def Oracle(prefix,
threshold,
maximum_distance,
endpoint_distance,
network_distance,
filtersize=0):
# get all of the candidates
positive_candidates = FindCandidates(prefix, threshold, maximum_distance,
endpoint_distance, network_distance,
'positive')
negative_candidates = FindCandidates(prefix, threshold, maximum_distance,
endpoint_distance, network_distance,
'negative')
candidates = positive_candidates + negative_candidates
# read in all relevant information
segmentation = dataIO.ReadSegmentationData(prefix)
gold = dataIO.ReadGoldData(prefix)
seg2gold_mapping = seg2gold.Mapping(segmentation, gold)
# create the union find data structure
max_value = np.amax(segmentation) + 1
union_find = [unionfind.UnionFindElement(iv) for iv in range(max_value)]
# iterate over all candidates and collapse edges
for candidate in candidates:
label_one = candidate.labels[0]
label_two = candidate.labels[1]
if not seg2gold_mapping[label_one] or not seg2gold_mapping[label_two]:
continue
if (seg2gold_mapping[label_one] == seg2gold_mapping[label_two]):
unionfind.Union(union_find[label_one], union_find[label_two])
# create a mapping for the labels
mapping = np.zeros(max_value, dtype=np.int64)
for iv in range(max_value):
mapping[iv] = unionfind.Find(union_find[iv]).label
segmentation = seg2seg.MapLabels(segmentation, mapping)
comparestacks.CremiEvaluate(segmentation,
gold,
dilate_ground_truth=1,
mask_ground_truth=True,
mask_segmentation=False,
filtersize=filtersize)
def CollapseGraph(prefix, segmentation, vertex_ones, vertex_twos,
maintained_edges, algorithm):
# get the number of edges
nedges = maintained_edges.shape[0]
# create the union find data structure and collapse the graph
max_label = np.amax(segmentation) + 1
union_find = [unionfind.UnionFindElement(iv) for iv in range(max_label)]
# go through all of the edges
for ie in range(nedges):
# skip if the edge should not collapse
if maintained_edges[ie]: continue
# merge these vertices
vertex_one = vertex_ones[ie]
vertex_two = vertex_twos[ie]
unionfind.Union(union_find[vertex_one], union_find[vertex_two])
# create the mapping and save the result
mapping = np.zeros(max_label, dtype=np.int64)
for iv in range(max_label):
mapping[iv] = unionfind.Find(union_find[iv]).label
# apply the mapping and save the result
seg2seg.MapLabels(segmentation, mapping)
rhoana_filename = 'rhoana/{}-{}.h5'.format(prefix, algorithm)
dataIO.WriteH5File(segmentation, rhoana_filename, 'main')
# spawn a new meta file
dataIO.SpawnMetaFile(prefix, rhoana_filename, 'main')
# get the variation of information for this result
new_prefix = rhoana_filename.split('/')[1][:-3]
# read in the new gold data
gold = dataIO.ReadGoldData(prefix)
rand_error, vi = comparestacks.VariationOfInformation(
new_prefix, segmentation, gold)
#adapted_rand = comparestacks.adapted_rand(prefix, segmentation, gold)
print 'Rand Error Full: {}'.format(rand_error[0] + rand_error[1])
print 'Rand Error Merge: {}'.format(rand_error[0])
print 'Rand Error Split: {}'.format(rand_error[1])
print 'Variation of Information Full: {}'.format(vi[0] + vi[1])
print 'Variation of Information Merge: {}'.format(vi[0])
print 'Variation of Information Split: {}'.format(vi[1])
#print 'Adapted Rand: {}'.format(adapted_rand)
# make sure that the options are either multicut or lifted-multicut
if 'lifted-multicut' in algorithm: output_folder = 'lifted-multicut'
elif 'multicut' in algorithm: output_folder = 'multicut'
elif 'graph-baseline' in algorithm: output_folder = 'graph-baselines'
else: assert (False)
with open('{}-results/{}-{}.txt'.format(output_folder, algorithm, prefix),
'w') as fd:
fd.write('Rand Error Full: {}\n'.format(rand_error[0] + rand_error[1]))
fd.write('Rand Error Merge: {}\n'.format(rand_error[0]))
fd.write('Rand Error Split: {}\n'.format(rand_error[1]))
fd.write('Variation of Information Full: {}\n'.format(vi[0] + vi[1]))
fd.write('Variation of Information Merge: {}\n'.format(vi[0]))
