def test_pixelwise_transform_3d(self):
frames = 10
img_list = []
for im in _generate_test_masks():
frame_list = []
for _ in range(frames):
frame_list.append(label(im))
img_stack = np.array(frame_list)
img_list.append(img_stack)
with self.cached_session():
K.set_image_data_format('channels_last')
# test single edge class
maskstack = np.vstack(img_list)
batch_count = maskstack.shape[0] // frames
new_shape = tuple([batch_count, frames] + list(maskstack.shape[1:]))
maskstack = np.reshape(maskstack, new_shape)
for i in range(maskstack.shape[0]):
img = maskstack[i, ...]
img = np.squeeze(img)
pw_img = transform_utils.pixelwise_transform(
img, data_format=None, separate_edge_classes=False)
pw_img_dil = transform_utils.pixelwise_transform(
img, dilation_radius=2,
data_format='channels_last',
separate_edge_classes=False)
self.assertEqual(pw_img.shape[-1], 3)
self.assertEqual(pw_img_dil.shape[-1], 3)
assert(np.all(np.equal(pw_img[..., 0] + pw_img[..., 1], img > 0)))
self.assertGreater(
pw_img_dil[..., 0].sum() + pw_img_dil[..., 1].sum(),
pw_img[..., 0].sum() + pw_img[..., 1].sum())
# test separate edge classes
maskstack = np.vstack(img_list)
batch_count = maskstack.shape[0] // frames
new_shape = tuple([batch_count, frames] + list(maskstack.shape[1:]))
maskstack = np.reshape(maskstack, new_shape)
for i in range(maskstack.shape[0]):
img = maskstack[i, ...]
img = np.squeeze(img)
pw_img = transform_utils.pixelwise_transform(
img, data_format=None, separate_edge_classes=True)
pw_img_dil = transform_utils.pixelwise_transform(
img, dilation_radius=2,
data_format='channels_last',
separate_edge_classes=True)
self.assertEqual(pw_img.shape[-1], 4)
self.assertEqual(pw_img_dil.shape[-1], 4)
assert(np.all(np.equal(pw_img[..., 0] + pw_img[..., 1] + pw_img[..., 2], img > 0)))
self.assertGreater(
pw_img_dil[..., 0].sum() + pw_img_dil[..., 1].sum(),
pw_img[..., 0].sum() + pw_img[..., 1].sum())
def test_pixelwise_transform_2d(self):
with self.cached_session():
K.set_image_data_format('channels_last')
# test single edge class
for img in _generate_test_masks():
img = label(img)
img = np.squeeze(img)
pw_img = transform_utils.pixelwise_transform(
img, data_format=None, separate_edge_classes=False)
pw_img_dil = transform_utils.pixelwise_transform(
img,
dilation_radius=1,
data_format='channels_last',
separate_edge_classes=False)
self.assertEqual(pw_img.shape[-1], 3)
self.assertEqual(pw_img_dil.shape[-1], 3)
assert (np.all(
np.equal(pw_img[..., 0] + pw_img[..., 1], img > 0)))
self.assertGreater(
pw_img_dil[..., 0].sum() + pw_img_dil[..., 1].sum(),
pw_img[..., 0].sum() + pw_img[..., 1].sum())
# test separate edge classes
for img in _generate_test_masks():
img = label(img)
img = np.squeeze(img)
pw_img = transform_utils.pixelwise_transform(
img, data_format=None, separate_edge_classes=True)
pw_img_dil = transform_utils.pixelwise_transform(
img,
dilation_radius=1,
data_format='channels_last',
separate_edge_classes=True)
self.assertEqual(pw_img.shape[-1], 4)
self.assertEqual(pw_img_dil.shape[-1], 4)
assert (np.all(
np.equal(pw_img[..., 0] + pw_img[..., 1] + pw_img[..., 2],
img > 0)))
self.assertGreater(
pw_img_dil[..., 0].sum() + pw_img_dil[..., 1].sum(),
pw_img[..., 0].sum() + pw_img[..., 1].sum())
def test_pixelwise_transform_2d(self):
with self.test_session():
K.set_image_data_format('channels_last')
