def predict(self, x, theta_true):
"""
Run the unbinned ML fit
"""
# Data
roodata = RooDataSet('data', 'data', RooArgSet(self.phistar))
for xval in x:
self.phistar.setVal(xval)
roodata.add(RooArgSet(self.phistar))
theta = RooRealVar('theta', 'theta', 0.5, self.theta_min, self.theta_max)
# The combined pdf
model = RooAddPdf('model', 'model',
RooArgList(self.pdfs['A'], self.pdfs['H']),
RooArgList(theta))
with stdout_redirected_to('%s/minuit_output.log' % self.outdir):
res = model.fitTo(roodata, Save(True))
nll = res.minNll()
fitted_theta = theta.getValV()
# Get Lambda(theta_true | theta_best)
with stdout_redirected_to():
logl = model.createNLL(roodata)
theta.setVal(theta_true)
nll_theta_true = logl.getValV()
nll_ratio = nll_theta_true - nll
return fitted_theta, nll, nll_ratio
def predict(self, x, theta_true):
"""
Run an unbinned ML fit to make predictions
"""
# Create RooDataSet
xs = self.scaler.transform(x)
preds = self.model.predict(xs)[:, 1]
min_nn_output_local, max_nn_output_local = np.min(preds), np.max(preds)
if min_nn_output_local < self.min_nn_output:
self.min_nn_output = min_nn_output_local
if max_nn_output_local > self.max_nn_output:
self.max_nn_output = max_nn_output_local
roodata = RooDataSet('data', 'data', RooArgSet(self.roopred))
for pred in preds:
self.roopred.setVal(pred)
roodata.add(RooArgSet(self.roopred))
# Fit
theta = RooRealVar('theta', 'theta', 0.5, self.theta_min, self.theta_max)
model = RooAddPdf('model', 'model',
RooArgList(self.pdfs['A'], self.pdfs['H']),
RooArgList(theta))
with stdout_redirected_to('%s/minuit_output.log' % self.outdir):
res = model.fitTo(roodata, Save(True))
nll = res.minNll()
fitstatus = res.status()
fitstatus |= (not subprocess.call(['grep', 'p.d.f value is less than zero', 'output_MLE_unbinned/minuit_output.log']))
fitted_theta = theta.getValV()
# Get Lambda(theta_true | theta_best)
logl = model.createNLL(roodata)
theta.setVal(theta_true)
nll_theta_true = logl.getValV()
nll_ratio = nll_theta_true - nll
return fitted_theta, nll, nll_ratio, fitstatus
def fit_mass(data,
column,
x,
sig_pdf=None,
bkg_pdf=None,
n_sig=None,
n_bkg=None,
blind=False,
nll_profile=False,
second_storage=None,
log_plot=False,
pulls=True,
sPlot=False,
bkg_in_region=False,
importance=3,
plot_importance=3):
"""Fit a given pdf to a variable distribution
Parameter
---------
data : |hepds_type|
The data containing the variable to fit to
column : str
The name of the column to fit the pdf to
sig_pdf : RooFit pdf
The signal Probability Density Function. The variable to fit to has
to be named 'x'.
bkg_pdf : RooFit pdf
The background Probability Density Function. The variable to fit to has
to be named 'x'.
n_sig : None or numeric
The number of signals in the data. If it should be fitted, use None.
n_bkg : None or numeric
The number of background events in the data.
If it should be fitted, use None.
blind : boolean or tuple(numberic, numberic)
If False, the data is fitted. If a tuple is provided, the values are
used as the lower (the first value) and the upper (the second value)
limit of a blinding region, which will be omitted in plots.
Additionally, no true number of signal will be returned but only fake.
nll_profile : boolean
If True, a Negative Log-Likelihood Profile will be generated. Does not
work with blind fits.
second_storage : |hepds_type|
A second data-storage that will be concatenated with the first one.
importance : |importance_type|
|importance_docstring|
plot_importance : |plot_importance_type|
|plot_importance_docstring|
Return
------
tuple(numerical, numerical)
Return the number of signals and the number of backgrounds in the
signal-region. If a blind fit is performed, the signal will be a fake
number. If no number of background events is required, -999 will be
returned.
