def train_classifier(X_train, y_train, base_learners, method, cl_names):
if method == 'AdaBoost':
clf = CostSensitiveAlgorithms(algorithm='AdaBoost',
n_estimators=base_learners)
clf.fit(X_train, y_train)
elif 'AdaCC' in method or 'AdaN-CC' in method:
clf = AdaCC(n_estimators=base_learners, algorithm=method)
clf.fit(X_train, y_train)
elif 'RareBoost' in method:
clf = RareBoost(n_estimators=base_learners)
clf.fit(X_train, y_train)
else:
counter_dict = Counter(list(y_train))
majority = max(counter_dict.items(), key=operator.itemgetter(1))[0]
minority = max(counter_dict.items(), key=operator.itemgetter(0))[0]
ratios = [1., 2., 3., 4., 5., 6., 7, 8., 9., 9.5, 9.9]
processes = []
if not os.path.exists('temp_preds_cdf'):
os.makedirs('temp_preds_cdf')
for ratio in ratios:
p = Process(target=train_competitors,
args=(X_train, y_train, base_learners, method,
majority, minority, ratio))
p.start()
processes.append(p)
for p in processes:
p.join()
best_score = -1
for ratio in ratios:
if os.path.exists('temp_preds_cdf/' + method + str(ratio)):
with open('temp_preds_cdf/' + method + str(ratio),
'rb') as filehandle:
temp = pickle.load(filehandle)
if temp[0] > best_score:
best_score = temp[0]
clf = temp[1]
if os.path.exists('temp_preds_cdf/' + method + str(ratio)):
os.remove('temp_preds_cdf/' + method + str(ratio))
with open('temp/' + method, 'wb') as filehandle:
pickle.dump(numpy.asarray(clf.get_confidence_scores(X_train)),
filehandle)
def train_classifier(X_train, y_train, base_learners, method, cl_names):
if method == 'AdaBoost':
clf = CostSensitiveAlgorithms(algorithm='AdaBoost',
n_estimators=base_learners)
clf.fit(X_train, y_train)
elif 'AdaCC' in method or 'AdaN-CC' in method:
clf = AdaCC(n_estimators=base_learners, algorithm=method)
clf.fit(X_train, y_train)
elif 'RareBoost' in method:
clf = RareBoost(n_estimators=base_learners)
clf.fit(X_train, y_train)
else:
counter_dict = Counter(list(y_train))
majority = max(counter_dict.items(), key=operator.itemgetter(1))[0]
minority = max(counter_dict.items(), key=operator.itemgetter(0))[0]
best = -1
ratios = [1., 2., 3., 4., 5., 6., 7, 8., 9., 10.]
for j in ratios:
try:
clf = CostSensitiveAlgorithms(n_estimators=base_learners,
algorithm=method,
class_weight={
minority: 1,
majority: j / 10.
})
clf.fit(X_train, y_train)
if clf.error == 1:
clf = None
else:
score = f1_score(y_train, clf.predict(X_train))
# score = balanced_accuracy_score(y_train, clf.predict(X_train))
if score >= best:
best = score
best_clf = clf
except:
pass
clf = best_clf
with open('temp_features/' + method, 'wb') as filehandle:
pickle.dump(clf, filehandle)
def train_and_predict(X_train, y_train, X_test, base_learners, method, cl_names):
if "AdaCC" not in method or "AdaN-CC" not in method:
X_train, X_valid, y_train, y_valid = train_test_split(X_train, y_train, test_size=0.4, stratify=y_train)
if method == 'AdaBoost':
t1_start = process_time()
clf = CostSensitiveAlgorithms(n_estimators=base_learners, algorithm=method)
clf.fit(X_train, y_train)
t1_stop = process_time()
oveall_time = t1_stop - t1_start
with open('temp_preds/stats_' + method, 'wb') as filehandle:
pickle.dump([oveall_time], filehandle)
with open('temp_preds/' + method, 'wb') as filehandle:
pickle.dump([clf.predict(X_test), clf.predict_proba(X_test)], filehandle)
return
elif 'AdaCC' in method:
t1_start = process_time()
clf = AdaCC(n_estimators=base_learners, algorithm=method)
clf.fit(X_train, y_train)
t1_stop = process_time()
oveall_time = t1_stop - t1_start
with open('temp_preds/stats_' + method, 'wb') as filehandle:
pickle.dump([oveall_time], filehandle)
with open('temp_preds/' + method, 'wb') as filehandle:
# pickle.dump(clf.predict(X_test), filehandle)
pickle.dump([clf.predict(X_test), clf.predict_proba(X_test)], filehandle)
return
elif 'AdaN-CC' in method:
t1_start = process_time()
clf = AdaCC(n_estimators=base_learners, algorithm=method.replace("N-", ""), amortised=False)
clf.fit(X_train, y_train)
t1_stop = process_time()
oveall_time = t1_stop - t1_start
with open('temp_preds/stats_' + method, 'wb') as filehandle:
pickle.dump([oveall_time], filehandle)
with open('temp_preds/' + method, 'wb') as filehandle:
# pickle.dump(clf.predict(X_test), filehandle)
pickle.dump([clf.predict(X_test), clf.predict_proba(X_test)], filehandle)
return
elif 'AdaMEC_Cal' in method:
counter_dict = Counter(list(y_train))
majority = max(counter_dict.items(), key=operator.itemgetter(1))[0]
minority = max(counter_dict.items(), key=operator.itemgetter(0))[0]
ratios = [1., 2., 3., 4., 5., 6., 7, 8., 9., 10.]
