Fix apfel benchmarks

[niclaurenti]

Fixing apfel benchmarks. Tomorrow I will run also the other benchmarks to see if they are broken or not

[felixhekhorn]

• this is a patch to First jets implementation #242 and to repeat what @niclaurenti said live: we should run more benchmarks in Github ...
• @niclaurenti : src/ekomark/benchmark/runner.py:223:8: W0612: Unused variable 'q2s' (unused-variable) from here
• @alecandido

  eko/src/ekomark/benchmark/runner.py

  Lines 249 to 250 in 9a72b85

  	for (q2, ref_pdfs), (q2a, res) in zip(ext["values"].items(), pdf_grid.items()):
  	np.testing.assert_allclose(q2, q2a)

  was the quickest fix I could come up with, but arguably this is not the best solution ... do you have a better idea?

[felixhekhorn]

Actually this is in conflict with #284 we need to decide a merge order - and since this is older: what are we missing to merge?

[niclaurenti]

Actually this is in conflict with #284 we need to decide a merge order - and since this is older: what are we missing to merge?

Sorry I have seen it now. The problem is in

eko/src/ekomark/benchmark/runner.py

Line 249 in 1313263

np.testing.assert_allclose(q2, q2a)

for some reason the q values are in wrong order, i.e. ext["values"] has first Q=31 and then Q=100 while in pdf_grid.items() it was the opposite

edit: your commit fixed it

[felixhekhorn]

edit: your commit fixed it

indeed - I like the fix in #284 better

[niclaurenti]

@felixhekhorn @alecandido In the apfel benchmarks (and also in banana) we still have QrefQED that is no longer present in eko. My proposal was to restore Qedref to make it compatible with the nnpdf runcards and use it to switch on and off the em_running

[alecandido]

@felixhekhorn @alecandido In the apfel benchmarks (and also in banana) we still have QrefQED that is no longer present in eko. My proposal was to restore Qedref to make it compatible with the nnpdf runcards and use it to switch on and off the em_running

Fine by me, for as long as it is not tracking EKO and yadism it is reasonable not to implement a third convention

[felixhekhorn]

@felixhekhorn @alecandido In the apfel benchmarks (and also in banana) we still have QrefQED that is no longer present in eko. My proposal was to restore Qedref to make it compatible with the nnpdf runcards and use it to switch on and off the em_running

didn't we have this behaviour already at some point of the 1M iterations? aem_running = QrefQED == Qref

[niclaurenti]

@felixhekhorn @alecandido In the apfel benchmarks (and also in banana) we still have QrefQED that is no longer present in eko. My proposal was to restore Qedref to make it compatible with the nnpdf runcards and use it to switch on and off the em_running

didn't we have this behaviour already at some point of the 1M iterations? aem_running = QrefQED == Qref

Yes in eko 0.12 it is there. I think that since eko 0.13 it has been removed. Now I will push my proposal and tell me what you think (obviously I have to fix also banana)

[niclaurenti]

Last commit shows what I had in mind: I changed nan with 0 since em_running will be enabled only if Qref==Qedref so 0 could be every number different from Qref

[niclaurenti]

Last commit shows what I had in mind: I changed nan with 0 since em_running will be enabled only if Qref==Qedref so 0 could be every number different from Qref

Moreover in this way both pineko and evolven3fit_new will use running alpha with theory 523 and fixed alpha with theory 522 without adding some lines that explicitly set em_running=True for example like in

https://github.com/NNPDF/nnpdf/blob/6bfc220a766777e31df9ce242cfd4c584b078a9a/n3fit/src/evolven3fit_new/eko_utils.py#L83

[felixhekhorn]

You need to fix (+ release + update) banana before merging, else we're breaking the benchmarks that are supposed to fix in this PR again

[felixhekhorn]

@niclaurenti you need to run poetry lock in order to get the lock file correct, since there is a hashing algorithm in the game ...

[niclaurenti]

for me we can merge

[felixhekhorn]

let's wait for https://github.com/NNPDF/eko/actions/runs/5280648705/jobs/9553070195 ...

PS: can you double check you can run APFEL still (i.e. with new banana)?

[niclaurenti]

let's wait for https://github.com/NNPDF/eko/actions/runs/5280648705/jobs/9553070195 ...

PS: can you double check you can run APFEL still (i.e. with new banana)?

It works