def __init__(self,
caps_directory,
subjects_visits_tsv,
diagnoses_tsv,
dwi_map,
tissue_type,
threshold,
output_dir,
fwhm=None,
n_threads=15,
n_iterations=100,
test_size=0.3,
grid_search_folds=10,
balanced=True,
c_range=np.logspace(-6, 2, 17),
splits_indices=None):
self._output_dir = output_dir # DTI_SVM folder to contain the outputf
self._n_threads = n_threads # number of threds is to use
self._n_iterations = n_iterations # number of iteration to train the training set
self._test_size = test_size # split the dataset into training and test dataset
self._grid_search_folds = grid_search_folds
self._balanced = balanced
self._c_range = c_range # the hyperparameteWr to tune
self._splits_indices = splits_indices
self._input = DWIVoxelBasedInput(caps_directory, subjects_visits_tsv,
diagnoses_tsv, dwi_map, tissue_type,
threshold, fwhm)
self._validation = None
self._algorithm = None
class DWI_VB_RepHoldOut_DualSVM(MLWorkflow):
"""
This is a class for DWI voxel based features classification. Currently, only DTI features have been included,
other model's feautures will be addded in the future. e.g. NODDI maps
"""
def __init__(self,
caps_directory,
subjects_visits_tsv,
diagnoses_tsv,
dwi_map,
tissue_type,
threshold,
output_dir,
fwhm=None,
n_threads=15,
n_iterations=100,
test_size=0.3,
grid_search_folds=10,
balanced=True,
c_range=np.logspace(-6, 2, 17),
splits_indices=None):
self._output_dir = output_dir # DTI_SVM folder to contain the outputf
self._n_threads = n_threads # number of threds is to use
self._n_iterations = n_iterations # number of iteration to train the training set
self._test_size = test_size # split the dataset into training and test dataset
self._grid_search_folds = grid_search_folds
self._balanced = balanced
self._c_range = c_range # the hyperparameteWr to tune
self._splits_indices = splits_indices
self._input = DWIVoxelBasedInput(caps_directory, subjects_visits_tsv,
diagnoses_tsv, dwi_map, tissue_type,
threshold, fwhm)
self._validation = None
self._algorithm = None
def run(self):
x = self._input.get_x()
y = self._input.get_y()
kernel = self._input.get_kernel() # what is a kernel here???
if y[0] == 0:
print 'The first label of diagnose is 0, it means the voxels with negative coefficients in the weight image are more likely to be classified as the first label in the diagnose tsv'
else:
print 'The first label of diagnose is 1, it means the voxels with positive coefficients in the weight image are more likely to be classified as the second label in the diagnose tsv'
self._algorithm = algorithm.DualSVMAlgorithm(
kernel,
y,
balanced=self._balanced,
grid_search_folds=self._grid_search_folds,
c_range=self._c_range,
n_threads=self._n_threads)
self._validation = validation.RepeatedHoldOut(
self._algorithm,
n_iterations=self._n_iterations,
test_size=self._test_size)
classifier, best_params, results = self._validation.validate(
y, n_threads=self._n_threads, splits_indices=self._splits_indices)
classifier_dir = path.join(self._output_dir, 'classifier')
if not path.exists(classifier_dir):
os.makedirs(classifier_dir)
self._algorithm.save_classifier(classifier, classifier_dir)
self._algorithm.save_parameters(best_params, classifier_dir)
weights = self._algorithm.save_weights(classifier, x, classifier_dir)
self._input.save_weights_as_nifti(weights, classifier_dir)
self._validation.save_results(self._output_dir)
class DWI_VB_RepHoldOut_DualSVM_FeatureSelectionNonNested(MLWorkflow):
"""
This is a class for DWI voxel based features classification. Currently, only DTI features have been included,
other model's feautures will be addded in the future. e.g. NODDI maps
"""
def __init__(self,
caps_directory,
subjects_visits_tsv,
diagnoses_tsv,
dwi_map,
tissue_type,
threshold,
output_dir,
fwhm=None,
n_threads=15,
n_iterations=100,
test_size=0.3,
grid_search_folds=10,
balanced=True,
c_range=np.logspace(-6, 2, 17),
splits_indices=None,
top_k=50,
feature_selection_method=None,
feature_rescaling_method=None):
self._output_dir = output_dir # DTI_SVM folder to contain the outputf
self._n_threads = n_threads # number of threds is to use
self._n_iterations = n_iterations # number of iteration to train the training set
self._test_size = test_size # split the dataset into training and test dataset
self._grid_search_folds = grid_search_folds
self._balanced = balanced
self._c_range = c_range # the hyperparameteWr to tune
self._splits_indices = splits_indices
self._top_k = top_k
self._feature_selection_method = feature_selection_method
self._feature_rescaling_method = feature_rescaling_method
self._input = DWIVoxelBasedInput(caps_directory, subjects_visits_tsv,
diagnoses_tsv, dwi_map, tissue_type,
threshold, fwhm)
self._validation = None
self._algorithm = None
def run(self):
x = self._input.get_x()
y = self._input.get_y()
### feature rescaling
if self._feature_rescaling_method == 'zscore':
selector = StandardScaler(with_std=True)
selector.fit(x)
x = selector.transform(x)
elif self._feature_rescaling_method == 'minmax':
selector = MinMaxScaler()
selector.fit(x)
x = selector.transform(x)
elif self._feature_rescaling_method == None:
pass
else:
raise Exception('Method has not been implemented')
## feature selection for all the data
if self._feature_selection_method == 'ANOVA':
selector = SelectPercentile(f_classif, percentile=self._top_k)
selector.fit(x, y)
elif self._feature_selection_method == 'RF':
clf = RandomForestClassifier(n_estimators=250,
random_state=0,
n_jobs=-1)
clf.fit(x, y)
selector = SelectFromModel(clf, threshold=self._top_k)
selector.fit(x, y)
elif self._feature_selection_method == 'PCA':
selector = PCA(n_components=self._top_k)
selector.fit(x)
elif self._feature_selection_method == 'RFE':
svc = SVR(kernel="linear")
selector = RFE(estimator=svc,
n_features_to_select=int(0.01 * self._top_k *
x.shape[1]),
step=0.5)
selector.fit(x, y)
else:
print('Method has not been implemented')
x_after_fs = selector.transform(x)
kernel = utils.gram_matrix_linear(x_after_fs)
if y[0] == 0:
print 'The first label of diagnose is 0, it means the voxels with negative coefficients in the weight image are more likely to be classified as the first label in the diagnose tsv'
else:
print 'The first label of diagnose is 1, it means the voxels with positive coefficients in the weight image are more likely to be classified as the second label in the diagnose tsv'
self._algorithm = algorithm_dwi.DualSVMAlgorithm(
kernel,
y,
balanced=self._balanced,
grid_search_folds=self._grid_search_folds,
c_range=self._c_range,
n_threads=self._n_threads)
self._validation = validation_dwi.RepeatedHoldOut(
self._algorithm,
n_iterations=self._n_iterations,
test_size=self._test_size)
classifier, best_params, results = self._validation.validate(
y, n_threads=self._n_threads, splits_indices=self._splits_indices)
classifier_dir = path.join(self._output_dir, 'classifier')
if not path.exists(classifier_dir):
os.makedirs(classifier_dir)
self._algorithm.save_classifier(classifier, classifier_dir)
self._algorithm.save_parameters(best_params, classifier_dir)
weights = self._algorithm.save_weights(classifier, x, classifier_dir)
self._input.save_weights_as_nifti(weights, classifier_dir)
self._validation.save_results(self._output_dir)