def trainSVM_models_1vs1(model_ubm, tr_data, modeldir_patient_svm, ParametersSVM):
"""
This function will produce the 1vs1 multiclass SVM models
It will first create the supervectors from the data and then feed them to make_svm_model for selecting c parameter
and producing a SVM model with best parameter
:param model_ubm: UBM model to adapt the sequences.
:param tr_data: training data
:param modeldir_patient_svm: model directory to save the trained SVM model
:param ParametersSVM: parameters to be used for selecting relevance factor, and supervector creation.
:return: nothing
"""
for i in range(1,5,1):
for j in range(i+1,5,1):
data1 = getattr(tr_data, str('data_g' + str(i)))
data2 = getattr(tr_data, str('data_g' + str(j)))
r_final = 5
# Select what ratio of each class should be present for training dataset.
# Make supervectors
if ParametersSVM.equal_number_of_sequences == 1:
if max(data1[0]) < max(data2[0]):
num_of_seq = int(max(data1[0]))
else:
num_of_seq = int(max(data2[0]))
super_vector1 = make_super_vector( data1[0], data1[1], r_final, num_of_seq, model_ubm, ParametersSVM)
super_vector2 = make_super_vector( data2[0], data2[1], r_final, num_of_seq, model_ubm, ParametersSVM)
else:
num_of_seq = int(max(data1[0]))
super_vector1 = make_super_vector( data1[0], data1[1], r_final, num_of_seq, model_ubm, ParametersSVM)
num_of_seq = int(max(data2[0]))
super_vector2 = make_super_vector( data2[0], data2[1], r_final, num_of_seq, model_ubm, ParametersSVM)
fulltrset_sv = np.concatenate([super_vector1, super_vector2], axis=1)
labels = np.concatenate([np.ones(np.size(super_vector1, 1)), np.zeros(np.size(super_vector2, 1))])
del super_vector1, super_vector2
remove_NaNs(fulltrset_sv, labels)
print('Building SVM models for grade', i, 'vs', j)
save_file_name_svm = path.join(modeldir_patient_svm, ('SVM_model_' + str(i) + '_' + str(j)))
model_svm = make_svm_model(fulltrset_sv, labels)
model_svm.rel_factor=r_final
fid = open(save_file_name_svm,'wb')
pickle.dump(model_svm, fid)
fid.close()
def classify_1vs1(data, model_ubm, modeldir_patient_svm, ParametersSVM):
prob_epoch = np.zeros((6, int(max(data[0]))))
dec_class = np.zeros((6, int(max(data[0]))))
a = 0
for i in range(1, 5, 1):
for j in range(i + 1, 5, 1):
load_file_name_svm = os.path.join(
modeldir_patient_svm, ('SVM_model_' + str(i) + '_' + str(j)))
fid = open(load_file_name_svm, 'rb')
model_svm = pickle.load(fid)
fid.close()
super_vectors = make_super_vector(data[0],
data[1],
model_svm.rel_factor,
int(max(data[0])),
model_ubm,
ParametersSVM,
test_flag=1)
Dzero = (np.where(super_vectors[1] == 0))
dec_class[a, :] = model_svm.predict(super_vectors[0].T)
prob_epoch[a, :] = model_svm.predict_proba(super_vectors[0].T)[:,
1]
prob_epoch[a, Dzero] = np.nan
dec_class[a, Dzero] = 0
a = a + 1
if ParametersSVM.prob_based_dec == 1:
final_class_epoch = max(prob_epoch)
else:
final_class_epoch = get_class(dec_class)
return final_class_epoch, dec_class, prob_epoch
def classify_1vs1(data, model_ubm, modeldir_patient_svm, ParametersSVM):
prob_epoch = np.zeros((6,int(max(data[0]))))
dec_class = np.zeros((6,int(max(data[0]))))
a=0
for i in range(1,5,1):
for j in range(i+1,5,1):
load_file_name_svm = os.path.join(modeldir_patient_svm, ('SVM_model_' + str(i) + '_' + str(j)))
fid = open(load_file_name_svm,'rb')
model_svm = pickle.load(fid)
fid.close()
super_vectors = make_super_vector(data[0], data[1], model_svm.rel_factor, int(max(data[0])), model_ubm, ParametersSVM, test_flag=1)
Dzero=(np.where(super_vectors[1]==0))
dec_class[a,:] = model_svm.predict(super_vectors[0].T)
prob_epoch[a,:] = model_svm.predict_proba(super_vectors[0].T)[:,1]
prob_epoch[a, Dzero] = np.nan
dec_class[a, Dzero] = 0
a=a+1
if ParametersSVM.prob_based_dec==1:
final_class_epoch = max(prob_epoch)
else:
final_class_epoch = get_class(dec_class)
return final_class_epoch, dec_class, prob_epoch
def trainSVM_models_crammer_singer(model_ubm, tr_data, modeldir_patient_svm,
ParametersSVM):
"""
This function will train a multiclass SVM based on crammers_singer implementation.