"""
if not (isinstance(column, str) or len(column) == 1):
raise ValueError("Fitting to several columns " + str(column) +
" not supported.")
if type(sig_pdf) == type(bkg_pdf) == None:
raise ValueError("sig_pdf and bkg_pdf are both None-> no fit possible")
if blind is not False:
lower_blind, upper_blind = blind
blind = True
n_bkg_below_sig = -999
# create data
data_name = data.name
data_array, _t1, _t2 = data.make_dataset(second_storage, columns=column)
del _t1, _t2
# double crystalball variables
min_x, max_x = min(data_array[column]), max(data_array[column])
# x = RooRealVar("x", "x variable", min_x, max_x)
# create data
data_array = np.array([i[0] for i in data_array.as_matrix()])
data_array.dtype = [('x', np.float64)]
tree1 = array2tree(data_array, "x")
data = RooDataSet("data", "Data", RooArgSet(x), RooFit.Import(tree1))
# # TODO: export somewhere? does not need to be defined inside...
# mean = RooRealVar("mean", "Mean of Double CB PDF", 5280, 5100, 5600)#, 5300, 5500)
# sigma = RooRealVar("sigma", "Sigma of Double CB PDF", 40, 0.001, 200)
# alpha_0 = RooRealVar("alpha_0", "alpha_0 of one side", 5.715)#, 0, 150)
# alpha_1 = RooRealVar("alpha_1", "alpha_1 of other side", -4.019)#, -200, 0.)
# lambda_0 = RooRealVar("lambda_0", "Exponent of one side", 3.42)#, 0, 150)
# lambda_1 = RooRealVar("lambda_1", "Exponent of other side", 3.7914)#, 0, 500)
#
# # TODO: export somewhere? pdf construction
# frac = RooRealVar("frac", "Fraction of crystal ball pdfs", 0.479, 0.01, 0.99)
#
# crystalball1 = RooCBShape("crystallball1", "First CrystalBall PDF", x,
# mean, sigma, alpha_0, lambda_0)
# crystalball2 = RooCBShape("crystallball2", "Second CrystalBall PDF", x,
# mean, sigma, alpha_1, lambda_1)
# doubleCB = RooAddPdf("doubleCB", "Double CrystalBall PDF",
# crystalball1, crystalball2, frac)
# n_sig = RooRealVar("n_sig", "Number of signals events", 10000, 0, 1000000)
# test input
if n_sig == n_bkg == 0:
raise ValueError("n_sig as well as n_bkg is 0...")
if n_bkg is None:
n_bkg = RooRealVar("n_bkg", "Number of background events", 10000, 0,
500000)
elif n_bkg >= 0:
n_bkg = RooRealVar("n_bkg", "Number of background events", int(n_bkg))
else:
raise ValueError("n_bkg is not >= 0 or None")
if n_sig is None:
n_sig = RooRealVar("n_sig", "Number of signal events", 1050, 0, 200000)
# START BLINDING
blind_cat = RooCategory("blind_cat", "blind state category")
blind_cat.defineType("unblind", 0)
blind_cat.defineType("blind", 1)
if blind:
blind_cat.setLabel("blind")
blind_n_sig = RooUnblindPrecision("blind_n_sig",
"blind number of signals",
"wasistdas", n_sig.getVal(),
10000, n_sig, blind_cat)
else:
# blind_cat.setLabel("unblind")
blind_n_sig = n_sig
print "n_sig value " + str(n_sig.getVal())
# raw_input("blind value " + str(blind_n_sig.getVal()))
# n_sig = blind_n_sig
# END BLINDING
elif n_sig >= 0:
n_sig = RooRealVar("n_sig", "Number of signal events", int(n_sig))
else:
raise ValueError("n_sig is not >= 0")
