def load_data_raw_images(serialization_path='data.pickle',
train_images_folder='data_road/roadC621/'):
"""Load color images (3 channels) and labels (as images).
Returns
-------
tuple : (featurevector list, label list)
"""
logging.info("Start loading data...")
data_source = serialization_path + ".npz"
if not os.path.exists(data_source):
# build lists of files which will be read
path_data = os.path.join(os.environ['DATA_PATH'],
train_images_folder,
"image_2/")
files_data = [os.path.join(path_data, f)
for f in sorted(os.listdir(path_data))
if f.endswith('.png')]
path_gt = os.path.join(os.environ['DATA_PATH'],
train_images_folder,
"gt_image_2/")
files_gt = [os.path.join(path_gt, f)
for f in sorted(os.listdir(path_gt))
if f.endswith('.png')]
if not os.path.isfile('training.pickle') or \
not os.path.isfile('testing.pickle'):
logging.info("Write training.pickle and testing.pickle")
write_files(files_data, files_gt)
filelist_tuples = read_filelist('training.pickle')
files_data, files_gt = [], []
for file_data, file_gt in filelist_tuples:
files_data.append(file_data)
files_gt.append(file_gt)
# read files (data first)
print("Start reading images: ", end='')
colored_image_features = []
for img_path in files_data:
print('.', end='')
ac = utils.load_color_image_features(img_path)
if(ac.shape[0] == 188): # TODO: Why is this skipped?
colored_image_features.append(ac)
print('')
xs_colored = np.array(colored_image_features, copy=False)
# read grayscale groundtruth
yl = []
for f in files_gt:
img = scipy.misc.imread(f)
if(img.shape[0] != 188): # TODO: Why is this skipped?
continue
new_img = np.zeros(img.shape)
for i, row in enumerate(img):
for j, pixel in enumerate(row):
new_img[i][j] = (105 == pixel)
yl.append(new_img)
yl = np.array(yl)
assert len(xs_colored) == len(yl), "len(xs_colored) != len(yl)"
for i, (X, y) in enumerate(zip(xs_colored, yl), start=1):
logging.info("Get labels (%i/%i)...", i, len(yl))
assert X.shape[:2] == y.shape, \
("X.shape[1:]=%s and y.shape=%s" %
(X.shape[:2], y.shape))
assert min(y.flatten()) == 0.0, \
("min(y)=%s" % str(min(y.flatten())))
assert max(y.flatten()) == 1.0, \
("max(y)=%s" % str(max(y.flatten())))
np.savez(serialization_path, xs_colored, yl)
else:
logging.info("!! Loaded pickled data" + "!" * 80)
logging.info("Data source: %s", data_source)
logging.info("This implies same test / training split as before.")
npzfile = np.load(data_source)
xs_colored = npzfile['arr_0']
yl = npzfile['arr_1']
return (xs_colored, yl)
zipped = list(zip(data_paths, label_paths))
random.shuffle(zipped)
assert 0.1 <= training_ratio <= 1.0, 'wrong training ratio'
split_to = int(len(zipped) * training_ratio)
training_files = zipped[:split_to]
testing_files = zipped[split_to:]
np.array(training_files).shape
training_data=[x[0] for x in training_files]
img_path = os.path.join(data_prefix,training_data[0])
ac = utils.load_color_image_features(img_path)
# read files (data first)
print("Start reading images: ", end='')
colored_image_features = []
for img_path in files_data:
print('.', end='')
ac = utils.load_color_image_features(img_path)
if(ac.shape[0] == 188): # TODO: Why is this skipped?
colored_image_features.append(ac)
print('')
xs_colored = np.array(colored_image_features, copy=False)
