def VarCrossCheck(Sig, Bkg, SigW, BkgW, name, xmin, xmax, bins):
hSig = ROOT.TH1F("hSig", name, bins, xmin, xmax)
fill_hist(hSig, Sig, weights=SigW)
hSig.SetLineColor(2)
hSig.SetLineWidth(3)
SetOverflow(hSig)
hSig.Scale(np.sum(BkgW) / hSig.Integral())
hBkg = ROOT.TH1F("hBkg", name, bins, xmin, xmax)
fill_hist(hBkg, Bkg, weights=BkgW)
hBkg.SetLineWidth(3)
SetOverflow(hBkg)
c1 = ROOT.TCanvas("c1", "c1", 800, 600)
ROOT.gStyle.SetOptStat(0)
hBkg.Draw("Hist")
hSig.Draw("SameHist")
if (hSig.GetMaximum() > hBkg.GetMaximum()):
hBkg.SetMaximum(int(round(hSig.GetMaximum() * 1.1)))
hBkg.GetXaxis().SetTitle("Jet multiplicity")
hBkg.GetYaxis().SetTitle("Yield")
leg = ROOT.TLegend(0.7, 0.7, 0.9, 0.9)
leg.AddEntry(hSig, "Sig (Yield: {:04.2f})".format(np.sum(SigW)))
leg.AddEntry(hBkg, "Bkg (Yield: {:04.2f})".format(np.sum(BkgW)))
leg.Draw()
c1.Update()
c1.SaveAs("./plots/VarCrossCeck.png")
def makePlot(variable):
data_var = np.array(data[variable])
data[
"totalWeight"] = data.evtWeight * data.lep1_frWeight * data.lep2_frWeight * data.lep3_frWeight * data.tau_frWeight
data_weights = np.array(data['totalWeight'])
data_tight_var = np.array(data_tight[variable])
data_tight[
"totalWeight"] = data_tight.evtWeight * data_tight.lep1_frWeight * data_tight.lep2_frWeight * data_tight.lep3_frWeight
data_tight_weights = np.array(data_tight['totalWeight'])
c1 = TCanvas()
c1.SetFillColor(10)
c1.SetBorderSize(2)
c1.SetLeftMargin(0.12)
c1.SetBottomMargin(0.12)
c1.SetRightMargin(0.05)
c1.SetLogy()
histogram_base = TH1F("histogram_base", "", 100, np.nanmin(data_var),
np.nanmax(data_var))
histogram_base.SetTitle("")
histogram_base.SetStats(False)
histogram_base.SetMinimum(0.001)
histogram_base.SetMaximum(10.0)
histogram_base.GetXaxis().SetTitle(variable)
histogram_base.GetYaxis().SetTitle("Events")
histogram_base.Draw("hist")
hist_loose = TH1F("hist_loose", "", 100, np.nanmin(data_var),
np.nanmax(data_var))
hist_tight = TH1F("hist_tight", "", 100, np.nanmin(data_var),
np.nanmax(data_var))
root_numpy.fill_hist(hist_loose, data_var, weights=data_weights)
root_numpy.fill_hist(hist_tight,
data_tight_var,
weights=data_tight_weights)
hist_loose.SetLineColor(2)
hist_tight.SetLineColor(4)
hist_loose.SetFillColor(2)
hist_tight.SetFillColor(4)
hist_loose.SetFillStyle(3004)
hist_tight.SetFillStyle(3005)
leg = TLegend(0.2, 0.65, 0.5, 0.9)
leg.SetBorderSize(0)
leg.SetFillColor(10)
leg.SetLineColor(0)
leg.SetFillStyle(0)
leg.SetTextSize(0.04)
leg.SetTextFont(42)
leg.AddEntry(hist_loose, "loose", "F")
leg.AddEntry(hist_tight, "loose_genTau_matched", "F")
hist_loose.DrawNormalized("histsame")
hist_tight.DrawNormalized("histsame")
leg.Draw()
c1.SaveAs("plots/" + variable + "_" + process + "_loose_vs_genTau.png")
def Make_Binned_ROC_histograms(title, DDT, kNN, pT, bins, sample_weights=None):
rt.gROOT.SetBatch(True)
N = len(DDT)
assert len(pT) == N and len(kNN) == N
if sample_weights is not None: assert len(sample_weights) == N
nbins = 100
DDT_hist_list = []
kNN_hist_list = []
for bin_ in range(len(bins) - 1):
DDT_hist_list.append(
rt.TH1D("DDT_" + str(bins[bin_]) + "_" + str(bins[bin_ + 1]),
"DDT_" + str(bins[bin_]) + "_" + str(bins[bin_ + 1]),
nbins, 0, 1))
kNN_hist_list.append(
rt.TH1D("kNN_" + str(bins[bin_]) + "_" + str(bins[bin_ + 1]),
"kNN_" + str(bins[bin_]) + "_" + str(bins[bin_ + 1]),
nbins, 0, 1))
condition = np.logical_and((pT > bins[bin_]), (pT < bins[bin_ + 1]))
root_numpy.fill_hist(DDT_hist_list[bin_],
DDT[condition],
weights=sample_weights)
root_numpy.fill_hist(kNN_hist_list[bin_],
kNN[condition],
weights=sample_weights)
tfile = rt.TFile("ROC_doublechecks/{}_ROC_histograms.root".format(title),
"recreate")
for hist in DDT_hist_list:
hist.Write()
for hist in kNN_hist_list:
hist.Write()
print "saved histograms in ROC_doublechecks/{}_ROC_histograms.root".format(
title)
def _diff_plot1D_numpy(self,
data,
bins,
weights=None,
option='',
**kwargs):
""" ... """
# Check(s)
if bins is None:
warning(
"You need to specify 'bins' when plotting a numpy-type input.")
return
if len(bins) < 2:
warning("Number of bins {} is not accepted".format(len(bins)))
return
# Fill histogram
h1 = ROOT.TH1F('h_num_{}'.format(id(data)), "", len(bins) - 1, bins)
h2 = ROOT.TH1F('h_den_{}'.format(id(data)), "", len(bins) - 1, bins)
fill_hist(h1, data[0], weights=weights[0])
fill_hist(h2, data[1], weights=weights[1])
return _diff_plot1D((h1, h2), option, **kwargs)
def test_project_3d_to_2d(self):
hist_3d = _get_hist(3)
# populate overflow bins to make sure that they are treated as expected
fill_hist(hist_3d, np.random.uniform(-1, 0, (100, 3)))
fill_hist(hist_3d, np.random.uniform(1, 2, (100, 3)))
val3d, err3d = hu.get_array(hist_3d), hu.get_array(hist_3d, errors=True)
hist_xy = hu.project(hist_3d, 'xy')
val, err = hu.get_array(hist_xy), hu.get_array(hist_xy, errors=True)
npt.assert_equal(val, np.sum(val3d, axis=2))
npt.assert_equal(err, np.sqrt(np.sum(err3d**2, axis=2)))
hist_yz = hu.project(hist_3d, 'yz')
val, err = hu.get_array(hist_yz), hu.get_array(hist_yz, errors=True)
npt.assert_equal(val, np.sum(val3d, axis=0))
npt.assert_equal(err, np.sqrt(np.sum(err3d**2, axis=0)))
hist_zx = hu.project(hist_3d, 'zx')
val, err = hu.get_array(hist_zx), hu.get_array(hist_zx, errors=True)
npt.assert_equal(val, np.sum(val3d, axis=1).T)
npt.assert_equal(err, np.sqrt(np.sum(err3d**2, axis=1)).T)
hist_yx = hu.project(hist_3d, 'yx')
val, err = hu.get_array(hist_yx), hu.get_array(hist_yx, errors=True)
npt.assert_equal(val, np.sum(val3d, axis=2).T)
npt.assert_equal(err, np.sqrt(np.sum(err3d**2, axis=2)).T)
def plotRatio(input_df, inputFile):
gStyle.SetOptStat(0)
#hPtRatioPredicted = TH1F('hPtRatioPredicted', 'NN prediction, RMS=0.135', 40, 0., 2.)
#hPtRatioCorrected = TH1F('hPtRatioPredicted', 'Colinear approximation, RMS=0.175', 40, 0., 2.)
rmse1 = ((input_df['bc_ptRatio_predictedGen'].mean() -
input_df['bc_ptRatio_predictedGen'])**2).mean()**.5
rmse2 = ((input_df['bc_ptRatio_correctedGen'].mean() -
input_df['bc_ptRatio_correctedGen'])**2).mean()**.5
label1 = "NN prediction, RMS=%3.3f" % rmse1
label2 = "Colinear correction, RMS=%3.3f" % rmse2
hPtRatioPredicted = TH1F('hPtRatioPredicted', label1, 40, 0., 2.)
hPtRatioCorrected = TH1F('hPtRatioPredicted', label2, 40, 0., 2.)
fill_hist(hPtRatioPredicted,
input_df['bc_ptRatio_predictedGen'].to_numpy())
fill_hist(hPtRatioCorrected,
input_df['bc_ptRatio_correctedGen'].to_numpy())
c1 = TCanvas('c1', 'c1', 700, 500)
hPtRatioPredicted.SetLineColor(ROOT.kAzure)
hPtRatioCorrected.SetLineColor(ROOT.kRed)
#profilePtGenVsPtRatioPredictedGen.SetTitle("")
gStyle.SetOptTitle(0)
hPtRatioPredicted.GetYaxis().SetTitle("Events/50 MeV")
hPtRatioPredicted.GetXaxis().SetTitle(
"pT_{corrected}(B_{c}^{+})/pT_{gen}(B_{c}^{+})")
hPtRatioPredicted.Draw("")
hPtRatioCorrected.Draw('same')
gPad.BuildLegend()
c1.SaveAs(plotsDir + inputFile + '_ratio_predicted_gen.pdf')
def plotProfile(input_df):
gStyle.SetOptStat(0)
hPtGenVsPtRatioPredictedGen = TH2F('hPtGenVsPtRatioPredictedGen',
'NN prediction', 80, 0., 80., 40, 0.,
2.)
fill_hist(hPtGenVsPtRatioPredictedGen,
input_df[['gen_b_pt', 'bc_ptRatio_predictedGen']].to_numpy())
profilePtGenVsPtRatioPredictedGen = hPtGenVsPtRatioPredictedGen.ProfileX()
profilePtGenVsPtRatioPredictedGen.SetMarkerStyle(ROOT.kFullCircle)
hPtGenVsPtRatioCorrectedGen = TH2F('hPtGenVsPtRatioCorrectedGen',
'Jonas correction', 80, 0., 80., 40, 0.,
2.)
