def InnerCycle_Train(X, y, inject_drift, perc_train):
# get number of training samples
ntrain = int(perc_train * X.shape[0])
if inject_drift:
# pick a point between 0.7 and 0.9 of the stream
dpoints = Driftpoints(X)
dpoints["cleanrun"] = dpoints["row"] - ntrain
# contaminate X after that point
X = Swapcols(df=X,
class_vec=y,
ids=dpoints["cols"],
t_change=dpoints["row"])
else:
dpoints = dict({"row": X.shape[0], "cols": 0})
# cast data as DataStream class
stream = DataStream(X, y)
stream.prepare_for_use()
# call incr model (main classifier, teacher model)
stream_clf = ARF(n_estimators=25) #,
#drift_detection_method=None,
#warning_detection_method=None
#)
# get training data... first ntrain rows
Xtrain, ytrain = stream.next_sample(ntrain)
# partial fit of the incre model using training data
stream_clf.fit(Xtrain, ytrain, classes=stream.target_values)
yhat_train = stream_clf.predict(Xtrain)
yhat_train_prob = stream_clf.predict_proba(
Xtrain) ### needs warnings!!!!!!!!!
yhat_tr_max_prob = np.array([np.max(x) for x in yhat_train_prob])
# fit student model
student_clf = ARF(n_estimators=25) #,
#drift_detection_method=None,
#warning_detection_method=None)
student_clf.fit(Xtrain, yhat_train, classes=stream.target_values)
student_regr = RHT()
student_regr.fit(Xtrain, yhat_tr_max_prob)
results = dict()
results["Teacher"] = stream_clf
results["Student"] = student_clf
results["StudentRegression"] = student_regr
results["Driftpoints"] = dpoints
results["n"] = ntrain
results["Stream"] = stream
results["Xtrain"] = Xtrain
return (results)
alpha=90.0, position=N_SAMPLES / 2)
stream.name = 'LED ABRUPBT'
STREAMS.append(stream)
"""Evaluate on ARSLVQ, SAM and HAT"""
# TODO NB and ARSLVQ working
for stream in STREAMS:
print('{}:\n'.format(stream.name))
f = open(res_file, 'a+')
f.write('{}:\n'.format(stream.name))
f.close()
rrslvq = RRSLVQ(prototypes_per_class=2,confidence=1e-10)
high_dim_test(copy.copy(stream), copy.copy(rrslvq), N_SAMPLES)
low_dim_test(copy.copy(stream), copy.copy(rrslvq), N_SAMPLES)
arslvq = RSLVQ(gradient_descent='Adadelta')
high_dim_test(copy.copy(stream), copy.copy(arslvq), N_SAMPLES)
low_dim_test(copy.copy(stream), copy.copy(arslvq), N_SAMPLES)
samknn = SAMKNN(max_window_size=5000,stm_size_option=None)
high_dim_test(copy.copy(stream), copy.copy(samknn), N_SAMPLES)
low_dim_test(copy.copy(stream), copy.copy(samknn), N_SAMPLES)
arf = ARF()
high_dim_test(copy.copy(stream), copy.copy(arf), N_SAMPLES)
low_dim_test(copy.copy(stream), copy.copy(arf), N_SAMPLES)
# labels = []
# while 1:
# line = f.readline()
# if line == '': break
# arr = np.array(line.split(','), dtype='float64')
# labels.append(arr[1])
# f.close()
# HIGH-DIM
X, y = data[:, :-1], data[:, -1]
clfs = [
RSLVQ(prototypes_per_class=2, gradient_descent="Adadelta"),
RRSLVQ(prototypes_per_class=2, confidence=1e-10),
ARF(),
SAMKNN()
]
for clf in clfs:
acc_fold = []
kappa_fold = []
time_fold = []
for _ in range(5):
_clf = copy.deepcopy(clf)
start_time = time.time()
y_true = []
y_pred = []
x = data[0, :-1].reshape(1, 1000)
def InnerCycle(X, y, inject_drift, perc_train, window, delta, pval,
prob_instance, inst_delay):
# get number of training samples
ntrain = int(perc_train * X.shape[0])
if inject_drift:
# pick a point between 0.7 and 0.9 of the stream
dpoints = Driftpoints(X)
dpoints["cleanrun"] = dpoints["row"] - ntrain
# contaminate X after that point
X = Swapcols(df=X,
class_vec=y,
ids=dpoints["cols"],
t_change=dpoints["row"])
else:
dpoints = dict({"row": X.shape[0], "cols": 0})
# cast data as DataStream class
stream = DataStream(X, y)
stream.prepare_for_use()
# call incr model (main classifier, teacher model)
stream_clf = ARF(n_estimators=25,
drift_detection_method=None,
warning_detection_method=None)
# get training data... first ntrain rows
Xtrain, ytrain = stream.next_sample(ntrain)
# partial fit of the incre model using training data
stream_clf.fit(Xtrain, ytrain, classes=stream.target_values)
yhat_train = stream_clf.predict(Xtrain)
yhat_train_prob = stream_clf.predict_proba(
Xtrain) ### needs warnings!!!!!!!!!
