def PersonWorker(person):
print('starting on person: ', str(person))
#data = 40 videos x 32 alpha(csp channel)
(X_train, y_train, X_test, y_test) = DL.loadPersonEpochDimRedu(person=person,
featureFunc = featureFunc,
)
#http://stackoverflow.com/questions/26963454/lda-ignoring-n-components => only 1 feature :(
print(np.shape(X_train))
svm = LinearSVC()
svm.fit(X_train, y_train)
y = svm.predict(X_train)
y = label_binarize(y, classes=[0, 1, 2, 3])
train_auc = UT.auc(y, y_train)
y = svm.predict(X_test)
y = label_binarize(y, classes=[0, 1, 2, 3])
test_auc = UT.auc(y, y_test)
print('person: ', person,
' - train auc: ', str(train_auc),
' - test auc: ' , str(test_auc)
)
return [train_auc, test_auc]
def PersonWorker(person):
max_k = 4#len(X_train[0])
#load data
X_train, y_train, X_test, y_test = loadPerson(
person = person,
classFunc = valClassFunc,
featureFunc = featureFunc,
plots = False
)
#init academic loop to optimize k param
k = 1
anova_filter = SelectKBest(f_regression)
lda = LinearDiscriminantAnalysis()
anova_lda = Pipeline([
('anova', anova_filter),
('lda', lda)
])
anova_lda.set_params(anova__k=k)
K_CV = KFold(n=len(X_train),
n_folds=len(X_train),
random_state=17, #fixed randomseed ensure that the sets are always the same
shuffle=False
) #leave out one validation
predictions, truths = [], []
for train_index, CV_index in K_CV: #train index here is a part of the train set
#train
anova_lda.fit(X_train[train_index], y_train[train_index])
#predict
pred = anova_lda.predict(X_train[CV_index])
#save for metric calculations
predictions.extend(pred)
truths.extend(y_train[CV_index])
#optimization metric:
best_acc = UT.accuracy(predictions, truths)
best_k = k
#now try different k values
for k in range(2,max_k):
anova_filter = SelectKBest(f_regression)
lda = LinearDiscriminantAnalysis()
anova_lda = Pipeline([
('anova', anova_filter),
('lda', lda)
])
#set k param
anova_lda.set_params(anova__k=k)
#leave one out validation to determine how good the k value performs
K_CV = KFold(n=len(X_train),
n_folds=len(X_train),
random_state=17, #fixed randomseed ensure that the sets are always the same
shuffle=False
)
predictions, truths = [], []
for train_index, CV_index in K_CV: #train index here is a part of the train set
#train
anova_lda.fit(X_train[train_index], y_train[train_index])
#predict
pred = anova_lda.predict(X_train[CV_index])
#save for metric calculations
predictions.extend(pred)
truths.extend(y_train[CV_index])
#optimization metric:
curr_acc = UT.accuracy(predictions, truths)
if curr_acc > best_acc:
best_acc = curr_acc
best_k = k
#now the k param is optimized and stored in best_k
#create classifier and train it on all train data
anova_filter = SelectKBest(f_regression)
lda = LinearDiscriminantAnalysis()
anova_lda = Pipeline([
('anova', anova_filter),
('lda', lda)
])
#set k param
anova_lda.set_params(anova__k=best_k)
anova_lda.fit(X_train, y_train)
predictions = anova_lda.predict(X_test)
acc = UT.accuracy(predictions, y_test)
(tpr,tnr,fpr,fnr) = UT.tprtnrfprfnr(predictions, y_test)
auc = UT.auc(predictions, y_test)
print('person: ', person,
' - k: ' , str(best_k),
' - acc: ', str(acc),
' - tpr: ' , str(tpr),
' - tnr: ' , str(tnr),
' - auc: ', str(auc),
'used features', anova_lda.named_steps['anova'].get_support()
)
retArr = [best_k, acc,tpr,tnr,fpr,fnr,auc]
retArr.extend(anova_lda.named_steps['anova'].get_support())
'''
print('person: ', person,
' - k: ' , str(best_k),
' - acc: ', str(acc),
'used features', getUsedFeatures(anova_lda.named_steps['anova'].get_support())
#anova_lda.named_steps['anova'].get_support()
)
classCorr = UT.classCorrect(predictions, y_test)
dimCorr = UT.dimCorrect(predictions, y_test)
returnArr = [best_k, acc ]
returnArr.extend(classCorr)
returnArr.extend(dimCorr)
returnArr.extend(anova_lda.named_steps['anova'].get_support())
return returnArr
'''
return retArr
def PersonWorker(person):
print('starting on person: ', str(person))
#data = 40 videos x 32 alpha(csp channel)
X_train, y_train, X_test, y_test, csp = DL.loadPerson(person=person,
featureFunc = featureFunc,
use_csp=False,
use_median = False
)
C = 1
clf = LinearSVC(C=C,random_state=40)
K_CV = KFold(n=len(X_train), n_folds=len(X_train), random_state=17, shuffle=False) #leave out one validation
predictions, truths = [], []
