Saving new python pseudodata during fit

[siranipour]

As mentioned in #1328 (comment), the current fitting::savepseudodata flag doesn't save the python make_replica, but rather the C++ make_replica.

This blocks #1328, because I need to create a new test fit for the tests to pass there.

@Zaharid, is there anyway to change the name of the file that @table uses? Currently it's not within keeping of the convention of training.dat and validation.dat so

nnpdf/validphys2/src/validphys/pseudodata.py

Line 32 in b636ae0

def read_fit_pseudodata(fitcontext, context_index):

doesn't work with new style fits.

[Zaharid]

@siranipour Sorry I am a bit at loss: could you describe a bit more what it does and what should it do?

[siranipour]

Previously if you wanted to save the fit pseudodata, you would have fitting::savepseudodata: true which would append the following action

nnpdf/validphys2/src/validphys/n3fit_data.py

Line 358 in b636ae0

def pseudodata_table(replicas_exps_pseudodata, replicas, experiments_index):

to the N3FIT_DEFAULT_ACTIONS.

But this used the C++ make_replica, which is now no longer valid seeing as how we changed it to use the python make_replica recently. This PR addresses that issue by adding the two new functions training_pseudodata and validation_pseudodata to the list of actions, which in turn call the new python make_replica.

If you set the flag to true and run n3fit <runcard.yml> 1, you can see in nnfit/replica_1 there will be datacuts_theory_fitting_training.dat and datacuts_theory_fitting_validation.dat files which save the training and validation data for replica 1.

I was wondering if there was anyway to change the name of the saved .dat files because currently the read_fit_pseudodata function, which reads in the saved pseudodata (i.e the reverse of this PR), expects them to be called training.dat and validation.dat.

[Zaharid]

Right I think one can add a parameter to override the path, but do we insist on keeping the old behaviour?

[siranipour]

I'm not overly insistent on keeping the old behaviour, I guess it would be useful for backwards compatibility since I know @RosalynLP for example has used this functionality in the past. It's more so that I'm not the biggest fan of datacuts_theory_fitting_training.dat being the file name.

[RosalynLP]

@siranipour I've used the functionality but I might be the only one! In any case I won't be using it any more because I'm going to leave very soon. So don't keep it just because of me 😄 .

[Zaharid]

Superficially this looks good to me. @scarlehoff would you mind merging this?

[scarlehoff]

If you set the flag to true and run n3fit <runcard.yml> 1

Sorry, which flag? If it is a runcard flag, could it be used to the Basic Runcard with an explanation (and to the docs somewhere, probably here https://docs.nnpdf.science/n3fit/runcard_detailed.html#other-options?)

Incidentally, in https://docs.nnpdf.science/n3fit/runcard_detailed.html#other-options there has been a problem (I guess coming from a merge) and a section is duplicated so if it is indeed a runcard parameter and you need to add a section there, please fix that as well (otherwise I'll do so later)

[siranipour]

If you set the flag to true and run n3fit <runcard.yml> 1

Sorry, which flag? If it is a runcard flag, could it be used to the Basic Runcard with an explanation (and to the docs somewhere, probably here https://docs.nnpdf.science/n3fit/runcard_detailed.html#other-options?)

Incidentally, in https://docs.nnpdf.science/n3fit/runcard_detailed.html#other-options there has been a problem (I guess coming from a merge) and a section is duplicated so if it is indeed a runcard parameter and you need to add a section there, please fix that as well (otherwise I'll do so later)

Ah good idea, will do that. It is indeed a runcard flag fitting::savepseudodata : true, but this had always been an option

[scarlehoff]

lgtm