use error_members instead of rawdata

[scarlehoff]

This is straightforward for MC PDFs (they are the same thing now and this change will make the change from all_members to only_replicas very simple) but for Hessian PDFs is instead very dangerous since the error members are the rawdata-1.

Making this as a separate PR since it makes many changes in the code but are almost a sed (a manual one nonetheless) and because I will be running the fit bot quite a bit I guess.

[scarlehoff]

@Zaharid are you using the paramfits package? Is it tested / are there runcards that I can use as an example that things don't break? (the mean, std and raw .data is used a lot there)

[github-actions]

Greetings from your nice fit 🤖 !
I have good news for you, I just finished my tasks:

• Fit Name: NNBOT-3614a0b64-2022-02-14
• Fit Report: https://vp.nnpdf.science/_IzljX4FT6mJEQP-NPKHJw==
• Fit Data: https://data.nnpdf.science/fits/NNBOT-3614a0b64-2022-02-14.tar.gz

Check the report carefully, and please buy me a ☕ , or better, a GPU 😉!

[scarlehoff]

oh, that's a pontetially important one, didn't realise it was paramfits stuff.

Maybe using Stats everywhere will create the parabolas :P

[Zaharid]

Sorry, this one https://vp.nnpdf.science/dKVAidYgQRW9hPBGQ9z3Rw==/input/runcard.yaml

[scarlehoff]

The error seems to be reportengine not knowning what to do with a numpy array here: https://github.com/NNPDF/reportengine/blob/fe736351c94b1f0a24b9c695da5bffafa1e5e98b/src/reportengine/floatformatting.py#L80 I think this case should be taken into account (maybe with a nicely-formatted error) since it seems to accept them.

That said, the fact that now they are arrays (of one element) when they used to be floats is a side-effect I was not expecting so I'm actually happy it failed.

[scarlehoff]

Sorry, this one https://vp.nnpdf.science/dKVAidYgQRW9hPBGQ9z3Rw==/input/runcard.yaml

Ok, for now it seems to work and generate the same thing. Thanks!

[Zaharid]

The error seems to be reportengine not knowning what to do with a numpy array here: https://github.com/NNPDF/reportengine/blob/fe736351c94b1f0a24b9c695da5bffafa1e5e98b/src/reportengine/floatformatting.py#L80 I think this case should be taken into account (maybe with a nicely-formatted error) since it seems to accept them.

That said, the fact that now they are arrays (of one element) when they used to be floats is a side-effect I was not expecting so I'm actually happy it failed.

How did you trigger that?

[scarlehoff]

How did you trigger that?

from reportengine.floatformatting import ValueErrorTuple
import numpy as np 
aa = ValueErrorTuple(np.array([3]), np.array([1]))
print(aa)

I don't think throwing an exception here is a bug as it did help me, but naively I would think it is a common enough situation that warrants its own exception.

[Zaharid]

I'd say it is behaving as expected. Not adding that special case.

[Zaharid]

This looks good as well. Not testing too much manually but then again hopefully we have automated tests for most stuff.