for train_index, CV_index in K_CV:
#train
clf.fit(X_train[train_index], y_train[train_index])
#predict
pred = clf.predict(X_train[CV_index])
#save for metric calculations
predictions.extend(pred)
truths.extend(y_train[CV_index])
#optimization metric:
best_metric = UT.auc(predictions, truths)
best_C = C
#try other C values
for C in [0.01,0.03,0.1,0.3,3,10]:
clf = LinearSVC(C=C,random_state=40)
K_CV = KFold(n=len(X_train), n_folds=len(X_train), random_state=17, shuffle=True) #leave out one validation
predictions, truths = [], []
for train_index, CV_index in K_CV:
#train
clf.fit(X_train[train_index], y_train[train_index])
#predict
pred = clf.predict(X_train[CV_index])
#save for metric calculations
predictions.extend(pred)
truths.extend(y_train[CV_index])
#optimization metric:
metric = UT.auc(predictions, truths)
if metric > best_metric:
best_metric = metric
best_C = C
#C param is now optimized, its value is stored in best_C
#calculate all performance metrics on testset, using the optimal classifier
clf = LinearSVC(C=C,random_state=40)
clf.fit(X_train,y_train) #fit all training data
#print("coef ", clf.coef_)
predictions = clf.predict(X_test)
acc = UT.accuracy(predictions, y_test)
(tpr,tnr,fpr,fnr) = UT.tprtnrfprfnr(predictions, y_test)
auc = UT.auc(predictions, y_test)
print('person: ', person,
' - acc: ', str(acc),
' - tpr: ' , str(tpr),
' - tnr: ' , str(tnr),
' - auc: ', str(auc)
)
return [acc,tpr,tnr,fpr,fnr,auc]
def PersonWorker(person):
print('starting on person: ', str(person))
#data = 40 videos x 32 alpha(csp channel)
X_train, y_train, X_test, y_test, csp = DL.loadPerson(person=person,
featureFunc = featureFunc,
use_median=False,
use_csp=True,
prefilter=False
)
#store weights of upper CSP channel for topoplots
csp.write_filters()
#optimize channelPairs with leave-one out validation
#prior probabilities
pos_prior = np.sum(y_train)
neg_prior = 40 - pos_prior
pos_prior /= float(40)
neg_prior /= float(40)
#academic loop start with 1 channelPair
channelPairs = 1
#filter out the channel pairs
X = np.zeros((len(X_train),channelPairs * 2,))
top_offset = channelPairs * 2 - 1
for j, k in zip(range(channelPairs), range(31,31-channelPairs,-1)):
X[:,j] = X_train[:,j]
X[:,top_offset -j] = X_train[:,k]
#LDA
lda = LinearDiscriminantAnalysis(priors=[neg_prior, pos_prior])
K_CV = KFold(n=len(X), n_folds=len(X), random_state=17, shuffle=False) #leave out one validation
predictions, truths = [], []
for train_index, CV_index in K_CV:
#train
lda = lda.fit(X[train_index], y_train[train_index])
#predict
pred = lda.predict(X[CV_index])
#save for metric calculations
predictions.extend(pred)
truths.extend(y_train[CV_index])
#optimization metric:
best_metric = UT.accuracy(predictions, truths)
best_channelPairs = channelPairs
#try other channel pairs
for channelPairs in range(2,17):
#filter out the channel pairs
X = np.zeros((len(X_train),channelPairs * 2,))
top_offset = channelPairs * 2 - 1
for j, k in zip(range(channelPairs), range(31,31-channelPairs,-1)):
X[:,j] = X_train[:,j]
X[:,top_offset -j] = X_train[:,k]
#LDA
lda = LinearDiscriminantAnalysis(priors=[neg_prior, pos_prior])
K_CV = KFold(n=len(X), n_folds=len(X), random_state=17, shuffle=True) #leave out one validation
predictions, truths = [], []
for train_index, CV_index in K_CV:
#train
lda = lda.fit(X[train_index], y_train[train_index])
#predict
pred = lda.predict(X[CV_index])
#save for metric calculations
predictions.extend(pred)
truths.extend(y_train[CV_index])
#optimization metric:
metric = UT.accuracy(predictions, truths)
if metric > best_metric:
best_metric = metric
best_channelPairs = channelPairs
#channel pairs are now optimized, its value is stored in best_channelPairs
#calculate all performance metrics on testset, using the optimal classifier
lda = LinearDiscriminantAnalysis(priors=[neg_prior, pos_prior])
lda = lda.fit(X_train,y_train) #fit all training data
predictions = lda.predict(X_test)
acc = UT.accuracy(predictions, y_test)
(tpr,tnr,fpr,fnr) = UT.tprtnrfprfnr(predictions, y_test)
auc = UT.auc(predictions, y_test)
print('person: ', person,
' - channelPairs: ', str(best_channelPairs),
' - acc: ', str(acc),
' - tpr: ' , str(tpr),
' - tnr: ' , str(tnr),
' - auc: ', str(auc)
)
return [best_channelPairs, acc,tpr,tnr,fpr,fnr,auc]
