def create_patch_index_list(index_list,
image_shape,
patch_shape,
patch_overlap,
patch_start_offset=None):
"""
Function that returns an array of all the patches - of a given patch shape - in a given image.
:param index_list:
:param image_shape:
:param patch_shape:
:param patch_overlap:
:param patch_start_offset:
:return:
"""
patch_index = list()
for index in index_list:
if patch_start_offset is not None:
random_start_offset = np.negative(
get_random_nd_index(patch_start_offset))
patches = compute_patch_indices(image_shape,
patch_shape,
overlap=patch_overlap,
start=random_start_offset)
else:
patches = compute_patch_indices(image_shape,
patch_shape,
overlap=patch_overlap)
patch_index.extend(itertools.product([index], patches))
return patch_index
def create_patch_index_list(index_list,
image_shape,
patch_shape,
patch_overlap,
patch_start_offset=None):
patch_index = list()
for index in index_list:
if patch_start_offset is not None:
random_start_offset = np.negative(
get_random_nd_index(patch_start_offset))
patches = compute_patch_indices(image_shape,
patch_shape,
overlap=patch_overlap,
start=random_start_offset)
else:
patches = compute_patch_indices(image_shape,
patch_shape,
overlap=patch_overlap)
patch_index.extend(itertools.product([index], patches))
return patch_index
def test_reconstruct_from_patches(self):
patch_shape = (32, 32, 32)
patch_overlap = 0
patch_indices = compute_patch_indices(self.image.shape, patch_shape,
patch_overlap)
patches = [
get_patch_from_3d_data(self.image.get_data(), patch_shape, index)
for index in patch_indices
]
reconstruced_data = reconstruct_from_patches(patches, patch_indices,
self.image.shape)
# noinspection PyTypeChecker
self.assertTrue(np.all(self.image.get_data() == reconstruced_data))
def test_reconstruct_with_overlapping_patches2(self):
image_shape = (144, 144, 144)
data = np.arange(0, image_shape[0] * image_shape[1] *
image_shape[2]).reshape(image_shape)
patch_overlap = 16
patch_shape = (64, 64, 64)
patch_indices = compute_patch_indices(data.shape, patch_shape,
patch_overlap)
patches = [
get_patch_from_3d_data(data, patch_shape, index)
for index in patch_indices
]
no_overlap_indices = compute_patch_indices(data.shape, patch_shape, 32)
patch_indices = np.concatenate([patch_indices, no_overlap_indices])
patches.extend([
get_patch_from_3d_data(data, patch_shape, index)
for index in no_overlap_indices
])
reconstruced_data = reconstruct_from_patches(patches, patch_indices,
data.shape)
# noinspection PyTypeChecker
self.assertTrue(np.all(data == reconstruced_data))
def test_reconstruct_with_overlapping_patches(self):
patch_overlap = 0
patch_shape = (32, 32, 32)
patch_indices = compute_patch_indices(self.image.shape, patch_shape,
patch_overlap)
patches = [
get_patch_from_3d_data(self.image.get_data(), patch_shape, index)
for index in patch_indices
]
# extend patches with modified patches that are 2 lower than the original patches
patches.extend([patch - 2 for patch in patches])
patch_indices = np.concatenate([patch_indices, patch_indices], axis=0)
reconstruced_data = reconstruct_from_patches(patches, patch_indices,
self.image.shape)
# The reconstructed data should be 1 lower than the original data as 2 was subtracted from half the patches.
# The resulting reconstruction should be the average.
# noinspection PyTypeChecker
self.assertTrue(
np.all((self.image.get_data() - 1) == reconstruced_data))
def test_reconstruct_with_multiple_channels(self):
image_shape = (144, 144, 144)
n_channels = 4
data = np.arange(
0, image_shape[0] * image_shape[1] * image_shape[2] *
n_channels).reshape([n_channels] + list(image_shape))
patch_overlap = 16
patch_shape = (64, 64, 64)
patch_indices = compute_patch_indices(image_shape, patch_shape,
patch_overlap)
patches = [
get_patch_from_3d_data(data, patch_shape, index)
for index in patch_indices
]
self.assertEqual(patches[0].shape, tuple([4] + list(patch_shape)))
reconstruced_data = reconstruct_from_patches(patches, patch_indices,
data.shape)
# noinspection PyTypeChecker
self.assertTrue(np.all(data == reconstruced_data))
def patch_wise_prediction(model,
data,
overlap=0,
batch_size=10,
permute=False):
"""
:param batch_size:
:param model:
:param data:
:param overlap:
:return:
"""
patch_shape = tuple([int(dim) for dim in model.input.shape[-3:]])
predictions = list()
print("Patch_shape prediction:", patch_shape)
print("Overlap_ :", overlap)
indices = compute_patch_indices(data.shape[-3:],
patch_size=patch_shape,
overlap=overlap)
batch = list()
i = 0
while i < len(indices):
while len(batch) < batch_size and i < len(indices):
patch = get_patch_from_3d_data(data[0],
patch_shape=patch_shape,
patch_index=indices[i])
batch.append(patch)
i += 1
prediction = predict(model, np.asarray(batch), permute=permute)
batch = list()
if isinstance(prediction, list):
prediction = prediction[-1]
for predicted_patch in prediction:
predictions.append(predicted_patch)
if isinstance(model.output, list):
output_shape = [int(model.output[-1].shape[1])] + list(data.shape[-3:])
else:
output_shape = [int(model.output.shape[1])] + list(data.shape[-3:])
return reconstruct_from_patches(predictions,
patch_indices=indices,
data_shape=output_shape)