data-manipulation-utilities 0.0.9

No project description provided

D(ata) M(anipulation) U(tilities)

These are tools that can be used for different data analysis tasks.

For LHCb specific tools (not useful for people outside LHCb) see this

GIT

Pushing

From the root directory of a version controlled project (i.e. a directory with the .git subdirectory) using a pyproject.toml file, run:

publish

such that:

1. The pyproject.toml file is checked and the version of the project is extracted.
2. If a tag named as the version exists move to the steps below.
3. If it does not, make a new tag with the name as the version

Then, for each remote it pushes the tags and the commits.

Why?

1. Tags should be named as the project's version
2. As soon as a new version is created, that version needs to be tagged.
3. In GitHub, one can configure actions to publish projects when the commits are tagged.

Generic

This section describes generic tools that could not be put in a specific category, but tend to be useful.

Timer

In order to benchmark functions do:

import dmu.generic.utilities as gut

# Needs to be turned on, it's off by default
gut.TIMER_ON=True
@gut.timeit
def fun():
    sleep(3)

fun()

JSON dumper

The following lines will dump data (dictionaries, lists, etc) to a JSON file:

import dmu.generic.utilities as gut

data = [1,2,3,4]

gut.dump_json(data, '/tmp/list.json')

Physics

Truth matching

In order to compare the truth matching efficiency and distributions after it is performed in several samples, run:

check_truth -c configuration.yaml

where the config file, can look like:

# ---------
max_entries : 1000
samples:
  # Below are the samples for which the methods will be compared
  sample_a:
    file_path : /path/to/root/files/*.root
    tree_path : TreeName
    methods :
        #Below we specify the ways truth matching will be carried out
        bkg_cat : B_BKGCAT == 0 || B_BKGCAT == 10 || B_BKGCAT == 50
        true_id : TMath::Abs(B_TRUEID) == 521 && TMath::Abs(Jpsi_TRUEID) == 443 && TMath::Abs(Jpsi_MC_MOTHER_ID) == 521 && TMath::Abs(L1_TRUEID) == 11 && TMath::Abs(L2_TRUEID) == 11 && TMath::Abs(L1_MC_MOTHER_ID) == 443 && TMath::Abs(L2_MC_MOTHER_ID) == 443 && TMath::Abs(H_TRUEID) == 321 && TMath::Abs(H_MC_MOTHER_ID) == 521
    plot:
      # Below are the options used by Plottter1D (see plotting documentation below)
      definitions:
          mass : B_nopv_const_mass_M[0]
      plots:
          mass :
              binning    : [5000, 6000, 40]
              yscale     : 'linear'
              labels     : ['$M_{DTF-noPV}(B^+)$', 'Entries']
              normalized : true
      saving:
        plt_dir : /path/to/directory/with/plots

Math

PDFs

Printing PDFs

One can print a zfit PDF by doing:

from dmu.stats.utilities   import print_pdf

print_pdf(pdf)

this should produce an output that will look like:

PDF: SumPDF
OBS: <zfit Space obs=('m',), axes=(0,), limits=(array([[-10.]]), array([[10.]])), binned=False>
Name                                                        Value            Low           HighFloating               Constraint
--------------------
fr1                                                     5.000e-01      0.000e+00      1.000e+00    1                     none
fr2                                                     5.000e-01      0.000e+00      1.000e+00    1                     none
mu1                                                     4.000e-01     -5.000e+00      5.000e+00    1                     none
mu2                                                     4.000e-01     -5.000e+00      5.000e+00    1                     none
sg1                                                     1.300e+00      0.000e+00      5.000e+00    1                     none
sg2                                                     1.300e+00      0.000e+00      5.000e+00    1                     none

showing basic information on the observable, the parameter ranges and values, whether they are Gaussian constrained and floating or not. One can add other options too:

from dmu.stats.utilities   import print_pdf

# Constraints, uncorrelated for now
d_const = {'mu1' : [0.0, 0.1], 'sg1' : [1.0, 0.1]}
#-----------------
# simplest printing to screen
print_pdf(pdf)

# Will not show certain parameters
print_pdf(pdf,
          blind   = ['sg.*', 'mu.*'])

# Will add constraints
print_pdf(pdf,
          d_const = d_const,
          blind   = ['sg.*', 'mu.*'])
#-----------------
# Same as above but will dump to a text file instead of screen
#-----------------
print_pdf(pdf,
          txt_path = 'tests/stats/utilities/print_pdf/pdf.txt')

print_pdf(pdf,
          blind    =['sg.*', 'mu.*'],
          txt_path = 'tests/stats/utilities/print_pdf/pdf_blind.txt')

print_pdf(pdf,
          d_const  = d_const,
          txt_path = 'tests/stats/utilities/print_pdf/pdf_const.txt')

Fits

The Fitter class is a wrapper to zfit, use to make fitting easier.

