def apply():
if len(sys.argv) < 4:
print("Usage: ", sys.argv[0], " apply datafile weightfile")
return 1
datafile = sys.argv[2]
weightfile = sys.argv[3]
X, y, w = readDataFile(datafile)
forest = FastBDT.Classifier()
forest.load(weightfile)
analyse(forest, X, y)
return 0
def worker(reduced, data, workerID=None):
"""Worker process
reduced:
Feature subset
data:
Data
"""
# Make a copy of data
data = data.copy()
bdt = FastBDT.Classifier()
bdt.fit(data[reduced], data.isSignal)
data['mva'] = bdt.predict(data[reduced])
auc = roc_auc_score(data.isSignal, data.mva)
if workerID is not None:
print("Worker %d finished the job!" % workerID)
return {'auc': auc}
def output():
if len(sys.argv) < 4:
print("Usage: ", sys.argv[0], " output datafile weightfile")
return 1
datafile = sys.argv[2]
weightfile = sys.argv[3]
X, y, w = readDataFile(datafile)
forest = FastBDT.Classifier()
forest.load(weightfile)
p = forest.predict(X)
for i, p in enumerate(p):
print(int(y[i] == 1), p)
return 0
def train():
if len(sys.argv) < 4:
print(
"Usage: ", sys.argv[0],
" train datafile weightfile [nCuts=4] [nTrees=100] [nLevels=3] [shrinkage=0.1] [randRatio=0.5]"
)
return 1
datafile = sys.argv[2]
weightfile = sys.argv[3]
forest = FastBDT.Classifier(*sys.argv[4:])
X, y, w = readDataFile(datafile)
forest.fit(X, y, w)
forest.save(weightfile)
analyse(forest, X, y)
def backward_selection(features, data, min_auc=0.9975):
# Make a copy of features so we do not accidentally modify them
features = features.copy()
print('Fitting the model with all the features...')
bdt = FastBDT.Classifier()
bdt.fit(data[features], data.isSignal)
data['mva'] = bdt.predict(data[features])
init_auc = roc_auc_score(data.isSignal, data.mva)
print('Initial AUC:', init_auc)
print('Minimum AUC to continue the search: %.4f' % min_auc)
current_auc = init_auc
best = pd.DataFrame()
best = best.append({
'n_features': len(features),
'best_auc': current_auc
},
ignore_index=True)
while current_auc >= min_auc:
print("Trying to remove one feature...")
result = pd.DataFrame()
n_workers = 10
pool = multiprocessing.Pool(n_workers)
worker_results = []
for i, _ in enumerate(features):
reduced = [features[j] for j in range(len(features)) if j != i]
worker_results.append(pool.apply_async(worker, [reduced, data]))
print('Worker[%d]: Trying to remove %s' % (i, features[i]))
print("Waiting for the workers to finish...")
for i, res in enumerate(worker_results):
print("Return value from worker %d is %r" % (i, res.get()))
result = result.append({
'feature': features[i],
**res.get()
},
ignore_index=True)
print(result)
result = result.sort_values(by='auc', ascending=False)
current_auc = result.iloc[0].auc
best = best.append(
{
'n_features': len(features) - 1,
'best_auc': current_auc
},
ignore_index=True)
features = result.iloc[1:].feature.values
print("Highest AUC = %.4f" % current_auc)
print("%s will be removed" % result.iloc[0].feature)
print('Features left (%d):' % (len(features)))
for f in features:
print(f)
print("AUC table begins".center(40, '='))
print(result)
print("AUC table ends".center(40, '='))
fig = plot_auc_chart(init_auc, result)
fig.savefig('../models/multiproc/auc_%d.png' % (len(features)),
bbox_inches="tight")
return best
for i in range(len(mean)):
for j in range(i + 1, len(mean)):
cov[j][i] = cov[i][j]
N_train, N_test = 100000, 2000
data = np.random.multivariate_normal(mean, cov, N_train + N_test)
X_train, y_train = data[:N_train, 1:], data[:N_train, 0] > 0
X_test, y_test = data[N_train:, 1:], data[N_train:, 0] > 0
# First variable is the variable we want to have independent of our network output
prior = Prior(X_train[y_train == 1, 0], X_train[y_train == 0, 0])
p_prior = prior.get_prior(X_test[:, 0])