# if not blind:
# blind_n_sig = n_sig
# # create bkg-pdf
# lambda_exp = RooRealVar("lambda_exp", "lambda exp pdf bkg", -0.00025, -1., 1.)
# bkg_pdf = RooExponential("bkg_pdf", "Background PDF exp", x, lambda_exp)
if blind:
comb_pdf = RooAddPdf("comb_pdf", "Combined DoubleCB and bkg PDF",
RooArgList(sig_pdf, bkg_pdf),
RooArgList(blind_n_sig, n_bkg))
else:
comb_pdf = RooAddPdf("comb_pdf", "Combined DoubleCB and bkg PDF",
RooArgList(sig_pdf, bkg_pdf),
RooArgList(n_sig, n_bkg))
# create test dataset
# mean_gauss = RooRealVar("mean_gauss", "Mean of Gaussian", 5553, -10000, 10000)
# sigma_gauss = RooRealVar("sigma_gauss", "Width of Gaussian", 20, 0.0001, 300)
# gauss1 = RooGaussian("gauss1", "Gaussian test dist", x, mean_gauss, sigma_gauss)
# lambda_data = RooRealVar("lambda_data", "lambda exp data", -.002)
# exp_data = RooExponential("exp_data", "data example exp", x, lambda_data)
# frac_data = RooRealVar("frac_data", "Fraction PDF of data", 0.15)
#
# data_pdf = RooAddPdf("data_pdf", "Data PDF", gauss1, exp_data, frac_data)
# data = data_pdf.generate(RooArgSet(x), 30000)
# data.printValue()
# xframe = x.frame()
# data_pdf.plotOn(xframe)
# print "n_cpu:", meta_config.get_n_cpu()
# input("test")
# comb_pdf.fitTo(data, RooFit.Extended(ROOT.kTRUE), RooFit.NumCPU(meta_config.get_n_cpu()))
# HACK to get 8 cores in testing
c5 = TCanvas("c5", "RooFit pdf not fit vs " + data_name)
c5.cd()
x_frame1 = x.frame()
# data.plotOn(x_frame1)
# comb_pdf.pdfList()[1].plotOn(x_frame1)
if __name__ == "__main__":
n_cpu = 8
else:
n_cpu = meta_config.get_n_cpu()
print "n_cpu = ", n_cpu
# HACK
# n_cpu = 8
result_fit = comb_pdf.fitTo(data, RooFit.Minos(ROOT.kTRUE),
RooFit.Extended(ROOT.kTRUE),
RooFit.NumCPU(n_cpu))
# HACK end
if bkg_in_region:
x.setRange("signal", bkg_in_region[0], bkg_in_region[1])
bkg_pdf_fitted = comb_pdf.pdfList()[1]
int_argset = RooArgSet(x)
# int_argset = x
# int_argset.setRange("signal", bkg_in_region[0], bkg_in_region[1])
integral = bkg_pdf_fitted.createIntegral(int_argset,
RooFit.NormSet(int_argset),
RooFit.Range("signal"))
bkg_cdf = bkg_pdf_fitted.createCdf(int_argset, RooFit.Range("signal"))
bkg_cdf.plotOn(x_frame1)
# integral.plotOn(x_frame1)
n_bkg_below_sig = integral.getVal(int_argset) * n_bkg.getVal()
x_frame1.Draw()
if plot_importance >= 3:
c2 = TCanvas("c2", "RooFit pdf fit vs " + data_name)
c2.cd()
x_frame = x.frame()
# if log_plot:
# c2.SetLogy()
# x_frame.SetTitle("RooFit pdf vs " + data_name)
x_frame.SetTitle(data_name)
if pulls:
pad_data = ROOT.TPad("pad_data", "Pad with data and fit", 0, 0.33,
1, 1)
pad_pulls = ROOT.TPad("pad_pulls", "Pad with data and fit", 0, 0,
1, 0.33)
pad_data.SetBottomMargin(0.00001)
pad_data.SetBorderMode(0)
if log_plot:
pad_data.SetLogy()
pad_pulls.SetTopMargin(0.00001)
pad_pulls.SetBottomMargin(0.2)
pad_pulls.SetBorderMode(0)
pad_data.Draw()
pad_pulls.Draw()
pad_data.cd()
else:
if log_plot:
c2.SetLogy()
if blind:
# HACK
column = 'x'
# END HACK
x.setRange("lower", min_x, lower_blind)
x.setRange("upper", upper_blind, max_x)