# test single edge class
maskstack = np.array([label(i) for i in _generate_test_masks()])
pw_maskstack = transform_utils.pixelwise_transform(
maskstack, data_format=None, separate_edge_classes=False)
pw_maskstack_dil = transform_utils.pixelwise_transform(
maskstack,
dilation_radius=1,
data_format='channels_last',
separate_edge_classes=False)
self.assertEqual(pw_maskstack.shape[-1], 3)
self.assertEqual(pw_maskstack_dil.shape[-1], 3)
self.assertGreater(
pw_maskstack_dil[..., 0].sum() +
pw_maskstack_dil[..., 1].sum(),
pw_maskstack[..., 0].sum() + pw_maskstack[..., 1].sum())
# test separate edge classes
maskstack = np.array([label(i) for i in _generate_test_masks()])
pw_maskstack = transform_utils.pixelwise_transform(
maskstack, data_format=None, separate_edge_classes=True)
pw_maskstack_dil = transform_utils.pixelwise_transform(
maskstack,
dilation_radius=1,
data_format='channels_last',
separate_edge_classes=True)
self.assertEqual(pw_maskstack.shape[-1], 4)
self.assertEqual(pw_maskstack_dil.shape[-1], 4)
self.assertGreater(
pw_maskstack_dil[..., 0].sum() +
pw_maskstack_dil[..., 1].sum(),
pw_maskstack[..., 0].sum() + pw_maskstack[..., 1].sum())
def _transform_masks(y, transform, data_format=None, **kwargs):
"""Based on the transform key, apply a transform function to the masks.
Refer to :mod:`deepcell.utils.transform_utils` for more information about
available transforms. Caution for unknown transform keys.
Args:
y (numpy.array): Labels of ndim 4 or 5
transform (str): Name of the transform, one of
{"deepcell", "disc", "watershed", None}
data_format (str): One of 'channels_first', 'channels_last'.
kwargs (dict): Optional transform keyword arguments.
Returns:
numpy.array: the output of the given transform function on y
Raises:
ValueError: Rank of y is not 4 or 5.
ValueError: Channel dimension of y is not 1.
ValueError: Transform is invalid value.
"""
valid_transforms = {
'deepcell', # deprecated for "pixelwise"
'pixelwise',
'disc',
'watershed', # deprecated for "outer-distance"
'watershed-cont', # deprecated for "outer-distance"
'inner-distance',
'outer-distance',
'centroid', # deprecated for "inner-distance"
'fgbg'
}
if data_format is None:
data_format = K.image_data_format()
if y.ndim not in {4, 5}:
raise ValueError('`labels` data must be of ndim 4 or 5. Got', y.ndim)
channel_axis = 1 if data_format == 'channels_first' else -1
if y.shape[channel_axis] != 1:
raise ValueError('Expected channel axis to be 1 dimension. Got',
y.shape[1 if data_format == 'channels_first' else -1])
if isinstance(transform, str):
transform = transform.lower()
if transform not in valid_transforms and transform is not None:
raise ValueError('`{}` is not a valid transform'.format(transform))
if transform in {'pixelwise', 'deepcell'}:
if transform == 'deepcell':
warnings.warn('The `{}` transform is deprecated. Please use the '
'`pixelwise` transform insetad.'.format(transform),
DeprecationWarning)
dilation_radius = kwargs.pop('dilation_radius', None)
separate_edge_classes = kwargs.pop('separate_edge_classes', False)
edge_class_shape = 4 if separate_edge_classes else 3
if data_format == 'channels_first':
y_transform = np.zeros(tuple([y.shape[0]] + [edge_class_shape] + list(y.shape[2:])))
else:
y_transform = np.zeros(tuple(list(y.shape[0:-1]) + [edge_class_shape]))
for batch in range(y_transform.shape[0]):
if data_format == 'channels_first':
mask = y[batch, 0, ...]
