def validate(in_path,
in_labels_path,
val_path,
val_labels_path,
c_exponent_range,
gamma_exponent_range,
image_size,
cell_size,
block_size,
bin_count,
extension=""):
'''
Performs cross validation for ranges of the classifier parameters.
Arguments:
in_path -- The path to the input images.
labels_path -- The path to the class labels.
fold_count -- The number of folds for the cross validation.
c_exponent_range -- The range of values for n where parameter C=10^n.
gamma_exponent_range -- The range of values for n where parameter gamma=10^n.
image_size -- The size to which to resize the images.
cell_size -- The cell size for HOG feature calculation.
block_size -- The block size for HOG feature calculation.
bin_count -- The number of bins for HOG feature calculation.
extension -- The file type extension of the images to load.
'''
_, in_images, in_labels = inp.load_images(in_path,
in_labels_path,
image_size,
extension=extension)
in_features = inp.compute_features(in_images, cell_size, block_size,
bin_count)
_, val_images, val_labels = inp.load_images(val_path,
val_labels_path,
image_size,
extension=extension)
val_features = inp.compute_features(val_images, cell_size, block_size,
bin_count)
best_parameters = (None, None)
best_error = float('inf')
for c_exponent in range(c_exponent_range[0], c_exponent_range[1] + 1):
for gamma_exponent in range(gamma_exponent_range[0],
gamma_exponent_range[1] + 1):
c = 10**c_exponent
gamma = 10**gamma_exponent
error = validate_once(in_features, in_labels, val_features,
val_labels, c, gamma)
if error < best_error:
best_error = error
best_parameters = (c, gamma)
print(best_parameters, best_error)
def train(in_path, labels_path, model_path, image_size, c, gamma, cell_size, block_size, bin_count, extension = ""):
'''
Trains and writes to file a multi-class classifier.
Arguments:
in_path -- The path to the directory containing the dataset's images.
labels_path -- The path to the JSON file containing label information.
model_path -- The path to the file where to store the pickled model.
image_size -- The size to which to resize the images.
c -- The C parameter for the SVM classifier.
gamma -- The gamma parameter for the SVM classifier.
cell_size -- The cell size for HOG feature calculation.
block_size -- The block size for HOG feature calculation.
bin_count -- The number of bins for HOG feature calculation.
extension -- The file type extension of the images to load.
'''
_, images, labels = inp.load_images(in_path, labels_path, image_size, extension = extension)
features = inp.compute_features(images, cell_size, block_size, bin_count)
classifier = model.train(features, labels, c, gamma)
print("Error: ", model.test(features, labels, classifier).error)
classifier.hog_image_size = image_size
classifier.hog_cell_size = cell_size
classifier.hog_block_size = block_size
classifier.hog_bin_count = bin_count
with open(model_path, 'wb') as model_file: pickle.dump(classifier, model_file)
def diff(in_path,
labels_path,
model_path1,
model_path2,
report_path,
extension=""):
'''
Outputs the difference in error of two classifier models on a given dataset.
Arguments:
in_path -- The path to the directory containing the dataset's images.
labels_path -- The path to the JSON file containing label information.
model_path1 -- The path to the file containing the first pickled model.
model_path2 -- The path to the file containing the second pickled model.
report_path -- The path to the file where to store the error report.
extension -- The file type extension of the images to load.
'''
with open(model_path1, 'rb') as model_file:
classifier1 = pickle.load(model_file)
with open(model_path2, 'rb') as model_file:
classifier2 = pickle.load(model_file)
_, images, labels = inp.load_images(in_path,
labels_path,
classifier1.hog_image_size,
extension=extension)
features = inp.compute_features(images, classifier1.hog_cell_size,
classifier1.hog_block_size,
classifier1.hog_bin_count)
report1 = model.test(features, labels, classifier1)
report2 = model.test(features, labels, classifier2)
report_diff = report1 - report2
print("Error: ", report_diff.error)
print()
print(
report_diff.get_sparse_string(report_diff.confusion_matrix_relative,
0.01))
with open(report_path, 'w') as report_file:
report_diff.dump(report_file)