evaluation("Prior", X_test, y_test, p_prior, p_prior)
p = FastBDT.Classifier().fit(X=X_train, y=y_train).predict(X_test)
evaluation("Full", X_test, y_test, p, p_prior)
p = FastBDT.Classifier().fit(X=X_train[:, 1:],
y=y_train).predict(X_test[:, 1:])
evaluation("Restricted", X_test, y_test, p, p_prior)
boost_p = FastBDT.Classifier().fit(
X=numpy.r_[X_train[:, 1:], X_train[:, 1:]],
y=numpy.r_[numpy.ones(N_train),
numpy.zeros(N_train)],
weights=prior.get_boost_weights(X_train[:,
0])).predict(X_train[:,
1:])
p = FastBDT.Classifier().fit(X=X_train[:, 1:],
[0.0, 0.0, 1.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 1.0]]
for i in range(len(mean)):
for j in range(i + 1, len(mean)):
cov[j][i] = cov[i][j]
N_train, N_test = 100000, 2000
data = np.random.multivariate_normal(mean, cov, N_train + N_test)
X_train, y_train = data[:N_train, 1:], data[:N_train, 0] > 0
X_test, y_test = data[N_train:, 1:], data[N_train:, 0] > 0
# First variable is the variable we want to have independent of our network output
prior = Prior(X_train[y_train == 1, 0], X_train[y_train == 0, 0])
p_prior = prior.get_prior(X_test[:, 0])
evaluation("Prior", X_test, y_test, p_prior, p_prior)
p = FastBDT.Classifier(flatnessLoss=1.0,
numberOfFlatnessFeatures=1).fit(
X=np.c_[X_train[:, 1:], X_train[:, 0]],
y=y_train).predict(X_test[:, 1:])
print(p)
evaluation("UBoost", X_test, y_test, p, p_prior)
p = FastBDT.Classifier().fit(X=X_train, y=y_train).predict(X_test)
evaluation("Full", X_test, y_test, p, p_prior)
p = FastBDT.Classifier().fit(X=X_train[:, 1:],
y=y_train).predict(X_test[:, 1:])
evaluation("Restricted", X_test, y_test, p, p_prior)
mean = [0.0, 1.0, 2.0, 3.0, 4.0, 5.0]
cov = [[1.0, 0.8, 0.4, 0.2, 0.1, 0.0], [0.0, 1.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 1.0]]
for i in range(len(mean)):
for j in range(i + 1, len(mean)):
cov[j][i] = cov[i][j]
N_train, N_test = 10000, 10000
data = np.random.multivariate_normal(mean, cov, N_train + N_test)
X_train, y_train = data[:N_train, 1:], data[:N_train, 0] > 0
X_test, y_test = data[N_train:, 1:], data[N_train:, 0] > 0
# Train FastBDT using its PythonInterface, which is based on the SKLearn classifiers
clf = FastBDT.Classifier(purityTransformation=1)
clf.fit(X=X_train, y=y_train)
p = clf.predict(X_test)
global_auc = sklearn.metrics.roc_auc_score(y_test, p)
print("Global AUC", global_auc)
# Intern feature importance is calculated using the sum of the information gains
# provided by each feature in all decision trees
print("Intern Feature Importance")
print(clf.internFeatureImportance())
# Extern feature importance is calculated using the drop in the area under the receiver operating characteristics curve
# if the most important feature is left out recursively
print("Extern Feature Importance")
print(
clf.externFeatureImportance(X_train, y_train, None, X_test, y_test,
from PyFastBDT import FastBDT
import pandas
import numpy as np
import sklearn.metrics
if __name__ == '__main__':
data = np.arange(100000)
X = (data % 100).reshape((100000, 1))
y = (data % 2) == 1
clf = FastBDT.Classifier(nTrees=1, depth=1, shrinkage=0.1, subsample=1.0, purityTransformation=[False]).fit(X=X, y=y)
p = clf.predict(X)
print('No Purity Transformation', sklearn.metrics.roc_auc_score(y, p))
clf = FastBDT.Classifier(nTrees=1, depth=1, shrinkage=0.1, subsample=1.0, purityTransformation=[True]).fit(X=X, y=y)
p = clf.predict(X)
print('With Purity Transformation', sklearn.metrics.roc_auc_score(y, p))
[0.0, 0.0, 1.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 0.0, 1.0]]
for i in range(len(mean)):
for j in range(i+1, len(mean)):
cov[j][i] = cov[i][j]
N_train, N_test = 10000, 10000
data = np.random.multivariate_normal(mean, cov, N_train + N_test)
