Making theory order local per dataset

[jacoterh]

This PR removes the pQCD order key in the runcard and replaces it by locally specifying it per dataset, e.g:

datasets:

  ATLAS_CMS_SSinc_RunI: {"order": "NLO_QCD"}
  ATLAS_CMS_tt_AC_8TeV: {"order": "NLO_QCD"}
  ATLAS_SSinc_RunII: {"order": "NLO_QCD"}
  ATLAS_STXS_runII_13TeV: {"order": "NLO_QCD"}
  ATLAS_WH_Hbb_13TeV: {"order": "NLO_QCD"}

The reason we need this feature is because we now have different higher order corrections that we like to distinguish: NLO_QCD, NLO_EW, NLO_QCD_EW. The theory files need updating accordingly, which is done in PR#25 on smefit_database.

[LucaMantani]

@jacoterh Correct me if wrong, this will be fully compatible with current tables as long as we specify "NLO" or "LO" for each dataset, right?

[jacoterh]

Yes, that's right, the only thing that mattes is that the order specified for each dataset also appears in the theory json. This can be any key as you point out, but PR#25 on smefit_database actually updates the keys in the jsons to our new convention (i.e. NLO_QCD, NLO_EW, etc)

[arossia94]

@jacoterh, is there any default behavior for the case in which one forgets to specify the order for a dataset? That could be useful in case one reuses an old runcard.
I'd print a warning to the user and either exclude the dataset or use it with LO (assuming all datasets will have at least LO).

[arossia94]

Minor requests regarding default and limit behaviors.
Should we also update the documentation within this same PR before forgetting about this change?

[jacoterh]

Should be ready to go. Updated behaviour

datasets:
  - dataset_1
  - dataset_2: {"order": "NLO_QCD"}
  - dataset_3: {"order": "NLO_EW"}

where the predictions of dataset_1 are LO (default behaviour), dataset_2 NLO_QCD and dataset_3 at NLO_EW

[jacoterh]

Updated syntax (no mixed types anymore)

datasets:
  - name: dataset_1
  - name: dataset_2
    order: NLO_QCD
  - name: dataset_3
    order: "NLO_EW

This syntax is future proof as it supports naturally additional properties we might want to specify per dataset, i.e. cuts.

[LucaMantani]

Updated syntax (no mixed types anymore)

datasets:
  - name: dataset_1
  - name: dataset_2
    order: NLO_QCD
  - name: dataset_3
    order: "NLO_EW

This syntax is future proof as it supports naturally additional properties we might want to specify per dataset, i.e. cuts.

So now you made it a list of dictionaries, right? it's more similar to the NNPDF format. I am in favour of this but I thought we wanted to keep the functioning of the old runcards?

[jacoterh]

That's right, it's a list of dictionaries, so it'd be equivalent to doing

datasets:
  - {name: dataset_1}
  - {name: dataset_2, order: NLO_QCD}
  - {name: dataset_3, order: "NLO_EW

This breaks compatibility with the old runcards, but for this we can provide a short script to convert them. In any case it's a small difference.

[LucaMantani]

I have one comment on the order datasets are looped through but for the rest it looks fine.

[LucaMantani]

Here and in similar loops (the one in load_rge_mat for example) datasets are looped without being put in alphabetical order now? I am a bit worried about having mismatches, are we guaranteed that things are ordered correctly?

[jacoterh]

I see your concern, but datasets is a list which has a well defined ordering, so there's no need to sort anything (as opposed to dictionaries). But it was a list also before, and yet it got sorted. Perhaps @giacomomagni can comment if/why this was necessary at the time?

[LucaMantani]

I see, so the order is always the one put in the card basically.

[giacomomagni]

I guess the unique was there just to be sure a dataset is not loaded twice, but yes from the DataTuple you can always infer the order.

[giacomomagni]

why do you need to have a default "LO" ? Do you need to load a dataset without any theory being loaded?

[jacoterh]

This is just the default theory order that gets loaded if one doesn't specify it for a particular dataset. Of course, the default can be set to NLO as well, but not all predictions are available at NLO.

[giacomomagni]

sure, but sorry to be picky why do you want to have a default ?
I mean if you load, you need to specify an order

[giacomomagni]

I guess the unique was there just to be sure a dataset is not loaded twice, but yes from the DataTuple you can always infer the order.