class PixelClassificationSerializer(AppletSerializer):
"""Encapsulate the serialization scheme for pixel classification
workflow parameters and datasets.
"""
def __init__(self, operator, projectFileGroupName):
self._serialClassifierSlot = SerialClassifierSlot(operator.Classifier,
operator.classifier_cache,
name="ClassifierForests")
slots = [SerialListSlot(operator.LabelNames,
transform=str),
SerialListSlot(operator.LabelColors, transform=lambda x: tuple(x.flat)),
SerialListSlot(operator.PmapColors, transform=lambda x: tuple(x.flat)),
SerialBlockSlot(operator.LabelImages,
operator.LabelInputs,
operator.NonzeroLabelBlocks,
name='LabelSets',
subname='labels{:03d}',
selfdepends=False,
shrink_to_bb=True),
SerialPickledSlot(operator.ClassifierFactory),
self._serialClassifierSlot ]
super(PixelClassificationSerializer, self).__init__(projectFileGroupName, slots, operator)
def _deserializeFromHdf5(self, topGroup, groupVersion, hdf5File, projectFilePath):
"""
Override from AppletSerializer.
Implement any additional deserialization that wasn't already accomplished by our list of serializable slots.
"""
# If this is an old project file that didn't save the label names to the project,
# create some default names.
if not self.operator.LabelNames.ready() or len(self.operator.LabelNames.value) == 0:
# How many labels are there?
# We have to count them.
# This is slow, but okay for this special backwards-compatibilty scenario.
# For each image
all_labels = set()
for image_index, group in enumerate(topGroup['LabelSets'].values()):
# For each label block
for block in group.values():
data = block[:]
all_labels.update( numpy.unique(data) )
max_label = max(all_labels)
label_names = []
for i in range(max_label):
label_names.append( "Label {}".format( i+1 ) )
self.operator.LabelNames.setValue( label_names )
# Make some default colors, too
default_colors = [(255,0,0),
(0,255,0),
(0,0,255),
(255,255,0),
(255,0,255),
(0,255,255),
(128,128,128),
(255, 105, 180),
(255, 165, 0),
(240, 230, 140) ]
colors = []
for i, _ in enumerate(label_names):
colors.append( default_colors[i] )
self.operator.LabelColors.setValue( colors )
self.operator.PmapColors.setValue( colors )
# Now RE-deserialize the classifier, so it isn't marked dirty
self._serialClassifierSlot.deserialize(topGroup)
class PixelClassificationSerializer(AppletSerializer):
"""Encapsulate the serialization scheme for pixel classification
workflow parameters and datasets.
"""
def __init__(self, operator, projectFileGroupName):
self._serialClassifierSlot = SerialClassifierSlot(operator.Classifier,
operator.classifier_cache,
name="ClassifierForests")
slots = [SerialListSlot(operator.LabelNames,
transform=str),
SerialListSlot(operator.LabelColors, transform=lambda x: tuple(x.flat)),
SerialListSlot(operator.PmapColors, transform=lambda x: tuple(x.flat)),
SerialPickledValueSlot(operator.Bookmarks),
SerialBlockSlot(operator.LabelImages,
operator.LabelInputs,
operator.NonzeroLabelBlocks,
name='LabelSets',
subname='labels{:03d}',
selfdepends=False,
shrink_to_bb=True),
SerialClassifierFactorySlot(operator.ClassifierFactory),
self._serialClassifierSlot ]
super(PixelClassificationSerializer, self).__init__(projectFileGroupName, slots, operator)
def _deserializeFromHdf5(self, topGroup, groupVersion, hdf5File, projectFilePath):
"""
Override from AppletSerializer.
Implement any additional deserialization that wasn't already accomplished by our list of serializable slots.
"""
# If this is an old project file that didn't save the label names to the project,
# create some default names.
if (not self.operator.LabelNames.ready() or len(self.operator.LabelNames.value) == 0)\
and 'LabelSets' in topGroup:
# How many labels are there?
# We have to count them.
# This is slow, but okay for this special backwards-compatibilty scenario.
# For each image
all_labels = set()
for image_index, group in enumerate(topGroup['LabelSets'].values()):
# For each label block
for block in group.values():
data = block[:]
all_labels.update( vigra.analysis.unique(data) )
if all_labels:
max_label = max(all_labels)
else:
max_label = 0
label_names = []
for i in range(max_label):
label_names.append( "Label {}".format( i+1 ) )
self.operator.LabelNames.setValue( label_names )
# Make some default colors, too
default_colors = [(255,0,0),
(0,255,0),
(0,0,255),
(255,255,0),
(255,0,255),
(0,255,255),
(128,128,128),
(255, 105, 180),
(255, 165, 0),
(240, 230, 140) ]
colors = []
for i, _ in enumerate(label_names):
colors.append( default_colors[i] )
self.operator.LabelColors.setValue( colors )
self.operator.PmapColors.setValue( colors )
# Now RE-deserialize the classifier, so it isn't marked dirty
self._serialClassifierSlot.deserialize(topGroup)
# SPECIAL CLEANUP for backwards compatibility:
# Due to a bug, it was possible for a project to be saved with a classifier that was
# trained with more label classes than the project file saved in the end.
# That can cause a crash. So here, we inspect the restored classifier and remove it if necessary.
if not self.operator.classifier_cache._dirty:
restored_classifier = self.operator.classifier_cache._value
if hasattr(restored_classifier, 'known_classes'):
num_classifier_classes = len(restored_classifier.known_classes)
num_saved_label_classes = len(self.operator.LabelNames.value)
if num_classifier_classes > num_saved_label_classes:
# Delete the classifier from the operator
logger.info( "Resetting classifier... will be forced to retrain" )
self.operator.classifier_cache.resetValue()
class PixelClassificationSerializer(AppletSerializer):
"""Encapsulate the serialization scheme for pixel classification
workflow parameters and datasets.
"""
def __init__(self, operator, projectFileGroupName):
self._serialClassifierSlot = SerialClassifierSlot(operator.Classifier,
operator.classifier_cache,
name="ClassifierForests",
subname="Forest{:04d}")
slots = [SerialListSlot(operator.LabelNames,
transform=str),
SerialListSlot(operator.LabelColors, transform=lambda x: tuple(x.flat)),
SerialListSlot(operator.PmapColors, transform=lambda x: tuple(x.flat)),
SerialBlockSlot(operator.LabelImages,
operator.LabelInputs,
operator.NonzeroLabelBlocks,
name='LabelSets',
subname='labels{:03d}',
selfdepends=False,
shrink_to_bb=True),
self._serialClassifierSlot ]
super(PixelClassificationSerializer, self).__init__(projectFileGroupName, slots, operator)
def _deserializeFromHdf5(self, topGroup, groupVersion, hdf5File, projectFilePath):
"""
Override from AppletSerializer.
Implement any additional deserialization that wasn't already accomplished by our list of serializable slots.
"""
# If this is an old project file that didn't save the label names to the project,
# create some default names.
if not self.operator.LabelNames.ready():
# How many labels are there?
max_label = 0
for op in self.operator.opLabelPipeline:
max_label = max(max_label, op.opLabelArray.maxLabel.value)
label_names = []
for i in range(max_label):
label_names.append( "Label {}".format( i+1 ) )
self.operator.LabelNames.setValue( label_names )
# Make some default colors, too
default_colors = [(255,0,0),
(0,255,0),
(0,0,255),
(255,255,0),
(255,0,255),
(0,255,255),
(128,128,128),
(255, 105, 180),
(255, 165, 0),
(240, 230, 140) ]
colors = []
for i, _ in enumerate(label_names):
colors.append( default_colors[i] )
self.operator.LabelColors.setValue( colors )
self.operator.PmapColors.setValue( colors )
# Now RE-deserialize the classifier, so it isn't marked dirty
self._serialClassifierSlot.deserialize(topGroup)