Simplest fit

from dmu.stats.fitter      import Fitter

obj = Fitter(pdf, dat)
res = obj.fit()

Customizations

In order to customize the way the fitting is done one would pass a configuration dictionary to the fit(cfg=config) function. This dictionary can be represented in YAML as:

# The strategies below are exclusive, only can should be used at a time
strategy      :
      # This strategy will fit multiple times and retry the fit until either
      # ntries is exhausted or the pvalue is reached.
      retry   :
          ntries        : 4    #Number of tries
          pvalue_thresh : 0.05 #Pvalue threshold, if the fit is better than this, the loop ends
          ignore_status : true #Will pick invalid fits if this is true, otherwise only valid fits will be counted
      # This will fit smaller datasets and get the value of the shape parameters to allow
      # these shapes to float only around this value and within nsigma
      # Fit can be carried out multiple times with larger and larger samples to tighten parameters
      steps   :
          nsteps   : [1e3, 1e4] #Number of entries to use
          nsigma   : [5.0, 2.0] #Number of sigmas for the range of the parameter, for each step
          yields   : ['ny1', 'ny2'] # in the fitting model ny1 and ny2 are the names of yields parameters, all the yield need to go in this list
# The lines below will split the range of the data [0-10] into two subranges, such that the NLL is built
# only in those ranges. The ranges need to be tuples
ranges        :
      - !!python/tuple [0, 3]
      - !!python/tuple [6, 9]
#The lines below will allow using contraints for each parameter, where the first element is the mean and the second
#the width of a Gaussian constraint. No correlations are implemented, yet.
constraints   :
      mu : [5.0, 1.0]
      sg : [1.0, 0.1]
#After each fit, the parameters spciefied below will be printed, for debugging purposes
print_pars    : ['mu', 'sg']
likelihood :
    nbins : 100 #If specified, will do binned likelihood fit instead of unbinned

Arrays

Scaling by non-integer

Given an array representing a distribution, the following lines will increase its size by fscale, where this number is a float, e.g. 3.4.

from dmu.arrays.utilities import repeat_arr

arr_val = repeat_arr(arr_val = arr_inp, ftimes = fscale)

in such a way that the output array will be fscale larger than the input one, but will keep the same distribution.

Functions

The project contains the Function class that can be used to:

• Store (x,y) coordinates.
• Evaluate the function by interpolating
• Storing the function as a JSON file
• Loading the function from the JSON file

It can be used as:

import numpy
from dmu.stats.function    import Function

x    = numpy.linspace(0, 5, num=10)
y    = numpy.sin(x)

path = './function.json'

# By default the interpolation is 'cubic', this uses scipy's interp1d
# refer to that documentation for more information on this.
fun  = Function(x=x, y=y, kind='cubic')
fun.save(path = path)

fun  = Function.load(path)

xval = numpy.lispace(0, 5, num=100)
yval = fun(xval)

Machine learning

Classification

To train models to classify data between signal and background, starting from ROOT dataframes do:

from dmu.ml.train_mva      import TrainMva

rdf_sig = _get_rdf(kind='sig')
rdf_bkg = _get_rdf(kind='bkg')
cfg     = _get_config()

obj= TrainMva(sig=rdf_sig, bkg=rdf_bkg, cfg=cfg)
obj.run()

where the settings for the training go in a config dictionary, which when written to YAML looks like:

training :
    nfold    : 10
    features : [w, x, y, z]
    hyper    :
      loss              : log_loss
      n_estimators      : 100
      max_depth         : 3
      learning_rate     : 0.1
      min_samples_split : 2
saving:
    path : 'tests/ml/train_mva/model.pkl'
plotting:
    val_dir : 'tests/ml/train_mva'
    features:
        saving:
            plt_dir : 'tests/ml/train_mva/features'
        plots:
          w :
            binning : [-4, 4, 100]
            yscale  : 'linear'
            labels  : ['w', '']
          x :
            binning : [-4, 4, 100]
            yscale  : 'linear'
            labels  : ['x', '']
          y :
            binning : [-4, 4, 100]
            yscale  : 'linear'
            labels  : ['y', '']
          z :
            binning : [-4, 4, 100]
            yscale  : 'linear'
            labels  : ['z', '']

the TrainMva is just a wrapper to scikit-learn that enables cross-validation (and therefore that explains the nfolds setting).