It will give all the data from all the grades and then finally train a SVM model.
:param model_ubm: UBM model to adapt the sequences.
:param tr_data: training data
:param modeldir_patient_svm: model directory to save the trained SVM model
:param ParametersSVM: parameters to be used for selecting relevance factor, and supervector creation.
:return: nothing
"""
targets = np.arange(1, ParametersSVM.num_of_classes + 1, 1)
# Concatenate the data of all classes
for lv in range(0, ParametersSVM.num_of_classes, 1):
data_temp = getattr(tr_data, str('data_g' + str(targets[lv])))
if lv == 0:
data = np.asarray(data_temp)
else:
data[0] = np.concatenate([data[0], data_temp[0] + np.max(data[0])],
axis=0)
data[1] = np.concatenate([data[1], data_temp[1]], axis=1)
num_of_seq = int(max(data[0]))
r_final = 5
fulltrset_sv = make_super_vector(data[0], data[1], r_final, num_of_seq,
model_ubm, ParametersSVM)
labels = np.concatenate([
np.ones((int(max(tr_data.data_g1[0])))), 2 * np.ones(
(int(max(tr_data.data_g2[0])))), 3 * np.ones(
(int(max(tr_data.data_g3[0])))), 4 * np.ones(
(int(max(tr_data.data_g4[0]))))
])
# Do k-fold internal cross validation to select best C parameter
parameters = [{'C': [1, 10, 100, 1000]}]
kf = StratifiedKFold(labels, n_folds=3)
gscv = GridSearchCV(LinearSVC(multi_class='crammer_singer'),
parameters,
cv=kf)
model_svm = gscv.fit(fulltrset_sv.T, labels)
print('best score =', gscv.best_score_, 'with C ', gscv.best_params_)
save_file_name_svm = path.join(modeldir_patient_svm, ('SVM_model_CS'))
model_svm.rel_factor = r_final
fid = open(save_file_name_svm, 'wb')
pickle.dump(model_svm, fid)
fid.close()
def classify_builtin_1vs1(data, model_ubm, modeldir_patient_svm, ParametersSVM):
result=result_class(4,int(max(data[0])))
load_file_name_svm = os.path.join(modeldir_patient_svm, ('SVM_model_CS'))
fid = open(load_file_name_svm,'rb')
model_svm = pickle.load(fid)
fid.close()
super_vectors = make_super_vector(data[0], data[1], model_svm.rel_factor, int(max(data[0])), model_ubm, ParametersSVM, test_flag=1)
result.final_dec_epoch = model_svm.predict(super_vectors[0].T)
result.dis_epoch = model_svm.decision_function(super_vectors[0].T)
Dzero=(np.where(super_vectors[1]==0))[0]
result.final_dec_epoch[Dzero]=0
prob_epoch=0
return result
def classify_crammer_singer(data, model_ubm, modeldir_patient_svm, ParametersSVM):
load_file_name_svm = os.path.join(modeldir_patient_svm, ('SVM_model_CS'))
fid = open(load_file_name_svm,'rb')
model_svm = pickle.load(fid)
fid.close()
super_vectors = make_super_vector(data[0], data[1], model_svm.rel_factor, int(max(data[0])), model_ubm, ParametersSVM, test_flag=1)
final_class_epoch = model_svm.predict(super_vectors[0].T)
dec_dist = model_svm.decision_function(super_vectors[0].T)
Dzero=(np.where(super_vectors[1]==0))
final_class_epoch[Dzero]=0
prob_epoch=0
return final_class_epoch, dec_dist, prob_epoch