range_str = "lower,upper"
lower_cut_str = str(
min_x) + "<=" + column + "&&" + column + "<=" + str(lower_blind)
upper_cut_str = str(
upper_blind) + "<=" + column + "&&" + column + "<=" + str(max_x)
sideband_cut_str = "(" + lower_cut_str + ")" + "||" + "(" + upper_cut_str + ")"
n_entries = data.reduce(
sideband_cut_str).numEntries() / data.numEntries()
# raw_input("n_entries: " + str(n_entries))
if plot_importance >= 3:
data.plotOn(x_frame, RooFit.CutRange(range_str),
RooFit.NormRange(range_str))
comb_pdf.plotOn(
x_frame, RooFit.Range(range_str),
RooFit.Normalization(n_entries, RooAbsReal.Relative),
RooFit.NormRange(range_str))
if pulls:
# pull_hist(pull_frame=x_frame, pad_data=pad_data, pad_pulls=pad_pulls)
x_frame_pullhist = x_frame.pullHist()
else:
if plot_importance >= 3:
data.plotOn(x_frame)
comb_pdf.plotOn(x_frame)
if pulls:
pad_pulls.cd()
x_frame_pullhist = x_frame.pullHist()
pad_data.cd()
comb_pdf.plotOn(x_frame,
RooFit.Components(sig_pdf.namePtr().GetName()),
RooFit.LineStyle(ROOT.kDashed))
comb_pdf.plotOn(x_frame,
RooFit.Components(bkg_pdf.namePtr().GetName()),
RooFit.LineStyle(ROOT.kDotted))
# comb_pdf.plotPull(n_sig)
if plot_importance >= 3:
x_frame.Draw()
if pulls:
pad_pulls.cd()
x_frame.SetTitleSize(0.05, 'Y')
x_frame.SetTitleOffset(0.7, 'Y')
x_frame.SetLabelSize(0.04, 'Y')
# c11 = TCanvas("c11", "RooFit\ pulls" + data_name)
# c11.cd()
# frame_tmp = x_frame
frame_tmp = x.frame()
# frame_tmp.SetTitle("significance")
frame_tmp.SetTitle("Roofit\ pulls\ " + data_name)
frame_tmp.addObject(x_frame_pullhist)
frame_tmp.SetMinimum(-5)
frame_tmp.SetMaximum(5)
# frame_tmp.GetYaxis().SetTitle("significance")
frame_tmp.GetYaxis().SetNdivisions(5)
frame_tmp.SetTitleSize(0.1, 'X')
frame_tmp.SetTitleOffset(1, 'X')
frame_tmp.SetLabelSize(0.1, 'X')
frame_tmp.SetTitleSize(0.1, 'Y')
frame_tmp.SetTitleOffset(0.5, 'Y')
frame_tmp.SetLabelSize(0.1, 'Y')
frame_tmp.Draw()
# raw_input("")
if not blind and nll_profile:
# nll_range = RooRealVar("nll_range", "Signal for nLL", n_sig.getVal(),
# -10, 2 * n_sig.getVal())
sframe = n_sig.frame(RooFit.Bins(20), RooFit.Range(1, 1000))
# HACK for best n_cpu
lnL = comb_pdf.createNLL(data, RooFit.NumCPU(8))
# HACK end
lnProfileL = lnL.createProfile(ROOT.RooArgSet(n_sig))
lnProfileL.plotOn(sframe, RooFit.ShiftToZero())
c4 = TCanvas("c4", "NLL Profile")
c4.cd()
# input("press ENTER to show plot")
sframe.Draw()
if plot_importance >= 3:
pass
params = comb_pdf.getVariables()
params.Print("v")
# print bkg_cdf.getVal()
if sPlot:
sPlotData = ROOT.RooStats.SPlot(
"sPlotData",
"sPlotData",
data, # variable fitted to, RooDataSet
comb_pdf, # fitted pdf
ROOT.RooArgList(
n_sig,
n_bkg,
# NSigB0s
))
sweights = np.array([
sPlotData.GetSWeight(i, 'n_sig') for i in range(data.numEntries())
])
return n_sig.getVal(), n_bkg_below_sig, sweights
if blind:
return blind_n_sig.getVal(), n_bkg_below_sig, comb_pdf
else:
return n_sig.getVal(), n_bkg_below_sig, comb_pdf
def rooFit601():
print ">>> setup pdf and likelihood..."
