def get_knn_probability(k, training_data, training_data_class, test_data):
""" Use knn to compute probabilities of test_data belonging to training_data_class
k: number of nearest neighbors
training_data: training training_data instances
training_data_class: classes of each training training_data instance
test_data: test_data to classify
returns: probabilities of each input instance belonging to each class
"""
num_inputs = np.shape(test_data)[0]
unique_classes = np.unique(training_data_class)
unique_class_to_index = {}
for i in range(len(unique_classes)):
unique_class_to_index[unique_classes[i]] = i
num_classes = len(unique_classes)
if False:
print 'num_inputs =', num_inputs
print 'num_classes =', num_classes
print 'training_data_class =', training_data_class
print 'unique_classes =', unique_classes
print 'unique_class_to_index =', unique_class_to_index
exit()
print 'training_data =', training_data
print 'test_data =', test_data
probabilites = np.zeros((num_inputs,num_classes),dtype = 'f')
if USE_KD_TREE:
print 'Training kd tree'
kd_tree = KDTree(training_data)
print 'Done training kd, tree'
for n in range(num_inputs):
if USE_KD_TREE:
distances, indices = kd_tree.query(test_data[n,:], k=k)
else:
distances = np.sqrt(np.sum((training_data - test_data[n,:])**2, axis = 1))
indices = np.argsort(distances, axis = 0)
#print 'i =', test_data[n,:]
#print 'd =', distances
#print 'indices =', indices
classes = training_data_class[indices[:k]]
if False:
print 'classes =', classes
class_totals = np.zeros(num_classes)
for i in range(classes.shape[0]):
class_totals[unique_class_to_index[classes[i]]] += 1
#print 'class_totals =', class_totals
for i in range(class_totals.shape[0]):
probabilites[n,i] = class_totals[i]/classes.shape[0]
return unique_classes, probabilites
def get_knn(k, training_data_class, test_data, kd_tree):
""" Naive implementation of k nearest neighbours
k: number of nearest neighbours
training_data_class: classes of each training training_data instance
test_data: test_data to classify
returns: classes of each input instance
"""
num_inputs = np.shape(test_data)[0]
if True:
print 'k =', k
print 'num_inputs =', num_inputs
print 'training_data_class =', training_data_class[:20]
closest = np.zeros(num_inputs)
for n in range(num_inputs):
retval = kd_tree.query(test_data[n,:], k=k)
#print 'kd_tree.query returned', retval
distances, indices = retval
if k == 1:
indices = np.array([indices])
distances = np.array([distances])
#print 'k =', k
#print 'i =', test_data[n,:]
#print 'd =', distances
#print 'indices =', indices
classes = training_data_class[indices[:k]]
#print 'classes =', classes
classes = np.unique(classes)
#print 'unique classes =', classes
if len(classes) == 1:
closest[n] = np.unique(classes)
else:
#print 'x'*10
counts = np.zeros(max(classes) + 1)
for i in range(k):
counts[training_data_class[indices[i]]] += 1
#print 'counts =', counts
closest[n] = np.argmax(counts)
return closest