X_train, y_train = data[:N_train, 1:], data[:N_train, 0] > 0
X_test, y_test = data[N_train:, 1:], data[N_train:, 0] > 0
# Train FastBDT using its PythonInterface, which is based on the SKLearn classifiers
clf = FastBDT.Classifier()
clf.fit(X=X_train, y=y_train)
p = clf.predict(X_test)
global_auc = sklearn.metrics.roc_auc_score(y_test, p)
print("Global AUC", global_auc)
# Intern feature importance is calculated using the sum of the information gains
# provided by each feature in all decision trees
print("Intern Feature Importance")
print(clf.internFeatureImportance())
# Extern feature importance is calculated using the drop in the area under the receiver operating characteristics curve
# if the most important feature is left out recursively
print("Extern Feature Importance")
print(clf.externFeatureImportance(X_train, y_train, None, X_test, y_test, None))
mean = [0.0, 1.0, 2.0, 3.0, 4.0, 5.0]
cov = [[1.0, 0.8, 0.4, 0.2, 0.1, 0.0], [0.0, 1.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 1.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 1.0]]
for i in range(len(mean)):
for j in range(i + 1, len(mean)):
cov[j][i] = cov[i][j]
N_train, N_test = 10000, 10000
data = np.random.multivariate_normal(mean, cov, N_train + N_test)
X_train, y_train = data[:N_train, 1:], data[:N_train, 0] > 0
X_test, y_test = data[N_train:, 1:], data[N_train:, 0] > 0
# Train FastBDT using its PythonInterface, which is based on the SKLearn classifiers
clf = FastBDT.Classifier()
clf.fit(X=X_train, y=y_train)
p = clf.predict(X_test)
global_auc = sklearn.metrics.roc_auc_score(y_test, p)
print("Global AUC", global_auc)
# Intern feature importance is calculated using the sum of the information gains
# provided by each feature in all decision trees
print("Intern Feature Importance")
print(clf.internFeatureImportance())
# Extern feature importance is calculated using the drop in the area under the receiver operating characteristics curve
# if the most important feature is left out recursively
print("Extern Feature Importance")
print(
clf.externFeatureImportance(X_train, y_train, None, X_test, y_test,
weightfile = sys.argv[3]
X, y, w = readDataFile(datafile)
forest = FastBDT.Classifier()
forest.load(weightfile)
p = forest.predict(X)
for i, p in enumerate(p):
print(int(y[i] == 1), p)
return 0
if __name__ == '__main__':
FastBDT.PrintVersion()
if len(sys.argv) <= 1:
print("Usage ", sys.argv[0], " [train|apply|output]")
sys.exit(1)
if sys.argv[1] == 'train':
ret = train()
elif sys.argv[1] == 'apply':
ret = apply()
elif sys.argv[1] == 'output':
ret = output()
else:
print("Unkown option", sys.argv[1])
ret = 1
[0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 1.0]]
for i in range(len(mean)):
for j in range(i + 1, len(mean)):
cov[j][i] = cov[i][j]
N_train, N_test = 100000, 2000
data = np.random.multivariate_normal(mean, cov, N_train + N_test)
X_train, y_train = data[:N_train, 1:], data[:N_train, 0] > 0
X_test, y_test = data[N_train:, 1:], data[N_train:, 0] > 0
# First variable is the variable we want to have independent of our network output
prior = Prior(X_train[y_train == 1, 0], X_train[y_train == 0, 0])
p_prior = prior.get_prior(X_test[:, 0])
evaluation("Prior", X_test, y_test, p_prior, p_prior)
p = FastBDT.Classifier(flatnessLoss=10.0).fit(X=np.c_[X_train[:, 1:],
X_train[:, 0]],
y=y_train,
nSpectators=1).predict(
X_test[:, 1:])
print(p)
evaluation("UBoost", X_test, y_test, p, p_prior)
p = FastBDT.Classifier().fit(X=X_train, y=y_train).predict(X_test)
evaluation("Full", X_test, y_test, p, p_prior)
p = FastBDT.Classifier().fit(X=X_train[:, 1:],
y=y_train).predict(X_test[:, 1:])
evaluation("Restricted", X_test, y_test, p, p_prior)