Remove need for _template.yaml

[evagroenendijk]

Remove the need for _template.yaml as in issue #202

[evagroenendijk]

https://github.com/NNPDF/pineko/blob/f4097f29dec964e8b2c17818b502575ffa3444f0/tests/test_evolve.py#L55C1-L55C43

@scarlehoff Could I maybe just remove this test? I could make it check that the theory cards' Q0 is the same as the default one, but to me it seems a bit superfluous

[scarlehoff]

Sure, shuffling around the _template and the operators might change which tests are needed.

[felixhekhorn]

some quick things I noticed last week

[evagroenendijk]

@felixhekhorn The regression benchmark comparing convolved predictions failed, but the maximal relative difference is 0.03497747. Would you say it's close enough?

[scarlehoff]

The template name should be removed from here:

pineko/src/pineko/configs.py

Line 19 in 89304e2

"operator_card_template_name",

also the NEEDED_FILE (and then needs to be removed from scaffold.py)

[scarlehoff]

This small of a relative difference is ok I'd say.

[scarlehoff]

These should be sourced from the theory card and only from there.

[evagroenendijk]

Is the inversion method also set in the theory card? I can't find it in the current theory cards

[scarlehoff]

@felixhekhorn what about the inversion method?
I think for the evolution of the PDF we simply fix

truncated -> expanded
iterate-exact -> exact

Should we fix the same here? Or is this not relevant and we can use expanded in both cases?

[evagroenendijk]

For now I put

if "inversion_method" in tcard:
opconf = operators_card["configs"]
opconf["inversion_method"] = tcard["inversion_method"]
else:
opconf["inversion_method"] = opcard_template.CONSTANTS["configs"]["inversion_method"]
logger.warning(
"Inversion method not set in theory card, it's being set to default value "
f"{opcard_template.CONSTANTS['configs']['inversion_method']}. "
"Check if that is what you want."
)

But I can change it

[scarlehoff]

Let's see what Felix says.

Fwiw, in python this can be shortened to:

opconf["inversion_method"] = tcard.get("inversion_method", opcard_template.CONSTANTS["configs"]["inversion_method"])

(also because I think right now you are overwritting the whole of opconf with `operator_card["configs"] I think, is that what you want?)

If the inversion method is irrelevant and can be set constant, I think the warning is not necessary.

[felixhekhorn]

truncated -> expanded
iterate-exact -> exact

this is a choice which seems consistent, but is not necessary. I think we can assume this mapping as default here

[evagroenendijk]

Okay, I'll change it to that then!

[scarlehoff]

I meant to request changes 😅

[felixhekhorn]

@felixhekhorn The regression benchmark comparing convolved predictions failed, but the maximal relative difference is 0.03497747. Would you say it's close enough?

I'm quite worried by this

Your theory has a different evolution method than the default(iterate-exact vs truncated).The evolution method will be set to the default value iterate-exact, check if this is what you want
Warning! The number of iteration in the theory and template is different, (60 vs 1).The number of iterations will be set to default value 60, check if this is what you want

this sounds like we can no longer run actually truncated - i.e. we can no longer reproduce 4.0. If that would be true that must be changed. This might well explain the difference we observe. We want to enforce consistent settings, i.e. "if method=truncated then iterations=1", but we don't want to enforce method=exact everywhere.

[scarlehoff]

This has to do with the comment I left. We just want to ensure that the evolution settings are coming from the theory not from a constant setting.

[evagroenendijk]

@felixhekhorn The regression benchmark comparing convolved predictions failed, but the maximal relative difference is 0.03497747. Would you say it's close enough?

I'm quite worried by this

Your theory has a different evolution method than the default(iterate-exact vs truncated).The evolution method will be set to the default value iterate-exact, check if this is what you want
Warning! The number of iteration in the theory and template is different, (60 vs 1).The number of iterations will be set to default value 60, check if this is what you want

this sounds like we can no longer run actually truncated - i.e. we can no longer reproduce 4.0. If that would be true that must be changed. This might well explain the difference we observe. We want to enforce consistent settings, i.e. "if method=truncated then iterations=1", but we don't want to enforce method=exact everywhere.

You're right! I think here I ran the benchmark while using iterate-exact (which I hardcoded into template.py). But as Juan suggested I will now let the evolution method be totally determined by the theory card. Then it would also do truncated evolution

[felixhekhorn]

I think this can be removed entirely?

[evagroenendijk]

At first I did that, but then I thought it would be useful to leave it in case at any point there is another adjustment to the code that needs some files. It's what I thought, but let me know if you think differently

[scarlehoff]

I would remove unnecessary code and then re-add it if needed given past experiences with having a very big graveyard of code

[felixhekhorn]

I'm a friend of early exits in code, since they safe indents - thus I suggests

if "ModEv" not in tcard:
    raise ValueError("blub")
# no explicit `else` needed nor the associated indentation

same for "IterEv" I would say

[felixhekhorn]

I think this can never happen now with this structure

[evagroenendijk]

Good point

[felixhekhorn]

Can you maybe add a # TODO there saying something like "we impose a default now, but we may at later point allow the user to choose (e.g. via CLI)"

[evagroenendijk]

It's now already possible to determine these from the CLI (when doing "pineko theory opcards ..." and when doing "pineko opcard ... ")

edit: ah no, I'm sorry I was confused and talking about the other settings! I can add the to do