Dispatching Python covariance matrices

[siranipour]

We now dispatch covmats purely from python. Addresses various points including #1476 (comment) and #1475 (comment)

Let's see what the tests do

[scarlehoff]

I've rebased your commit on top of #1430 to see whether the test passes also there https://github.com/NNPDF/nnpdf/runs/4333530141?check_suite_focus=true

[siranipour]

Looks like the linux tests are passing on the rebase

[Zaharid]

Where has all this code and documentation gone? How are the norm_threshold and use_weights_in_covmat arguments handled?

[siranipour]

The weights are used in the lower level functions

nnpdf/validphys2/src/validphys/covmats.py

Line 35 in eeae19b

use_weights_in_covmat=True,

[Zaharid]

I am a bit worried that this is going to obscure --help for various things a bit too much. Admittedly that is because reportengine sucks in that regard. That could possibly be remedied with a tutorial on chi²s which we clearly need for other reasons.

[Zaharid]

It would be good to have tests that cover the various options passed to the covmat. E.g. the thing with the norm_threshold clearly should have failed something.

[siranipour]

I am a bit worried that this is going to obscure --help for various things a bit too much. Admittedly that is because reportengine sucks in that regard. That could possibly be remedied with a tutorial on chi²s which we clearly need for other reasons.

Yeah I couldn't figure out how to do this without having a dummy function which the explicit_node could dispatch on. Let me know if you can think of a better alternative, I was thinking perhaps I've missed something obvious

It would be good to have tests that cover the various options passed to the covmat. E.g. the thing with the norm_threshold clearly should have failed something.

will add now

[Zaharid]

One thing I'd like to do someday is have a more restricted version of explict_node that looks like

produce_covmat_t0_considered = dispatch_node(on="use_t0", {True: covmats.t0_covmat_from_systematics, False: covmat_from_systematics})

which does the same thing but additionally knows how to follow the possibilities and produce help.

[siranipour]

Yes that would be nice, since it could then dispatch another dispatch_node which is the problem I was having here.

[scarlehoff]

Should I / Can I rebase #1430 on top of this one already or are more changes needed?

[siranipour]

I’m done with it if @Zaharid is happy On 29 Nov 2021, at 10:14, Juan M. Cruz-Martinez ***@***.***> wrote:  Should I / Can I rebase #1430<#1430> on top of this one already or are more changes needed? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#1477 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AKMAJEG3OFDACMHS4FZTF4TUONHAJANCNFSM5I2PCJUQ>.

[Zaharid]

It seems ok. I would hope that the documentation challenges are addressed by #1478.

[siranipour]

Shall we merge this?

[scarlehoff]

I'm in favour of merging since everything seems to work in this one and also down the line (#1430 )