def get_um_loss_gradient(mst, dist, gt_seg, alpha):
'''Compute the ultra-metric loss gradient given an MST and segmentation.
Args:
mst (Tensor, shape ``(3, n-1)``): u, v indices and distance of edges of
the MST spanning n nodes.
dist (Tensor, shape ``(n-1)``): The distances of the edges. This
argument will be ignored, it is used only to communicate to
tensorflow that there is a dependency on distances. The distances
actually used are the ones in parameter ``mst``.
gt_seg (Tensor, arbitrary shape): The label of each node. Will be
flattened. The indices in mst should be valid indices into this
array.
alpha (Tensor, single float): The margin value of the quadrupel loss.
Returns:
A Tensor containing the gradient on the distances.
'''
# We don't use 'dist' here, it is already contained in the mst. It is
# passed here just so that tensorflow knows there is dependecy to the
# ouput.
(_, gradient, _, _, _, _) = mala.um_loss(mst, gt_seg.flatten(), alpha)
return gradient.astype(np.float32)
def get_um_loss(mst, dist, gt_seg, alpha):
'''Compute the ultra-metric loss given an MST and segmentation.
Args:
mst (Tensor, shape ``(3, n-1)``): u, v indices and distance of edges of
the MST spanning n nodes.
dist (Tensor, shape ``(n-1)``): The distances of the edges. This
argument will be ignored, it is used only to communicate to
tensorflow that there is a dependency on distances. The distances
actually used are the ones in parameter ``mst``.
gt_seg (Tensor, arbitrary shape): The label of each node. Will be
flattened. The indices in mst should be valid indices into this
array.
alpha (Tensor, single float): The margin value of the quadrupel loss.
Returns:
A tuple::
(loss, ratio_pos, ratio_neg)
Except for ``loss``, each entry is a tensor of shape ``(n-1,)``,
corresponding to the edges in the MST. ``ratio_pos`` and ``ratio_neg``
are the ratio of positive and negative pairs that share an edge, of the
total number of positive and negative pairs.
'''
# We don't use 'dist' here, it is already contained in the mst. It is
# passed here just so that tensorflow knows there is dependecy to the
# ouput.
(loss, _, ratio_pos, ratio_neg, num_pairs_pos,
num_pairs_neg) = mala.um_loss(mst, gt_seg.flatten(), alpha)
return (np.float32(loss), ratio_pos.astype(np.float32),
ratio_neg.astype(np.float32), np.float32(num_pairs_pos),
np.float32(num_pairs_neg))