fill_hist(hPtGenVsPtRatioCorrectedGen,
input_df[['gen_b_pt', 'bc_ptRatio_correctedGen']].to_numpy())
profilePtGenVsPtRatioCorrectedGen = hPtGenVsPtRatioCorrectedGen.ProfileX()
profilePtGenVsPtRatioCorrectedGen.SetMarkerStyle(ROOT.kFullSquare)
c1 = TCanvas('c1', 'c1', 700, 500)
profilePtGenVsPtRatioPredictedGen.SetLineColor(ROOT.kAzure)
profilePtGenVsPtRatioCorrectedGen.SetLineColor(ROOT.kOrange)
#profilePtGenVsPtRatioPredictedGen.SetTitle("")
gStyle.SetOptTitle(0)
profilePtGenVsPtRatioPredictedGen.GetXaxis().SetTitle("pT_{gen}(Bc) [GeV]")
profilePtGenVsPtRatioPredictedGen.GetYaxis().SetTitle(
"pT_{corrected}(B_{c}^{+})/pT_{gen}(B_{c}^{+})")
profilePtGenVsPtRatioPredictedGen.Draw("")
profilePtGenVsPtRatioCorrectedGen.Draw('same')
gPad.BuildLegend()
c1.SaveAs(plotsDir + 'profile.png')
return 0
def Plot_variable_from_data_2D(out_file,
data_path,
x_title,
x_var,
x_bins,
x_range,
y_title,
y_var,
y_bins,
y_range,
test=False,
weights=None):
if test:
data, features, _ = load_data(data_path, test_full_signal=True)
else:
data, features, _ = load_data(data_path, train_full_signal=True)
if weights is not None:
weights = data[weights]
f1 = ROOT.TFile(out_file, "RECREATE")
hist = ROOT.TH2D('hist', 'hist', x_bins, x_range[0], x_range[1], y_bins,
y_range[0], y_range[1])
X = data[x_var]
Y = data[y_var]
root_numpy.fill_hist(hist, np.vstack((X, Y)).T, weights=weights)
canv = ROOT.TCanvas('canv', 'canv', 600, 600)
hist.SetContour(256)
hist.GetXaxis().SetTitle(x_title)
hist.GetYaxis().SetTitle(y_title)
hist.Draw("COLZ")
canv.Write()
f1.Close()
def create_TH1D(x,
name='h',
title=None,
binning=[None, None, None],
weights=None,
h2clone=None):
if title is None:
title = name
if h2clone == None:
if binning[1] is None:
binning[1] = min(x)
if binning[2] is None:
if ((np.percentile(x, 95) - np.percentile(x, 50)) < 0.2 *
(max(x) - np.percentile(x, 95))):
binning[2] = np.percentile(x, 90)
else:
binning[2] = max(x)
if binning[0] is None:
bin_w = 4 * (np.percentile(x, 75) -
np.percentile(x, 25)) / (len(x))**(1. / 3.)
binning[0] = int((binning[2] - binning[1]) / bin_w)
h = rt.TH1D(name, title, binning[0], binning[1], binning[2])
else:
h = h2clone.Clone(name)
h.SetTitle(title)
h.Reset()
rtnp.fill_hist(h, x, weights=weights)
h.binning = binning
return h
def scatterplot(dfevt, nvar1, nvar2, nbins1, min1, max1, nbins2, min2, max2):
hmult1_mult2 = TH2F(nvar1 + nvar2, nvar1 + nvar2, nbins1, min1, max1,
nbins2, min2, max2)
dfevt_rd = dfevt[[nvar1, nvar2]]
arr2 = dfevt_rd.values
fill_hist(hmult1_mult2, arr2)
return hmult1_mult2
def create_TH2D(sample,
name='h',
title=None,
binning=[None, None, None, None, None, None],
weights=None,
axis_title=['', '', '']):
if title is None:
title = name
if (sample.shape[0] == 0):
for i in range(len(binning)):
if binning[i] == None:
binning[i] = 1
else:
if binning[1] is None:
binning[1] = min(sample[:, 0])
if binning[2] is None:
binning[2] = max(sample[:, 0])
if binning[0] is None:
bin_w = 4 * (np.percentile(sample[:, 0], 75) - np.percentile(
sample[:, 0], 25)) / (len(sample[:, 0]))**(1. / 3.)
if bin_w == 0:
bin_w = 0.5 * np.std(sample[:, 0])
if bin_w == 0:
bin_w = 1
binning[0] = int((binning[2] - binning[1]) / bin_w)
if binning[4] is None:
binning[4] = min(sample[:, 1])
if binning[5] == None:
binning[5] = max(sample[:, 1])
if binning[3] == None:
bin_w = 4 * (np.percentile(sample[:, 1], 75) - np.percentile(
sample[:, 1], 25)) / (len(sample[:, 1]))**(1. / 3.)
if bin_w == 0:
bin_w = 0.5 * np.std(sample[:, 1])
if bin_w == 0:
bin_w = 1
binning[3] = int((binning[5] - binning[4]) / bin_w)
if len(binning) == 6:
h = rt.TH2D(name, title, binning[0], binning[1], binning[2],
binning[3], binning[4], binning[5])
else:
h = rt.TH2D(name, title, binning[-2] - 1,
array('f', binning[:binning[-2]]), binning[-1] - 1,
array('f', binning[binning[-2]:-2]))
#for i in range(len(sample)):
# if weights is None:
# h.Fill(sample[i,0],sample[i,1])
# else:
# h.Fill(sample[i,0],sample[i,1],weights[i])
rtnp.fill_hist(h, sample, weights=weights)
h.SetXTitle(axis_title[0])
h.SetYTitle(axis_title[1])
h.SetZTitle(axis_title[2])
h.binning = binning
return h
def SigBkgHist(Sample, XTitle, bins, xmin, xmax, tag=''):
x = Sample.Events
Sig = x[Sample.OutTrue == 1] # Signal values of each event
wSig = Sample.Weights[Sample.OutTrue == 1] # Signal weights of each event
Bkg = x[Sample.OutTrue == 0] # Background values of each event
wBkg = Sample.Weights[Sample.OutTrue ==
0] # Background weights of each event
c1 = ROOT.TCanvas("c1", "Canvas", 800, 600)
ROOT.gStyle.SetOptStat(0)
hSig = ROOT.TH1F("hSig", "", bins, xmin, xmax)
fill_hist(hSig, Sig, weights=wSig)
SetOverflow(hSig)
SetUnderflow(hSig)
hBkg = ROOT.TH1F("hBkg", "", bins, xmin, xmax)
fill_hist(hBkg, Bkg, weights=wBkg)
SetOverflow(hBkg)
SetUnderflow(hBkg)
hBkg.SetLineColor(2)
if (hSig.GetMaximum() > hBkg.GetMaximum()):
hSig.GetYaxis().SetRangeUser(0, hSig.GetMaximum() * 1.4)
else:
hSig.GetYaxis().SetRangeUser(0, hBkg.GetMaximum() * 1.4)
hSig.GetXaxis().SetTitle(XTitle)
hSig.Draw("Hist")
hBkg.Draw("SameHist")
c1.SaveAs("./plots/SigBkg" + tag + ".png")
def makefill1dhist(df_, h_name, h_tit, arrayx, nvar1):
"""
Create a TH1F histogram and fill it with one variables from a dataframe.
"""
histo = buildhisto(h_name, h_tit, arrayx)
fill_hist(histo, df_[nvar1])
return histo
def fillHistograms(self, cut, path, category, weight = Weight("1.0",[]) ):
data_content = self.read_data_content(cut, path, weight)
ff_weights = self.getFFWeights(data_content)
FFHistos = {}
for i,uncert in enumerate(self.uncerts):
if not "jetFakes" in uncert:
name = self.convertSystematicName(uncert)
else:
name = uncert
FFHistos[name] = getEmptyHist(name, self.variable, category)
FFHistos[name].Sumw2(True)
if len(data_content):
rn.fill_hist( FFHistos[name], array = data_content[self.variable.getBranches()].values,
weights = ff_weights[i].values )
FFHistos[name] = self.unroll2D(FFHistos[name])
data_content.drop( data_content.index, inplace = True )
return copy.deepcopy(FFHistos)
def setup(self):
if os.path.exists(self.output_dir):
os.system("rm -rf %s" % (self.output_dir + "/*" + self.tag + "*"))
else:
os.system("mkdir %s" % self.output_dir)
# Initialize workspace
self.w = ROOT.RooWorkspace("w")
self.rooVar = "mgg"
# Initialize and fill histogram
self.h = ROOT.TH1F("h_mgg", "h_mgg", 320, 100, 180)
self.h.Sumw2()
root_numpy.fill_hist(self.h, self.events["ggMass"], weights = self.events["weight"])
# Convert to RooDataHist
self.d = ROOT.RooDataHist("d_mgg_" + self.tag, "", ROOT.RooArgList(self.w.var(self.rooVar)), self.h, 1)
self.norm = self.d.sumEntries()
self.rooVarNorm = ROOT.RooRealVar(self.tag + "_norm", "", self.norm)
self.pdf = ROOT.RooExtendPdf(self.tag + "_pdf", "", self.w.pdf(self.tag), self.rooVarNorm)
if not self.resonant:
self.w.var(self.rooVar).setRange("SL", 100, 120)
self.w.var(self.rooVar).setRange("SU", 130, 180)
self.w.var(self.rooVar).setRange("full", 100, 180)
return
def process_histomass(self):
myfile = TFile.Open(self.n_filemass, "recreate")
for ipt in range(self.p_nptfinbins):
bin_id = self.bin_matching[ipt]
df = pickle.load(openfile(self.lpt_recodecmerged[bin_id], "rb"))
df = df.query(self.l_selml[bin_id])
if self.s_evtsel is not None:
df = df.query(self.s_evtsel)
if self.s_trigger is not None:
df = df.query(self.s_trigger)
df = seldf_singlevar(df, self.v_var_binning, \
self.lpt_finbinmin[ipt], self.lpt_finbinmax[ipt])
for ibin2 in range(len(self.lvar2_binmin)):
suffix = "%s%d_%d_%.2f%s_%.2f_%.2f" % \
(self.v_var_binning, self.lpt_finbinmin[ipt],
self.lpt_finbinmax[ipt], self.lpt_probcutfin[bin_id],
self.v_var2_binning, self.lvar2_binmin[ibin2], self.lvar2_binmax[ibin2])
h_invmass = TH1F("hmass" + suffix, "", self.p_num_bins,
self.p_mass_fit_lim[0], self.p_mass_fit_lim[1])
df_bin = seldf_singlevar(df, self.v_var2_binning,
self.lvar2_binmin[ibin2], self.lvar2_binmax[ibin2])
fill_hist(h_invmass, df_bin.inv_mass)
myfile.cd()
h_invmass.Write()
if "pt_jet" in df_bin.columns:
zarray = z_calc(df_bin.pt_jet, df_bin.phi_jet, df_bin.eta_jet,
df_bin.pt_cand, df_bin.phi_cand, df_bin.eta_cand)
h_zvsinvmass = TH2F("hzvsmass" + suffix, "", 5000, 1.00, 6.00, 2000, -0.5, 1.5)
zvsinvmass = np.vstack((df_bin.inv_mass, zarray)).T
fill_hist(h_zvsinvmass, zvsinvmass)
h_zvsinvmass.Write()
def plot_hist(self, values, bins, **params):
name = None
if 'name' in params:
name = params['name']
if ('root' in params):
histo = r.TH1F(name, name, len(bins), np.amin(bins), np.amax(bins))
rnp.fill_hist(histo, values)
histo.Scale(1/histo.Integral(), 'width')
histo.Write()
fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(15, 10))
label=None
if 'label' in params:
label=params['label']
ax.legend()
norm = False
if 'norm' in params:
norm = True
if 'ylog' in params:
if norm: ax.set_yscale('log')
else: ax.set_yscale('symlog')
if 'x_label' in params:
ax.set_xlabel(params['x_label'])
ax.hist(values, bins, histtype='step', lw=1.5, label=label, normed=norm)
self.pdf.savefig(bbox_inches='tight')
plt.close()
def _plot1D_numpy(self, data, bins, weights=None, option='', **kwargs):
""" ... """
# Check(s)
if bins is None:
warning(
"You need to specify 'bins' when plotting a numpy-type input.")