else:
mask = y[batch, ..., 0]
y_transform[batch] = transform_utils.pixelwise_transform(
mask, dilation_radius, data_format=data_format,
separate_edge_classes=separate_edge_classes)
elif transform in {'outer-distance', 'watershed', 'watershed-cont'}:
if transform in {'watershed', 'watershed-cont'}:
warnings.warn('The `{}` transform is deprecated. Please use the '
'`outer-distance` transform insetad.'.format(transform),
DeprecationWarning)
bins = kwargs.pop('distance_bins', None)
erosion = kwargs.pop('erosion_width', 0)
by_frame = kwargs.pop('by_frame', True)
if data_format == 'channels_first':
y_transform = np.zeros(tuple([y.shape[0]] + list(y.shape[2:])))
else:
y_transform = np.zeros(y.shape[0:-1])
if y.ndim == 5:
if by_frame:
_distance_transform = transform_utils.outer_distance_transform_movie
else:
_distance_transform = transform_utils.outer_distance_transform_3d
else:
_distance_transform = transform_utils.outer_distance_transform_2d
for batch in range(y_transform.shape[0]):
if data_format == 'channels_first':
mask = y[batch, 0, ...]
else:
mask = y[batch, ..., 0]
y_transform[batch] = _distance_transform(
mask, bins=bins, erosion_width=erosion)
y_transform = np.expand_dims(y_transform, axis=-1)
if bins is None:
pass
else:
# convert to one hot notation
y_transform = to_categorical(y_transform, num_classes=bins)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform == 'inner-distance':
if transform == 'centroid':
warnings.warn('The `{}` transform is deprecated. Please use the '
'`inner-distance` transform insetad.'.format(transform),
DeprecationWarning)
bins = kwargs.pop('distance_bins', None)
erosion = kwargs.pop('erosion_width', 0)
by_frame = kwargs.pop('by_frame', True)
alpha = kwargs.pop('alpha', 0.1)
beta = kwargs.pop('beta', 1)
if data_format == 'channels_first':
y_transform = np.zeros(tuple([y.shape[0]] + list(y.shape[2:])))
else:
y_transform = np.zeros(y.shape[0:-1])
if y.ndim == 5:
if by_frame:
_distance_transform = transform_utils.inner_distance_transform_movie
else:
_distance_transform = transform_utils.inner_distance_transform_3d
else:
_distance_transform = transform_utils.inner_distance_transform_2d
for batch in range(y_transform.shape[0]):
if data_format == 'channels_first':
mask = y[batch, 0, ...]
else:
mask = y[batch, ..., 0]
y_transform[batch] = _distance_transform(mask, bins=bins,
erosion_width=erosion,
alpha=alpha, beta=beta)
y_transform = np.expand_dims(y_transform, axis=-1)
if bins is None:
pass
else:
# convert to one hot notation
y_transform = to_categorical(y_transform, num_classes=bins)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform == 'disc':
y_transform = to_categorical(y.squeeze(channel_axis))
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform == 'fgbg':
y_transform = np.where(y > 1, 1, y)
# convert to one hot notation
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, 1, y.ndim)
y_transform = to_categorical(y_transform)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform is None:
y_transform = to_categorical(y.squeeze(channel_axis))
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
return y_transform
def _transform_masks(y, transform, data_format=None, **kwargs):
"""Based on the transform key, apply a transform function to the masks.
More detailed description. Caution for unknown transform keys.
Args:
y (numpy.array): Labels of ndim 4 or 5
transform (str): Name of the transform, one of
{"deepcell", "disc", "watershed", None}
data_format (str): One of 'channels_first', 'channels_last'.
kwargs (dict): Optional transform keyword arguments.
Returns:
numpy.array: the output of the given transform function on y
Raises:
ValueError: Rank of y is not 4 or 5.
ValueError: Channel dimension of y is not 1.
ValueError: Transform is invalid value.
"""
valid_transforms = {
'pixelwise',
'disc',
'watershed',
'centroid',
'fgbg'
}
if data_format is None:
data_format = K.image_data_format()
if y.ndim not in {4, 5}:
raise ValueError('`labels` data must be of ndim 4 or 5. Got', y.ndim)
channel_axis = 1 if data_format == 'channels_first' else -1
if y.shape[channel_axis] != 1:
raise ValueError('Expected channel axis to be 1 dimension. Got',
y.shape[1 if data_format == 'channels_first' else -1])
if isinstance(transform, str):
transform = transform.lower()
if transform == 'deepcell':
warnings.warn('The `deepcell` transform is deprecated. '
'Please use the`pixelwise` transform insetad.',
DeprecationWarning)
transform = 'pixelwise'
if transform not in valid_transforms:
raise ValueError('`{}` is not a valid transform'.format(transform))
if transform == 'pixelwise':
dilation_radius = kwargs.pop('dilation_radius', None)
separate_edge_classes = kwargs.pop('separate_edge_classes', False)
y_transform = transform_utils.pixelwise_transform(
y, dilation_radius,
data_format=data_format,
separate_edge_classes=separate_edge_classes)
elif transform == 'watershed':
distance_bins = kwargs.pop('distance_bins', 4)
erosion = kwargs.pop('erosion_width', 0)
if data_format == 'channels_first':
y_transform = np.zeros(tuple([y.shape[0]] + list(y.shape[2:])))
else:
y_transform = np.zeros(y.shape[0:-1])
if y.ndim == 5:
_distance_transform = transform_utils.distance_transform_3d
else:
_distance_transform = transform_utils.distance_transform_2d
for batch in range(y_transform.shape[0]):
if data_format == 'channels_first':
mask = y[batch, 0, ...]
