def cla_filter():
aux = []
resul1 = [[], [], [], [], [], [], []]
resul2 = [[], [], [], [], [], [], []]
resul3 = [[], [], [], [], [], [], []]
resul4 = [[], [], [], [], [], [], []]
resul5 = [[], [], [], [], [], [], []]
resul6 = [[], [], [], [], [], [], []]
resul7 = [[], [], [], [], [], [], []]
resul8 = [[], [], [], [], [], [], []]
resul9 = [[], [], [], [], [], [], []]
roc_m_1 = [[], [], [], [], [], [], []]
roc_m_2 = [[], [], [], [], [], [], []]
roc_m_3 = [[], [], [], [], [], [], []]
roc_m_4 = [[], [], [], [], [], [], []]
roc_m_5 = [[], [], [], [], [], [], []]
roc_m_6 = [[], [], [], [], [], [], []]
roc_m_7 = [[], [], [], [], [], [], []]
roc_m_8 = [[], [], [], [], [], [], []]
roc_m_9 = [[], [], [], [], [], [], []]
SMILaMax = [simpleMIL(), {'type': 'max'}, 'MIL max', resul1, roc_m_1]
SMILaMin = [simpleMIL(), {'type': 'min'}, 'MIL min', resul2, roc_m_2]
SMILaExt = [
simpleMIL(), {
'type': 'extreme'
}, 'MIL Extreme', resul3, roc_m_3
]
BOW_clas = [
BOW(), {
'k': 90,
'covar_type': 'diag',
'n_iter': 20
}, 'BOW', resul4, roc_m_4
]
CKNN_cla = [
CKNN(), {
'references': 3,
'citers': 5
}, 'CKNN', resul5, roc_m_5
]
maxDD_cl = [maxDD(), {}, 'DIVERSE DENSITY', resul6, roc_m_6]
EMDD_cla = [EMDD(), {}, 'EM-DD', resul7, roc_m_7]
MILB_cla = [MILBoost(), {}, 'MILBOOST', resul8, roc_m_8]
MILES_cl = [MILES(), {}, 'MILES', resul9, roc_m_9]
aux.append(SMILaMax)
# aux.append(SMILaMin)
# aux.append(SMILaExt)
aux.append(BOW_clas)
# aux.append(CKNN_cla)
aux.append(maxDD_cl)
# aux.append(EMDD_cla)
# aux.append(MILB_cla)
# aux.append(MILES_cl)
return aux
def cla_filter_ipf():
aux = []
resul1 = [[],[],[],[],[],[],[]]
roc_m_1 = [[],[],[],[],[],[],[]]
SMILaMax = [simpleMIL(),{'type': 'max'},'MIL max',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
SMILaMin = [simpleMIL(),{'type': 'min'},'MIL min',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
SMILaExt = [simpleMIL(),{'type': 'extreme'},'MIL Extreme',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
BOW_clas = [BOW(),{'k':90,'covar_type':'diag','n_iter':20},'BOW',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
CKNN_cla = [CKNN(),{'references': 3, 'citers': 5},'CKNN',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
maxDD_cl = [maxDD(),{},'DIVERSE DENSITY',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
EMDD_cla = [EMDD(),{},'EM-DD',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
MILB_cla = [MILBoost(),{},'MILBOOST',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
# MILES_cl = [MILES(),{},'MILES',copy.deepcopy(resul1),copy.deepcopy(roc_m_1)]
aux.append(SMILaMax)
aux.append(SMILaMin)
aux.append(SMILaExt)
aux.append(BOW_clas)
aux.append(CKNN_cla)
aux.append(maxDD_cl)
aux.append(EMDD_cla)
aux.append(MILB_cla)
# aux.append(MILES_cl)
return aux
def mil_cv_filter_ipf(bags_f,labels_f,folds,votacion,clasificador_):
# print('\t\t\tFiltrando...')
error = 0.01
toStop = 3
stop = True
countToStop = 0
if len(labels_f) < folds:
folds = len(labels_f)
skf = StratifiedKFold(n_splits=folds)
totalNoisyLabel = 0
while stop:
bags_f,labels_f = shuffle(bags_f, labels_f, random_state=rand.randint(0, len(labels_f)-1))
isCorrectLabel = np.ones((folds, len(labels_f)), dtype=bool)
fold = 0
for train_index, test_index in skf.split(bags_f, labels_f.reshape(len(labels_f))):
X_train = [bags_f[i] for i in train_index]
Y_train = labels_f[train_index]
# print('\t\t\t=>FOLD : '+str(fold))
try:
if len(clasificador_[1]) > 0:
clasificador_[0].fit(X_train, Y_train, **clasificador_[1])
else:
clasificador_[0].fit(bags_f, labels_f)
predictions = clasificador_[0].predict(X_train)
if (isinstance(predictions, tuple)):
predictions = predictions[0]
except:
print('Fallo, segundo intento')
try:
if len(clasificador_[1]) > 0:
clasificador_[0].fit(X_train, Y_train, **clasificador_[1])
else:
clasificador_[0].fit(bags_f, labels_f)
predictions = clasificador_[0].predict(X_train)
if (isinstance(predictions, tuple)):
predictions = predictions[0]
print('OK')
except:
print('Posible fallo en bolsa...')