x = RooRealVar("x", "x", -20, 20)
mean = RooRealVar("mean", "mean of g1 and g2", 0)
sigma1 = RooRealVar("sigma1", "width of g1", 3)
sigma2 = RooRealVar("sigma2", "width of g2", 4, 3.0,
6.0) # intentional strong correlations
gauss1 = RooGaussian("gauss1", "gauss1", x, mean, sigma1)
gauss2 = RooGaussian("gauss2", "gauss2", x, mean, sigma2)
frac = RooRealVar("frac", "frac", 0.5, 0.0, 1.0)
model = RooAddPdf("model", "model", RooArgList(gauss1, gauss2),
RooArgList(frac))
print ">>> generate to data..."
data = model.generate(RooArgSet(x), 1000) # RooDataSet
print ">>> construct unbinned likelihood of model wrt data..."
nll = model.createNLL(data) # RooAbsReal
print ">>> interactive minimization and error analysis with MINUIT interface object..."
minuit = RooMinuit(nll)
print ">>> set avtive verbosity for logging of MINUIT parameter space stepping..."
minuit.setVerbose(kTRUE)
print ">>> call MIGRAD to minimize the likelihood..."
minuit.migrad()
print "\n>>> parameter values and error estimates that are back propagated from MINUIT:"
model.getParameters(RooArgSet(x)).Print("s")
print "\n>>> disable verbose logging..."
minuit.setVerbose(kFALSE)
print ">>> run HESSE to calculate errors from d2L/dp2..."
minuit.hesse()
print ">>> value of and error on sigma2 (back propagated from MINUIT):"
sigma2.Print()
print "\n>>> run MINOS on sigma2 parameter only..."
minuit.minos(RooArgSet(sigma2))
print "\n>>> value of and error on sigma2 (back propagated from MINUIT after running MINOS):"
sigma2.Print()
print "\n>>> saving results, contour plots..."
# Save a snapshot of the fit result. This object contains the initial
# fit parameters, the final fit parameters, the complete correlation
# matrix, the EDM, the minimized FCN , the last MINUIT status code and
# the number of times the RooFit function object has indicated evaluation
# problems (e.g. zero probabilities during likelihood evaluation)
result = minuit.save() # RooFitResult
# Make contour plot of mx vs sx at 1,2,3 sigma
frame1 = minuit.contour(frac, sigma2, 1, 2, 3) # RooPlot
frame1.SetTitle("RooMinuit contour plot")
# Print the fit result snapshot
result.Print("v")
print "\n>>> change value of \"mean\" parameter..."
mean.setVal(0.3)
# Rerun MIGRAD,HESSE
print ">>> rerun MIGRAD, HESSE..."
minuit.migrad()
minuit.hesse()
print ">>> value on and error of frac:"
frac.Print()
print "\n>>> fix value of \"sigma\" parameter (setConstant)..."
sigma2.setConstant(kTRUE)
print ">>> rerun MIGRAD, HESSE..."
