UL: Corrections

[mcremone]

First of all, pull the current most up to date version from the master branch:

https://github.com/mcremone/decaf/blob/master/analysis/utils/corrections.py

then replace the corrections one by one with the ones recommended for UL. Please be mindful of a couple of things:

1. Use 'correctionlib': https://github.com/cms-nanoAOD/correctionlib
2. Use the coffea lookup tools to interface with correctionlib. If need help, ask Nick Smith.
3. Comment on each single correction, including links from where correction files have been taken etc.
4. Clean up the 'analysis/data' folder from non-UL files, keep only the ones that are used. With correctionlib, they should be reduced to just a bunch of json files

[mcremone]

An additional comment, we need to double check one by one against the corrections implemented by KIT. Still, we want to implement ours using correctionlib.

[ParticleChef]

Electron id, photon id, MET phi corrections, and pu weight are included at correction.py file with json files (https://gitlab.cern.ch/cms-nanoAOD/jsonpog-integration)
The electron trigger weight and reco sf should be included.

[ParticleChef]

For nlo ewk scale factor, the root files are made by monojet analysis. At the last meeting, we discussed that this scale factor can be used. https://github.com/ParticleChef/decaf/blob/master/analysis/utils/corrections.py#L220

[mcremone]

Electron id, photon id, MET phi corrections, and pu weight are included at correction.py file with json files (https://gitlab.cern.ch/cms-nanoAOD/jsonpog-integration)
The electron trigger weight and reco sf should be included.

Can you point me to the part of your code where these are used? Also, how about muon isolation weights? For what concerns trigger weight, most likely you're also missing single muon and MET trigger, am I right?

[ParticleChef]

Yes I'm not uploaded muon trigger, isolation weight yet.
And I pointed the part of current codes.
Electron ID sf : https://github.com/ParticleChef/decaf/blob/UL/analysis/utils/corrections.py#L32
Photon ID sf: https://github.com/ParticleChef/decaf/blob/UL/analysis/utils/corrections.py#L32
pu weight: https://github.com/ParticleChef/decaf/blob/UL/analysis/utils/corrections.py#L32
MET correction: https://github.com/ParticleChef/decaf/blob/UL/analysis/utils/corrections.py#L32

And I included the MET trigger and nlo sf from previous corrections.py file.
nlo sf: https://github.com/ParticleChef/decaf/blob/UL/analysis/utils/corrections.py#L154
met trigger: https://github.com/ParticleChef/decaf/blob/UL/analysis/utils/corrections.py#L15

The btag weight part is modifying now.

[ParticleChef]

I make quick test with json file and btageff.merged file existed.

# generate 20 dummy jet features
jet_pt    = np.random.exponential(50., 15) 
jet_eta   = np.random.uniform(0.0, 2.4, 15) 
jet_flav  = np.random.choice([0, 4, 5], 15) 
jet_discr = np.random.uniform(0.0, 1.0, 15) 

# separate light and b/c jets
light_jets = np.where(jet_flav == 0)
bc_jets    = np.where(jet_flav != 0)

btag = load('hists/btageff2017.merged')
bpass = btag[tagger].integrate('dataset').integrate('wp',workingpoint).integrate('btag', 'pass').values()[()]
ball = btag[tagger].integrate('dataset').integrate('wp',workingpoint).integrate('btag').values()[()]
nom = bpass / np.maximum(ball, 1.) 
eff = lookup_tools.dense_lookup.dense_lookup(nom, [ax.edges() for ax in btag[tagger].axes()[3:]])

btvjson = correctionlib.CorrectionSet.from_file('data/BtagSF/'+year+'_UL/btagging.json.gz')
sf_nom = btvjson["deepJet_comb"].evaluate('central','M', jet_flav[bc_jets], jet_eta[bc_jets], jet_pt[bc_jets])
print('sf_nom: ', sf_nom, len(sf_nom))

def P(eff):
    weight = eff.ones_like()
    weight[istag] = eff[istag]
    weight[~istag] = (1 - eff[~istag])
    return weight.prod()

eff = eff(jet_pt, jet_eta, jet_flav)                                                                                                                       
print('extract eff:', eff, len(eff))

eff_data_nom  = np.minimum(1., sf_nom*eff)
nnom = P(eff_data_nom)/P(eff)
print('P(eff_data_nom)/P(eff)', nnom)

I printed the values and I got error like this

sf_nom:  [0.94694163 0.95233112 0.9551299  0.95522698 0.95875001 0.94341749
 0.95456105 0.94572292 0.94499175 0.95435803 0.94332464] 11
extract eff: [0.9375     0.9375     0.9375     0.9375     0.9375     0.61748634
 0.9375     0.9375     0.91052632 0.91052632 0.91052632 0.91052632
 0.9375     0.9375     0.61748634] 15
Traceback (most recent call last):
  File "utils/cortest.py", line 89, in <module>
    eff_data_nom  = np.minimum(1., sf_nom*eff)
ValueError: operands could not be broadcast together with shapes (11,) (15,) 

Which part should I fix to solve this error?

[mcremone]

To avoid this shape mismatch you can use real data/MC in the test. My suggestion is that we finish first implementing all corrections with correctionlib (when possible). I'll then do a quick review of the code and then we structure a test.

[ParticleChef]

I finished the modifying btag. How you implement the jec? Other than jec, I modified all corrections I need.

[mcremone]

For jet you can follow this:

https://github.com/nsmith-/boostedhiggs/blob/master/boostedhiggs/build_jec.py

[ParticleChef]

I update the corrections.py and jet energy correction files.

[ParticleChef]

When run the correction.py, the error is accured at import uproot_methods

Traceback (most recent call last):
  File "utils/corrections.py", line 6, in <module>
    import uproot, uproot_methods
  File "/uscms/home/jhong/.local/lib/python3.6/site-packages/uproot_methods/__init__.py", line 5, in <module>
    from uproot_methods.classes.TVector2 import TVector2, TVector2Array
  File "/uscms/home/jhong/.local/lib/python3.6/site-packages/uproot_methods/classes/TVector2.py", line 8, in <module>
    import awkward.array.jagged
ModuleNotFoundError: No module named 'awkward.array'

Instead of this, update error lines to uproot3.

And separate the '2016' to '2016preVFP' and '2016postVFP' at btag part.
(https://github.com/ParticleChef/decaf/blob/UL/analysis/utils/corrections.py#L465)
(https://github.com/ParticleChef/decaf/blob/UL/analysis/utils/common.py#L19)

[mcremone]

Hi think you want to do the other way around, which means using the latest awkward version, changing the code lines to use what the latest awkward version wants you to use.

[ParticleChef]

Then, is there no need to change the version of Awkward?

[mcremone]

In general you want to use the latest version of everything, both awkward and uproot. If we need to change the code a bit to adjust to the format the new versions may want, that's what I would do.

[ParticleChef]

The current correction.py works in awkward version 1.9.0. I checked my current setup and latest version of some module.

        (current / latest)
awkward ( 1.9.0 / 2.4.10 )
uproot  ( 4.3.7 / 5.1.2 )
uproot3 ( 3.14.4 / 3.14.4 )
numpy  ( 1.17.0 / 1.26.0 )

Is it okay to change the version in current coffea (0.7.12)?