yhat_tr_max_prob = np.array([np.max(x) for x in yhat_train_prob])
# fit student model
student_clf = ARF(n_estimators=25,
drift_detection_method=None,
warning_detection_method=None)
student_clf.fit(Xtrain, yhat_train, classes=stream.target_values)
student_regr = RHT()
student_regr.fit(Xtrain, yhat_tr_max_prob)
####### Call drift detectors
## Supervised
# Supervised with ADWIN
S_ADWIN = ADWIN() #(delta=delta)
S_ADWIN_alarms = []
# Supervised with PHT
S_PHT = PHT() #(min_instances=window,delta=delta)
S_PHT_alarms = []
# Delayed Supervised with ADWIN
DS_ADWIN = ADWIN() #(delta=delta)
DS_ADWIN_alarms = []
# Delayed Supervised with PHT
DS_PHT = PHT() #(min_instances=window,delta=delta)
DS_PHT_alarms = []
## Semi-supervised
# Semi-Supervised with ADWIN
WS_ADWIN = ADWIN() #(delta=delta)
WS_ADWIN_alarms = []
# Supervised with PHT
WS_PHT = PHT() #(min_instances=window,delta=delta)
WS_PHT_alarms = []
# Delayed Supervised with ADWIN
DWS_ADWIN = ADWIN() #(delta=delta)
DWS_ADWIN_alarms = []
# Delayed Supervised with PHT
DWS_PHT = PHT() #(min_instances=window,delta=delta)
DWS_PHT_alarms = []
##### Unsupervised
# Student with ADWIN
U_ADWIN = ADWIN() #(delta=delta)
U_ADWIN_alarms = []
# Student with PHT
U_PHT = PHT() #(min_instances=window,delta=delta)
U_PHT_alarms = []
# Student with ADWIN
UR_ADWIN = ADWIN() #(delta=delta)
UR_ADWIN_alarms = []
# Student with PHT
UR_PHT = PHT() #(min_instances=window,delta=delta)
UR_PHT_alarms = []
# WRS with output
WRS_Output = HypothesisTestDetector(method="wrs", window=window, thr=pval)
WRS_Output_alarms = []
# WRS with class prob
WRS_Prob = HypothesisTestDetector(method="wrs", window=window, thr=pval)
WRS_Prob_alarms = []
# TT with output
TT_Output = HypothesisTestDetector(method="tt", window=window, thr=pval)
TT_Output_alarms = []
# TT with class prob
TT_Prob = HypothesisTestDetector(method="tt", window=window, thr=pval)
TT_Prob_alarms = []
# KS with output
KS_Output = HypothesisTestDetector(method="ks", window=window, thr=pval)
KS_Output_alarms = []
# KS with class prob
KS_Prob = HypothesisTestDetector(method="ks", window=window, thr=pval)
KS_Prob_alarms = []
Driftmodels = [
S_ADWIN, S_PHT, DS_ADWIN, DS_PHT, WS_ADWIN, WS_PHT, DWS_ADWIN, DWS_PHT,
U_ADWIN, U_PHT, UR_ADWIN, UR_PHT, WRS_Output, TT_Output, KS_Output,
WRS_Prob, TT_Prob, KS_Prob
]
Driftmodels_alarms = [
S_ADWIN_alarms, S_PHT_alarms, DS_ADWIN_alarms, DS_PHT_alarms,
WS_ADWIN_alarms, WS_PHT_alarms, DWS_ADWIN_alarms, DWS_PHT_alarms,
U_ADWIN_alarms, U_PHT_alarms, UR_ADWIN_alarms, UR_PHT_alarms,
WRS_Output_alarms, TT_Output_alarms, KS_Output_alarms, WRS_Prob_alarms,
TT_Prob_alarms, KS_Prob_alarms
]
S_driftmodels = Driftmodels[0:2]
DS_driftmodels = Driftmodels[2:4]
