correct wrong extras setting of scale_factors
During data generation inspection, noticed that automatic scale_factor
extras attribute for kkr_imp_wc
was not correct.
Created method check_scale_factor_extras()
for it to find those mislabelings for large numbers of workchains (original commit for kkr_imp_wc
, extension to kkr_scf_wc
).
Noticed glitches:
- Step4: impurity embedding
kkr_imp_wc
's: don't remember, but many non-1.0
kkr_imp_wc
had1.0
asscale_factor
extra. - Step3: host-in-host GF writeout
kkr_imp_wc
's: About half of thefailed
non-1.0
kkr_imp_wc
had1.0
asscale_factor
extra. None of thefinished_ok
ones. - Step2: host
kkr_scf_wc
's: Some of them had noscale_factor
extra (but their ancestor nodes did).
Check the data_generation
loops for where these scale_factor
extras are set / should be set, and correct it, so these glitches do not appear again when new data is generated. Can reuse the above fixer method (e.g. split into application to a single workchain and a group of workchains).
The strange thing is that I had already let the checker function run over the imp
groups a few weeks before and had set fix_mismatches=True
, which should fix the wrong extras (success confirmed by rerun without fix got same number of mismatches, rerun with fix got zero number of mismatches), but having run it again now, they have reappeared, as if the extras updates had not been saved in the database. In the time since, I have not touched nor used the data generation pipelines. So either I don't understand how extras work correctly (do I have to call .store()
so they get permanently updated?), or some other code I have used in the meantime has reset the extras to wrong 1.0
for all kkr_imp_wc
. This goes actually for all the glitches listed above, they all should have been already fixed but had reappeared.
So find out where this is coming from, and stamp it out.
Update 2021-08-13
The issue was in wrong setting of scale factors in the dg_submit
HostLoop
/ GFWriteoutLoop
/ ImpLoop
. This has been resolved in one of the commits from the previous week.