minuit.migrad()
minuit.hesse()
frac.Print()
def predict_and_plot(self, x):
"""
Do the same as predict(), but add plots
Return -logL ratio, for external plotting
"""
rcParams['xtick.major.pad'] = 12
rcParams['ytick.major.pad'] = 12
roodata = RooDataSet('data', 'data', RooArgSet(self.phistar))
for xval in x:
self.phistar.setVal(xval)
roodata.add(RooArgSet(self.phistar))
theta = RooRealVar('theta', 'theta', self.theta_min, self.theta_max)
model = RooAddPdf('model', 'model',
RooArgList(self.pdfs['A'], self.pdfs['H']),
RooArgList(theta))
#with stdout_redirected_to():
print '\n\nPHISTAR FIT'
model.fitTo(roodata)
fitted_theta = theta.getValV()
# Histogram binning for data points
nbins = 10
xvals = np.linspace(0, 2*pi, 200)
yvals_H = []
yvals_A = []
# Get points for pdf curves
for xval in xvals:
self.phistar.setVal(xval)
yvals_H.append(self.pdfs['H'].getValV(RooArgSet(self.phistar)))
yvals_A.append(self.pdfs['A'].getValV(RooArgSet(self.phistar)))
yvals_H = np.array(yvals_H)
yvals_A = np.array(yvals_A)
# Plot pdfs by themselves
#print 'integral H =', np.trapz(yvals_H, dx=1.0/200.0)
fig = plt.figure()
plt.plot(xvals, yvals_H, color=self.colors['H'], label=r'$p_{H}(\varphi^{*})$')
plt.plot(xvals, yvals_A, color=self.colors['A'], label=r'$p_{A}(\varphi^{*})$')
plt.fill_between(xvals, 0, yvals_H, color=self.colors['H'], alpha=0.2)
plt.fill_between(xvals, 0, yvals_A, color=self.colors['A'], alpha=0.2)
plt.xlim([0, 2*pi])
plt.ylim([0, 0.295])
plt.xlabel(r'$\varphi^*$')
plt.ylabel('Probability density')
plt.legend(loc='upper right')
plt.tight_layout()
fig.show()
fig.savefig('phistar_pdfs.pdf')
# Scale to event yield
yvals_H *= roodata.numEntries()*(1-fitted_theta)/float(nbins)*2*pi
yvals_A *= roodata.numEntries()*(fitted_theta)/float(nbins)*2*pi
yvals_sum = yvals_H + yvals_A
# Make plot
fig, ax = plt.subplots(1)
histentries, binedges = np.histogram(x, bins=nbins, range=(0, 2*pi))
bincenters = (binedges[:-1] + binedges[1:])/2.0
yerr = np.sqrt(histentries)
plt.errorbar(bincenters, histentries, xerr=np.diff(binedges)*0.5, yerr=yerr, linestyle='None', ecolor='black', label='Data')
plt.plot(xvals, yvals_H, color=self.colors['H'], label=r'$p_{H}(\varphi^{*})$')
plt.plot(xvals, yvals_A, color=self.colors['A'], label=r'$p_{A}(\varphi^{*})$')
plt.plot(xvals, yvals_sum, color=self.colors['model'], label=r'$p(\varphi^{*} \,|\, \alpha = %.2f)$' % fitted_theta)
plt.fill_between(xvals, 0, yvals_H, color=self.colors['H'], alpha=0.2)
plt.fill_between(xvals, 0, yvals_A, color=self.colors['A'], alpha=0.2)
# Set correct legend order
handles, labels = ax.get_legend_handles_labels()
handles = [handles[3], handles[2], handles[0], handles[1]]
labels = [labels[3], labels[2], labels[0], labels[1]]
ax.legend(handles, labels, loc='upper right')
plt.xlabel(r'$\varphi^{*}$')
plt.ylabel('Events / %.2f' % ((2*pi)/nbins))
axes = plt.gca()
axes.set_xlim([0, 2*pi])
axes.set_ylim([0, max(histentries)*1.8])
plt.tight_layout()
fig.show()
fig.savefig('phistar_fit.pdf')
# Create likelihood curve
logl = model.createNLL(roodata)
# Extract curve
xvals = np.linspace(0, 1, 200)
yvals = []
ymin = 999.