else:
mask = y[batch, ..., 0]
y_transform[batch] = _distance_transform(
mask, distance_bins, erosion)
# convert to one hot notation
y_transform = np.expand_dims(y_transform, axis=-1)
y_transform = to_categorical(y_transform, num_classes=distance_bins)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform == 'disc':
y_transform = to_categorical(y.squeeze(channel_axis))
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform == 'fgbg':
y_transform = np.where(y > 1, 1, y)
# convert to one hot notation
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, 1, y.ndim)
y_transform = to_categorical(y_transform)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform is None:
y_transform = to_categorical(y.squeeze(channel_axis))
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
return y_transform
def _transform_masks(y, transform, data_format=None, **kwargs):
"""Based on the transform key, apply a transform function to the masks.
Refer to :mod:`deepcell.utils.transform_utils` for more information about
available transforms. Caution for unknown transform keys.
Args:
y (numpy.array): Labels of ``ndim`` 4 or 5
transform (str): Name of the transform, one of
``{"deepcell", "disc", "watershed", None}``.
data_format (str): A string, one of ``channels_last`` (default)
or ``channels_first``. The ordering of the dimensions in the
inputs. ``channels_last`` corresponds to inputs with shape
``(batch, height, width, channels)`` while ``channels_first``
corresponds to inputs with shape
``(batch, channels, height, width)``.
kwargs (dict): Optional transform keyword arguments.
Returns:
numpy.array: the output of the given transform function on ``y``.
Raises:
ValueError: Rank of ``y`` is not 4 or 5.
ValueError: Channel dimension of ``y`` is not 1.
ValueError: ``transform`` is invalid value.
"""
valid_transforms = {
'deepcell', # deprecated for "pixelwise"
'pixelwise',
'disc',
'watershed', # deprecated for "outer-distance"
'watershed-cont', # deprecated for "outer-distance"
'inner-distance',
'inner_distance',
'outer-distance',
'outer_distance',
'centroid', # deprecated for "inner-distance"
'fgbg'
}
if data_format is None:
data_format = K.image_data_format()
if y.ndim not in {4, 5}:
raise ValueError('`labels` data must be of ndim 4 or 5. Got', y.ndim)
channel_axis = 1 if data_format == 'channels_first' else -1
if y.shape[channel_axis] != 1:
raise ValueError('Expected channel axis to be 1 dimension. Got',
y.shape[1 if data_format == 'channels_first' else -1])
if isinstance(transform, str):
transform = transform.lower()
if transform not in valid_transforms and transform is not None:
raise ValueError('`{}` is not a valid transform'.format(transform))
if transform in {'pixelwise', 'deepcell'}:
if transform == 'deepcell':
warnings.warn(
'The `{}` transform is deprecated. Please use the '
'`pixelwise` transform instead.'.format(transform),
DeprecationWarning)
dilation_radius = kwargs.pop('dilation_radius', None)
separate_edge_classes = kwargs.pop('separate_edge_classes', False)
edge_class_shape = 4 if separate_edge_classes else 3
if data_format == 'channels_first':
shape = tuple([y.shape[0]] + [edge_class_shape] +
list(y.shape[2:]))
else:
shape = tuple(list(y.shape[0:-1]) + [edge_class_shape])
# using uint8 since should only be 4 unique values.
y_transform = np.zeros(shape, dtype=np.uint8)
for batch in range(y_transform.shape[0]):
if data_format == 'channels_first':
mask = y[batch, 0, ...]