def classify_1vsRest(data, model_ubm, modeldir_patient_svm, ParametersSVM):
result = result_class(4, int(max(data[0])))
a = 0
for i in range(1, 5, 1):
load_file_name_svm = os.path.join(modeldir_patient_svm,
('SVM_model_' + str(i) + '_rest'))
fid = open(load_file_name_svm, 'rb')
model_svm = pickle.load(fid)
fid.close()
super_vectors = make_super_vector(data[0],
data[1],
model_svm.rel_factor,
int(max(data[0])),
model_ubm,
ParametersSVM,
test_flag=1)
Dzero = np.where(super_vectors[1] == 0)[0]
result.dec_epoch[a, :] = model_svm.predict(super_vectors[0].T)
result.dis_epoch[a, :] = model_svm.decision_function(
super_vectors[0].T)
result.prob_epoch[a, :] = model_svm.predict_proba(
super_vectors[0].T)[:, 1]
result.dec_epoch[a, np.where(result.dec_epoch == 0)] = 5
result.dec_epoch[a, np.where(result.dec_epoch == 1)] = i
result.prob_epoch[a, Dzero] = np.nan
result.dec_epoch[a, Dzero] = 0
result.dis_epoch[a, Dzero] = np.nan
a = a + 1
if ParametersSVM.final_dec is 'prob':
result.final_dec_epoch = np.argmax(result.prob_epoch, axis=0) + 1
result.final_dec_epoch[Dzero] = 0
elif ParametersSVM.final_dec is 'distance':
result.final_dec_epoch = np.argmin(result.dis_epoch, axis=0) + 1
result.final_dec_epoch[Dzero] = 0
return result
def trainSVM_models_builtin_1vs1(model_ubm, tr_data, modeldir_patient_svm, ParametersSVM):
"""
This function will train a multiclass SVM based on crammers_singer implementation.
It will give all the data from all the grades and then finally train a SVM model.
:param model_ubm: UBM model to adapt the sequences.
:param tr_data: training data
:param modeldir_patient_svm: model directory to save the trained SVM model
:param ParametersSVM: parameters to be used for selecting relevance factor, and supervector creation.
:return: nothing
"""
targets = np.arange(1,ParametersSVM.num_of_classes+1,1)
# Concatenate the data of all classes
for lv in range(0,ParametersSVM.num_of_classes,1):
data_temp=getattr(tr_data, str('data_g' + str(targets[lv])))
if lv == 0:
data=np.asarray(data_temp)
else:
data[0] = np.concatenate([data[0], data_temp[0]+np.max(data[0])], axis=0)
data[1] = np.concatenate([data[1], data_temp[1]], axis=1)
num_of_seq = int(max(data[0]))
r_final=5
fulltrset_sv = make_super_vector(data[0], data[1], r_final, num_of_seq, model_ubm, ParametersSVM)
labels = np.concatenate([np.ones((int(max(tr_data.data_g1[0])))), 2*np.ones((int(max(tr_data.data_g2[0])))),
3*np.ones((int(max(tr_data.data_g3[0])))), 4*np.ones((int(max(tr_data.data_g4[0]))))])
# Do k-fold internal cross validation to select best C parameter
parameters = [{'C': [1, 10, 100, 1000]}]
kf = StratifiedKFold(labels, n_folds=3)
gscv = GridSearchCV(SVC(probability=True), parameters, cv=kf)
model_svm = gscv.fit(fulltrset_sv.T, labels)
print('best score =', gscv.best_score_ , 'with C ', gscv.best_params_)
save_file_name_svm = path.join(modeldir_patient_svm, ('SVM_model_CS'))
model_svm.rel_factor=r_final
