def main_function(self, i, label_all_perm, feature_all):
"""The training data, validation data and test data are randomly splited
"""
print(f"Permutaion {i}...\n")
# KFold Cross Validation
accuracy, sensitivity, specificity, AUC = np.array([]), np.array(
[]), np.array([]), np.array([])
kf = KFold(n_splits=self.cv, shuffle=True, random_state=0)
for i, (tr_ind, te_ind) in enumerate(kf.split(feature_all)):
feature_train = feature_all[tr_ind, :]
label_train = label_all_perm[tr_ind]
feature_test = feature_all[te_ind, :]
label_test = label_all_perm[te_ind]
# normalization
prep = el_preprocessing.Preprocessing(
data_preprocess_method='StandardScaler',
data_preprocess_level='group')
feature_train, feature_test = prep.data_preprocess(
feature_train, feature_test)
# dimension reduction
if self.is_dim_reduction:
feature_train, feature_test, model_dim_reduction = el_dimreduction.pca_apply(
feature_train, feature_test, self.components)
# train
model = self.training(feature_train, label_train)
# test
pred, dec = self.testing(model, feature_test)
# Evaluating classification performances
acc, sens, spec, auc = eval_performance(label_test,
pred,
dec,
accuracy_kfold=None,
sensitivity_kfold=None,
specificity_kfold=None,
AUC_kfold=None,
verbose=0,
is_showfig=0)
accuracy = np.append(accuracy, acc)
sensitivity = np.append(sensitivity, sens)
specificity = np.append(specificity, spec)
AUC = np.append(AUC, auc)
# return np.mean(accuracy),np.mean(sensitivity), np.mean(specificity), np.mean(AUC)
return accuracy, sensitivity, specificity, AUC
def main_svc_rfe_cv(sel):
print('Training model and testing...\n')
# Load data
uid_550, feature_550, label_550 = sel._load_data(sel.dataset_our_center_550)
uid_206, feature_206, label_206 = sel._load_data(sel.dataset_206)
uid_COBRE, feature_COBRE, label_COBRE = sel._load_data(sel.data_COBRE)
uid_UCAL, feature_UCAL, label_UCAL = sel._load_data(sel.data_UCAL)
uid_all = np.concatenate([uid_550, uid_206, uid_COBRE, uid_UCAL])
feature_all = [feature_550, feature_206, feature_COBRE, feature_UCAL]
sel.label_all = [label_550, label_206, label_COBRE, label_UCAL]
name = ['550','206','COBRE','UCLA']
# Leave one site CV
n_site = len(sel.label_all)
test_index = np.array([], dtype=np.int16)
sel.decision = np.array([], dtype=np.int16)
sel.prediction = np.array([], dtype=np.int16)
sel.accuracy = np.array([], dtype=np.float16)
sel.sensitivity = np.array([], dtype=np.float16)
sel.specificity = np.array([], dtype=np.float16)
sel.AUC = np.array([], dtype=np.float16)
sel.coef = []
for i in range(n_site):
print('-'*40)
print(f'{i+1}/{n_site}: test dataset is {name[i]}...')
feature_train, label_train = feature_all.copy(), sel.label_all.copy()
feature_test, label_test = feature_train.pop(i), label_train.pop(i)
feature_train = np.concatenate(feature_train, axis=0)
label_train = np.concatenate(label_train, axis=0)
# Resampling training data
# feature_train, label_train = sel.re_sampling(feature_train, label_train)
# Normalization
prep = el_preprocessing.Preprocessing(data_preprocess_method='StandardScaler', data_preprocess_level='subject')
feature_train, feature_test = prep.data_preprocess(feature_train, feature_test)
# Dimension reduction
if sel.is_dim_reduction:
feature_train, feature_test, model_dim_reduction = el_dimreduction.pca_apply(
feature_train, feature_test, sel.components
)
print(f'After dimension reduction, the feature number is {feature_train.shape[1]}')
else:
print('No dimension reduction perfromed\n')
# Train and test
print('training and testing...\n')
model = sel.training(feature_train, label_train, sel.cv)
if sel.is_dim_reduction:
sel.coef.append(model_dim_reduction.inverse_transform(model.coef_)) # save coef
else:
sel.coef.append(clf.coef_) # save coef
pred, dec = sel.testing(model, feature_test)
sel.prediction = np.append(sel.prediction, np.array(pred))
sel.decision = np.append(sel.decision, np.array(dec))
# Evaluating classification performances
acc, sens, spec, auc = eval_performance(label_test, pred, dec,
accuracy_kfold=None, sensitivity_kfold=None, specificity_kfold=None, AUC_kfold=None,
verbose=1, is_showfig=0)
sel.accuracy = np.append(sel.accuracy, acc)
sel.sensitivity = np.append(sel.sensitivity, sens)
sel.specificity = np.append(sel.specificity, spec)
sel.AUC = np.append(sel.AUC, auc)
print(f'performances = {acc, sens, spec,auc}')
sel.label_all = np.concatenate(sel.label_all)
sel.special_result = np.concatenate( [uid_all, sel.label_all, sel.decision, sel.prediction], axis=0).reshape(4, -1).T
return sel
def main_function(self):
"""
This function is the main function.