return
if len(data) != len(bins) and len(bins) < 2:
warning("Number of bins {} is not accepted".format(len(bins)))
return
# Fill histogram
if len(data) == len(bins):
# Assuming 'data' and 'bins' are sets of (x,y)-points
h = ROOT.TGraph(len(bins), np.array(bins, dtype=np.float),
np.array(data, dtype=np.float))
else:
h = ROOT.TH1F('h_{:d}'.format(int(time.time() * 1E+06)), "",
len(bins) - 1, np.array(bins, dtype=np.float))
if len(data) == len(bins) - 1:
# Assuming 'data' are bin values
array2hist(data, h)
else:
# Assuming 'data' are values to be filled
fill_hist(h, data, weights=weights)
pass
pass
# Plot histogram
return self._plot1D(h, option, **kwargs)
def _analyze(self, file):
tree = ur.open(file)["Events"]
arrays = tree.arrays(
['Jet_pt', 'nJet', 'Jet_eta', 'Jet_phi', 'MET_pt'])
jet_pt = arrays[b"Jet_pt"]
jet_eta = arrays[b"Jet_eta"]
jet_phi = arrays[b"Jet_phi"]
met_pt = arrays[b"MET_pt"]
n_jet = arrays[b"nJet"]
# Select events with more than one jet
mask = (n_jet > 1)
jet_pt = jet_pt[mask]
jet_phi = jet_phi[mask]
jet_eta = jet_eta[mask]
jet_ht = calculate_jet_ht(jet_pt)
fill_hist(self._histos["ptj"], jet_pt.flatten())
fill_hist(self._histos["ht"], calculate_jet_ht(jet_pt).flatten())
fill_hist(self._histos["lead_jet_eta"], jet_eta[:, 0])
fill_hist(self._histos["trail_jet_eta"], jet_eta[:, 1])
twod = jet_eta[:, 0:2].tolist()
fill_hist(self._histos["jet_eta_2d"], twod)
def write_score_hists(f, mass, scores_list, hist_template, no_neg_bins=True):
sys_hists = {}
for samp, scores_dict in scores_list:
for sys_term, (scores, weights) in scores_dict.items():
if sys_term == 'NOMINAL':
suffix = ''
else:
suffix = '_' + '_'.join(sys_term)
hist = hist_template.Clone(
name=samp.name + ('_{0}'.format(mass)) + suffix)
fill_hist(hist, scores, weights)
if sys_term not in sys_hists:
sys_hists[sys_term] = []
sys_hists[sys_term].append(hist)
f.cd()
for sys_term, hists in sys_hists.items():
bad_bins = []
if no_neg_bins:
# check for negative bins over all systematics and zero them out
# negative bins cause lots of problem in the limit setting
# negative bin contents effectively means
# the same as "no events here..."
total_hist = sum(hists)
for bin, content in enumerate(total_hist):
if content < 0:
log.warning("Found negative bin %d (%f) for "
"systematic %s" % (
bin, content, sys_term))
bad_bins.append(bin)
for hist in hists:
for bin in bad_bins:
# zero out bad bins
hist[bin] = 0.
hist.Write()
def efficiency_graph(pass_function,
function_inputs,
xs,
bins=None,
error=0.005):
pass_results = pass_function(function_inputs)
if bins is None: # Automatic binning
# Compute the number of bins such that the error on the efficiency is equal to 'error' in each bin
# The calculation is based on binomial errors and assumes that the efficiency is flat (that the distributions of all and selected events are the same)
k = float(np.count_nonzero(pass_results))
n = float(len(pass_results))
percentiles = [0., 100.]
if k > 0:
nbins = (error * n)**2 / k / (1 - k / n)
# Compute the bin bounaries with the same number of events in all bins
percentiles = np.arange(0., 100., 100. / nbins)
percentiles[-1] = 100.
bins = np.unique(np.percentile(xs, percentiles))
# Fill histograms of selected and all events and compute efficiency
histo_pass = Hist(bins)
histo_total = Hist(bins)
fill_hist(histo_pass, xs, pass_results)
fill_hist(histo_total, xs)
efficiency = Graph()
efficiency.Divide(histo_pass, histo_total)
return efficiency
def main ():
inputdir = '/eos/atlas/user/a/asogaard/Analysis/2016/BoostedJetISR/StatsInputs/2017-06-28/'
outputdir = '/eos/atlas/user/a/asogaard/Analysis/2016/BoostedJetISR/StatsInputs/2017-07-10/'
inputpaths = glob.glob(inputdir + '/ISRgamma_*.root')
outputpaths = [p.replace(inputdir, outputdir).replace('ISRgamma', 'hist_ISRgamma') for p in inputpaths]
for inputpath, outputpath in zip(inputpaths,outputpaths):
print "Processing '%s'" % inputpath
infile = ROOT.TFile(inputpath, 'READ')
outfile = ROOT.TFile(outputpath, 'RECREATE')
categories = [key.GetName() for key in infile.GetListOfKeys()]
for category in categories:
print "-- '%s'" % category
tree = infile.Get(category)
array = tree2array(tree)
#hist = ROOT.TH1F(category, "", 30, 100, 250)
hist = ROOT.TH1F(category, "", 32, 100, 260)
fill_hist(hist, array['mJ'], weights=array['weight'])
# TF shape/norm ...
outfile.cd()
hist.Write()
pass
outfile.Write()
outfile.Close()
infile.Close()
pass
return
def fillHistos(self, content, histname, cat, cut, add_systematics):
binning = self.var.bins(int(cat))
tmpCont = content.query(cut)
tmpHist = R.TH1D(histname, histname, *binning)
tmpHist.GetXaxis().SetTitle(self.var.name)
tmpHist.Sumw2()
rn.fill_hist(tmpHist,
array=tmpCont[self.var.name].values,
weights=tmpCont["event_weight"].values)
self.DCfile.cd(self.d(self.target_names[int(cat)]))
tmpHist.Write()
if add_systematics:
for rw in self.systematics:
rwname = rw.replace("reweight", histname).replace(
"CHAN", self.channel).replace("CAT",
self.target_names[int(cat)])
tmpHist = R.TH1D(rwname, rwname, *binning)
rn.fill_hist(tmpHist,
array=tmpCont[self.var.name].values,
weights=tmpCont.eval(self.systematics[rw]).values)
tmpHist.Write()
def create_TH2D(sample,
name='h',
title=None,
binning=[None, None, None, None, None, None],
weights=None,
axis_title=['', '']):
if title is None:
title = name
if binning[1] is None:
binning[1] = min(sample[:, 0])
if binning[2] is None:
binning[2] = max(sample[:, 0])
if binning[0] is None:
bin_w = 4 * (np.percentile(sample[:, 0], 75) - np.percentile(
sample[:, 0], 25)) / (len(sample[:, 0]))**(1. / 3.)
binning[0] = int((binning[2] - binning[1]) / bin_w)
if binning[4] is None:
binning[4] = min(sample[:, 1])
if binning[5] == None:
binning[5] = max(sample[:, 1])
if binning[3] == None:
bin_w = 4 * (np.percentile(sample[:, 1], 75) - np.percentile(
sample[:, 1], 25)) / (len(sample[:, 1]))**(1. / 3.)
binning[3] = int((binning[5] - binning[4]) / bin_w)
h = rt.TH2D(name, title, binning[0], binning[1], binning[2], binning[3],
binning[4], binning[5])
rtnp.fill_hist(h, sample, weights=weights)
h.SetXTitle(axis_title[0])
h.SetYTitle(axis_title[1])
h.binning = binning
return h
def create_histogram(var, hist_sett, **kwargs):
"""
Create a ROOT histogram from the passed variable(s)
Args:
var (np.array): Array with maximum of 3 columns containing the variables
to plot.
hist_set (tuple): Histogram settings, that are directly unpacked into
the constructor of the ROOT histogram
Keyword Args:
name (str, optional): Name to be used for the histogram
weights (np.array, optional): weight array with the same number of
events as the var array. Each entry corresponds to the weight of
the event
{x,y,z}_axis (str): axis labels to be set for the histogram
Returns:
ROOT.TH{1,2,3}D: The histogram with the dimension corresponding to the
number of columns of var
"""
name = kwargs.pop('name', '')
if not name:
name = create_random_str()
# use the number of dimensions from the var to determine which sort of
# histogram to use
ndim = var.shape
if len(ndim) == 1:
ndim = 1
else:
ndim = ndim[1]
if ndim > 3 or ndim < 0:
logging.error('Dimension of histogram is {}. Cannot create histogram'
.format(ndim))
raise TypeError('Invalid number of dimensions in create_histograms')
hist_type = 'TH{}D'.format(ndim)
try:
hist = getattr(r, hist_type)(name, '', *hist_sett)
except TypeError as exc:
logging.error('Could not construct TH{}D with passed hist_sett: {}'
.format(ndim, hist_sett))
raise exc
set_hist_opts(hist)
# set axis labels
xax, yax, zax = (kwargs.pop(a, '') for a in ['x_axis', 'y_axis', 'z_axis'])
if xax:
hist.SetXTitle(xax)
if yax:
hist.SetYTitle(yax)
if zax:
hist.SetZTitle(zax)
fill_hist(hist, var, weights=kwargs.pop('weights', None))
return hist
def HistNW(hist, Events, Weights, Color):
fill_hist(hist, Events, weights=Weights)
hist.SetLineColor(Color)
hist.SetLineWidth(1)
SetOverflow(hist)
SetUnderflow(hist)
if (hist.Integral() != 0):
hist.Scale(1 / hist.Integral())
def fill2dhist(df_, histo, nvar1, nvar2):
"""
Fill a TH2 histogram with two variables from a dataframe.
