Skip to content

Instantly share code, notes, and snippets.

@kbarbary
Last active September 8, 2015 17:26
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save kbarbary/2a169680e52bd7089950 to your computer and use it in GitHub Desktop.
Save kbarbary/2a169680e52bd7089950 to your computer and use it in GitHub Desktop.
Ellipsoid refinement code for Nestle
def ellipsoid_dist_sq(x, ell):
"""Return the square of each point's distance from the ellipsoid center,
relative to ellipsoid boundary.
Refered to in Feroz, Hobson & Bridges (2009) as the Mahalanobis distance.
Parameters
----------
x : `~numpy.ndarray`
Points, array with shape (npoints, ndim).
ell : Ellipsoid
Ellipsoid to compare to.
"""
delta = x - ell.ctr
# fast way to compute delta[i] @ A @ delta[i] for all i.
return np.einsum('...i, ...i', np.tensordot(delta, ell.a, axes=1), delta)
def refine_ellipsoids(x, label, pointvol, xs, ells):
"""Reassign points between two clusters so as to minimize total volume
of the ellipsoids bounding each cluster.
This is as described in Feroz, Hobson & Bridges (2009), Algorithm 1.
There, the distance measure is referred to as the "Mahalanobis distance".
Parameters
----------
x : `~numpy.ndarray`
All points in either cluster.
label : `~numpy.ndarray`
1-d integer array of 0 or 1 giving cluster membership of each point
in x.
pointvol : float
Expected volume corresponding to each point.
"""
hs = [None, None]
ndim = x.shape[1]
for it in range(10): # limit to 10 iterations
for k in (0, 1):
hs[k] = ellipsoid_dist_sq(x, ells[k])
hs[k] *= ells[k].vol / (len(xs[k]) * pointvol) # actual/expected vol
# reassign each point to the cluster that gives it the smallest h.
# if hs[1] < hs[0] -> True -> (cast to int) -> 1 (label cluster 1)
# if hs[0] < hs[1] -> False -> (cast to int) -> 0 (label cluster 0)
newlabel = (hs[1] < hs[0]).astype(np.int)
# If no points were reassigned, exit the loop.
if np.all(newlabel == label):
break
# update labels, calculate new member points
label = newlabel
new_xs = [x[label == k, :] for k in (0, 1)]
# if either cluster doesn't have enough points, return previous
# clusters
if new_xs[0].shape[0] <= ndim or new_xs[1].shape[0] <= ndim:
return xs, ells
for k in (0, 1):
xs[k] = new_xs[k]
ells[k] = bounding_ellipsoid(xs[k], pointvol=pointvol,
minvol=True)
@ipashchenko
Copy link

I think there is typo in line 39: hs instead of h...

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment