In [2]:
!free -m
In [1]:
%pylab inline
In [2]:
import pandas
Bdata = pandas.read_csv('models/Bdata_tracks_SS.csv')
In [3]:
Bdata.head()
Out[3]:
In [4]:
# for Nan put 1 as non influence factor
Bdata.ix[numpy.isnan(Bdata.track_relation_prob.values), 'track_relation_prob'] = 1.
In [5]:
relation_prob = Bdata['track_relation_prob'].values
Bprob = relation_prob / (1 + relation_prob)
Bweight = Bdata.Bweight.values
Bsign = Bdata.Bsign.values
In [6]:
Bprob[~numpy.isfinite(Bprob)] = 0.5
In [7]:
from utils import calibrate_probs
Bprob_calibrated, (iso_reg1, iso_reg2) = calibrate_probs(Bsign, Bweight, Bprob,
symmetrize=True, return_calibrator=True)
In [8]:
Bprob_calibrated = Bprob_calibrated + numpy.random.normal(size=len(Bprob_calibrated)) * 0.001
In [9]:
figure(figsize=(15, 5))
subplot(1,2,1)
hist(Bprob[Bsign == 1], weights=Bweight[Bsign == 1], bins=60, alpha=0.2, normed=True, label='$B^+$')
hist(Bprob[Bsign == -1], weights=Bweight[Bsign == -1], bins=60, alpha=0.2, normed=True, label='$B^-$')
legend(), title('B probs')
subplot(1,2,2)
hist(Bprob_calibrated[Bsign == 1], weights=Bweight[Bsign == 1], bins=60, alpha=0.2,
normed=True, range=(0, 1), label='$B^+$')
hist(Bprob_calibrated[Bsign == -1], weights=Bweight[Bsign == -1], bins=60, alpha=0.2,
normed=True, range=(0, 1), label='$B^-$')
legend(), title('B probs calibrated')
plt.savefig('img/Bprob_iso_calibrated_SS.png' , format='png')
In [10]:
from utils import calculate_auc_with_and_without_untag_events
from sklearn.metrics import roc_curve
auc, auc_full = calculate_auc_with_and_without_untag_events(Bsign, Bprob_calibrated, Bweight)
print 'AUC for tagged:', auc, 'AUC with untag:', auc_full
fpr, tpr, _ = roc_curve(Bsign, Bprob_calibrated, sample_weight=Bweight)
plot(fpr, tpr)
plot([0, 1], [0, 1], 'k--')
ylim(0, 1), xlim(0, 1)
Out[10]:
In [11]:
figsize(12, 8)
for sign in [-1, 1]:
hist(sign * (Bprob[Bsign == sign] - 0.5), bins=101, normed=True, alpha=0.2,
weights=Bweight[Bsign == sign], range=(-0.5, 0.5), label='$B^-$' if sign == -1 else '$B^+$')
legend(), title('Symmetry of $p(B^+)$ for $B^+$ and $B^-$, before calibration')
Out[11]:
In [12]:
fpr, tpr, _ = roc_curve(Bsign, (Bprob - 0.5) * Bsign, sample_weight=Bweight)
In [13]:
'KS distance', max(abs(fpr - tpr))
Out[13]:
In [14]:
figsize(6, 5)
plot(fpr, tpr), grid()
xlim(0, 1), ylim(0, 1)
Out[14]:
In [15]:
figsize(12, 8)
for sign in [-1, 1]:
hist(sign * (Bprob_calibrated[Bsign == sign] - 0.5), bins=101, normed=True, alpha=0.2,
weights=Bweight[Bsign == sign], range=(-0.5, 0.5), label='$B^-$' if sign == -1 else '$B^+$')
legend(), title('Symmetry of $p(B^+)$ for $B^+$ and $B^-$, after calibration')
Out[15]:
In [16]:
fpr, tpr, _ = roc_curve(Bsign, (Bprob_calibrated - 0.5) * Bsign, sample_weight=Bweight)
In [17]:
'KS distance', max(abs(fpr - tpr))
Out[17]:
In [18]:
figsize(6, 5)
plot(fpr, tpr), grid()
xlim(0, 1), ylim(0, 1)
Out[18]:
In [19]:
from utils import get_N_B_events, bootstrap_calibrate_prob, result_table
N_B_passed = Bweight.sum()
tagging_efficiency = N_B_passed / get_N_B_events()
tagging_efficiency_delta = numpy.sqrt(N_B_passed) / get_N_B_events()
D2, aucs = bootstrap_calibrate_prob(Bsign, Bweight, Bprob, symmetrize=True)
print 'AUC', numpy.mean(aucs), numpy.var(aucs)
result = result_table(tagging_efficiency, tagging_efficiency_delta, D2, auc_full, 'Inclusive tagging, PID less')
In [20]:
result
Out[20]:
In [21]:
result.to_csv('img/eff_tracking_SS.csv', header=True, index=False)
In [22]:
x = numpy.linspace(0, 1, 100)
plot(x, -(iso_reg1.transform((1-x)) + iso_reg2.transform((1-x))) / 2 + 1, label='isotonic transformation reverse')
plot(x, (iso_reg1.transform(x) + iso_reg2.transform(x)) / 2, label='isotonic transformation')
legend(loc='best')
plot([0, 1], [0, 1], "k--")
xlabel('B prob'), ylabel('B prob calibrated')
plt.savefig('img/iso_transformation_SS.png' , format='png')
In [23]:
from utils import get_N_B_events, compute_mistag
In [24]:
bins = [0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45]
percentile_bins = [10, 20, 30, 40, 50, 60, 70, 80, 90]
In [25]:
figsize(12, 10)
compute_mistag(Bprob, Bsign, Bweight, Bsign > -100, label="$B$", bins=bins)
compute_mistag(Bprob, Bsign, Bweight, Bsign == 1, label="$B^+$", bins=bins)
compute_mistag(Bprob, Bsign, Bweight, Bsign == -1, label="$B^-$", bins=bins)
legend(loc='best')
title('B prob, uniform bins'), xlabel('mistag probability'), ylabel('true mistag probability')
plt.savefig('img/Bprob_calibration_check_uniform_SS.png' , format='png')
In [26]:
compute_mistag(Bprob, Bsign, Bweight, Bsign > -100, label="$B$", uniform=False, bins=percentile_bins)
p1 = compute_mistag(Bprob, Bsign, Bweight, Bsign == 1, label="$B^+$", uniform=False, bins=percentile_bins)
compute_mistag(Bprob, Bsign, Bweight, Bsign == -1, label="$B^-$", uniform=False, bins=percentile_bins)
legend(loc='best')
title('B prob, percentile bins'), xlabel('mistag probability'), ylabel('true mistag probability')
plt.savefig('img/Bprob_calibration_check_percentile_SS.png' , format='png')
In [27]:
compute_mistag(Bprob_calibrated, Bsign, Bweight, Bsign > -100, label="$B$", bins=bins)
compute_mistag(Bprob_calibrated, Bsign, Bweight, Bsign == 1, label="$B^+$", bins=bins)
compute_mistag(Bprob_calibrated, Bsign, Bweight, Bsign == -1, label="$B^-$", bins=bins)
legend(loc='best')
title('B prob isotonic calibrated, uniform bins'), xlabel('mistag probability'), ylabel('true mistag probability')
plt.savefig('img/Bprob_calibration_check_iso_uniform_SS.png' , format='png')
In [28]:
figsize(12, 10)
compute_mistag(Bprob_calibrated, Bsign, Bweight, Bsign > -100, label="$B$", uniform=False,
bins=percentile_bins)
compute_mistag(Bprob_calibrated, Bsign, Bweight, Bsign == 1, label="$B^+$", uniform=False,
bins=percentile_bins)
compute_mistag(Bprob_calibrated, Bsign, Bweight, Bsign == -1, label="$B^-$", uniform=False,
bins=percentile_bins)
legend(loc='best'), xlabel('mistag probability'), ylabel('true mistag probability')
title('B prob isotonic calibrated, percentile bins')
plt.savefig('img/Bprob_calibration_check_iso_percentile_SS.png' , format='png')
In [29]:
print numpy.average((2*(Bprob - 0.5))**2, weights=Bweight) * tagging_efficiency * 100
print numpy.average((2*(Bprob_calibrated - 0.5))**2, weights=Bweight) * Bweight.sum() / get_N_B_events() * 100