else:
mask = y[batch, ..., 0]
y_transform[batch] = transform_utils.pixelwise_transform(
mask,
dilation_radius,
data_format=data_format,
separate_edge_classes=separate_edge_classes)
elif transform in {
'outer-distance', 'outer_distance', 'watershed', 'watershed-cont'
}:
if transform in {'watershed', 'watershed-cont'}:
warnings.warn(
'The `{}` transform is deprecated. Please use the '
'`outer-distance` transform instead.'.format(transform),
DeprecationWarning)
by_frame = kwargs.pop('by_frame', True)
bins = kwargs.pop('distance_bins', None)
distance_kwargs = {
'bins': bins,
'erosion_width': kwargs.pop('erosion_width', 0),
}
# If using 3d transform, pass in scale arg
if y.ndim == 5 and not by_frame:
distance_kwargs['sampling'] = kwargs.pop('sampling',
[0.5, 0.217, 0.217])
if data_format == 'channels_first':
shape = tuple([y.shape[0]] + list(y.shape[2:]))
else:
shape = y.shape[0:-1]
y_transform = np.zeros(shape, dtype=K.floatx())
if y.ndim == 5:
if by_frame:
_distance_transform = transform_utils.outer_distance_transform_movie
else:
_distance_transform = transform_utils.outer_distance_transform_3d
else:
_distance_transform = transform_utils.outer_distance_transform_2d
for batch in range(y_transform.shape[0]):
if data_format == 'channels_first':
mask = y[batch, 0, ...]
else:
mask = y[batch, ..., 0]
y_transform[batch] = _distance_transform(mask, **distance_kwargs)
y_transform = np.expand_dims(y_transform, axis=-1)
if bins is not None:
# convert to one hot notation
# uint8's max value of255 seems like a generous limit for binning.
y_transform = to_categorical(y_transform,
num_classes=bins,
dtype=np.uint8)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform in {'inner-distance', 'inner_distance', 'centroid'}:
if transform == 'centroid':
warnings.warn(
'The `{}` transform is deprecated. Please use the '
'`inner-distance` transform instead.'.format(transform),
DeprecationWarning)
by_frame = kwargs.pop('by_frame', True)
bins = kwargs.pop('distance_bins', None)
distance_kwargs = {
'bins': bins,
'erosion_width': kwargs.pop('erosion_width', 0),
'alpha': kwargs.pop('alpha', 0.1),
'beta': kwargs.pop('beta', 1)
}
# If using 3d transform, pass in scale arg
if y.ndim == 5 and not by_frame:
distance_kwargs['sampling'] = kwargs.pop('sampling',
[0.5, 0.217, 0.217])
if data_format == 'channels_first':
shape = tuple([y.shape[0]] + list(y.shape[2:]))
else:
shape = y.shape[0:-1]
y_transform = np.zeros(shape, dtype=K.floatx())
if y.ndim == 5:
if by_frame:
_distance_transform = transform_utils.inner_distance_transform_movie
else:
_distance_transform = transform_utils.inner_distance_transform_3d
else:
_distance_transform = transform_utils.inner_distance_transform_2d
for batch in range(y_transform.shape[0]):
if data_format == 'channels_first':
mask = y[batch, 0, ...]
else:
mask = y[batch, ..., 0]
y_transform[batch] = _distance_transform(mask, **distance_kwargs)
y_transform = np.expand_dims(y_transform, axis=-1)
if distance_kwargs['bins'] is not None:
# convert to one hot notation
# uint8's max value of255 seems like a generous limit for binning.
y_transform = to_categorical(y_transform,
num_classes=bins,
dtype=np.uint8)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform == 'disc' or transform is None:
dtype = K.floatx() if transform == 'disc' else np.int32
y_transform = to_categorical(y.squeeze(channel_axis), dtype=dtype)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
elif transform == 'fgbg':
y_transform = np.where(y > 1, 1, y)
# convert to one hot notation
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, 1, y.ndim)
# using uint8 since should only be 2 unique values.
y_transform = to_categorical(y_transform, dtype=np.uint8)
if data_format == 'channels_first':
y_transform = np.rollaxis(y_transform, y.ndim - 1, 1)
return y_transform