"""
df_rd = df_[[nvar1, nvar2]]
arr2 = df_rd.values
fill_hist(histo, arr2)
return histo
def draw_hist(hist, xarr, fields, weight=None):
warr = xarr[weight] if weight else None
if len(fields) == 1:
return rnp.fill_hist(hist=hist, array=xarr[fields[0]], weights=warr)
else:
varr = np.array([xarr[f] for f in fields])
varr = varr.transpose()
return rnp.fill_hist(hist=hist, array=varr, weights=warr)
def create_TH1D(x,
name='h',
title=None,
binning=[None, None, None],
weights=None,
h2clone=None,
axis_title=['', ''],
opt='',
color=0):
if title is None:
title = name
if (x.shape[0] == 0):
print('Empty sample')
h = rt.TH1D(name, title, 1, 0, 1)
elif not h2clone is None:
h = h2clone.Clone(name)
h.SetTitle(title)
h.Reset()
elif isinstance(binning, np.ndarray):
h = rt.TH1D(name, title, len(binning) - 1, binning)
elif len(binning) == 3:
if binning[1] is None:
binning[1] = min(x)
if binning[2] is None:
if ((np.percentile(x, 95) - np.percentile(x, 50)) < 0.2 *
(max(x) - np.percentile(x, 95))):
binning[2] = np.percentile(x, 90)
else:
binning[2] = max(x)
if binning[0] is None:
bin_w = 4 * (np.percentile(x, 75) -
np.percentile(x, 25)) / (len(x))**(1. / 3.)
if bin_w == 0:
bin_w = 0.5 * np.std(x)
if bin_w == 0:
bin_w = 1
binning[0] = int((binning[2] - binning[1]) / bin_w) + 5
h = rt.TH1D(name, title, binning[0], binning[1], binning[2])
else:
print('Binning not recognized')
raise
if 'underflow' in opt:
m = h.GetBinCenter(1)
x = np.copy(x)
x[x < m] = m
if 'overflow' in opt:
M = h.GetBinCenter(h.GetNbinsX())
x = np.copy(x)
x[x > M] = M
rtnp.fill_hist(h, x, weights=weights)
h.SetXTitle(axis_title[0])
h.SetYTitle(axis_title[1])
h.SetLineColor(color)
h.binning = binning
return h
def check_sample_category(analysis, sample, category):
clf = analysis.get_clf(category, mass=125, load=True)
scores, weights = sample.scores(
clf, category, TARGET_REGION,
systematics=False)['NOMINAL']
hist = Hist(20, scores.min() - 1E-5, scores.max() + 1E-5)
fill_hist(hist, scores, weights)
assert_almost_equal(sample.events(category, TARGET_REGION)[1].value,
hist.integral(), 3)
def makefill2dhist(df_, titlehist, arrayx, arrayy, nvar1, nvar2):
"""
Create a TH2F histogram and fill it with two variables from a dataframe.
"""
histo = build2dhisto(titlehist, arrayx, arrayy)
df_rd = df_[[nvar1, nvar2]]
arr2 = df_rd.to_numpy()
fill_hist(histo, arr2)
return histo
def GetScoreTH(ClassNum, OutPreOther, Weights, MultiClass, xmin, xmax):
Pre = OutPreOther[MultiClass == ClassNum]
Weight = Weights[MultiClass == ClassNum]
hist = ROOT.TH1F("h" + str(ClassNum), "", 20, xmin, xmax)
fill_hist(hist, Pre, weights=Weight)
hist.Scale(1 / hist.GetSumOfWeights() / Getdx(hist))
return hist
def cutBasedAMS(self):
sig_hist = r.TH1F('sigHistTemp', ';m_{ee};Entries', 160, 110, 150)
fill_hist(sig_hist, self.sigMass, weights=self.sigWeights)
N_sig = 0.68 * sig_hist.Integral()
sig_width = self.getRealSigma(sig_hist)
bkg_hist = r.TH1F('bkgHistTemp', ';m_{ee};Entries', 160, 110, 150)
fill_hist(bkg_hist, self.bkgMass, weights=self.bkgWeights)
N_bkg = self.computeBkg(bkg_hist, sig_width)
return self.getAMS(N_sig, N_bkg)
def get_background_signal(nbins=100, scale=1, sig_events=1000, bkg_events=1000):
bkg_scores = transform(np.random.normal(-.1, .2, size=bkg_events),
scale=scale)
sig_scores = transform(np.random.normal(.1, .2, size=sig_events),
scale=scale)
bkg_hist = Hist(nbins, -1, 1)
sig_hist = Hist(nbins, -1, 1)
fill_hist(bkg_scores)
fill_hist(sig_scores)
return bkg_hist, sig_hist
def fillHistFromNumpy(np_arr, histList=[], histname=""):
"""
Fill a 1D or 2D hist from a np array
Inputs: The numpy array and a list with [nbins, min, max] for 1D or
[nbinsX, minX, maxX, nbinsY, minY, maxY]
Return: The hist
"""
from ROOT import TH1F, TH2F
from root_numpy import fill_hist
hist = 0
if histname =="" : histname="hist"
if len(histList) == 0 : pass
elif len(hist)==3 : hist = ROOT.TH1F(histname, histname, int(histList[0]),histList[1],histList[2])
elif len(hist)==6 : hist = ROOT.TH2F(histname, histname, int(histList[0]),histList[1],histList[2], int(histList[3]),histList[4],histList[5])
info('(fillHistFromNumpy) filling hist with name %s' % histname)
fill_hist(hist, np_arr)
return hist
def plot_hists(self, values, bins, **params):
fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(15, 10))
label=None
norm = False
if 'norm' in params:
norm = True
if 'ylog' in params:
if norm: ax.set_yscale('log')
else: ax.set_yscale('symlog')
if 'x_label' in params:
ax.set_xlabel(params['x_label'])
labels = None
if 'labels' in params:
labels = params['labels']
label_loc=0
if 'label_loc' in params:
label_loc=params['label_loc']
for x_arr, label in izip(values, labels):
ax.hist(x_arr, bins, histtype='step', lw=1.5, normed=norm, label=label)
if labels: ax.legend(framealpha=0.0, frameon=False, loc=label_loc)
self.pdf.savefig(bbox_inches='tight')
plt.close()
if ('root' in params):
for x_arr, label in izip(values, labels):
histo = r.TH1F(label, label, len(bins), np.amin(bins), np.amax(bins))
rnp.fill_hist(histo, x_arr)
histo.Scale(1/histo.Integral(), 'width')
histo.Write()
def efficiency_graph(pass_function, function_inputs, xs, bins=None, error=0.005):
pass_results = pass_function(function_inputs)
if bins is None: # Automatic binning
# Compute the number of bins such that the error on the efficiency is equal to 'error' in each bin
# The calculation is based on binomial errors and assumes that the efficiency is flat (that the distributions of all and selected events are the same)
k = float(np.count_nonzero(pass_results))
n = float(len(pass_results))
percentiles = [0.,100.]
if k>0:
nbins = (error*n)**2/k / (1-k/n)
# Compute the bin boundaries with the same number of events in all bins
percentiles = np.arange(0., 100., 100./nbins)
percentiles[-1] = 100.
bins = np.unique(np.percentile(xs, percentiles))
# Fill histograms of selected and all events and compute efficiency
histo_pass = Hist(bins)
histo_total = Hist(bins)
fill_hist(histo_pass, xs, pass_results)
fill_hist(histo_total, xs)
efficiency = Graph()
efficiency.Divide(histo_pass, histo_total)
return efficiency
def calculateBkgRej(self, discriminant, signal_idx, bkg_idx, weights=None):
'''
This does essentially the same thing as the plotDiscriminant method, except that it does it for
an arbritrary discriminant and doesn't save the histograms. It just calculates the score.
'''
import ROOT as root
from ROOT import TH2D, TCanvas, TFile, TNamed, TH1F, TLegend
import numpy as np
from root_numpy import fill_hist
import functions as fn
import os
# stop showing plots to screen
root.gROOT.SetBatch(True)
bins = 100
# when creating the plots do it over the range of all probas (scores)
discriminant_bins = np.linspace(np.min(discriminant), np.max(discriminant), bins)
hist_bkg = TH1F("Background Discriminant","Discriminant",bins, np.min(discriminant), np.max(discriminant))
hist_sig = TH1F("Signal Discriminant","Discriminant",bins, np.min(discriminant), np.max(discriminant))
# fill the signal and background histograms
if weights is not None:
fill_hist(hist_bkg,discriminant[bkg_idx], weights[bkg_idx])
fill_hist(hist_sig,discriminant[signal_idx], weights[signal_idx])
else:
fill_hist(hist_bkg,discriminant[bkg_idx])
fill_hist(hist_sig,discriminant[signal_idx])
if hist_bkg.Integral() != 0:
hist_bkg.Scale(1/hist_bkg.Integral())
if hist_sig.Integral() != 0:
hist_sig.Scale(1/hist_sig.Integral())
# before deciding whether to do a left or right cut for the roc curve we have to find the median.
sig_median = np.median(discriminant[signal_idx])
bkg_median = np.median(discriminant[bkg_idx])
if sig_median > bkg_median:
roc_cut = 'R'
else:
roc_cut = 'L'
roc_graph = fn.RocCurve_SingleSided(hist_sig, hist_bkg, self.sig_eff,self.bkg_eff, roc_cut)
fpr_05 = fn.GetBGRej50(roc_graph)
if fpr_05 != 1:
return float(1/(1-fpr_05))
return -1.0
def check_events(analysis, sample, category, region):
clf = analysis.get_clf(category, mass=125, load=True)
scores, weights = sample.scores(
clf, category, region,
systematics=False)['NOMINAL']
rec = sample.merged_records(
category, region)
sample_events = sample.events(category, region)[1].value
hist = Hist(5, scores.min() - 1, scores.max() + 1)
fill_hist(hist, scores, weights)
clf_events = hist.integral()
# test events consistency
assert_equal(weights.shape[0], rec['weight'].shape[0])
assert_array_equal(weights, rec['weight'])
assert_almost_equal(clf_events, weights.sum(), 1)
assert_almost_equal(sample_events, rec['weight'].sum(), 1)
assert_almost_equal(sample_events, clf_events, 1)
# test draw_array
hist = Hist(1, -1000, 1000)
sample.draw_array({'tau1_charge': hist}, category, region)
assert_almost_equal(hist.integral(), sample_events, 1)
# test scaling
orig_scale = sample.scale
sample.scale *= 2.
scores, weights = sample.scores(
clf, category, region,
systematics=False)['NOMINAL']
hist.Reset()
fill_hist(hist, scores, weights)
scale_clf_events = hist.integral()
assert_almost_equal(scale_clf_events, weights.sum(), 1)
assert_almost_equal(scale_clf_events, 2. * clf_events, 1)
scale_sample_events = sample.events(category, region)[1].value
assert_almost_equal(scale_sample_events, 2. * sample_events, 1)
sample.scale = orig_scale
def test_fill_hist():
np.random.seed(0)
data1D = np.random.randn(1E6)
w1D = np.empty(1E6)
w1D.fill(2.)
data2D = np.random.randn(1E6, 2)
data3D = np.random.randn(1E4, 3)
a = TH1D('th1d', 'test', 1000, -5, 5)
rnp.fill_hist(a, data1D)
# one element lies beyond hist range; that's why it's not 1e6
assert_almost_equal(a.Integral(), 999999.0)
a_w = TH1D('th1dw', 'test', 1000, -5, 5)
rnp.fill_hist(a_w, data1D, w1D)
assert_almost_equal(a_w.Integral(), 999999.0 * 2)
b = TH2D('th2d', 'test', 100, -5, 5, 100, -5, 5)
rnp.fill_hist(b, data2D)
assert_almost_equal(b.Integral(), 999999.0)
c = TH3D('th3d', 'test', 10, -5, 5, 10, -5, 5, 10, -5, 5)
rnp.fill_hist(c, data3D)
assert_almost_equal(c.Integral(), 10000.0)
# array and weights lengths do not match
assert_raises(ValueError, rnp.fill_hist, c, data3D, np.ones(10))
# weights is not 1D
assert_raises(ValueError, rnp.fill_hist, c, data3D,
np.ones((data3D.shape[0], 1)))
# array not 2-d when filling 2D/3D histogram
for h in (b, c):
assert_raises(ValueError, rnp.fill_hist, h, np.random.randn(1E4))
# length of second axis does not match dimensionality of histogram
for h in (a, b, c):
assert_raises(ValueError, rnp.fill_hist, h, np.random.randn(1E4, 4))
# wrong type
h = list()
a = np.random.randn(100)
assert_raises(TypeError, rnp.fill_hist, h, a)
def test_fill_hist():
n_samples = 1000
data1D = RNG.randn(n_samples)
w1D = np.empty(n_samples)
w1D.fill(2.)