WS_driftmodels = Driftmodels[4:6]
DWS_driftmodels = Driftmodels[6:8]
Ustd_driftmodels = Driftmodels[8:10]
Ustdreg_driftmodels = Driftmodels[10:12]
Uoutput_driftmodels = Driftmodels[12:15]
Uprob_driftmodels = Driftmodels[15:18]
# always updated
S_clf = copy.deepcopy(stream_clf)
# always updated with delay
DS_clf = copy.deepcopy(stream_clf)
# updated immediately with some prob
WS_clf = copy.deepcopy(stream_clf)
# updated with delay with some prob
DWS_clf = copy.deepcopy(stream_clf)
# never updated
U_clf = copy.deepcopy(stream_clf)
i = ntrain
k = 0
DWS_yhat_hist = []
DS_yhat_hist = []
X_hist = []
y_hist = []
while (stream.has_more_samples()):
print(i)
#i=3000
Xi, yi = stream.next_sample()
y_hist.append(yi[0])
X_hist.append(Xi)
ext_Xi = np.concatenate([Xtrain[-10:], Xi])
U_prob = U_clf.predict_proba(ext_Xi)[-1]
U_yhat = U_clf.predict(ext_Xi)[-1]
S_yhat = S_clf.predict(ext_Xi)[-1]
WS_yhat = WS_clf.predict(ext_Xi)[-1]
DS_yhat = DS_clf.predict(ext_Xi)[-1]
DWS_yhat = DWS_clf.predict(ext_Xi)[-1]
DWS_yhat_hist.append(DWS_yhat)
DS_yhat_hist.append(DS_yhat)
if len(U_prob) < 2:
U_yhat_prob_i = U_prob[0]
elif len(U_prob) == 2:
U_yhat_prob_i = U_prob[1]
else:
U_yhat_prob_i = np.max(U_prob)
y_meta_hat_i = student_clf.predict(ext_Xi)[-1]
y_meta_prob = student_regr.predict(ext_Xi)[-1]
# Updating student model
student_clf.partial_fit(Xi, [U_yhat])
# Updating supervised model
S_clf.partial_fit(Xi, yi)
# Computing loss
S_err_i = int(yi[0] != S_yhat)
student_err_i = int(y_meta_hat_i != U_yhat)
student_prob_err_i = U_yhat_prob_i - y_meta_prob
for model in S_driftmodels:
model.add_element(S_err_i)
for model in Ustd_driftmodels:
model.add_element(student_err_i)
for model in Ustdreg_driftmodels:
model.add_element(student_prob_err_i)
for model in Uoutput_driftmodels:
model.add_element(U_yhat)
for model in Uprob_driftmodels:
model.add_element(U_yhat_prob_i)
put_i_available = np.random.binomial(1, prob_instance)
if k >= inst_delay:
DS_err_i = int(
y_hist[k - inst_delay] != DS_yhat_hist[k - inst_delay])
DS_clf.partial_fit(X_hist[k - inst_delay],
[y_hist[k - inst_delay]])
for model in DS_driftmodels:
model.add_element(DS_err_i)
if put_i_available > 0:
DWS_err_i = int(
y_hist[k - inst_delay] != DWS_yhat_hist[k - inst_delay])
DWS_clf.partial_fit(X_hist[k - inst_delay],
[y_hist[k - inst_delay]])
for model in DWS_driftmodels:
model.add_element(DWS_err_i)
if put_i_available > 0:
WS_err_i = int(yi[0] != WS_yhat)
WS_clf.partial_fit(Xi, yi)
for model in WS_driftmodels:
model.add_element(WS_err_i)
# detect changes
for j, model in enumerate(Driftmodels):
has_change = model.detected_change()
if has_change:
Driftmodels_alarms[j].append(i)
i += 1
k += 1
return ([Driftmodels_alarms, dpoints])