fid = open(save_file_name_svm,'wb')
pickle.dump(model_svm, fid)
fid.close()
def classify_builtin_1vs1(data, model_ubm, modeldir_patient_svm,
ParametersSVM):
result = result_class(4, int(max(data[0])))
load_file_name_svm = os.path.join(modeldir_patient_svm, ('SVM_model_CS'))
fid = open(load_file_name_svm, 'rb')
model_svm = pickle.load(fid)
fid.close()
super_vectors = make_super_vector(data[0],
data[1],
model_svm.rel_factor,
int(max(data[0])),
model_ubm,
ParametersSVM,
test_flag=1)
result.final_dec_epoch = model_svm.predict(super_vectors[0].T)
result.dis_epoch = model_svm.decision_function(super_vectors[0].T)
Dzero = (np.where(super_vectors[1] == 0))[0]
result.final_dec_epoch[Dzero] = 0
prob_epoch = 0
return result
def classify_crammer_singer(data, model_ubm, modeldir_patient_svm,
ParametersSVM):
load_file_name_svm = os.path.join(modeldir_patient_svm, ('SVM_model_CS'))
fid = open(load_file_name_svm, 'rb')
model_svm = pickle.load(fid)
fid.close()
super_vectors = make_super_vector(data[0],
data[1],
model_svm.rel_factor,
int(max(data[0])),
model_ubm,
ParametersSVM,
test_flag=1)
final_class_epoch = model_svm.predict(super_vectors[0].T)
dec_dist = model_svm.decision_function(super_vectors[0].T)
Dzero = (np.where(super_vectors[1] == 0))
final_class_epoch[Dzero] = 0
prob_epoch = 0
return final_class_epoch, dec_dist, prob_epoch
def classify_1vsRest(data, model_ubm, modeldir_patient_svm, ParametersSVM):
result=result_class(4,int(max(data[0])))
a=0
for i in range(1,5,1):
load_file_name_svm = os.path.join(modeldir_patient_svm, ('SVM_model_' + str(i) + '_rest'))
fid = open(load_file_name_svm,'rb')
model_svm = pickle.load(fid)
fid.close()
super_vectors = make_super_vector(data[0], data[1], model_svm.rel_factor, int(max(data[0])), model_ubm, ParametersSVM, test_flag=1)
Dzero = np.where(super_vectors[1]==0)[0]
result.dec_epoch[a,:] = model_svm.predict(super_vectors[0].T)
result.dis_epoch[a,:] = model_svm.decision_function(super_vectors[0].T)
result.prob_epoch[a,:] = model_svm.predict_proba(super_vectors[0].T)[:,1]
result.dec_epoch[a,np.where(result.dec_epoch==0)]=5
result.dec_epoch[a,np.where(result.dec_epoch==1)]=i
result.prob_epoch[a, Dzero] = np.nan
result.dec_epoch[a, Dzero] = 0
result.dis_epoch[a, Dzero] = np.nan
a=a+1
if ParametersSVM.final_dec is 'prob':
result.final_dec_epoch = np.argmax(result.prob_epoch, axis=0)+1
result.final_dec_epoch[Dzero]=0
elif ParametersSVM.final_dec is 'distance':
result.final_dec_epoch = np.argmin(result.dis_epoch, axis=0)+1
result.final_dec_epoch[Dzero]=0
return result
def trainSVM_models_1vsRest(model_ubm, tr_data, modeldir_patient_svm,
ParametersSVM):
"""
This function will produce the 1vsRest SVM models. It will first create the supervectors from the data and
then feed them to make_svm_model for selecting c parameter and producing a SVM model with best parameter
:param model_ubm: UBM model to adapt the sequences.
