def classificationValidation(self,test_list, kmeans_path, kernel, C, gamma):
'''
Main classification Validation function to validate model on
:param true_val_Vector: true values of test set.
:param test_list: list of paths where test images are held.
:param KmeansName: Load the Kmeans classifier for Binerization Task.
'''
if gamma == None:
clf = SVC(C=C,kernel=kernel)
else:
clf = SVC(C=C,gamma=gamma,kernel=kernel)
print "kernel: " + kernel
print "gamma: " + str(gamma)
print "C: " + str(C)
clf.fit(self.X,self.y)
results_vector = []
y_true = []
cl=0
k_means = joblib.load(kmeans_name)
[m,num_of_clusters] = np.shape(self.X)
for path in test_list:
for item in os.listdir(path):
p = path + "/" + item
im = cv.imread(p)
fe = FeatureExtractor(im)
feature_vector = np.zeros(num_of_clusters)
raw_vector = fe.computeFeatureVector()
Km_vector = k_means.predict(raw_vector)
for k in range(len(Km_vector)):
feature_vector[Km_vector[k]] = feature_vector[Km_vector[k]] + 1
res = clf.predict(feature_vector)
# Debugging
if res[0] == 1:
print p + " is not a foram!"
if res[0] == 0:
print p + " is a foram!"
y_true.append(cl)
results_vector.append(res[0])
cl = cl + 1
print "confusion_matrix"
print confusion_matrix(y_true,results_vector)
def createClassificationTrainingFromDataset(self, dataset_name, labels_list, path_list):
'''
Creates a new training set to work on from given path list and labels.
Notice path_list and path_labels are intended to be lists of the same length. see tests in __main__ for examples.
:param dataset_name: the name of the data set
:param path_list: a list of pathes frome which the images are collected.
:param labels_list: a list of labels to use for the images collected from corresponding path. (i.e. first label correspond to first path in the path list.)
'''
base_path = "binData/"
labels = []
trainingData = []
classes = []
cl = 0
### Building the feature matrix.
for i, path in enumerate(path_list):
labels.append(labels_list[i])
print labels_list[i]
for item in os.listdir(path):
p = path + "/" + item
print p # DEBUG
im = cv.imread(p)
fe = FeatureExtractor(im)
feature_vector = fe.computeFeatureVector()
if len(trainingData) == 0:
trainingData = feature_vector
else:
np.vstack((trainingData, feature_vector))
classes.append(cl)
print "vstack Kmeans Classifier: "
print np.shape(trainingData)
classes = np.array(classes)
cl = cl + 1
### DEBUG
print np.shape(trainingData)
print np.shape(classes)
### SAVING THE DATASETS TO NPZ FORMAT
np.savez(os.path.join(base_path, dataset_name), trainingData, labels, classes)
def createKmeansTrainingDataset(self,kmeans_data, dataset_name, kmeans_name, path_list, labels_list, num_of_clusters):
'''
Create Training for Kmeans With regression.
:param: KmeansData: the training matrix obtained using createClassificationTrainingFromDataset method on the HOLDOUT set.
:param: kmeansName: the name of the Kmeans classifier to be saved and pickeled.
:param: dataset_name: the name of the NEW dataset created using the Kmeans classifier on the training. i.e. clustering the feature vector.
:param: path_list: list of paths where the training set is at.
:param: labels_list: the list of labels for the samples in the training set.
:param: num_of_clusters: the number of clusters for the Kmeans classifier.
'''
npzfile = np.load(kmeans_data)
KmeansData = npzfile['arr_0']
Kmeanslabels = npzfile['arr_1']
Kmeansclasses = npzfile['arr_2']
k_means = cluster.KMeans(n_clusters=num_of_clusters)
k_means.fit(kmeans_data)
base_path = "binData/"
labels = labels_list
trainingData = []
classes = []
cl=0
### Building the feature matrix.
for i, path in enumerate(path_list):
print labels_list[i]
for item in os.listdir(path):
p = path + "/" + item
print p # DEBUG
im = cv.imread(p)
fe = FeatureExtractor(im)
feature_vector = np.zeros(num_of_clusters)
raw_vector = fe.computeFeatureVector()
Km_vector = k_means.predict(raw_vector)
for j in range(len(Km_vector)):
feature_vector[Km_vector[j]] = feature_vector[Km_vector[j]] + 1
trainingData.append(feature_vector)
classes.append(cl)
# Here we multiply the number of POSITIVE samples in the training set so that the 'unbalanced' problem of "Foram vs. Not-Foram"
# 'becomes balanced'.
if i == 0:
print "working on positive samples"
print "Original training size: (should be 68 by 10)"
print np.shape(trainingData)
print np.shape(classes)
for k in range(9):
trainingData = np.vstack((trainingData, trainingData))
classes = np.hstack((classes,classes))
print "After Multipling Positive Samples by 8"
print np.shape(trainingData)
print np.shape(classes)
trainingData = trainingData.tolist()
classes = classes.tolist()
cl = cl + 1
### DEBUG
print "final shape: (should be 54,000~ by 10):"
print np.shape(trainingData)
### SAVING THE DATASETS TO NPZ FORMAT
joblib.dump(k_means, os.path.join(base_path, kmeans_name), compress=9)
np.savez(os.path.join(base_path, dataset_name), trainingData, labels_list, classes)