fd.write('Variation of Information Split: {}\n'.format(vi[1]))
def CollapseGraph(segmentation, candidates, maintain_edges, probabilities,
output_filename):
ncandidates = len(candidates)
# get the ground truth and the predictions
labels = np.zeros(ncandidates, dtype=np.bool)
for iv in range(ncandidates):
labels[iv] = candidates[iv].ground_truth
# create an empty union find data structure
max_value = np.amax(segmentation) + 1
union_find = [unionfind.UnionFindElement(iv) for iv in range(max_value)]
# create adjacency sets for the elements in the segment
adjacency_sets = [set() for _ in range(max_value)]
for candidate in candidates:
label_one = candidate.labels[0]
label_two = candidate.labels[1]
adjacency_sets[label_one].add(label_two)
adjacency_sets[label_two].add(label_one)
# iterate over the candidates in order of decreasing probability
zipped = zip(probabilities, [ie for ie in range(ncandidates)])
for probability, ie in sorted(zipped, reverse=True):
# skip if the edge is not collapsed
if maintain_edges[ie]: continue
# skip if this creates a cycle
label_one, label_two = candidates[ie].labels
# get the parent of this label
label_two_union_find = unionfind.Find(union_find[label_two]).label
# make sure none of the other adjacent nodes already has this label
for neighbor_label in adjacency_sets[label_one]:
if neighbor_label == label_two: continue
if unionfind.Find(
union_find[neighbor_label]).label == label_two_union_find:
maintain_edges[ie] = True
# skip if the edge is no longer collapsed
if maintain_edges[ie]: continue
unionfind.Union(union_find[label_one], union_find[label_two])
print '\nBorder Constraints\n'
PrecisionAndRecall(labels, 1 - maintain_edges)
# for every edge, save if the edge is collapsed
with open(output_filename, 'wb') as fd:
fd.write(struct.pack('q', ncandidates))
for ie in range(ncandidates):
fd.write(struct.pack('?', maintain_edges[ie]))
mapping = np.zeros(max_value, dtype=np.int64)
for iv in range(max_value):
mapping[iv] = unionfind.Find(union_find[iv]).label
segmentation = seg2seg.MapLabels(segmentation, mapping)
gold = dataIO.ReadGoldData('SNEMI3D_train')
print comparestacks.adapted_rand(segmentation,
gold,
all_stats=False,
dilate_ground_truth=2,
filtersize=0)
def Forward(prefix,
model_prefix,
segmentation,
width,
radius,
subset,
evaluate=False,
threshold_volume=10368000):
# read in the trained model
model = model_from_json(open('{}.json'.format(model_prefix), 'r').read())
model.load_weights('{}-best-loss.h5'.format(model_prefix))
# get all of the examples
examples, npositives, nnegatives = CollectExamples(prefix, width, radius,
subset)
# get all of the large-small pairings
pairings = CollectLargeSmallPairs(prefix, width, radius, subset)
#assert (len(pairings) == examples.shape[0])
# get the threshold in terms of number of voxels
resolution = dataIO.Resolution(prefix)
threshold = int(threshold_volume /
(resolution[IB_Z] * resolution[IB_Y] * resolution[IB_X]))
# get the list of nodes over and under the threshold
small_segments, large_segments = FindSmallSegments(segmentation, threshold)
# get all of the probabilities
probabilities = model.predict_generator(NodeGenerator(examples, width),
examples.shape[0],
max_q_size=1000)
# save the probabilities to a file
output_filename = '{}-{}.probabilities'.format(model_prefix, prefix)
with open(output_filename, 'wb') as fd:
fd.write(struct.pack('q', examples.shape[0]))
for ie, (label_one, label_two) in enumerate(pairings):
fd.write(
struct.pack('qqd', label_one, label_two, probabilities[ie]))
# create the correct labels for the ground truth
ground_truth = np.zeros(npositives + nnegatives, dtype=np.bool)
for iv in range(npositives):
ground_truth[iv] = True
# get the results with labeled data
predictions = Prob2Pred(np.squeeze(probabilities[:npositives +
nnegatives]))
# print the confusion matrix
output_filename = '{}-{}-inference.txt'.format(model_prefix, prefix)
PrecisionAndRecall(ground_truth, predictions, output_filename)
# create a mapping
small_segment_predictions = dict()
for small_segment in small_segments:
small_segment_predictions[small_segment] = set()
# go through each pairing
for pairing, probability in zip(pairings, probabilities):
label_one, label_two = pairing
# make sure that either label one or two is small and the other is large
assert ((label_one in small_segments) ^ (label_two in small_segments))
if label_one in small_segments:
small_segment = label_one
large_segment = label_two
else:
small_segment = label_two
large_segment = label_one
small_segment_predictions[small_segment].add(
(large_segment, probability[0]))
# begin to map the small labels
max_label = np.amax(segmentation) + 1
mapping = [iv for iv in range(max_label)]
# look at seg2gold to see how many correct segments are merged
seg2gold_mapping = seg2gold.Mapping(prefix)
ncorrect_merges = 0
nincorrect_merges = 0
# go through all of the small segments
for small_segment in small_segments:
best_probability = -1
best_large_segment = -1
# go through all the neighboring large segments
for large_segment, probability in small_segment_predictions[
small_segment]:
if probability > best_probability:
best_probability = probability
best_large_segment = large_segment
# this should almost never happen but if it does just continue
if best_large_segment == -1 or best_probability < 0.5:
mapping[small_segment] = small_segment
continue
# get all of the best large segments
else:
mapping[small_segment] = best_large_segment
# don't consider undetermined locations
if seg2gold_mapping[small_segment] < 1 or seg2gold_mapping[
best_large_segment] < 1:
continue
if seg2gold_mapping[small_segment] == seg2gold_mapping[
best_large_segment]:
ncorrect_merges += 1
else:
nincorrect_merges += 1
print '\nResults:'
print ' Correctly Merged: {}'.format(ncorrect_merges)
print ' Incorrectly Merged: {}'.format(nincorrect_merges)
with open(output_filename, 'a') as fd:
fd.write('\nResults:\n')
fd.write(' Correctly Merged: {}\n'.format(ncorrect_merges))
fd.write(' Incorrectly Merged: {}\n'.format(nincorrect_merges))
# save the node mapping in the cache for later
end2end_mapping = [mapping[iv] for iv in range(max_label)]
# initiate the mapping to eliminate small segments
seg2seg.MapLabels(segmentation, mapping)
# reduce the labels and map again
mapping, _ = seg2seg.ReduceLabels(segmentation)
seg2seg.MapLabels(segmentation, mapping)
# update the end to end mapping with the reduced labels
for iv in range(max_label):
end2end_mapping[iv] = mapping[end2end_mapping[iv]]
# get the model name (first component is architecture and third is node-)
model_name = model_prefix.split('/')[1]
output_filename = 'rhoana/{}-reduced-{}.h5'.format(prefix, model_name)
dataIO.WriteH5File(segmentation, output_filename, 'main')
# spawn a new meta file
dataIO.SpawnMetaFile(prefix, output_filename, 'main')
# save the end to end mapping in the cache
mapping_filename = 'cache/{}-reduced-{}-end2end.map'.format(
prefix, model_name)
with open(mapping_filename, 'wb') as fd:
fd.write(struct.pack('q', max_label))
for label in range(max_label):
fd.write(struct.pack('q', end2end_mapping[label]))
if evaluate:
gold = dataIO.ReadGoldData(prefix)
# run the evaluation framework
rand_error, vi = comparestacks.VariationOfInformation(
segmentation, gold)
# write the output file
with open('node-results/{}-reduced-{}.txt'.format(prefix, model_name),
'w') as fd:
fd.write('Rand Error Full: {}\n'.format(rand_error[0] +
rand_error[1]))
fd.write('Rand Error Merge: {}\n'.format(rand_error[0]))
fd.write('Rand Error Split: {}\n'.format(rand_error[1]))
fd.write('Variation of Information Full: {}\n'.format(vi[0] +
vi[1]))
fd.write('Variation of Information Merge: {}\n'.format(vi[0]))
fd.write('Variation of Information Split: {}\n'.format(vi[1]))
def Agglomerate(prefix, model_prefix, threshold=0.5):
# read the segmentation data
segmentation = dataIO.ReadSegmentationData(prefix)
# get the multicut filename (with graph weights)
multicut_filename = 'multicut/{}-{}.graph'.format(model_prefix, prefix)
# get the maximum segmentation value
max_value = np.amax(segmentation) + 1
# create union find data structure
union_find = [UnionFind.UnionFindElement(iv) for iv in range(max_value)]
# read in all of the labels and merge the result
with open(multicut_filename, 'rb') as fd:
# read the number of vertices and edges
nvertices, nedges, = struct.unpack('QQ', fd.read(16))
# read in all of the edges
for ie in range(nedges):
# read in both labels
label_one, label_two, = struct.unpack('QQ', fd.read(16))
# skip over the reduced labels
fd.read(16)
# read in the edge weight
edge_weight, = struct.unpack('d', fd.read(8))
# merge label one and label two in the union find data structure
if (edge_weight > threshold):
UnionFind.Union(union_find[label_one], union_find[label_two])
# create a mapping
mapping = np.zeros(max_value, dtype=np.int64)
# update the segmentation
for iv in range(max_value):
label = UnionFind.Find(union_find[iv]).label
mapping[iv] = label
# update the labels
agglomerated_segmentation = seg2seg.MapLabels(segmentation, mapping)
gold_filename = 'gold/{}_gold.h5'.format(prefix)
# TODO fix this code temporary filename
agglomeration_filename = 'multicut/{}-agglomerate.h5'.format(prefix)
# temporary - write h5 file
dataIO.WriteH5File(agglomerated_segmentation, agglomeration_filename,
'stack')
import time
start_time = time.time()
print 'Agglomeration - {}:'.format(threshold)
# create the command line
command = '~/software/PixelPred2Seg/comparestacks --stack1 {} --stackbase {} --dilate1 1 --dilatebase 1 --relabel1 --relabelbase --filtersize 100 --anisotropic'.format(
agglomeration_filename, gold_filename)
# execute the command
os.system(command)
print time.time() - start_time
def MergeGroundTruth(prefix, model_prefix):
# read the segmentation data
segmentation = dataIO.ReadSegmentationData(prefix)
# get the multicut filename (with graph weights)
multicut_filename = 'multicut/{}-{}.graph'.format(model_prefix, prefix)
# read the gold data
gold = dataIO.ReadGoldData(prefix)
# read in the segmentation to gold mapping
mapping = seg2gold.Mapping(segmentation, gold)
# get the maximum segmentation value
max_value = np.amax(segmentation)
# create union find data structure
union_find = [UnionFind.UnionFindElement(iv) for iv in range(max_value)]
# read in all of the labels
with open(multicut_filename, 'rb') as fd:
# read the number of vertices and edges
nvertices, nedges, = struct.unpack('QQ', fd.read(16))
# read in all of the edges
for ie in range(nedges):
# read in the two labels
label_one, label_two, = struct.unpack('QQ', fd.read(16))
# skip over the reduced labels and edge weight
fd.read(24)
# if the labels are the same and the mapping is non zero
if mapping[label_one] == mapping[label_two] and mapping[label_one]:
UnionFind.Union(union_find[label_one], union_find[label_two])
# create a mapping
mapping = np.zeros(max_value, dtype=np.int64)
# update the segmentation
for iv in range(max_value):
label = UnionFind.Find(union_find[iv]).label
mapping[iv] = label
merged_segmentation = seg2seg.MapLabels(segmentation, mapping)
gold_filename = 'gold/{}_gold.h5'.format(prefix)
# TODO fix this code temporary filename
truth_filename = 'multicut/{}-truth.h5'.format(prefix)
# temporary write h5 file
dataIO.WriteH5File(merged_segmentation, truth_filename, 'stack')
import time
start_time = time.time()
print 'Ground truth: '
# create the command line
command = '~/software/PixelPred2Seg/comparestacks --stack1 {} --stackbase {} --dilate1 1 --dilatebase 1 --relabel1 --relabelbase --filtersize 100 --anisotropic'.format(
truth_filename, gold_filename)
# execute the command
os.system(command)
print time.time() - start_time