Caveats

When training on real data, several things might go wrong and the code will try to deal with them in the following ways:

• Repeated entries: Entire rows with features might appear multiple times. When doing cross-validation, this might mean that two identical entries will end up in different folds. The tool checks for wether a model is evaluated for an entry that was used for training and raise an exception. Thus, repeated entries will be removed before training.

• NaNs: Entries with NaNs will break the training with the scikit GradientBoostClassifier base class. Thus, we also remove them from the training.

Application

Given the models already trained, one can use them with:

from dmu.ml.cv_predict     import CVPredict

#Build predictor with list of models and ROOT dataframe with data
cvp     = CVPredict(models=l_model, rdf=rdf)

#This will return an array of probabilibies
arr_prb = cvp.predict()

If the entries in the input dataframe were used for the training of some of the models, the model that was not used will be automatically picked for the prediction of a specific sample.

The picking process happens through the comparison of hashes between the samples in rdf and the training samples. The hashes of the training samples are stored in the pickled model itself; which therefore is a reimplementation of GradientBoostClassifier, here called CVClassifier.

If a sample exist, that was used in the training of every model, no model can be chosen for the prediction and an CVSameData exception will be risen.

Caveats

When evaluating the model with real data, problems might occur, we deal with them as follows:

• Repeated entries: When there are repeated features in the dataset to be evaluated we assign the same probabilities, no filtering is used.
• NaNs: Entries with NaNs will break the evaluation. These entries will be patched with zeros and evaluated. However, before returning, the probabilities will be saved as -1. I.e. entries with NaNs will have probabilities of -1.

Rdataframes

These are utility functions meant to be used with ROOT dataframes.

Adding a column from a numpy array

For this do:

import dmu.rdataframe.utilities as ut

arr_val = numpy.array([10, 20, 30])
rdf     = ut.add_column(rdf, arr_val, 'values')

the add_column function will check for:

1. Presence of a column with the same name
2. Same size for array and existing dataframe

and return a dataframe with the added column

Attaching attributes

Use case When performing operations in dataframes, like Filter, Range etc; a new instance of the dataframe will be created. One might want to attach attributes to the dataframe, like the name of the file or the tree, etc. Those attributes will thus be dropped. In order to deal with this one can do:

from dmu.rdataframe.atr_mgr import AtrMgr
# Pick up the attributes
obj = AtrMgr(rdf)

# Do things to dataframe
rdf = rdf.Filter(x, y)
rdf = rdf.Define('a', 'b')

# Put back the attributes
rdf = obj.add_atr(rdf)

The attributes can also be saved to JSON with:

obj = AtrMgr(rdf)
...
obj.to_json('/path/to/file.json')

Logging

The LogStore class is an interface to the logging module. It is aimed at making it easier to include a good enough logging tool. It can be used as:

from dmu.logging.log_store import LogStore

LogStore.backend = 'logging' # This line is optional, the default backend is logging, but logzero is also supported
log = LogStore.add_logger('msg')
LogStore.set_level('msg', 10)

log.debug('debug')
log.info('info')
log.warning('warning')
log.error('error')
log.critical('critical')

Plotting from ROOT dataframes

1D plots

Given a set of ROOT dataframes and a configuration dictionary, one can plot distributions with:

from dmu.plotting.plotter_1d import Plotter1D as Plotter

ptr=Plotter(d_rdf=d_rdf, cfg=cfg_dat)
ptr.run()

where the config dictionary cfg_dat in YAML would look like:

selection:
    #Will do at most 50K random entries. Will only happen if the dataset has more than 50K entries
    max_ran_entries : 50000
    cuts:
    #Will only use entries with z > 0
      z : 'z > 0'
saving:
    #Will save lots to this directory
    plt_dir : tests/plotting/high_stat
definitions:
    #Will define extra variables
    z : 'x + y'
#Settings to make histograms for differen variables
plots:
    x :
        binning    : [0.98, 0.98, 40] # Here bounds agree => tool will calculate bounds making sure that they are the 2% and 98% quantile
        yscale     : 'linear' # Optional, if not passed, will do linear, can be log
        labels     : ['x', 'Entries'] # Labels are optional, will use varname and Entries as labels if not present
        title      : 'some title can be added for different variable plots'
        name       : 'plot_of_x' # This will ensure that one gets plot_of_x.png as a result, if missing x.png would be saved
    y :
        binning    : [-5.0, 8.0, 40]
        yscale     : 'linear'
        labels     : ['y', 'Entries']
    z :
        binning    : [-5.0, 8.0, 40]
        yscale     : 'linear'
        labels     : ['x + y', 'Entries']
        normalized : true #This should normalize to the area

it's up to the user to build this dictionary and load it.