:param tr_data: training data
:param modeldir_patient_svm: model directory to save the trained SVM model
:param ParametersSVM: parameters to be used for selecting relevance factor, and supervector creation.
:return: nothing
"""
targets = np.arange(1, ParametersSVM.num_of_classes + 1, 1)
for i in range(1, 5, 1):
data1 = getattr(tr_data, str('data_g' + str(i)))
# Concatenate the data of all classes other then the class(i)
classes_rest = np.where(targets != i)[0]
for lv in range(0, np.size(classes_rest), 1):
data2_temp = getattr(
tr_data, str('data_g' + str(targets[classes_rest[lv]])))
if lv == 0:
data2 = np.asarray(data2_temp)
else:
data2[0] = np.concatenate(
[data2[0], data2_temp[0] + np.max(data2[0])], axis=0)
data2[1] = np.concatenate([data2[1], data2_temp[1]], axis=1)
del data2_temp
r_final = 5
# Select what ratio of each class should be present for training dataset.
# Make supervectors
if ParametersSVM.equal_number_of_sequences == 1:
if max(data1[0]) < max(data2[0]):
num_of_seq = int(max(data1[0]))
else:
num_of_seq = int(max(data2[0]))
data_shuffle = 1
super_vector1 = make_super_vector(
data1[0],
data1[1],
r_final,
num_of_seq,
model_ubm,
ParametersSVM,
data_shuffle,
test_flag=0,
)
super_vector2 = make_super_vector(data2[0], data2[1], r_final,
num_of_seq, model_ubm,
ParametersSVM)
else:
num_of_seq = int(max(data1[0]))
super_vector1 = make_super_vector(data1[0], data1[1], r_final,
num_of_seq, model_ubm,
ParametersSVM)
num_of_seq = int(max(data2[0]))
super_vector2 = make_super_vector(data2[0], data2[1], r_final,
num_of_seq, model_ubm,
ParametersSVM)
fulltrset_sv = np.concatenate([super_vector1, super_vector2], axis=1)
labels = np.concatenate([
np.ones(np.size(super_vector1, 1)),
np.zeros(np.size(super_vector2, 1))
])
del super_vector1, super_vector2
remove_NaNs(fulltrset_sv, labels)
print('Building SVM models for grade', i, 'vs_rest')
save_file_name_svm = path.join(modeldir_patient_svm,
('SVM_model_' + str(i) + '_rest'))
model_svm = make_svm_model_1vsRest(fulltrset_sv, labels)
model_svm.rel_factor = r_final
fid = open(save_file_name_svm, 'wb')
pickle.dump(model_svm, fid)
fid.close()
def trainSVM_models_1vs1(model_ubm, tr_data, modeldir_patient_svm,
ParametersSVM):
"""
This function will produce the 1vs1 multiclass SVM models
It will first create the supervectors from the data and then feed them to make_svm_model for selecting c parameter
and producing a SVM model with best parameter
:param model_ubm: UBM model to adapt the sequences.
:param tr_data: training data
:param modeldir_patient_svm: model directory to save the trained SVM model
:param ParametersSVM: parameters to be used for selecting relevance factor, and supervector creation.
:return: nothing
"""
for i in range(1, 5, 1):
for j in range(i + 1, 5, 1):
data1 = getattr(tr_data, str('data_g' + str(i)))
data2 = getattr(tr_data, str('data_g' + str(j)))