t1_start = process_time()
clf = AdaMEC_Cal(n_estimators=base_learners, algorithm=method)
clf.fit(X_train, y_train)
best_score = -1
best_idx = 0
for idx, cost in enumerate(ratios):
class_weight = {minority: 1, majority: cost / 10.}
clf.set_costs(y_train, class_weight)
score = f1_score(y_valid, clf.predict(X_valid))
if best_score < score:
best_idx = idx
best_score = score
class_weight = {minority: 1, majority: ratios[best_idx] / 10.}
clf.set_costs(y_train, class_weight)
t1_stop = process_time()
oveall_time = t1_stop - t1_start
with open('temp_preds/stats_' + method, 'wb') as filehandle:
pickle.dump([oveall_time, best_score, ratios[best_idx]], filehandle)
with open('temp_preds/' + method, 'wb') as filehandle:
# pickle.dump(clf.predict(X_test), filehandle)
pickle.dump([clf.predict(X_test), clf.predict_proba(X_test)], filehandle)
return
elif 'AdaMEC' in method:
counter_dict = Counter(list(y_train))
majority = max(counter_dict.items(), key=operator.itemgetter(1))[0]
minority = max(counter_dict.items(), key=operator.itemgetter(0))[0]
ratios = [1., 2., 3., 4., 5., 6., 7, 8., 9., 10.]
t1_start = process_time()
clf = AdaMEC(n_estimators=base_learners, algorithm=method)
clf.fit(X_train, y_train)
best_score = -1
best_idx = 0
for idx, cost in enumerate(ratios):
class_weight = {minority: 1, majority: cost / 10.}
clf.set_costs(class_weight)
score = f1_score(y_valid, clf.predict(X_valid))
if best_score < score:
best_idx = idx
best_score = score
class_weight = {minority: 1, majority: ratios[best_idx] / 10.}
clf.set_costs(class_weight)
t1_stop = process_time()
oveall_time = t1_stop - t1_start
with open('temp_preds/stats_' + method, 'wb') as filehandle:
pickle.dump([oveall_time, best_score, ratios[best_idx]], filehandle)
with open('temp_preds/' + method, 'wb') as filehandle:
# pickle.dump(clf.predict(X_test), filehandle)
pickle.dump([clf.predict(X_test), clf.predict_proba(X_test)], filehandle)
return
elif method == 'RareBoost':
t1_start = process_time()
clf = RareBoost(n_estimators=base_learners)
clf.fit(X_train, y_train)
t1_stop = process_time()
oveall_time = t1_stop - t1_start
with open('temp_preds/stats_' + method, 'wb') as filehandle:
pickle.dump([oveall_time], filehandle)
with open('temp_preds/' + method, 'wb') as filehandle:
# pickle.dump(clf.predict(X_test), filehandle)
pickle.dump([clf.predict(X_test), clf.predict_proba(X_test)], filehandle)
return
else:
counter_dict = Counter(list(y_train))
majority = max(counter_dict.items(), key=operator.itemgetter(1))[0]
minority = max(counter_dict.items(), key=operator.itemgetter(0))[0]
ratios = [1., 2., 3., 4., 5., 6., 7, 8., 9., 10.]