for xval in xvals:
theta.setVal(xval)
yvals.append(logl.getValV())
if yvals[-1] < ymin:
ymin = yvals[-1]
# Shift minimum to zero
yvals = np.array(yvals)
yvals -= ymin
# Return points for the NLL curve
return xvals, yvals
def predict_and_plot(self, x):
"""
Do the same as predict(), but add plots
Return -logL ratio, for external plotting
"""
rcParams['xtick.major.pad'] = 12
rcParams['ytick.major.pad'] = 12
xs = self.scaler.transform(x)
preds = self.model.predict(xs)[:, 1]
roodata = RooDataSet('data', 'data', RooArgSet(self.roopred))
for pred in preds:
self.roopred.setVal(pred)
roodata.add(RooArgSet(self.roopred))
theta = RooRealVar('theta', 'theta', 0.5, self.theta_min, self.theta_max)
model = RooAddPdf('model', 'model',
RooArgList(self.pdfs['A'], self.pdfs['H']),
RooArgList(theta))
#with stdout_redirected_to():
print '\n\nNEURAL NETWORK FIT'
res = model.fitTo(roodata, PrintLevel(10))
fitted_theta = theta.getValV()
# Histogram binning for data points
nbins = 14
xvals = np.linspace(0, 1, 300)
yvals_H = []
yvals_A = []
# Get points for pdf curves
for xval in xvals:
self.roopred.setVal(xval)
yvals_H.append(self.pdfs['H'].getValV(RooArgSet(self.roopred)))
yvals_A.append(self.pdfs['A'].getValV(RooArgSet(self.roopred)))
yvals_H = np.array(yvals_H)
yvals_A = np.array(yvals_A)
# Plot pdfs by themselves
fig = plt.figure()
plt.plot(xvals, yvals_H, color=self.colors["H"], label=r'$p_{H}(y)$')
plt.plot(xvals, yvals_A, color=self.colors["A"], label=r'$p_{A}(y)$')
plt.fill_between(xvals, 0, yvals_H, color=self.colors["H"], alpha=0.2)
plt.fill_between(xvals, 0, yvals_A, color=self.colors["A"], alpha=0.2)
plt.xlim([0, 1])
plt.ylim([0, 11])
plt.xlabel(r'Network output ($y$)')
plt.ylabel('Probability density')
plt.legend(loc='upper right')
plt.tight_layout()
fig.show()
fig.savefig('mle_nn_pdfs.pdf')
# Scale to event yield
yvals_H *= roodata.numEntries()*(1-fitted_theta)/float(nbins)
yvals_A *= roodata.numEntries()*(fitted_theta)/float(nbins)
yvals_sum = yvals_H + yvals_A
# Make plot of fit to data
fig, ax = plt.subplots(1)
histentries, binedges = np.histogram(preds, bins=nbins, range=(0, 1))
bincenters = (binedges[:-1] + binedges[1:])/2.0
yerr = np.sqrt(histentries)
plt.errorbar(bincenters, histentries, xerr=np.diff(binedges)*0.5, yerr=yerr, linestyle='None', ecolor='black', label='Data')
plt.plot(xvals, yvals_H, color=self.colors["H"], label=r'$p_{H}(y)$')
plt.plot(xvals, yvals_A, color=self.colors["A"], label=r'$p_{A}(y)$')
plt.plot(xvals, yvals_sum, color=self.colors["model"], label=r'$p(y \,|\, \alpha = %.2f)$' % fitted_theta)
plt.fill_between(xvals, 0, yvals_H, color=self.colors["H"], alpha=0.2)
plt.fill_between(xvals, 0, yvals_A, color=self.colors["A"], alpha=0.2)
# Set correct legend order
handles, labels = ax.get_legend_handles_labels()
handles = [handles[3], handles[2], handles[0], handles[1]]
labels = [labels[3], labels[2], labels[0], labels[1]]
ax.legend(handles, labels, loc='upper right')
plt.xlabel(r'Network output ($y$)')
plt.ylabel('Events / %.2f' % (1.0/nbins))
axes = plt.gca()
axes.set_xlim([0, 1])
axes.set_ylim([0, max(histentries)*2.3])
plt.tight_layout()
fig.show()
fig.savefig('mle_nn_fit.pdf')
# Create likelihood curve
logl = model.createNLL(roodata)
xvals = np.linspace(0, 1, 200)
yvals = []
ymin = 999.
for xval in xvals:
theta.setVal(xval)
yvals.append(logl.getValV())
if yvals[-1] < ymin:
ymin = yvals[-1]
yvals = np.array(yvals)
yvals -= ymin
# Return points for the NLL curve
return xvals, yvals