r_final = 5
# Select what ratio of each class should be present for training dataset.
# Make supervectors
if ParametersSVM.equal_number_of_sequences == 1:
if max(data1[0]) < max(data2[0]):
num_of_seq = int(max(data1[0]))
else:
num_of_seq = int(max(data2[0]))
super_vector1 = make_super_vector(data1[0], data1[1], r_final,
num_of_seq, model_ubm,
ParametersSVM)
super_vector2 = make_super_vector(data2[0], data2[1], r_final,
num_of_seq, model_ubm,
ParametersSVM)
else:
num_of_seq = int(max(data1[0]))
super_vector1 = make_super_vector(data1[0], data1[1], r_final,
num_of_seq, model_ubm,
ParametersSVM)
num_of_seq = int(max(data2[0]))
super_vector2 = make_super_vector(data2[0], data2[1], r_final,
num_of_seq, model_ubm,
ParametersSVM)
fulltrset_sv = np.concatenate([super_vector1, super_vector2],
axis=1)
labels = np.concatenate([
np.ones(np.size(super_vector1, 1)),
np.zeros(np.size(super_vector2, 1))
])
del super_vector1, super_vector2
remove_NaNs(fulltrset_sv, labels)
print('Building SVM models for grade', i, 'vs', j)
save_file_name_svm = path.join(
modeldir_patient_svm, ('SVM_model_' + str(i) + '_' + str(j)))
model_svm = make_svm_model(fulltrset_sv, labels)
model_svm.rel_factor = r_final
fid = open(save_file_name_svm, 'wb')
pickle.dump(model_svm, fid)
fid.close()
def trainSVM_models_1vsRest(model_ubm, tr_data, modeldir_patient_svm, ParametersSVM):
"""
This function will produce the 1vsRest SVM models. It will first create the supervectors from the data and
then feed them to make_svm_model for selecting c parameter and producing a SVM model with best parameter
:param model_ubm: UBM model to adapt the sequences.
:param tr_data: training data
:param modeldir_patient_svm: model directory to save the trained SVM model
:param ParametersSVM: parameters to be used for selecting relevance factor, and supervector creation.
:return: nothing
"""
targets = np.arange(1,ParametersSVM.num_of_classes+1,1)
for i in range(1,5,1):
data1 = getattr(tr_data, str('data_g' + str(i)))
# Concatenate the data of all classes other then the class(i)
classes_rest = np.where(targets != i)[0]
for lv in range(0,np.size(classes_rest),1):
data2_temp=getattr(tr_data, str('data_g' + str(targets[classes_rest[lv]])))
if lv == 0:
data2=np.asarray(data2_temp)
else:
data2[0] = np.concatenate([data2[0], data2_temp[0]+np.max(data2[0])], axis=0)
data2[1] = np.concatenate([data2[1], data2_temp[1]], axis=1)
del data2_temp
r_final = 5
# Select what ratio of each class should be present for training dataset.
# Make supervectors
if ParametersSVM.equal_number_of_sequences == 1:
if max(data1[0]) < max(data2[0]):
num_of_seq = int(max(data1[0]))
else:
num_of_seq = int(max(data2[0]))
data_shuffle=1
super_vector1 = make_super_vector(data1[0], data1[1], r_final, num_of_seq, model_ubm, ParametersSVM, data_shuffle, test_flag=0,)
super_vector2 = make_super_vector(data2[0], data2[1], r_final, num_of_seq, model_ubm, ParametersSVM)
else:
num_of_seq = int(max(data1[0]))
super_vector1 = make_super_vector(data1[0], data1[1], r_final, num_of_seq, model_ubm, ParametersSVM)
num_of_seq = int(max(data2[0]))
super_vector2 = make_super_vector(data2[0], data2[1], r_final, num_of_seq, model_ubm, ParametersSVM)
fulltrset_sv = np.concatenate([super_vector1, super_vector2], axis=1)
labels = np.concatenate([np.ones(np.size(super_vector1, 1)), np.zeros(np.size(super_vector2, 1))])
del super_vector1, super_vector2
remove_NaNs(fulltrset_sv, labels)
print('Building SVM models for grade', i, 'vs_rest')
save_file_name_svm = path.join(modeldir_patient_svm, ('SVM_model_' + str(i) + '_rest'))
model_svm = make_svm_model_1vsRest(fulltrset_sv, labels)
model_svm.rel_factor = r_final
fid = open(save_file_name_svm,'wb')
pickle.dump(model_svm, fid)
fid.close()