def test_image_is_the_right_dtype(self):
"""Make sure the images are of type uint8 i.e. [0,255] range
"""
faux_data = 100 * np.ones(shape=(200, 200, 3), dtype=np.float32)
with self.assertRaises(Exception):
pipe_utils.data_is_ok(faux_data, use='patch', raise_exception=True)
def test_image_has_alpha_channel(self):
"""Make sure that an image with an alpha channel doesn't pass the data
check
"""
faux_data = np.ones(shape=(200, 200, 4), dtype=np.uint8)
with self.assertRaises(Exception):
pipe_utils.data_is_ok(faux_data, use='patch', raise_exception=True)
def test_image_is_in_wrong_range(self):
"""See that if a image mistakenly normalized to [0,1), and then converted
to uint8 raises an alert
"""
faux_data = np.ones(shape=(200, 200, 3), dtype=np.uint8)
with self.assertRaises(Exception):
pipe_utils.data_is_ok(faux_data, use='patch', raise_exception=True)
def test_wrongly_shaped_image_raises_exception(self):
"""Make sure an exception is raised if the data in not of the correct
shape
"""
faux_data = 100 * np.ones(shape=(300, 300, 3), dtype=np.uint8)
with self.assertRaises(Exception):
pipe_utils.data_is_ok(faux_data, use='patch', raise_exception=True)
def save_as_retinanet_data(ds,
image_save_dir,
annotations_save_dir,
percent_test_set=.2,
include_labels=LABELS_TO_INCLUDE):
"""Save the data in the csv format expected by RetinaNet
Args
----
ds (list : ObjImage) : Image data set
image_save_dir (str) :
annotations_save_dir (str) :
percent_test_set (float) : value in [0, 1), percent of data to use as test
"""
test_annot_path = os.path.join(annotations_save_dir, "test_annot.csv")
train_annot_path = os.path.join(annotations_save_dir, "train_annot.csv")
if not os.path.exists(image_save_dir):
print("Creating directory to save processed images")
os.makedirs(image_save_dir)
if not os.path.exists(annotations_save_dir):
print("Creating directory to save annotation file")
os.makedirs(annotations_save_dir)
with open(test_annot_path, 'a') as test_file, open(train_annot_path,
'a') as train_file:
test_writer = csv.writer(test_file)
train_writer = csv.writer(train_file)
for img in ds:
# skip this image if the data isn't in the right format
if not pipe_utils.data_is_ok(img.data):
continue
image_path = os.path.join(image_save_dir, img.image_id + '.png')
skio.imsave(image_path, img.data)
# Randomly determine if the image is for training or testing
if np.random.uniform() < .2:
writer = test_writer
else:
writer = train_writer
# Make a note of every feature and bounding location in the image
all_features = img.get_features()
for label, locations in all_features.items():
label = MAP_TO_LOCAL_LABELS[label]
if label in include_labels:
for location in locations:
row = [image_path, *location, label]
writer.writerow(row)
# Write a blank row if no features in this image
if not all_features:
row = [image_path, '', '', '', '', '']
writer.writerow(row)
def test_data_check_doesnt_have_to_raise_exception(self):
"""Repeat test_wrongly_shaped_image_raises_exception just to make sure
that an exception doesnt have to be raise, and that a 'False' is
returned instead
"""
faux_data = 100 * np.ones(shape=(300, 300, 3), dtype=np.uint8)
self.assertFalse(pipe_utils.data_is_ok(faux_data, use='patch'))
def read_raw_image(path, report=True, check_data=False):
"""Import a raw image
This could be as 8 bit or 16 bit... or 10 bit like some of the files...
Args
----
path (str) : path to the image file
report (bool) : output a short log on the imported data
Raises
------
ImportError : If the image passed does not have a file extension that's expected
AssertionError : if data_is_ok fails when check_data=True
Returns
-------
numpy array of raw image
"""
ext = os.path.splitext(path)[1]
if ext in [".jpg", ".png"]:
img = imread(path)
elif ext == ".tif":
img = tifffile.imread(path)
else:
raise ImportError("{} Not a supported extension".format(ext))
if report:
print("Image {} loaded".format(os.path.basename(path)))
print("\tShape: ", img.shape)
print("\tdtype: ", img.dtype)
print("Values: ({:.2f},{:.2f})".format(img.min(), img.max())),
if check_data:
pipe_utils.data_is_ok(img, raise_exception=True)
return img
def test_patch_data_shape_is_ok(self):
"""Make sure data check returns true for an obj data shape
"""
faux_data = 100 * np.ones(shape=(200, 200, 3), dtype=np.uint8)
self.assertTrue(pipe_utils.data_is_ok(faux_data, use='patch'))