tracking-parametrisation-tuner/parameterisations/train_matching_ghost_mlps_electron.py

# flake8: noqaq

import os
import argparse
import ROOT
from ROOT import TMVA, TList, TTree


def train_matching_ghost_mlp(
    input_file: str = "data/ghost_data_B.root",
    tree_name: str = "PrMatchNN.PrMCDebugMatchToolNN/MVAInputAndOutput",
    exclude_electrons: bool = False,
    only_electrons: bool = True,
    n_train_signal: int = 20e3,  # 50e3
    n_train_bkg: int = 50e3,  # 500e3
    n_test_signal: int = 10e3,
    n_test_bkg: int = 20e3,
    prepare_data: bool = True,
    outdir: str = "nn_electron_training",
):
    """Trains an MLP to classify the match between Velo and Seed track.

    Args:
        input_file (str, optional): Defaults to "data/ghost_data.root".
        tree_name (str, optional): Defaults to "PrMatchNN.PrMCDebugMatchToolNN/Tuple".
        exclude_electrons (bool, optional): Defaults to False.
        only_electrons (bool, optional): Signal only of electrons, but bkg of all particles. Defaults to True.
        n_train_signal (int, optional): Number of true matches to train on. Defaults to 200e3.
        n_train_bkg (int, optional): Number of fake matches to train on. Defaults to 200e3.
        n_test_signal (int, optional): Number of true matches to test on. Defaults to 75e3.
        n_test_bkg (int, optional): Number of fake matches to test on. Defaults to 75e3.
        prepare_data (bool, optional): Split data into signal and background file. Defaults to False.
    """

    if prepare_data:
        rdf = ROOT.RDataFrame(tree_name, input_file)
        if exclude_electrons:
            rdf_signal = rdf.Filter(
                "mc_quality == 1",  # -1 elec, 0 ghost, 1 all part wo elec
                "Signal is defined as one label (excluding electrons)",
            )
            rdf_bkg = rdf.Filter(
                "mc_quality == 0",
                "Ghosts are defined as zero label",
            )
        else:
            if only_electrons:
                rdf_signal = rdf.Filter(
                    "mc_quality == -1",  # electron that is true match but mlp said no match
                    "Signal is defined as negative one label (only electrons)",
                )
            else:
                rdf_signal = rdf.Filter(
                    "abs(mc_quality) > 0",
                    "Signal is defined as non-zero label",
                )
            rdf_bkg = rdf.Filter(
                "mc_quality == 0",
                "Ghosts are defined as zero label",
            )

        rdf_signal.Snapshot(
            "Signal",
            outdir + "/" + input_file.strip(".root") + "_matching_signal.root",
        )
        rdf_bkg.Snapshot(
            "Bkg",
            outdir + "/" + input_file.strip(".root") + "_matching_bkg.root",
        )

    signal_file = ROOT.TFile.Open(
        outdir + "/" + input_file.strip(".root") + "_matching_signal.root",
        "READ",
    )
    signal_tree = signal_file.Get("Signal")

    bkg_file = ROOT.TFile.Open(
        outdir + "/" + input_file.strip(".root") + "_matching_bkg.root"
    )
    bkg_tree = bkg_file.Get("Bkg")

    os.chdir(outdir + "/result")
    output = ROOT.TFile(
        "matching_ghost_mlp_training.root",
        "RECREATE",
    )

    factory = TMVA.Factory(
        "TMVAClassification",
        output,
        "V:!Silent:Color:DrawProgressBar:AnalysisType=Classification",
    )
    factory.SetVerbose(True)
    dataloader = TMVA.DataLoader("MatchNNDataSet")

    dataloader.AddVariable("mc_chi2", "F")
    dataloader.AddVariable("mc_teta2", "F")
    dataloader.AddVariable("mc_distX", "F")
    dataloader.AddVariable("mc_distY", "F")
    dataloader.AddVariable("mc_dSlope", "F")
    dataloader.AddVariable("mc_dSlopeY", "F")
    dataloader.AddVariable("mc_zMag", "F")

    dataloader.AddSignalTree(signal_tree, 1.0)
    dataloader.AddBackgroundTree(bkg_tree, 1.0)

    # these cuts are also applied in the algorithm
    preselectionCuts = ROOT.TCut(
        # "chi2<30 && distX<500 && distY<500 && dSlope<2.0 && dSlopeY<0.15",  #### ganz raus für elektronen
    )
    dataloader.PrepareTrainingAndTestTree(
        preselectionCuts,
        f"SplitMode=random:V:nTrain_Signal={n_train_signal}:nTrain_Background={n_train_bkg}:nTest_Signal={n_test_signal}:nTest_Background={n_test_bkg}",
        # normmode default is EqualNumEvents
    )

    factory.BookMethod(
        dataloader,
        TMVA.Types.kMLP,
        "matching_mlp",
        "!H:V:TrainingMethod=BP:NeuronType=ReLU:EstimatorType=CE:VarTransform=Norm:NCycles=700:HiddenLayers=N+2,N:TestRate=50:Sampling=1.0:SamplingImportance=1.0:LearningRate=0.02:DecayRate=0.01:!UseRegulator",
    )
    factory.TrainAllMethods()
    factory.TestAllMethods()
    factory.EvaluateAllMethods()
    output.Close()


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--input-file",
        type=str,
        help="Path to the input file",
        required=False,
    )
    parser.add_argument(
        "--exclude_electrons",
        action="store_true",
        help="Excludes electrons from training.",
        required=False,
    )
    parser.add_argument(
        "--only_electrons",
        action="store_true",
        help="Only electrons for signal training.",
        required=False,
    )
    parser.add_argument(
        "--n-train-signal",
        type=int,
        help="Number of training tracks for signal.",
        required=False,
    )
    parser.add_argument(
        "--n-train-bkg",
        type=int,
        help="Number of training tracks for bkg.",
        required=False,
    )
    parser.add_argument(
        "--n-test-signal",
        type=int,
        help="Number of testing tracks for signal.",
        required=False,
    )
    parser.add_argument(
        "--n-test-bkg",
        type=int,
        help="Number of testing tracks for bkg.",
        required=False,
    )

    args = parser.parse_args()
    args_dict = {arg: val for arg, val in vars(args).items() if val is not None}

    train_matching_ghost_mlp(**args_dict)