adding validation tool (KS test)

[lguzzi]

PR to add the vaidation tool. The script loads a list of files into a RDataFrame, splits the dataframe into different chunks and then runs a Kolmogorov-Smirnov test between the first chunk and the rest to check the compatibility of some distributions.

The distributions considered for the KS test are:

• lepton_gen_match, and:
  • tau_pt and tau_eta for each gen match value
• sample_type distribution, and:
  • tau_pt and tau_eta for each sample type value
• dataset_group_id distribution, and:
  • tau_pt, tau_eta for each group id value
  • dataset_id for each group id value

The script creates a root file storing the chunk histograms, a json file with the p-values and the pdf files of the KS comparison.
You can find and example at /afs/cern.ch/user/l/lguzzi/public/TauPOG/validation_tool_test (run on /eos/cms/store/group/phys_tau/TauML/prod_2018_v1/full_tuples/WJetsToLNu_HT-2500ToInf/eventTuple_1-1.root)

Time tests

Here are some timing test I run on an lxplus machine:

• root-version: 6.20/02
• dataset: WJetsToLNu_HT-1200To2500
  • size: 42 GB
  • events: 15890103 (15 milion)

source /cvmfs/sft.cern.ch/lcg/views/LCG_97apython3/x86_64-centos7-clang10-opt/setup.sh
python -m cProfile validation_tool.py --input "/eos/cms/store/group/phys_tau/TauML/prod_2018_v1/full_tuples/WJetsToLNu_HT-1200To2500/*.root" --output test-v5 --legend

Run time after a fresh login (subsequent calls are about twice as fast):
1 thread: 106 s
4 thread: 47 s
10 thread: 19 s

The most time consuming function is load_histograms, which takes about 7s per call running on one thread, 2s per call running on four threads and 1s per call on ten threads (13 calls in total).

[kandrosov]

Thank you @lguzzi! The current version code looks good. To close this PR, could you please:

• Post here timing to run the tool on a relatively large dataset (at least O(10 GB)) for the future references.
• Add instructions on how to run the validation into README.md