Skip to content

High level API for aux task plots#

To set up the inputs for the plots, have a look here. In general, the input structure is the same for aux task plots, but a separate container with track-level aux task outputs is required to be present in the files (by default "tracks"). For now, vertexing and track origin prediction aux tasks are supported.

The following examples use the dummy data which is described here

The high level API for aux tasks matches the general high level API in terms of structure. It can be used to produce vertexing performance plots.

Initialising the taggers#

Compared to the Results object, an AuxResults object is initialized without definition of a signal class (rather this is passed to the plotting functions directly). Otherwise initialization proceeds in the same fashion. Taggers can be added to these objects in an analogous way to Results, except that each tagger should be initialized with a list of available aux tasks (by default "vertexing" and "track_origin"). Relevant information for these aux tasks will then be read in from the provided file, assuming the information is available and the specific aux task is properly supported within puma.

from __future__ import annotations

from puma.hlplots import AuxResults, Tagger
from puma.utils import get_dummy_tagger_aux, logger

# The line below generates dummy data which is similar to a NN output
fname, file = get_dummy_tagger_aux()

# define jet selections
cuts = [("n_truth_promptLepton", "==", 0)]

# define the tagger
GN2 = Tagger(
    name="GN2",
    label="dummy GN2",
    colour="#4477AA",
    reference=True,
)

# create the AuxResults object
aux_results = AuxResults(sample="dummy")

# load tagger from the file object
logger.info("Loading taggers.")
aux_results.load_taggers_from_file(
    [GN2],
    fname,
    cuts=cuts,
    num_jets=len(file["jets"]),
)

aux_results.atlas_second_tag = (
    "$\\sqrt{s}=13$ TeV, dummy jets \n$t\\bar{t}$, $20$ GeV $< p_{T} <250$ GeV"

Vertexing performance#

Vertexing performance plots can be produced for a specified jet flavour as shown in 

logger.info("Plotting vertexing performance.")
aux_results.plot_var_vtx_perf(vtx_flavours=["bjets"], no_vtx_flavours=["ujets"])

# Track origin confusion matrix
Here vtx_flavours defines a list of flavours for which secondary vertices are expected (e.g. b-jets) and no_vertex_flavours defines a list where secondary vertices are not expected (e.g. l-jets). Different plots are produced in each case (see below). In general, this plotting function handles all considerations for vertexing performance. This includes processing truth vertex indices by removing vertices containing tracks not from HF and reco vertices by removing the vertex most consistent with the reconstructed PV (if a tagger has the capability to identify tracks from a PV via track origin classification). See here for more information about truth track origin and vertex definitions. If inclusive vertexing is enabled, all HF vertices are merged into a single truth vertex. For reconstructed vertices in a tagger with track origin classification, all vertices with at least one HF track are merged and all others are removed. If track origin classification is not available, but inclusive vertexing is enabled, then all vertices are merged. After this cleaning procedure, a 1:1 greedy matching procedure between truth and reconstructed vertices is performed, with which all the relevant performance metrics for plots are calculated. In total, 4 plots are produced for each jet flavour with expected SVs and 1 is produced for each flavour with no expected SVs (all plotted against specific performance variable):

  • Vertexing efficiency: defined as number of vertices matched divided by number of true vertices (expected SVs)
  • Vertexing purity: defined as number of vertices matched divided by number of reconstructed vertices (expected SVs)
  • Track-vertex association efficiency: defined as number of tracks in matched vertex common between truth and reco vertices divided by number of tracks in true vertex (expected SVs)
  • Track-vertex association purity: defined as number of tracks in matched vertex common between truth and reco vertices divided by number of tracks in reco vertex (expected SVs)
  • Vertexing fake rate: fraction of jets where at least one SV is found (no expected SVs)

Note that by default the vertex matching algorithm enforces purity criteria requiring track association efficiency > 0.65 and purity > 0.5

Track Origin Performances#

The Track Origin auxiliary task is a multiclass classification task, in which each track is associated with its most probable belonging origin, chosen between: - Pileup; - Fake; - Primary; - FromB; - FromBC; - FromC; - FromTau; - OtherSecondary;

A good metric to evaluate the performances of the classifier is the Confusion Matrix, which can be plotted using the method plot_track_origin_confmat.

The normalization of the tagger's confusion matrix can be chosen among the ones allowed by the confusion_matrix possible normalizations, by specifying the argument normalize in the plot_track_origin_confmat function. By default, the matrix's rows are normalized.

If the ATLAS subtext is too long and overflows over the matrix, more space can be allocated for it by increasing the value of the function's argument atlas_offset, which is 1.5 by default.