data2D = RNG.randn(n_samples, 2)
data3D = RNG.randn(n_samples, 3)
a = TH1D('th1d', 'test', 100, -5, 5)
rnp.fill_hist(a, data1D)
assert_almost_equal(a.Integral(), n_samples)
a_w = TH1D('th1dw', 'test', 100, -5, 5)
rnp.fill_hist(a_w, data1D, w1D)
assert_almost_equal(a_w.Integral(), n_samples * 2)
b = TH2D('th2d', 'test', 100, -5, 5, 100, -5, 5)
rnp.fill_hist(b, data2D)
assert_almost_equal(b.Integral(), n_samples)
c = TH3D('th3d', 'test', 10, -5, 5, 10, -5, 5, 10, -5, 5)
rnp.fill_hist(c, data3D)
assert_almost_equal(c.Integral(), n_samples)
# array and weights lengths do not match
assert_raises(ValueError, rnp.fill_hist, c, data3D, np.ones(10))
# weights is not 1D
assert_raises(ValueError, rnp.fill_hist, c, data3D,
np.ones((data3D.shape[0], 1)))
# array not 2-d when filling 2D/3D histogram
for h in (b, c):
assert_raises(ValueError, rnp.fill_hist, h, RNG.randn(10))
# length of second axis does not match dimensionality of histogram
for h in (a, b, c):
assert_raises(ValueError, rnp.fill_hist, h, RNG.randn(10, 4))
# wrong type
h = list()
a = RNG.randn(10)
assert_raises(TypeError, rnp.fill_hist, h, a)
def decisionFunctionCanvas(self):
'''
Create two histograms which are then drawn onto the same canvas.
This is only really defined for the BDT, not AGILE NN since that doesn't
give a "score".
'''
import ROOT as root
from ROOT import TH2D, TCanvas, TFile, TNamed, TH1F, TLegend
import numpy as np
from root_numpy import fill_hist
import functions as fn
import os
# check that the decision function output was set
if len(self.decision_function) == 0:
return False
df_sig = TH1F("Signal Decision Function", "Score", 100, -1.0, 1.0)
df_bkg = TH1F("Background Decision Function", "Score", 100, -1.0, 1.0)
# fill the histograms with the df
if self.df_weights is not None:
fill_hist(df_sig,self.decision_function[self.df_sig_idx],self.df_weights[self.df_sig_idx])
fill_hist(df_bkg,self.decision_function[self.df_bkg_idx],self.df_weights[self.df_bkg_idx])
else:
fill_hist(df_sig,self.decision_function[self.df_sig_idx])
fill_hist(df_bkg,self.decision_function[self.df_bkg_idx])
# normalise
if df_sig.Integral() != 0:
df_sig.Scale(1./df_sig.Integral())
if df_bkg.Integral() != 0:
df_bkg.Scale(1./df_bkg.Integral())
# set up drawing options and colours
df_sig.SetLineColor(4); df_sig.SetFillStyle(3004)
df_bkg.SetLineColor(2); df_bkg.SetFillStyle(3005)
# set the y axis
max_y = max(df_sig.GetMaximum(), df_bkg.GetMaximum())
df_sig.SetMaximum(max_y*1.2)
df_bkg.SetMaximum(max_y*1.2)
# clone these things
self.df_sig = df_sig.Clone(); self.df_sig.SetDirectory(0)
self.df_bkg = df_bkg.Clone(); self.df_bkg.SetDirectory(0)
return True
def hist_scores(hist, scores, systematic="NOMINAL"):
for sample, scores_dict in scores:
scores, weight = scores_dict[systematic]
fill_hist(hist, scores, weight)
def plot_clf(
background_scores,
category,
signal_scores=None,
signal_scale=1.0,
data_scores=None,
name=None,
draw_histograms=True,
draw_data=False,
save_histograms=False,
hist_template=None,
bins=10,
min_score=0,
max_score=1,
signal_colors=cm.spring,
systematics=None,
unblind=False,
**kwargs
):
if hist_template is None:
if hasattr(bins, "__iter__"):
# variable width bins
hist_template = Hist(bins)
min_score = min(bins)
max_score = max(bins)
else:
hist_template = Hist(bins, min_score, max_score)
bkg_hists = []
for bkg, scores_dict in background_scores:
hist = hist_template.Clone(title=bkg.label)
scores, weight = scores_dict["NOMINAL"]
fill_hist(hist, scores, weight)
hist.decorate(**bkg.hist_decor)
hist.systematics = {}
for sys_term in scores_dict.keys():
if sys_term == "NOMINAL":
continue
sys_hist = hist_template.Clone()
scores, weight = scores_dict[sys_term]
fill_hist(sys_hist, scores, weight)
hist.systematics[sys_term] = sys_hist
bkg_hists.append(hist)
if signal_scores is not None:
sig_hists = []
for sig, scores_dict in signal_scores:
sig_hist = hist_template.Clone(title=sig.label)
scores, weight = scores_dict["NOMINAL"]
fill_hist(sig_hist, scores, weight)
sig_hist.decorate(**sig.hist_decor)
sig_hist.systematics = {}
for sys_term in scores_dict.keys():
if sys_term == "NOMINAL":
continue
sys_hist = hist_template.Clone()
scores, weight = scores_dict[sys_term]
fill_hist(sys_hist, scores, weight)
sig_hist.systematics[sys_term] = sys_hist
sig_hists.append(sig_hist)
else:
sig_hists = None
if data_scores is not None and draw_data and unblind is not False:
data, data_scores = data_scores
if isinstance(unblind, float):
if sig_hists is not None:
# unblind up to `unblind` % signal efficiency
sum_sig = sum(sig_hists)
cut = efficiency_cut(sum_sig, 0.3)
data_scores = data_scores[data_scores < cut]
data_hist = hist_template.Clone(title=data.label)
data_hist.decorate(**data.hist_decor)
fill_hist(data_hist, data_scores)
if unblind >= 1 or unblind is True:
log.info("Data events: %d" % sum(data_hist))
log.info("Model events: %f" % sum(sum(bkg_hists)))
for hist in bkg_hists:
log.info("{0} {1}".format(hist.GetTitle(), sum(hist)))
log.info("Data / Model: %f" % (sum(data_hist) / sum(sum(bkg_hists))))
else:
data_hist = None
if draw_histograms:
output_name = "event_bdt_score"
if name is not None:
output_name += "_" + name
for logy in (False, True):
draw(
data=data_hist,
model=bkg_hists,
signal=sig_hists,
signal_scale=signal_scale,
category=category,
name="BDT Score",
output_name=output_name,
show_ratio=data_hist is not None,
model_colors=None,
signal_colors=signal_colors,
systematics=systematics,
logy=logy,
**kwargs
)
return bkg_hists, sig_hists, data_hist
def histogram_scores(hist_template, scores,
min_score=None, max_score=None,
inplace=False):
if not inplace:
hist = hist_template.Clone(name=hist_template.name + "_scores")
hist.Reset()
else:
hist = hist_template
if min_score is not None:
log.info("cutting out scores below %f" % min_score)
if max_score is not None:
log.info("cutting out scores above %f" % max_score)
if isinstance(scores, np.ndarray):
if min_score is not None:
scores = scores[scores > min_score]
if max_score is not None:
scores = scores[scores < max_score]
fill_hist(hist, scores)
elif isinstance(scores, tuple):
# data
scores, weight = scores
if min_score is not None:
scores_idx = scores > min_score
scores = scores[scores_idx]
weight = weight[scores_idx]
if max_score is not None:
scores_idx = scores < max_score
scores = scores[scores_idx]
weight = weight[scores_idx]
assert (weight == 1).all()
fill_hist(hist, scores)
elif isinstance(scores, dict):
# non-data with possible systematics
# nominal case:
nom_scores, nom_weight = scores['NOMINAL']
if min_score is not None:
scores_idx = nom_scores > min_score
nom_scores = nom_scores[scores_idx]
nom_weight = nom_weight[scores_idx]
if max_score is not None:
scores_idx = nom_scores < max_score
nom_scores = nom_scores[scores_idx]
nom_weight = nom_weight[scores_idx]
fill_hist(hist, nom_scores, nom_weight)
# systematics
sys_hists = {}
for sys_term, (sys_scores, sys_weight) in scores.items():
if sys_term == 'NOMINAL':
continue
if min_score is not None:
scores_idx = sys_scores > min_score
sys_scores = sys_scores[scores_idx]
sys_weight = sys_weight[scores_idx]
if max_score is not None:
scores_idx = sys_scores < max_score
sys_scores = sys_scores[scores_idx]
sys_weight = sys_weight[scores_idx]
sys_hist = hist.Clone(
name=hist.name + "_" + systematic_name(sys_term))
sys_hist.Reset()
fill_hist(sys_hist, sys_scores, sys_weight)
sys_hists[sys_term] = sys_hist
hist.systematics = sys_hists
else:
raise TypeError("scores not an np.array, tuple or dict")
return hist
#plot error on evaluation_data without normalization
evaluation_error_no_norm=[x*max_outputs[0] for x in evaluation_error]
evaluation_errors_last_epoch_no_norm=[x*max_outputs[0] for x in evaluation_errors_last_epoch]
titlename="eta="+str(eta)+", mini_batch_size="+str(mini_batch_size)+ ", lambda="+str(lm_da)+",\n Cost="+cost_function+", weight_initialization="+weight_initialization
titlename_no_norm=titlename+"_without_normalization"
plt.figure(2)
plt.title(titlename_no_norm)
plt.xlabel("epochs")