processes = []
for ratio in ratios:
p = Process(target=train_competitors,
args=(X_train, y_train, X_valid, y_valid, base_learners, method, majority, minority, ratio))
p.start()
processes.append(p)
for p in processes:
p.join()
best_score = -1
best_ratio = 0
oveall_time = 0
for ratio in ratios:
if os.path.exists('temp_preds/' + method + str(ratio)):
with open('temp_preds/' + method + str(ratio), 'rb') as filehandle:
temp = pickle.load(filehandle)
oveall_time += temp[2]
if temp[0] > best_score:
best_ratio = ratio
best_score = temp[0]
clf = temp[1]
if os.path.exists('temp_preds/' + method + str(ratio)):
os.remove('temp_preds/' + method + str(ratio))
with open('temp_preds/stats_' + method, 'wb') as filehandle:
pickle.dump([oveall_time, best_score, best_ratio], filehandle)
with open('temp_preds/' + method, 'wb') as filehandle:
# pickle.dump(clf.predict(X_test), filehandle)
pickle.dump([clf.predict(X_test), clf.predict_proba(X_test)], filehandle)
return
def train_and_predict(X_train, y_train, base_learners, method):
if method == 'AdaBoost':
clf = CostSensitiveAlgorithms(algorithm='AdaBoost',
n_estimators=base_learners,
debug=True)
clf.fit(X_train, y_train)
elif 'AdaCC' in method:
clf = AdaCC(n_estimators=base_learners, algorithm=method, debug=True)
clf.fit(X_train, y_train)
elif 'AdaN-AC' in method:
clf = AdaCC(n_estimators=base_learners,
algorithm=method.replace("N-", ""),
debug=True,
amortised=False)
clf.fit(X_train, y_train)
elif 'RareBoost' in method:
clf = RareBoost(n_estimators=base_learners, debug=True)
clf.fit(X_train, y_train)
with open('temp_preds_AdaCC/' + method, 'wb') as filehandle:
pickle.dump([
clf._class_weights_pos, clf._class_weights_neg,
clf.training_error, clf.estimator_weights_pos,
clf.estimator_weights_neg
], filehandle)
return
else:
counter_dict = Counter(list(y_train))
majority = max(counter_dict.items(), key=operator.itemgetter(1))[0]
minority = max(counter_dict.items(), key=operator.itemgetter(0))[0]
ratios = [1., 2., 3., 4., 5., 6., 7, 8., 9., 10.]
processes = []
for ratio in ratios:
p = Process(target=train_competitors,
args=(X_train, y_train, base_learners, method,
majority, minority, ratio))
p.start()
processes.append(p)
for p in processes:
p.join()
best_score = -1
for ratio in ratios:
if os.path.exists('temp_preds_AdaCC/' + method + str(ratio)):
with open('temp_preds_AdaCC/' + method + str(ratio),
'rb') as filehandle:
temp = pickle.load(filehandle)
if temp[0] > best_score:
best_score = temp[0]
clf = temp[1]
if os.path.exists('temp_preds_AdaCC/' + method + str(ratio)):
os.remove('temp_preds_AdaCC/' + method + str(ratio))
with open('temp_preds_AdaCC/' + method, 'wb') as filehandle:
pickle.dump([
clf._class_weights_pos, clf._class_weights_neg, clf.training_error,
clf.estimator_alphas_
], filehandle)
def train_and_predict(X_train, y_train, base_learners, method):
if method == 'AdaBoost':
clf = CostSensitiveAlgorithms(algorithm='AdaBoost',
n_estimators=base_learners)
clf.fit(X_train, y_train)
elif 'AdaCC' in method:
clf = AdaCC(n_estimators=base_learners, algorithm=method)
clf.fit(X_train, y_train)
elif 'AdaMEC' in method:
counter_dict = Counter(list(y_train))
majority = max(counter_dict.items(), key=operator.itemgetter(1))[0]
minority = max(counter_dict.items(), key=operator.itemgetter(0))[0]
ratios = [1., 2., 3., 4., 5., 6., 7, 8., 9., 10.]
clf = AdaMEC_Cal(n_estimators=base_learners, algorithm=method)
clf.fit(X_train, y_train)
best_score = -1
best_idx = 0
for idx, cost in enumerate(ratios):
class_weight = {minority: 1, majority: cost / 10.}
clf.set_costs(y_train, class_weight)
score = f1_score(y_train, clf.predict(X_train))
# score = balanced_accuracy_score(y_train, clf.predict(X_train))
if best_score < score:
best_idx = idx
best_score = score
class_weight = {minority: 1, majority: ratios[best_idx] / 10.}
clf.set_costs(y_train, class_weight)
elif 'RareBoost' in method:
clf = RareBoost(n_estimators=base_learners)
clf.fit(X_train, y_train)
else:
counter_dict = Counter(list(y_train))
majority = max(counter_dict.items(), key=operator.itemgetter(1))[0]
minority = max(counter_dict.items(), key=operator.itemgetter(0))[0]
ratios = [1., 2., 3., 4., 5., 6., 7, 8., 9., 9.99]
processes = []
for ratio in ratios:
p = Process(target=train_competitors,
args=(X_train, y_train, base_learners, method,
majority, minority, ratio))
p.start()
processes.append(p)
for p in processes:
p.join()
best_ratio = -1
predictor = None
for ratio in ratios:
if os.path.exists('boundary_temp_preds/' + method + str(ratio)):
with open('boundary_temp_preds/' + method + str(ratio),
'rb') as filehandle:
temp = pickle.load(filehandle)
if temp[0] > best_ratio:
best_ratio = temp[0]
predictor = temp[1]
if os.path.exists('boundary_temp_preds/' + method + str(ratio)):
os.remove('boundary_temp_preds/' + method + str(ratio))
with open('boundary_temp_preds/' + method, 'wb') as filehandle:
joblib.dump(predictor, filehandle)
return
with open('boundary_temp_preds/' + method, 'wb') as filehandle:
joblib.dump(clf, filehandle)