def generate_skels(prefix, output_file_path, segmentation_file_path, resolution):
if not os.path.exists("../meta"):
os.mkdir("../meta")
meta_path = "../meta/" + prefix + ".meta"
segmentation = dataIO.ReadH5File(segmentation_file_path, "main").astype(np.int64)
grid = [len(segmentation), len(segmentation[0]), len(segmentation[0][0])]
with open(meta_path, "w") as meta_file:
meta_file.write("# resolution in nm\n")
meta_file.write(f"{resolution[0]}x{resolution[1]}x{resolution[2]}\n")
meta_file.write("# segmentation filename\n")
meta_file.write(f"{segmentation_file_path} main\n")
meta_file.write("# grid size\n")
meta_file.write(f"{grid[0]}x{grid[1]}x{grid[2]}\n")
seg2seg.DownsampleMapping(prefix, segmentation)
generate_skeletons.TopologicalThinning(prefix, segmentation)
generate_skeletons.FindEndpointVectors(prefix)
skeletons = ReadSkeletons(prefix)
with open(output_file_path, 'wb') as fp:
pickle.dump(skeletons, fp)
def CollectExamples(prefix, width, radius, subset):
# get the parent directory with all of the featuers
parent_directory = 'features/biological/edges-{}nm-{}x{}x{}'.format(
radius, width[IB_Z + 1], width[IB_Y + 1], width[IB_X + 1])
positive_filename = '{}/{}/positives/{}-examples.h5'.format(
parent_directory, subset, prefix)
if os.path.exists(positive_filename):
positive_examples = dataIO.ReadH5File(positive_filename, 'main')
else:
positive_examples = np.zeros(
(0, width[IB_Z + 1], width[IB_Y + 1], width[IB_X + 1]))
negative_filename = '{}/{}/negatives/{}-examples.h5'.format(
parent_directory, subset, prefix)
if os.path.exists(negative_filename):
negative_examples = dataIO.ReadH5File(negative_filename, 'main')
else:
negative_examples = np.zeros(
(0, width[IB_Z + 1], width[IB_Y + 1], width[IB_X + 1]))
unknowns_filename = '{}/{}/unknowns/{}-examples.h5'.format(
parent_directory, subset, prefix)
if os.path.exists(unknowns_filename):
unknowns_example = dataIO.ReadH5File(unknowns_filename, 'main')
else:
unknowns_example = np.zeros(
(0, width[IB_Z + 1], width[IB_Y + 1], width[IB_X + 1]))
# concatenate all of the examples together
examples = np.concatenate(
(positive_examples, negative_examples, unknowns_example), axis=0)
# add in information needed for forward inference [regions masked out for training and validation]
forward_positive_filename = '{}/forward/positives/{}-examples.h5'.format(
parent_directory, prefix)
if os.path.exists(forward_positive_filename):
forward_positive_examples = dataIO.ReadH5File(
forward_positive_filename, 'main')
examples = np.concatenate((examples, forward_positive_examples),
axis=0)
forward_negative_filename = '{}/forward/negatives/{}-examples.h5'.format(
parent_directory, prefix)
if os.path.exists(forward_negative_filename):
forward_negative_examples = dataIO.ReadH5File(
forward_negative_filename, 'main')
examples = np.concatenate((examples, forward_negative_examples),
axis=0)
forward_unknowns_filename = '{}/forward/unknowns/{}-examples.h5'.format(
parent_directory, prefix)
if os.path.exists(forward_unknowns_filename):
forward_unknowns_examples = dataIO.ReadH5File(
forward_unknowns_filename, 'main')
examples = np.concatenate((examples, forward_unknowns_examples),
axis=0)
return examples, positive_examples.shape[0], negative_examples.shape[0]
def EdgeGenerator(parameters, width, radius, subset):
# get the directories corresponding to this radius and subset
positive_directory = 'features/biological/edges-{}nm-{}x{}x{}/{}/positives'.format(
radius, width[IB_Z + 1], width[IB_Y + 1], width[IB_X + 1], subset)
negative_directory = 'features/biological/edges-{}nm-{}x{}x{}/{}/negatives'.format(
radius, width[IB_Z + 1], width[IB_Y + 1], width[IB_X + 1], subset)
# get all the positive candidate filenames
positive_filenames = os.listdir(positive_directory)
positive_candidates = []
for positive_filename in positive_filenames:
if not positive_filename[-3:] == '.h5': continue
positive_candidates.append(
dataIO.ReadH5File(
'{}/{}'.format(positive_directory, positive_filename), 'main'))
positive_candidates = np.concatenate(positive_candidates, axis=0)
# get all the negative candidate filenames
negative_filenames = os.listdir(negative_directory)
negative_candidates = []
for negative_filename in negative_filenames:
if not negative_filename[-3:] == '.h5': continue
negative_candidates.append(
dataIO.ReadH5File(
'{}/{}'.format(negative_directory, negative_filename), 'main'))
negative_candidates = np.concatenate(negative_candidates, axis=0)
# create easy access to the numbers of candidates
npositive_candidates = positive_candidates.shape[0]
nnegative_candidates = negative_candidates.shape[0]
batch_size = parameters['batch_size']
examples = np.zeros((batch_size, width[0], width[IB_Z + 1],
width[IB_Y + 1], width[IB_X + 1]),
dtype=np.float32)
labels = np.zeros(batch_size, dtype=np.float32)
positive_order = range(npositive_candidates)
negative_order = range(nnegative_candidates)
random.shuffle(positive_order)
random.shuffle(negative_order)
positive_index = 0
negative_index = 0
while True:
for iv in range(batch_size // 2):
positive_candidate = positive_candidates[
positive_order[positive_index]]
negative_candidate = negative_candidates[
negative_order[negative_index]]
examples[2 * iv, :, :, :, :] = AugmentFeature(
positive_candidate, width)
labels[2 * iv] = True
examples[2 * iv + 1, :, :, :, :] = AugmentFeature(
negative_candidate, width)
labels[2 * iv + 1] = False
positive_index += 1
if positive_index == npositive_candidates:
random.shuffle(positive_order)
positive_index = 0
negative_index += 1
if negative_index == nnegative_candidates:
random.shuffle(negative_order)
negative_index = 0
yield (examples, labels)