plt.ylabel("total error validation data")
x_range=[x+1 for x in range(0,epochs)]
plt.plot(x_range,evaluation_error_no_norm)
plt.savefig("error_on_val_data_"+titlename_no_norm+".png")
print "total error on validation_data set after last training"
print evaluation_error_no_norm[-1]
hist_no_norm=TH1D('hist_no_norm',titlename_no_norm,50,-100,100)
fill_hist(hist_no_norm,evaluation_errors_last_epoch_no_norm)
canvas=TCanvas();
hist_no_norm.GetXaxis().SetTitle("desired Chi^2- outputted Chi^2");
hist_no_norm.Draw()
canvas.SaveAs('error_on_val_data_'+titlename_no_norm+'_hist.png')
#-----------------------------------------------------
def apply_tri_selection(rec, lumi):
electron_p = rec['electron_p']
electron_px = rec['electron_px']
electron_py = rec['electron_py']
electron_chi2 = rec['electron_chi2']
electron_has_l1 = rec['electron_has_l1']
electron_pt = np.sqrt(np.power(electron_px, 2) + np.power(electron_py, 2))
positron_p = rec['positron_p']
positron_px = rec['positron_px']
positron_py = rec['positron_py']
positron_chi2 = rec['positron_chi2']
positron_d0 = rec['positron_d0']
positron_has_l1 = rec['positron_has_l1']
positron_has_l2 = rec['positron_has_l2']
positron_pt = np.sqrt(np.power(positron_px, 2) + np.power(positron_py, 2))
top_cluster_time = rec['top_cluster_time']
bot_cluster_time = rec['bot_cluster_time']
cluster_time_diff = top_cluster_time - bot_cluster_time
top_time = rec['top_time']
bot_time = rec['bot_time']
mass = rec['invariant_mass']
v0_p = rec["v0_p"]
top_track_cluster_dt = top_cluster_time - top_time
abs_top_track_cluster_dt = np.absolute(top_track_cluster_dt - 43)
bot_track_cluster_dt = bot_cluster_time - bot_time
abs_bot_track_cluster_dt = np.absolute(bot_track_cluster_dt - 43)
track_cluster_dt_cut = ((abs_top_track_cluster_dt < 4.5)
& (abs_bot_track_cluster_dt < 4.5))
asym = (electron_pt - positron_pt)/(electron_pt + positron_pt)
#
# Define cuts
#
cuts = collections.OrderedDict()
# Base cuts used to reduce accidentals
cuts['Radiative cut'] = v0_p > 0.8*1.056 # GeV
cuts['abs(Ecal clust time - trk time) - 43 ns < 4.5'] = track_cluster_dt_cut
cuts['$p(V_0) < 1.2 E_{beam}$'] = v0_p < 1.2*1.056 # GeV
cuts['trk $\chi^2$ < 40'] = (electron_chi2 < 40) & (positron_chi2 < 40)
cuts['Ecal clust pair dt < 2 ns'] = np.absolute(cluster_time_diff) < 2
cuts['l1 & l2 hit'] = (positron_has_l1 == 1) & (positron_has_l2 == 1)
cuts['$d_{0}(e^+) < 1.1$'] = positron_d0 < 1.1
cuts['$p_t(e^-) - p_t(e^+)/p_t(e^-) + p_t(e^+)$'] = asym < .47
labels = ['Opp. Ecal clusters, trk-cluster match $\chi^2 < 10$, $p(e^-)<0.75E_{beam}$']
clust_dt_arr = [cluster_time_diff]
v0_p_arr = [v0_p]
electron_p_arr = [electron_p]
positron_d0_arr = [positron_d0]
electron_chi2_arr = [electron_chi2]
asym_arr = [asym]
abs_top_cluster_dt_arr = [abs_top_track_cluster_dt]
cut = np.ones(len(v0_p), dtype=bool)
for key, value in cuts.iteritems():
cut = cut & value
clust_dt_arr.append(cluster_time_diff[cut])
v0_p_arr.append(v0_p[cut])
electron_p_arr.append(electron_p[cut])
positron_d0_arr.append(positron_d0[cut])
electron_chi2_arr.append(electron_chi2[cut])
abs_top_cluster_dt_arr.append(abs_top_track_cluster_dt[cut])
asym_arr.append(asym[cut])
labels.append(key)
plt = Plotter.Plotter('trident_selection.pdf')
plt.plot_hists(clust_dt_arr,
np.linspace(-10, 10, 201),
labels=labels,
x_label='Top cluster time - Bottom cluster time (ns)',
label_loc=2,
norm=True,
ylog=True,
root=True)
plt.plot_hists(v0_p_arr,
np.linspace(0, 1.5, 151),
labels=labels,
label_loc=2,
x_label='$V_{0}(p)$ (GeV)',
ylog=True)
plt.plot_hists(electron_p_arr,
np.linspace(0, 1.5, 151),
labels=labels,
x_label='$p(e^-)$ (GeV)',
norm=True,
label_loc=4,
ylog=True)
plt.plot_hists(positron_d0_arr,
np.linspace(-20, 20, 201),
labels=labels,
x_label='$d_{0}(e^{+})$',
norm=True,
label_loc=2,
ylog=True)
plt.plot_hists(electron_chi2_arr,
np.linspace(0, 100, 201),
labels=labels,
x_label='Track $\chi^2$',
norm=True,
label_loc=1,
ylog=True)
plt.plot_hists(asym_arr,
np.linspace(-1, 1, 201),
labels=labels,
x_label='$p_t(e^-) - p_t(e^+)/p_t(e^-) + p_t(e^+)$',
norm=True,
label_loc=2,
ylog=True)
plt.plot_hists(abs_top_cluster_dt_arr,
np.linspace(0, 60, 121),
labels=labels,
x_label='abs(ECal cluster time - track time - 43) ns',
norm=True,
label_loc=1,
ylog=True)
plt.close()
file = r.TFile("invariant_mass.root", "recreate")
mass_histo = r.TH1F("invariant_mass", "invariant_mass", 2000, 0., 0.1)
#mass_histo = r.TH1F("invariant_mass", "invariant_mass", 50, 0., 0.1)
mass_histo.GetXaxis().SetTitle("m(e^+e^-) (GeV)")
mass_histo.GetYaxis().SetTitle("#sigma(#mub)")
bin_width = mass_histo.GetXaxis().GetBinWidth(1)
weights = np.empty(len(mass[cut]))
if lumi: weights.fill(1/(bin_width*float(lumi)))
else: weights.fill(1)
#rnp.fill_hist(mass_histo, mass[selection], weights=weights)
rnp.fill_hist(mass_histo, mass[cut], weights=weights)
mass_histo.Write()
file.Close()
def Hist_comp_ratios (dec, bkg) :
# weights = pickle.load( open( "weights_train.pck", "rb" ) )
# probas = pickle.load( open( "class_proba.pck", "rb" ) )
# mt_dec = pickle.load( open( "inputs_train.pck", "rb" ) )
# classes = pickle.load( open( "targets_train.pck", "rb" ) )
weights = pickle.load( open( "weights.pck", "rb" ) )
probas = pickle.load( open( "class_probaWholeSample.pck", "rb" ) )
mt_dec = pickle.load( open( "inputs.pck", "rb" ) )
classes = pickle.load( open( "classes.pck", "rb" ) )
#put mt_dec weights and the RELEVANT class probability together
class_num = (GetClassIndex(bkg)-1)
all_data = np.transpose(np.vstack((mt_dec[:,0],mt_dec[:,1], classes, weights, probas[:,class_num])))
all_data = np.array(filter(lambda x: x[2] == class_num, all_data))
#extract the relevant values for the specific decay channel
Filter = np.array(filter(lambda x: x[1] == dec, all_data))
MT = Filter[:,0]
Weight = Filter[:,3]
Prob_bkg = Filter[:,4]
SumProbXweight = np.multiply(Prob_bkg,Weight)
h1 = TH1D("h1","SumProbXweight_"+bkg+str(dec), 25 , 0.0 ,250.)
root_open("plots/ROOTfiles/H1_SumProbXweight_"+bkg+str(dec)+"Samptot.root", 'recreate')
fill_hist(h1,MT, weights=SumProbXweight)
h1.Write()
h2 = TH1D("h2","SumWeight_"+bkg+str(dec), 25 , 0.0 ,250.)
root_open("plots/ROOTfiles/H1_SumWeight_"+bkg+str(dec)+"Samptot.root", 'recreate')
fill_hist(h2,MT,weights=Weight)
h2.Write()
h3 = h1.Clone("h3")
h3.Divide(h2)
if (dec == 0.0) :
Filter2 = np.array(filter(lambda x: x[1] == 1.0, all_data))
MT2 = Filter2[:,0]
Weight2 = Filter2[:,3]
Prob_bkg2 = Filter2[:,4]
SumProbXweight2 = np.multiply(Prob_bkg2,Weight2)
h12 = TH1D("h12","SumProbXweight_"+bkg+str(1.0), 25 , 0.0 ,250.)
root_open("plots/ROOTfiles/H1_SumProbXweight_"+bkg+str(1.0)+"Samptot.root", 'recreate')
fill_hist(h12,MT2, weights=SumProbXweight2)
h12.Write()
h22 = TH1D("h22","SumWeight_"+bkg+str(1.0), 25 , 0.0 ,250.)
root_open("plots/ROOTfiles/H1_SumWeight_"+bkg+str(1.0)+"Samptot.root", 'recreate')
fill_hist(h22,MT2,weights=Weight2)
h22.Write()
h32 = h12.Clone("h32")
h32.Divide(h22)
h12.SetStats(0)
h22.SetStats(0)
h32.SetStats(0)
h12.SetLineColor(2)
h22.SetLineColor(2)
h32.SetLineColor(0)
h32.SetMarkerStyle(23)
h32.SetMarkerColor(2)
h32.SetMarkerSize(1.2)
#create Canvas and save the plot as png
c = Canvas()
c.Divide(2,2)
c.cd(1)
h1.SetStats(0)
if (dec == 0.0) :
if (h1.GetMaximum() > h12.GetMaximum()) :
print "h1"
h12.GetYaxis().SetRangeUser(0.,h1.GetMaximum())
else :
print "h12"
h12.GetYaxis().SetRangeUser(0.,h12.GetMaximum())
h12.Draw("HIST")
h1.Draw("HIST SAME")
c.cd(2)
h2.SetStats(0)
if (dec == 0.0) :
if (h2.GetMaximum() > h22.GetMaximum()) :
h22.GetYaxis().SetRangeUser(0.,h2.GetMaximum())
else :
h22.GetYaxis().SetRangeUser(0.,h22.GetMaximum())
h22.Draw("HIST")
h2.Draw("HIST SAME")
c.cd(3)
f1 = root_open(GetClassProbaPath(bkg))
#get the 2 histograms for the 2 decay channels: 1 track & 3 tracks
H1 = f1.Get("h_w_2d")
if (dec == 10.0) :
#3 tracks
h_data = Hist(list(H1.xedges()))
h_data[:] = H1[:,2]
else :
#1 track
h_data = Hist(list(H1.xedges()))
h_data[:] = H1[:,1]
h_data.GetXaxis().SetRangeUser(0.,250.)
h_data.GetYaxis().SetRangeUser(0.,1.)