try:
print('Cambiando clasificador..')
Cla_error = simpleMIL()
par_error = {'type': 'max'}
if len(par_error) > 0:
Cla_error.fit(X_train, Y_train, **par_error)
else:
Cla_error.fit(X_train, Y_train)
predictions = Cla_error.predict(X_train)
if (isinstance(predictions, tuple)):
predictions = predictions[0]
print('OK')
except:
predictions = np.ones((1, len(Y_train)), dtype=int)
predictions = predictions[0]
print('Fallo')
for l,p in enumerate(train_index):
try:
isCorrectLabel[fold][p] = (Y_train.T[0][l] == np.sign(predictions[l]))
except IndexError:
print("Fallo en ultimo indice!")
fold = fold + 1
if votacion == 'maxVotos':
noisyBags = []
for n in range(0,len(labels_f)):
aux = 0
for m in range(0,folds):
if not isCorrectLabel[m][n]:
aux = aux+1
if aux > folds/2:
noisyBags.append(n)
if votacion == 'consenso':
noisyBags = []
for n in range(0,len(labels_f)):
aux = True
for m in range(0,folds):
if aux:
if isCorrectLabel[m][n]:
aux = False
if aux:
noisyBags.append(n)
nonNoisyBags = []
cont = 0
if len(noisyBags) == 0:
for z in range(0,len(bags_f)):
nonNoisyBags.append(z)
else:
for z in range(0,len(bags_f)):
if cont < len(noisyBags) and noisyBags[cont] == z:
cont = cont + 1
else:
nonNoisyBags.append(z)
if len(noisyBags) < (len(bags_f)*error):
countToStop = countToStop + 1
else:
countToStop = 0
if countToStop == toStop:
stop = False
else:
bags_f = [bags_f[d] for d in nonNoisyBags]
labels_f = labels_f[nonNoisyBags]
if len(bags_f) < len(labels_f.reshape(len(labels_f))):
print('Número de bolsas, menor al número de etiquetas, no se puede continuar')
stop = False
#Comprobacion nueva 28/10/19
if len(labels_f) < 1:
stop = False
totalNoisyLabel+=len(noisyBags)
print('\t\t\t=>Elementos eliminados por '+clasificador_[2]+': '+str(totalNoisyLabel))
X_train_NoNy = bags_f
Y_train_NoNy = labels_f
return X_train_NoNy,Y_train_NoNy
def mil_cv_filter_ef(bags_f, labels_f, folds, votacion, num):
# print('\t\t\tFiltrando...')
if num == 1:
Clasificadores = cla_filter()
else:
Clasificadores = cla_filter2()
bags_f, labels_f = shuffle(bags_f,
labels_f,
random_state=rand.randint(0, 100))
if len(labels_f) < folds:
folds = len(labels_f)
skf = StratifiedKFold(n_splits=folds)
isCorrectLabel = np.ones((len(Clasificadores), len(labels_f)), dtype=bool)
for train_index, test_index in skf.split(bags_f,
labels_f.reshape(len(labels_f))):
X_train = [bags_f[i] for i in train_index]
Y_train = labels_f[train_index]
X_test = [bags_f[i] for i in test_index]
Y_test = labels_f[test_index]
for s, cl in enumerate(Clasificadores):
try:
if len(Clasificadores[s][1]) > 0:
Clasificadores[s][0].fit(X_train, Y_train,
**Clasificadores[s][1])
else:
Clasificadores[s][0].fit(bags_f, labels_f)
predictions = Clasificadores[s][0].predict(X_test)
if (isinstance(predictions, tuple)):
predictions = predictions[0]
except:
print('Fallo, segundo intento')
try:
if len(Clasificadores[s][1]) > 0:
Clasificadores[s][0].fit(X_train, Y_train,
**Clasificadores[s][1])
else:
Clasificadores[s][0].fit(bags_f, labels_f)
predictions = Clasificadores[s][0].predict(X_test)
if (isinstance(predictions, tuple)):
predictions = predictions[0]
print('OK')
except:
print('Posible fallo en bolsa...')
try:
if len(Clasificadores[s][1]) > 0:
Clasificadores[s][0].fit(X_train, Y_train,
**Clasificadores[s][1])
else:
Clasificadores[s][0].fit(X_train, Y_train)
predictions = Clasificadores[s][0].predict(X_test)
if (isinstance(predictions, tuple)):
predictions = predictions[0]
print('OK')
except:
try:
print('Cambiando clasificador..')