"""
# Load data and mask
data_all, label_all, self.orig_shape, self.mask_obj, self.mask_all = self._load_nii_and_gen_label(
)
# KFold Cross Validation
self.label_test_all = np.array([], dtype=np.int16)
train_index = np.array([], dtype=np.int16)
test_index = np.array([], dtype=np.int16)
self.decision = np.array([], dtype=np.int16)
self.prediction = np.array([], dtype=np.int16)
self.accuracy = np.array([], dtype=np.float16)
self.sensitivity = np.array([], dtype=np.float16)
self.specificity = np.array([], dtype=np.float16)
self.AUC = np.array([], dtype=np.float16)
self.coef = []
kf = KFold(n_splits=self.num_of_fold_outer,
shuffle=True,
random_state=0)
for i, (tr_ind, te_ind) in enumerate(kf.split(data_all)):
print(f'------{i+1}/{self.num_of_fold_outer}...------\n')
train_index = np.int16(np.append(train_index, tr_ind))
test_index = np.int16(np.append(test_index, te_ind))
feature_train = data_all[tr_ind, :]
label_train = label_all[tr_ind]
feature_test = data_all[te_ind, :]
label_test = label_all[te_ind]
self.label_test_all = np.int16(
np.append(self.label_test_all, label_test))
# Resampling training data
feature_train, label_train = self.re_sampling(
feature_train, label_train)
# data_preprocess
prep = elprep.Preprocessing(self.data_preprocess_method,
self.data_preprocess_level)
feature_train, feature_test = prep.data_preprocess(
feature_train, feature_test)
# dimension reduction using univariate feature selection
# feature_train, feature_test, mask_selected = self.dimReduction_filter(
# feature_train, label_train, feature_test, 0.05)
# Dimension reduction using PCA
if self.is_dim_reduction:
feature_train, feature_test, model_dim_reduction = self.dimReduction_PCA(
feature_train, feature_test, self.components)
print(
f'After dimension reduction, the feature number is {feature_train.shape[1]}'
)
else:
print('No dimension reduction perfromed\n')
print(f'The feature number is {feature_train.shape[1]}')
# Train: inner feature selection using RFECV
print('Training...\n')
model, weight = self.rfeCV_training(feature_train, label_train,
self.step,
self.num_fold_of_inner_rfeCV,
self.n_jobs)
if self.is_dim_reduction:
self.coef.append(model_dim_reduction.inverse_transform(weight))
else:
self.coef.append(weight)
# Testting
print('Testting...\n')
pred, dec = self.testing(model, feature_test)
self.prediction = np.append(self.prediction, np.array(pred))
self.decision = np.append(self.decision, np.array(dec))
# Evaluating classification performances
acc, sens, spec, auc = eval_performance(
label_test,
pred,
dec,
accuracy_kfold=None,
sensitivity_kfold=None,
specificity_kfold=None,
AUC_kfold=None,
verbose=1,
is_showfig=self.is_showfig_in_each_fold)
self.accuracy = np.append(self.accuracy, acc)
self.sensitivity = np.append(self.sensitivity, sens)
self.specificity = np.append(self.specificity, spec)
self.AUC = np.append(self.AUC, auc)
# Save results and fig to local path
self.save_results()
self._weight2nii(dimension_nii_data=(61, 73, 61))
self.save_fig()
print("--" * 10 + "Done!" + "--" * 10)
return self