-
Notifications
You must be signed in to change notification settings - Fork 6
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
Re-enable use_t0
flag.
#1599
Re-enable use_t0
flag.
#1599
Conversation
The point is that we wanted to have two separate options for using t0 in the sampling and/or in the fitting. This is why now the |
It is for things like vp reports where fitting/sampling don't apply. I'm actually wondering whether it shouldn't be |
If it is only for reports yes, most likely it should be true by default |
I'd say this should work as it has been documented (i.e. as in this PR), and the other change may have inadvertently corrupted results, e.g. anyone who has been computing chi2 vs experimenal chi2 recently (@RoyStegeman, @J-M-Moore ?). Similarly changing the default is going to be very disruptive for existing runcards, so please don't. I'd say |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'd prefer if the default was as it was.
fwwi this is the documented behavior https://docs.nnpdf.science/figuresofmerit/index.html#avoiding-bias-t0-method |
Uh, didn't realize (since for the fits it defaults to True). I'll roll it back (and I'll change the
I think only in the case: t0pdfset: someset
use_t0: false and only if the covmat used doesn't go through config. So i hope not |
Greetings from your nice fit 🤖 !
Check the report carefully, and please buy me a ☕ , or better, a GPU 😉! |
Ok, it works (compare with last report) https://vp.nnpdf.science/IzbRlZ4IQ8qbZddbo0B8dw==/ At some point we should really tackle #1025 in a systematic way. The current reference is way too old. |
Co-authored-by: Roy Stegeman <roystegeman@live.nl>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I don't believe the current behaviour is what we want either. If I say use_t0: False
, use_t0_sampling: True
I want the chi2 to be the experimental one (and the sampling flag to be ignored for that purpose). Also I'd think use_t0 should be the same as use_t0_fitting.
Maybe a better way to go about this, both making it nicer and maintaining compatibility is to remove |
Something something about not discussing policy #1528 (comment) In all seriousness, better to have the default behavior being "use t0 for the fit" given that the runcards for NNPDF4.0 are already "out there" and don't include |
Didn't understand that last comment (neither the serious part nor the quibble). Indeed I think the defaults should be as they were for 4.0, even if the whole thing is a bit convoluted. That means that n3fit implicitly defaults to using t0 (but we may want to make sure you can disable that, e.g. by having the script set Things that we should avoid are that the hundreds of runcards like https://vp.nnpdf.science/RGfy_dVZQa6xYT72MFW_7w==/ suddenly start failing screaming about some t0 pdf, or that the vp reports for fits with funny sampling configuration start computing the t0 chi2 inadvertently. |
The first version behavior is equal to the current version. To your specific point:
In this case I would say you want the
This is what this PR fixes (wouldn't that runcard fail now because of
The vp report should only look at |
The fundamental problem is that the thing that could be None to indicate that t0 was not used only makes sense if you can assume that t0 is either used or not in the whole namespace, which is not the case if we want to also have individual control over the sampling covmat. Hence why believe the design should be changed to explicit nodes (or see if we actually need the new functionality). |
We needed because we are sampling the replicas with python now. We didn't need it before because the replicas came from
Feel free to change it, but I think we should quickly fix the current buggy behaviour in master. |
I don't see how the new behaviour is not buggy. |
Then please fix it so it stops being buggy. |
The current version will raise a Please check if the current version satisfies all the interested parties as I would like to fix the current master (which is bugged for sure). If not please post a very specific example of something broken so I don't need to try out different combinations in order to guess what the problem is. For new designs please open a new issue where it can be discussed. |
@Zaharid any news here? |
Closed in favour of #1626 |
See #1489 (comment)
It is in principle a trivial change that should not do anything, adding @andreab1997 in case there was a reason why we wanted to eliminate
use_t0
that I've forgotten.