def load_test_data(self):
dirname = 'cifar-10-batches-py'
path = os.path.join(self.data_dir, dirname)
fpath = os.path.join(path, 'test_batch')
x_test, y_test = self._load_batch(fpath)
if K.image_data_format() == 'channels_last':
x_test = x_test.transpose(0, 2, 3, 1)
y_test_labels = one_hot_encoded(y_test, num_classes=self.n_classes)
return x_test, np.array(y_test), y_test_labels
def load_training_data(self):
dirname = 'cifar-10-batches-py'
path = os.path.join(self.data_dir, dirname)
n_train_batchs = 5
x_train = np.zeros((0, self.depth, self.width, self.height))
y_train = []
for batch in range(n_train_batchs):
fpath = os.path.join(path, 'data_batch_' + str(batch + 1))
cur_data, cur_labels = self._load_batch(fpath)
x_train = np.concatenate((cur_data, x_train), axis=0)
y_train = cur_labels + y_train
x_train = x_train.astype(np.uint8)
if K.image_data_format() == 'channels_last':
x_train = x_train.transpose(0, 2, 3, 1)
y_train_labels = one_hot_encoded(y_train, num_classes=self.n_classes)
return x_train, np.array(y_train), y_train_labels
def load_test_data(self):
dirname = 'cifar-100-python'
path = os.path.join(self.data_dir, dirname)
fpath = os.path.join(path, 'test')
x_test, y_test_fine = self._load_batch(fpath, 'fine_labels')
data_size = len(y_test_fine)
if K.image_data_format() == 'channels_last':
x_test = x_test.transpose(0, 2, 3, 1)
relevant_idxes = [i for i in range(data_size) if y_test_fine[i] in self.subsets_idxes]
x_test = x_test[relevant_idxes, :, :, :]
y_test = np.asarray(y_test_fine)[relevant_idxes]
y_test_values = sorted(list(set(y_test)))
assert(len(y_test_values) == self.n_classes)
map_dict = {val: i for i, val in enumerate(y_test_values)}
for i, y in enumerate(y_test):
y_test[i] = map_dict[y]
y_test_labels = one_hot_encoded(y_test, num_classes=self.n_classes)
return x_test, y_test, y_test_labels
def load_training_data(self):
dirname = 'cifar-100-python'
path = os.path.join(self.data_dir, dirname)
fpath = os.path.join(path, 'train')
x_train, y_train_fine = self._load_batch(fpath, 'fine_labels')
_, y_train_coarse = self._load_batch(fpath, 'coarse_labels')
if K.image_data_format() == 'channels_last':
x_train = x_train.transpose(0, 2, 3, 1)
curr_superclass_idxes = [i for i in range(len(y_train_fine)) if y_train_coarse[i] == self.superclass_idx]
x_train = x_train[curr_superclass_idxes]
y_train = np.asarray(y_train_fine)[curr_superclass_idxes]
y_train_values = sorted(list(set(y_train)))
assert(len(y_train_values) == self.n_classes)
map_dict = {val: i for i, val in enumerate(y_train_values)}
for i, y in enumerate(y_train):
y_train[i] = map_dict[y]
y_train_labels = one_hot_encoded(y_train, num_classes=self.n_classes)
return x_train, y_train, y_train_labels
