right it's w^2=1/sig^2, thanks. Last question: on what branch should I work? Is the DEV-v0.8.x business contained to the main ACE repo, or is there a similar split here? |
Yes you should start from dev-v0.8.x and merge back to that branch |
@cortner says this is in the latest IPFitting, but needs to be back-ported to the v0.10 (which is the ACE1-compatible successor to v0.4) to be available for ACE1, which I'd like. |
I think what's necessary is for me to tidy #22 and rebase those changes onto several places, including v0.10. @bernstei, how flexible should things be? Can I restrict the user to providing only weights or only inverse weights? For the per-config values, should there be an |
Certainly for any given config I'm happy to require that only one of weights and inverse weights are specified. That'd be OK for the entire set, too, but it's more restrictive and I'd be predisposed to not require that unless it really simplifies something. |
It would be nice if ipfitting could support (more or less) the syntax of
gap_fit
for its energy/force/virial error, interpreting the weight as the inverse of the error. For this, there would be a global set of default errors, and an optional per-configuration value in the xyz info field.The text was updated successfully, but these errors were encountered: