def deserialize_hdf5(cls, h5py_group):
model_str = h5py_group['serialized_model'][()]
model = iiboost.Booster()
model.deserialize(model_str)
known_labels = list(h5py_group['known_labels'][:])
feature_count = h5py_group['feature_count'][()]
return IIBoostLazyflowClassifier(model, known_labels, feature_count)
def create_and_train_pixelwise(self,
feature_images,
label_images,
axistags=None,
feature_names=None):
"""
feature_images: A sequence of ND images. See note above regarding required structure.
Last axis must be channel.
label_images: A sequence of ND label images. Last axis must be channel (size=1).
axistags: Optional. A vigra.AxisTags object describing ALL feature_images.
(Used to compute the anisotropy of the data.)
"""
logger.debug('training with IIBoost')
# Instantiate the classifier
model = iiboost.Booster()
# IIBoost requires both labels to be uint8, 3D only
converted_labels = []
for label_image in label_images:
assert len(
label_image.shape
) == 4, "IIBoost expects 4D data (including channel dimension)."
assert label_image.shape[
-1] == 1, "Expected label image to have only one channel."
converted = numpy.array(
numpy.asarray(label_image[..., 0], dtype=numpy.uint8))
converted_labels.append(converted)
# Save for future reference
flattened_labels = map(numpy.ndarray.flatten, converted_labels)
all_labels = numpy.concatenate(flattened_labels)
known_labels = numpy.unique(all_labels)
if known_labels[0] == 0:
known_labels = known_labels[1:]
assert set([1, 2]).issuperset(
known_labels), "IIBoost only accepts two label values: 1 and 2"
# We can't train if there not labels from both classes.
if set(known_labels) != set([1, 2]):
return None
# IIBoost requires raw images to be uint8, 3D only
# NOTE: we assume that the raw data can be found in channel 0.
raw_images = []
for image in feature_images:
assert len(image.shape
) == 4, "IIBoost expects 4D data (including channel)."
# (Even though copy=False here, we'll probably get a copy anyway. That's fine. Same for cases below, too.)
raw = image[..., 0].astype(dtype=numpy.uint8,
order='C',
copy=False)
raw_images.append(raw)
# Extract the hessian eigenvector (hev) images
hev_images = []
for image in feature_images:
hev_image = image[..., 1:10]
hev_image = hev_image.astype(dtype=numpy.float32,
order='C',
copy=False)
hev_image = hev_image.reshape(hev_image.shape[:-1] + (3, 3))
hev_images.append(hev_image)
integral_images = []
for image in feature_images:
assert len(
image.shape
) == 4, "IIBoost expects 4D data (including channel dimension)."
# Select integral channels (see note above about input data)
integral_image = image[..., 10:]
# Put channel first
integral_image = integral_image.transpose(3, 0, 1, 2)
# IIBoost requires filter images to be float32, in C-order
integral_image = integral_image.astype(dtype=numpy.float32,
copy=False,
order='C')
integral_images.append(integral_image)
# Calculate anisotropy factor.
z_anisotropy_factor = 1.0
if axistags:
x_tag = axistags['x']
z_tag = axistags['z']
if z_tag.resolution != 0.0 and x_tag.resolution != 0.0:
z_anisotropy_factor = z_tag.resolution / x_tag.resolution
model.trainWithChannels(raw_images, hev_images, converted_labels,
integral_images, z_anisotropy_factor,
self.num_stumps, *self._args, **self._kwargs)
return IIBoostLazyflowClassifier(model,
known_labels,
feature_count=len(integral_images[0]),
feature_names=feature_names)