Cla_error = simpleMIL()
par_error = {'type': 'max'}
if len(par_error) > 0:
Cla_error.fit(X_train, Y_train, **par_error)
else:
Cla_error.fit(X_train, Y_train)
predictions = Cla_error.predict(X_train)
if (isinstance(predictions, tuple)):
predictions = predictions[0]
print('OK')
except:
predictions = np.ones((1, len(Y_train)), dtype=int)
print('Fallo')
for l, p in enumerate(test_index):
try:
isCorrectLabel[s][p] = (Y_test.T[0][l] == np.sign(
predictions[l]))
except IndexError:
print("Fallo en ultimo indice!")
if votacion == 'maxVotos':
noisyBags = []
for n in range(0, len(labels_f)):
aux = 0
for m in range(0, len(Clasificadores)):
if not isCorrectLabel[m][n]:
aux = aux + 1
if aux > len(Clasificadores) / 2:
noisyBags.append(n)
if votacion == 'consenso':
noisyBags = []
for n in range(0, len(labels_f)):
aux = True
for m in range(0, len(Clasificadores)):
if aux:
if isCorrectLabel[m][n]:
aux = False
if aux:
noisyBags.append(n)
nonNoisyBags = []
cont = 0
if len(noisyBags) == 0:
for z in range(0, len(bags_f)):
nonNoisyBags.append(z)
else:
for z in range(0, len(bags_f)):
if cont < len(noisyBags) and noisyBags[cont] == z:
cont = cont + 1
else:
nonNoisyBags.append(z)
print('\t\t\t=>Elementos eliminados con Filter ' + str(num + 1) + ': ' +
str(len(noisyBags)))
X_train_NoNy = [bags_f[i] for i in nonNoisyBags]
Y_train_NoNy = labels_f[nonNoisyBags]
return X_train_NoNy, Y_train_NoNy
bags, labels = shuffle(bags, labels, random_state=rand.randint(0, 100))
#Number of Folds
folds = 5
bow_classifier = BOW()
#parameters_bow = {'k':100,'covar_type':'diag','n_iter':20}
parameters_bow = {'k': 10, 'covar_type': 'diag', 'n_iter': 20}
accuracie, results_accuracie, auc, results_auc = mil_cross_val(
bags=bags,
labels=labels,
model=bow_classifier,
folds=folds,
parameters=parameters_bow)
SMILa = simpleMIL()
parameters_smil = {'type': 'max'}
#En este me funciono maxDD porque no tiene problem con parametros
accuracie, results_accuracie, auc, results_auc, elapsed = mil_cross_val(
bags=bags,
labels=labels,
model=SMILa,
folds=folds,
parameters=parameters_smil,
timer=True)
parameters_smil = {'type': 'min'}
#En este me funciono maxDD porque no tiene problem con parametros
accuracie, results_accuracie, auc, results_auc = mil_cross_val(
bags=bags,
labels=labels,
train_labels,
k=10,
covar_type='diag',
n_iter=20)
predictions = bow_classifier.predict(test_bags)
accuracie = np.average(test_labels.T == np.sign(predictions))
print '\n Accuracy: %.2f%%' % (100 * accuracie)
fpr, tpr, thresholds = metrics.roc_curve(test_labels,
predictions,
pos_label=1.)
metrics.auc(fpr, tpr)
#####################
#simpleMIL [average]#
#####################
SMILa = simpleMIL()
SMILa.fit(train_bags, train_labels, type='average')
predictions = SMILa.predict(test_bags)
accuracie = np.average(test_labels.T == np.sign(predictions))
print '\n Accuracy: %.2f%%' % (100 * accuracie)
fpr, tpr, thresholds = metrics.roc_curve(test_labels,
predictions,
pos_label=1.)
metrics.auc(fpr, tpr)
#####################
#simpleMIL [extreme]#
#####################
SMILe = simpleMIL()
SMILe.fit(train_bags, train_labels, type='extreme')
predictions = SMILe.predict(test_bags)