h_data.fillstyle = '/'
h_data.fillcolor = (255,255,0) #yellow
h_data.SetStats(0)
h_data.Draw("HIST")
# h3.SetFillColor(4) #blue
# h3.SetFillStyle(3005)
h3.SetLineColor(0)
h3.SetMarkerStyle(21)
h3.SetMarkerColor(4)
h3.SetMarkerSize(1.2)
h3.SetStats(0)
h3.SetTitle(bkg+str(dec))
h3.GetXaxis().SetTitle("m_{T}")
h3.GetYaxis().SetTitle("Class probability")
h3.Draw("HIST P SAME")
if (dec == 0.0) :
h32.Draw("HIST P SAME")
c.Update()
if (dec == 0.0) :
legend = Legend(3, leftmargin=0.45, margin=0.3)
legend.AddEntry(h3, "training, no #pi^{0}", style='P')
legend.AddEntry(h32, "training, with #pi^{0}", style='P')
legend.AddEntry(h_data, "data", style='F')
legend.Draw()
else :
legend = Legend(2, leftmargin=0.45, margin=0.3)
legend.AddEntry(h3, "training", style='P')
legend.AddEntry(h_data, "data", style='F')
legend.Draw()
c.SaveAs("plots/H1_"+bkg+str(dec)+"_RatioCompSamptot.png")
#clf = linear_model.SGDRegressor()
clf = xgb.XGBRegressor(max_depth=4, learning_rate=0.1, n_estimators=400)
scores = cross_val_score(clf, data_reduced_n, data_target, cv=5, scoring='neg_mean_squared_error')
print scores
print("MSE: %0.3f (+/- %0.3f)" % (scores.mean(), scores.std() * 2))
predicted = cross_val_predict(clf, data_reduced_n, data_target, cv=5)
from sklearn import metrics
scores2 = metrics.mean_squared_error(data_target, predicted)
print scores
print("MSE 2: %0.3f (+/- %0.3f)" % (scores2.mean(), scores2.std() * 2))
fig, ax = plt.subplots()
ax.scatter(data_target, predicted)
ax.plot([data_target.min(), data_target.max()], [data_target.min(), data_target.max()], 'k--', lw=4)
ax.set_xlabel('True KE [GeV]')
ax.set_ylabel('Predicted KE [GeV]')
ax.set_xlim([data_target.min()*0.9, data_target.max()*1.1])
ax.set_ylim([data_target.min()*0.9, data_target.max()*1.1])
plt.savefig("xgb_cross_val_comparision.pdf")
ROOT.gStyle.SetOptStat(1)
hist = ROOT.TH1D("hist","hist", 100, -1, 1)
diff = (data_target - predicted)/data_target
fill_hist(hist, diff)
hist.GetXaxis().SetTitle("#DeltaE/E")
canvas = ROOT.TCanvas()
hist.Draw()
canvas.SaveAs("xgb_cross_val_DeltaE.pdf")
fig, axes = plt.subplots(nrows=1, ncols=2, figsize=(12, 6))
axes[0].set_yscale('log')
axes[1].set_yscale('log')
q .hist(histtype='step', bins = np.arange(loC, hiC + width, width), color='r', ax = axes[0])
#q_good.hist(histtype='step', bins = np.arange(loC, hiC + width, width), color='g', ax = axes[0])
q_ped .hist(histtype='step', bins = np.arange(loC, hiC + width, width), color='b', ax = axes[1])
axes[0].plot((threshold, threshold), (0, 1e5), 'k-')
axes[1].plot((threshold, threshold), (0, 1e5), 'k-')
from ROOT import TH1F, TFile
from root_numpy import fill_hist
q_file = TFile('plots/' + sys.argv[1] + '_' + sys.argv[2] + '_charge_spectrum.root', "recreate")
q_hist1 = TH1F('hist1', 'title', nBins, loC, hiC)
q_hist2 = TH1F('hist2', 'title', 55, -10, 100)
fill_hist(q_hist1, q.as_matrix())
fill_hist(q_hist2, max_voltages['minus_voltage'].as_matrix())
q_hist1.Write();
q_hist2.Write();
q_file.Close();
max_q = q.max()
# uncomment these lines if you want to plot the fitted Gaussian
#x = np.linspace(loC, hiC, nBins)
#pdf = n_events*norm.pdf(x, mu, std)/nBins
#plt.plot(x, pdf, 'k', linewidth = 2)
axes[0].set_xlabel("charge [pC]", fontsize = 20)
axes[0].set_ylabel("Entries / (%0.2f pC)" % width, fontsize = 20)
plt.ylabel("recoKE - trueKE [MeV]")
# In[12]:
plt.scatter(test_data_trueKE,(clf.predict(test_data_reduced_n)-test_data_trueKE)/test_data_trueKE)
plt.ylim((0,1))
plt.xlabel("trueKE [MeV]")
plt.ylabel("DeltaE/E")
res_twod_SGD = np.dstack((test_data_trueKE, (clf.predict(test_data_reduced_n)-test_data_trueKE)/test_data_trueKE))
# In[13]:
hist_SGD = ROOT.TH2D('name', 'title', 100, 0, 5000, 100, -1, 10)
fill_hist(hist_SGD, res_twod_SGD[0])
hist_SGD.Draw()
ROOT.gPad.Draw()
# In[14]:
profile_SGD = hist_SGD.ProfileX()
profile_SGD.SetLineColor(ROOT.kBlue)
profile_SGD.Draw()
ROOT.gPad.Draw()
# In[15]:
params = {'n_estimators': 1000, 'max_depth': 10, 'min_samples_split': 1,
for bootstrap_idx in range(100):
sys.stdout.write("bootstrap {0} ...\r".format(bootstrap_idx))
sys.stdout.flush()
# resample with replacement
# http://docs.scipy.org/doc/numpy-dev/reference/generated/numpy.random.choice.html
sample_idx = np.random.choice(len(array), size=len(array), replace=True)
array_bootstrapped = array[sample_idx]
# convert back to a TTree and write it out
tree_bootstrapped = array2tree(
array_bootstrapped,
name='bootstrap_{0}'.format(bootstrap_idx))
tree_bootstrapped.Write()
tree_bootstrapped.Delete()
# fill the ROOT histogram with the numpy array
hist.Reset()
fill_hist(hist, rec2array(array_bootstrapped))
hist.Draw()
hist.xaxis.title = 'x'
hist.yaxis.title = 'y'
hist.zaxis.title = 'Events'
hist.xaxis.limits = (-2.5, 2.5)
hist.yaxis.limits = (-2.5, 2.5)
hist.zaxis.range_user = (0, 60)
hist.xaxis.divisions = 5
hist.yaxis.divisions = 5
hist.zaxis.divisions = 5
canvas.Print('bootstrap.gif+50')
# loop the gif
canvas.Print('bootstrap.gif++')
output.Close()
def plotDiscriminant(self, discriminant, signal_idx, bkg_idx, weights = None, save_disc = True, rejection_power=True):
'''
Plot the discriminants and the resulting ROC curve derived from them.
Keyword args:
discriminant --- The score of the BDT (set in the setProbas method)
signal_idx --- The true indices of all signal events
bkg_idx ---The true indices of all background events
save_disc --- Flag indicating if the discriminant plots should be saved.
rejection_power --- Whether or not to calculate bkg power: 1/eff in addtion to 1-eff
'''
import ROOT as root
from ROOT import TH2D, TCanvas, TFile, TNamed, TH1F, TLegend
import numpy as np
from root_numpy import fill_hist
import functions as fn
import os
# stop showing plots to screen
root.gROOT.SetBatch(True)
if not os.path.exists(self.output_path):
os.makedirs(self.output_path)
fo = TFile.Open(self.output_path+"/"+self.output_prefix+str(self.job_id)+'.root','RECREATE')
bins = 100
# when creating the plots do it over the range of all probas (scores)
discriminant_bins = np.linspace(0,1,bins)#np.min(discriminant), np.max(discriminant), bins)
hist_bkg = TH1F("Background Discriminant","Discriminant",bins, np.min(discriminant), np.max(discriminant))
hist_sig = TH1F("Signal Discriminant","Discriminant",bins, np.min(discriminant), np.max(discriminant))
# fill the signal and background histograms
if weights is not None:
print 'weights is not none******************'
fill_hist(hist_bkg,discriminant[bkg_idx], weights[bkg_idx])
fill_hist(hist_sig,discriminant[signal_idx], weights[signal_idx])
else:
fill_hist(hist_bkg,discriminant[bkg_idx])
fill_hist(hist_sig,discriminant[signal_idx])
if hist_bkg.Integral() != 0:
hist_bkg.Scale(1/hist_bkg.Integral())
if hist_sig.Integral() != 0:
hist_sig.Scale(1/hist_sig.Integral())
hist_sig.SetLineColor(4)
hist_bkg.SetLineColor(2)
#hist_sig.SetFillColorAlpha(4, 0.5);
hist_sig.SetFillStyle(3004)
#hist_bkg.SetFillColorAlpha(2, 0.5);
hist_bkg.SetFillStyle(3005)
hist_sig.Write()
hist_bkg.Write()
c = TCanvas()
leg = TLegend(0.8,0.55,0.9,0.65);leg.SetFillColor(root.kWhite)
leg.AddEntry(hist_sig, "Signal","l")
leg.AddEntry(hist_bkg, "Background", "l")
max_y = max(hist_sig.GetMaximum(), hist_bkg.GetMaximum())
hist_sig.SetMaximum(max_y*1.2)
hist_bkg.SetMaximum(max_y*1.2)
hist_sig.Draw('hist')
hist_bkg.Draw('histsame')
c.Write()
if save_disc == True:
if not os.path.exists('disc_plots'):
os.makedirs('disc_plots')
c.SaveAs('disc_plots/discriminants_'+str(self.job_id)+'.png')
# before deciding whether to do a left or right cut for the roc curve we have to find the median.
sig_median = np.median(discriminant[signal_idx])
bkg_median = np.median(discriminant[bkg_idx])
if sig_median > bkg_median:
roc_cut = 'R'
else:
roc_cut = 'L'
# create the single sided roccurve with the code from Sam
self.roc_graph = fn.RocCurve_SingleSided(hist_sig, hist_bkg, self.sig_eff,self.bkg_eff, roc_cut)
self.roc_graph.SetName('BackgroundRejection')
self.roc_graph.SetTitle('BackgroundRejection')
self.roc_graph.Write()
# get teh background rejection power at 50% signal efficiency
# store the efficiencies first
self.ROC_sig_efficiency, self.ROC_bkg_rejection = fn.getEfficiencies(self.roc_graph)
self.bkgRejectionPower()
# write the roc score as a string to the output file
rej_string = 'rejection_power_'+str(self.ROC_rej_power_05)
rej_n = TNamed(rej_string,rej_string)
rej_n.Write()
if rejection_power:
c.SetLogy()
self.roc_graph_power = fn.RocCurve_SingleSided(hist_sig, hist_bkg, self.sig_eff,self.bkg_eff, roc_cut, rejection=False)
c.cd()
self.roc_graph_power.SetName('BackgroundPower')
self.roc_graph_power.SetTitle('BackgroundPower')
self.roc_graph_power.Write()
