Seed pseudodata

[siranipour]

Working on reconstructing the nth N3FIT replica pseudodata.

[siranipour]

@wilsonmr no need to take a look now still very much prototype stage, but one thing that may be difficult is that:

nnpdf/n3fit/src/n3fit/io/reader.py

Line 14 in 91a159b

def make_tr_val_mask(datasets, exp_name, seed):

still makes use of datasets. What are our plans for the future of this class? All the functionality we're implementing is either using CommonData or lists there-of.

Moreover, this function still uses the notion of experiment names, which has been deprecated since the data-kw.

[wilsonmr]

A few things:

• I think this function really wants to be an action that takes as input the initialize_seeds action* (or something similar) and then calls make_replica, you might find this considerably easier after both MakeReplica in python #866 and Add python t0 covmat #1052 are merged since the latter starts to add actions with are using the new python machinery.
• in the end we want to scrap most of reader.py in this instance I would say that you want some reproducible way to generate training validation splits. The splits themselves are easy to gen and can be copied from this func but you may wish to change how the splits are seeded if you don't like using exp name. I wouldn't get completely hung up on this right now but if you take data as input for now then you can extract the name from the data and use that in the hash to get a seed. Alternatively you could take group_name="data" as an argument like produce_data but having defaults being set all over the place is something we kind of want to avoid so if we were to go down that route we should use the lockfile mechanism to set the default.

*note its an n3fit action so would have to be moved in order to be accessed in vp