def test_edge_mask_2d(self):
np.random.seed(42)
edge_mask = np.random.random((len(self.offsets_2d), ) +
self.shape_2d) > 0.5
g1 = ngraph.undirectedLongRangeGridGraph(
list(self.shape_2d),
self.offsets_2d,
offsets_probabilities=self.offsets_probabilities_2d,
)
g2 = ngraph.undirectedLongRangeGridGraph(
list(self.shape_2d),
self.offsets_2d,
edge_mask=edge_mask,
)
def test_undirected_long_range_grid_graph_2d(self):
g = ngraph.undirectedLongRangeGridGraph(list(self.shape_2d),
self.offsets_2d)
self.assertTrue(g.numberOfNodes == 9)
self.assertTrue(g.numberOfEdges == 9)
# Call some extra methods:
edges_ID = g.projectEdgesIDToPixels()
nodes_ID = g.projectNodesIDToPixels()
edge_values = g.edgeValues(
np.random.random(self.shape_2d +
(len(self.offsets_2d), )).astype('float32'))
node_values = g.nodeValues(
np.random.random(self.shape_2d).astype('float32'))
def build_pixel_lifted_graph_from_offsets(image_shape,
offsets,
label_image=None,
GT_label_image=None,
offsets_probabilities=None,
offsets_weights=None,
strides=None,
nb_local_offsets=3,
downscaling_factor=None,
mask_used_edges=None):
# TODO: why label_image...?
"""
:param offsets: At the moment local offsets should be the first ones
:param nb_local_offsets: UPDATE AND GENERALIZE!
:param downscaling_factor: If a list [1,2,2] is given, then the image resolution is scaled down first
"""
raise DeprecationWarning("User updated function in GASP repo")
if downscaling_factor is not None:
raise NotImplementedError()
image_shape = tuple(image_shape) if not isinstance(image_shape,
tuple) else image_shape
is_local_offset = np.zeros(offsets.shape[0], dtype='bool')
is_local_offset[:nb_local_offsets] = True
warnings.warn(
"First {} offsets are assumed to be direct neighbors and the remaining ones long-range"
.format(nb_local_offsets))
if label_image is not None:
assert image_shape == label_image.shape
if offsets_weights is not None:
print("Offset weights ignored...!")
# TODO: change name offsets_probabilities
# print("Actually building graph...")
tick = time.time()
graph = ngraph.undirectedLongRangeGridGraph(
image_shape,
offsets,
is_local_offset,
offsets_probabilities=offsets_probabilities,
labels=label_image,
strides=strides,
mask_used_edges=mask_used_edges)
nb_nodes = graph.numberOfNodes
# print(label_image)
# print(is_local_offset)
# print(offsets_probabilities)
if label_image is None:
# print("Getting edge index...")
offset_index = graph.edgeOffsetIndex()
# print(np.unique(offset_index, return_counts=True))
is_local_edge = np.empty_like(offset_index, dtype='bool')
w = np.where(offset_index < nb_local_offsets)
warnings.warn(
"First {} offsets are assumed to be direct neighbors and the remaining ones long-range"
.format(nb_local_offsets))
is_local_edge[:] = 0
is_local_edge[w] = 1
# print("Nb. local edges: {} out of {}".format(is_local_edge.sum(), graph.numberOfEdges))
else:
# print("Took {} s!".format(time.time() - tick))
# print("Checking edge locality...")
is_local_edge = graph.findLocalEdges(label_image).astype('int32')
if offsets_weights is None or label_image is not None:
edge_sizes = np.ones(graph.numberOfEdges, dtype='int32')
else:
if isinstance(offsets_weights, (list, tuple)):
offsets_weights = np.array(offsets_weights)
assert offsets_weights.shape[0] == offsets.shape[0]
if offsets_weights.ndim == len(image_shape) + 1:
edge_sizes = graph.edgeValues(np.rollaxis(offsets_weights, 0, 4))
else:
if all([w >= 1.0 for w in offsets_weights]):
# Take the inverse:
offsets_weights = 1. / offsets_weights
else:
assert all([w <= 1.0 for w in offsets_weights]) and all(
[w >= 0.0 for w in offsets_weights])
# print("Edge weights...")
edge_sizes = offsets_weights[offset_index.astype('int32')]
if GT_label_image is None:
GT_labels_nodes = np.zeros(nb_nodes, dtype=np.int64)
else:
assert GT_label_image.shape == image_shape
GT_labels_image = GT_label_image.astype(np.uint64)
GT_labels_nodes = graph.nodeValues(GT_labels_image)
return graph, is_local_edge, GT_labels_nodes, edge_sizes
def build_pixel_lifted_graph_from_offsets(image_shape,
offsets,
label_image=None,
GT_label_image=None,
offsets_probabilities=None,
offsets_weights=None,
strides=None,
nb_local_offsets=3,
downscaling_factor=None,
mask_used_edges=None):
if downscaling_factor is not None:
raise NotImplementedError()
image_shape = tuple(image_shape) if not isinstance(image_shape,
tuple) else image_shape
is_local_offset = np.zeros(offsets.shape[0], dtype='bool')
is_local_offset[:nb_local_offsets] = True
if label_image is not None:
assert image_shape == label_image.shape
if offsets_weights is not None:
print("Offset weights ignored...!")
graph = ngraph.undirectedLongRangeGridGraph(
image_shape,
offsets,
is_local_offset,
offsets_probabilities=offsets_probabilities,
labels=label_image,
strides=strides,
mask_used_edges=mask_used_edges)
nb_nodes = graph.numberOfNodes
if label_image is None:
offset_index = graph.edgeOffsetIndex()
is_local_edge = np.empty_like(offset_index, dtype='bool')
w = np.where(offset_index < nb_local_offsets)
is_local_edge[:] = 0
is_local_edge[w] = 1
else:
is_local_edge = graph.findLocalEdges(label_image).astype('int32')
if offsets_weights is None or label_image is not None:
edge_sizes = np.ones(graph.numberOfEdges, dtype='int32')
else:
if isinstance(offsets_weights, (list, tuple)):
offsets_weights = np.array(offsets_weights)
assert offsets_weights.shape[0] == offsets.shape[0]
if offsets_weights.ndim == len(image_shape) + 1:
edge_sizes = graph.edgeValues(np.rollaxis(offsets_weights, 0, 4))
else:
if all([w >= 1.0 for w in offsets_weights]):
# Take the inverse:
offsets_weights = 1. / offsets_weights
else:
assert all([w <= 1.0 for w in offsets_weights]) and all(
[w >= 0.0 for w in offsets_weights])
edge_sizes = offsets_weights[offset_index.astype('int32')]
if GT_label_image is None:
GT_labels_nodes = np.zeros(nb_nodes, dtype=np.int64)
else:
assert GT_label_image.shape == image_shape
GT_labels_image = GT_label_image.astype(np.uint64)
GT_labels_nodes = graph.nodeValues(GT_labels_image)
return graph, is_local_edge, GT_labels_nodes, edge_sizes