# write the decision function to the root file as well, if it is defined.
if len(self.decision_function) > 0:
self.decisionFunctionCanvas()
# add the legends
leg2 = TLegend(0.8,0.55,0.9,0.65);leg2.SetFillColor(root.kWhite)
leg2.AddEntry(self.df_sig, "Signal","l")
leg2.AddEntry(self.df_bkg, "Background", "l")
# canvas to draw them on
c2 = TCanvas('Decision Functions')
self.df_sig.Draw('hist')
self.df_bkg.Draw('histsame')
leg2.Draw('same')
c2.Write()
# now write the df histograms as well
self.df_sig.Write()
self.df_bkg.Write()
self.hist_sig = hist_sig.Clone(); self.hist_sig.SetDirectory(0)
self.hist_bkg = hist_bkg.Clone(); self.hist_bkg.SetDirectory(0)
fo.Close()
def optimize_background_rejection_vs_ieta(effs, isolations, signalfile, signaltree, backgroundfile, backgroundtree, inputnames=['abs(ieta)','ntt'], targetname='iso', cut='et>10'):
ieta_binning = [0.5, 3.5, 6.5, 9.5, 13.5, 18.5, 24.5, 27.5]
# Compute signal efficiencies
ninputs = len(inputnames)
branches = copy.deepcopy(inputnames)
branches.append(targetname)
data = root2array(signalfile, treename=signaltree, branches=branches, selection=cut)
data = data.view((np.float64, len(data.dtype.names)))
inputs = data[:, range(ninputs)].astype(np.float32)
targets = data[:, [ninputs]].astype(np.float32).ravel()
xs = data[:, [0]].astype(np.float32).ravel()
# fill signal ieta histogram and normalize to 1
histo_signal = Hist(ieta_binning)
fill_hist(histo_signal, xs)
#histo_signal.Scale(1./histo_signal.integral(overflow=True))
# signal_efficiencies is a 2D array
# The first dimension corresponds to different ieta values
# The second dimension corresponds to different working points
signal_efficiencies = [graph2array(efficiency.efficiency_graph(pass_function=(lambda x:np.less(x[1],iso.predict(x[0]))), function_inputs=(inputs,targets), xs=xs, bins=ieta_binning))[:,[1]].ravel() for iso in isolations]
signal_efficiencies = np.column_stack(signal_efficiencies)
# Compute background efficiencies
ninputs = len(inputnames)
branches = copy.deepcopy(inputnames)
branches.append(targetname)
data = root2array(backgroundfile, treename=backgroundtree, branches=branches, selection=cut)
data = data.view((np.float64, len(data.dtype.names)))
inputs = data[:, range(ninputs)].astype(np.float32)
targets = data[:, [ninputs]].astype(np.float32).ravel()
xs = data[:, [0]].astype(np.float32).ravel()
# fill background ieta histogram and normalize to 1
histo_background = Hist(ieta_binning)
fill_hist(histo_background, xs)
#histo_background.Scale(1./histo_background.integral(overflow=True))
# background_efficiencies is a 2D array
# The first dimension corresponds to different ieta values
# The second dimension corresponds to different working points
background_efficiencies = [graph2array(efficiency.efficiency_graph(pass_function=(lambda x:np.less(x[1],iso.predict(x[0]))), function_inputs=(inputs,targets), xs=xs, bins=ieta_binning))[:,[1]].ravel() for iso in isolations]
background_efficiencies = np.column_stack(background_efficiencies)
signal_efficiencies_diff_graphs = []
background_efficiencies_diff_graphs = []
optimal_points_graphs = []
optimal_points = []
# compute best working point for each ieta (loop on ieta)
for i,(signal_effs,background_effs) in enumerate(zip(signal_efficiencies, background_efficiencies)):
# Compute the probability of signal in this ieta bin for the different efficiency points
# It is assumed that the cut is applied only in this bin, all the other bins keep the same number of entries
n_i = histo_signal[i+1].value
n_tot = histo_signal.integral(overflow=True)
proba_signal = np.array([n_i*eff/(n_tot-n_i*(1.-eff)) for eff in signal_effs])
# Same as above, but for background
n_i = histo_background[i+1].value
n_tot = histo_background.integral(overflow=True)
proba_background = np.array([n_i*eff/(n_tot-n_i*(1.-eff)) for eff in background_effs])
signal_efficiencies_diff_graph, background_efficiencies_diff_graph, optimal_points_graph, optimal_point = find_best_working_point(effs, signal_effs, background_effs, proba_signal, proba_background)
signal_efficiencies_diff_graph.SetName('efficiencies_signal_ieta_{}'.format(i))
background_efficiencies_diff_graph.SetName('efficiencies_background_ieta_{}'.format(i))
optimal_points_graph.SetName('signal_background_optimal_points_ieta_{}'.format(i))
signal_efficiencies_diff_graphs.append(signal_efficiencies_diff_graph)
background_efficiencies_diff_graphs.append(background_efficiencies_diff_graph)
optimal_points_graphs.append(optimal_points_graph)
optimal_points.append(optimal_point)
optimal_points_histo = Hist(ieta_binning)
array2hist(optimal_points, optimal_points_histo)
return signal_efficiencies_diff_graphs, background_efficiencies_diff_graphs, optimal_points_graphs, optimal_points_histo
def histfactory(self, sample, category, systematics=False,
rec=None, weights=None, mva=False,
uniform=False, nominal=None):
if not systematics:
return
if len(self.modes) != 1:
raise TypeError(
'histfactory sample only valid for single production mode')
if len(self.masses) != 1:
raise TypeError(
'histfactory sample only valid for single mass point')
# isolation systematic
sample.AddOverallSys(
'ATLAS_ANA_HH_{0:d}_Isolation'.format(self.year),
1. - 0.06,
1. + 0.06)
mode = self.modes[0]
if mode in ('Z', 'W'):
_uncert_mode = 'VH'
else:
_uncert_mode = self.MODES_WORKSPACE[mode]
if self.year == 2011:
energy = 7
elif self.year == 2012:
energy = 8
else:
raise ValueError(
"collision energy is unknown for year {0:d}".format(self.year))
# QCD_SCALE
for qcd_scale_term, qcd_scale_mode, qcd_scale_category, values in self.QCD_SCALE:
if qcd_scale_mode == _uncert_mode and qcd_scale_category == category.name:
high, low = map(float, values.split('/'))
sample.AddOverallSys(qcd_scale_term, low, high)
# UE UNCERTAINTY
for ue_term, ue_mode, ue_category, values in self.UE_UNCERT:
if ue_mode == _uncert_mode and ue_category == category.name:
high, low = map(float, values.split('/'))
sample.AddOverallSys(ue_term, low, high)
# PDF ACCEPTANCE UNCERTAINTY (OverallSys)
for pdf_term, pdf_mode, pdf_category, values in self.PDF_ACCEPT_NORM_UNCERT:
if pdf_mode == _uncert_mode and pdf_category == category.name:
high, low = map(float, values.split('/'))
sample.AddOverallSys(pdf_term, low, high)
sample_nom = sample.hist
# PDF ACCEPTANCE UNCERTAINTY (HistoSys) ONLY FOR MVA
if mva:
for pdf_term, pdf_mode, pdf_category, hist_names in self.PDF_ACCEPT_SHAPE_UNCERT:
if pdf_mode == _uncert_mode and pdf_category == category.name:
high_name, low_name = hist_names.format(energy).split('/')
high_shape, low_shape = self.PDF_ACCEPT_file[high_name], self.PDF_ACCEPT_file[low_name]
if len(high_shape) != len(sample.hist):
log.warning("skipping pdf acceptance shape systematic "
"since histograms are not compatible")
continue
high = sample_nom.Clone(shallow=True, name=sample_nom.name + '_{0}_UP'.format(pdf_term))
low = sample_nom.Clone(shallow=True, name=sample_nom.name + '_{0}_DOWN'.format(pdf_term))
high *= high_shape
low *= low_shape
histsys = histfactory.HistoSys(
pdf_term, low=low, high=high)
sample.AddHistoSys(histsys)
#mixing Norms
if self.SM:
log.info('adding norm factor')
sample.AddNormFactor('ATLAS_epsilon', 1., -200., 200., False)
elif self.BSM:
log.info('adding norm factor')
sample.AddNormFactor('ATLAS_epsilon_rejected', 1., -200., 200., False)
else:
log.info('no norms for {0}'.format(self.name))
# BR_tautau
_, (br_up, br_down) = yellowhiggs.br(
self.mass, 'tautau', error_type='factor')
sample.AddOverallSys('ATLAS_BR_tautau', br_down, br_up)
# <NormFactor Name="mu_BR_tautau" Val="1" Low="0" High="200" />
sample.AddNormFactor('mu_BR_tautau', 1., 0., 200., True)
#mu_XS[energy]_[mode]
#_, (xs_up, xs_down) = yellowhiggs.xs(
# energy, self.mass, self.MODES_DICT[self.mode][0],
# error_type='factor')
#sample.AddOverallSys(
# 'mu_XS{0:d}_{1}'.format(energy, self.MODES_WORKSPACE[self.mode]),
# xs_down, xs_up)
sample.AddNormFactor(
'mu_XS{0:d}_{1}'.format(energy, self.MODES_WORKSPACE[self.mode]),
1., 0., 200., True)
# https://twiki.cern.ch/twiki/bin/viewauth/AtlasProtected/HSG4Uncertainties
# pdf uncertainty
if mode == 'gg':
if energy == 8:
sample.AddOverallSys('pdf_Higgs_gg', 0.93, 1.08)
else: # 7 TeV
sample.AddOverallSys('pdf_Higgs_gg', 0.92, 1.08)
else:
if energy == 8:
sample.AddOverallSys('pdf_Higgs_qq', 0.97, 1.03)
else: # 7 TeV
sample.AddOverallSys('pdf_Higgs_qq', 0.98, 1.03)
# EWK NLO CORRECTION FOR VBF ONLY
if mode == 'VBF':
sample.AddOverallSys('NLO_EW_Higgs', 0.98, 1.02)
# QCDscale_ggH3in HistoSys ONLY FOR MVA
# also see ggH3in script
if mva and mode == 'gg' and category.name == 'vbf':
Rel_Error_2j = 0.215
Error_exc = 0.08613046469238815 # Abs error on the exclusive xsec
xsec_exc = 0.114866523583739 # Exclusive Xsec
Error_3j = sqrt(Error_exc**2 - (Rel_Error_2j*xsec_exc)**2)
rel_error = Error_3j / xsec_exc
dphi = rec['true_dphi_jj_higgs_no_overlap']
scores = rec['classifier']
idx_2j = ((pi - dphi) < 0.2) & (dphi >= 0)
idx_3j = ((pi - dphi) >= 0.2) & (dphi >= 0)
# get normalization factor
dphi_2j = weights[idx_2j].sum()
dphi_3j = weights[idx_3j].sum()
weight_up = np.ones(len(weights))
weight_dn = np.ones(len(weights))
weight_up[idx_2j] -= (dphi_3j / dphi_2j) * rel_error
weight_dn[idx_2j] += (dphi_3j / dphi_2j) * rel_error
weight_up[idx_3j] += rel_error
weight_dn[idx_3j] -= rel_error
weight_up *= weights
weight_dn *= weights
up_hist = nominal.clone(shallow=True, name=sample_nom.name + '_QCDscale_ggH3in_UP')
up_hist.Reset()
dn_hist = nominal.clone(shallow=True, name=sample_nom.name + '_QCDscale_ggH3in_DOWN')
dn_hist.Reset()
fill_hist(up_hist, scores, weight_up)
fill_hist(dn_hist, scores, weight_dn)
if uniform:
up_hist = uniform_hist(up_hist)
dn_hist = uniform_hist(dn_hist)
shape = histfactory.HistoSys('QCDscale_ggH3in',
low=dn_hist,
high=up_hist)
norm, shape = histfactory.split_norm_shape(shape, sample_nom)
sample.AddHistoSys(shape)
#printing#
# #
plt.figure(1)
plt.title("Costfunction of (modified) Training-data")
plt.xlabel("epochs")
plt.ylabel("cost function")
x_range=[x+1 for x in range(0,N_epochs)]
plt.plot(x_range,cost_training_data)
plt.savefig("cost_on_training_data.png")
plt.figure(2)
plt.title("f data")
plt.xlabel("epochs")
plt.ylabel("total error on validation data")
x_range=[x+1 for x in range(0,N_epochs)]
plt.plot(x_range,error_validation_data)
plt.savefig("error_on_val_data.png")
error_on_validation_data_after_training = diff_validation[-1].reshape((1,validation_set.get_N()))
hist=TH1D('hist',"Errors on val data after last training epoch",200,-10000,10000)
fill_hist(hist,error_on_validation_data_after_training[0])
canvas=TCanvas();
hist.GetXaxis().SetTitle("desired Chi^2- outputted Chi^2");
hist.Draw()
canvas.SaveAs('error_on_val_data_hist.png')