2D plots

For the 2D case it would look like:

from dmu.plotting.plotter_2d import Plotter2D as Plotter

ptr=Plotter(rdf=rdf, cfg=cfg_dat)
ptr.run()

where one would introduce only one dataframe instead of a dictionary, given that overlaying 2D plots is not possible. The config would look like:

saving:
    plt_dir : tests/plotting/2d
general:
    size : [20, 10]
plots_2d:
    # Column x and y
    # Name of column where weights are, null for not weights
    # Name of output plot, e.g. xy_x.png
    - [x, y, weights, 'xy_w']
    - [x, y,    null, 'xy_r']
axes:
    x :
        binning : [-5.0, 8.0, 40]
        label   : 'x'
    y :
        binning : [-5.0, 8.0, 40]
        label   : 'y'

Manipulating ROOT files

Getting trees from file

The lines below will return a dictionary with trees from the handle to a ROOT file:

import dmu.rfile.utilities   as rfut

ifile  = TFile("/path/to/root/file.root")

d_tree = rfut.get_trees_from_file(ifile)

Printing contents

The following lines will create a file.txt with the contents of file.root, the text file will be in the same location as the ROOT file.

from dmu.rfile.rfprinter import RFPrinter

obj = RFPrinter(path='/path/to/file.root')
obj.save()

Printing from the command line

This is mostly needed from the command line and can be done with:

print_trees -p /path/to/file.root

which would produce a /pat/to/file.txt file with the contents, which would look like:

Directory/Treename
    B_CHI2                        Double_t
    B_CHI2DOF                     Double_t
    B_DIRA_OWNPV                  Float_t
    B_ENDVERTEX_CHI2              Double_t
    B_ENDVERTEX_CHI2DOF           Double_t

Comparing ROOT files

Given two ROOT files the command below:

compare_root_files -f file_1.root file_2.root

will check if:

1. The files have the same trees. If not it will print which files are in the first file but not in the second and vice versa.
2. The trees have the same branches. The same checks as above will be carried out here.
3. The branches of the corresponding trees have the same values.

the output will also go to a summary.yaml file that will look like:

'Branches that differ for tree: Hlt2RD_BToMuE/DecayTree':
  - L2_BREMHYPOENERGY
  - L2_ECALPIDMU
  - L1_IS_NOT_H
'Branches that differ for tree: Hlt2RD_LbToLMuMu_LL/DecayTree':
  - P_CaloNeutralHcal2EcalEnergyRatio
  - P_BREMENERGY
  - Pi_IS_NOT_H
  - P_BREMPIDE
Trees only in file_1.root: []
Trees only in file_2.root:
  - Hlt2RD_BuToKpEE_MVA_misid/DecayTree
  - Hlt2RD_BsToPhiMuMu_MVA/DecayTree

Text manipulation

Transformations

Run:

transform_text -i ./transform.txt -c ./transform.toml

to apply a transformation to transform.txt following the transformations in transform.toml.

The tool can be imported from another file like:

from dmu.text.transformer import transformer as txt_trf

trf=txt_trf(txt_path=data.txt, cfg_path=data.cfg)
trf.save_as(out_path=data.out)

Currently the supported transformations are:

append

Which will apppend to a given line a set of lines, the config lines could look like:

[settings]
as_substring=true
format      ='--> {} <--'

[append]
'primes are'=['2', '3', '5']
'days are'=['Monday', 'Tuesday', 'Wednesday']

as_substring is a flag that will allow matches if the line in the text file only contains the key in the config e.g.:

the
first
primes are:
and
the first
days are:

format will format the lines to be inserted, e.g.:

the
first
primes are:
--> 2 <--
--> 3 <--
--> 5 <--
and
the first
days are:
--> Monday <--
--> Tuesday <--
--> Wednesday <--

coned

Utility used to edit SSH connection list, has the following behavior:

#Prints all connections
coned -p

#Adds a task name to a given server
coned -a server_name server_index task

#Removes a task name from a given server
coned -d server_name server_index task

the list of servers with tasks and machines is specified in a YAML file that can look like:

ihep:
    '001' :
        - checks
        - extractor
        - dsmanager
        - classifier
    '002' :
        - checks
        - hqm2
        - dotfiles
        - data_checks
    '003' :
        - setup
        - ntupling
        - preselection
    '004' :
        - scripts
        - tools
        - dmu
        - ap
lxplus:
    '984' :
        - ap

and should be placed in $HOME/.config/dmu/ssh/servers.yaml