def discretize(segs, n_class, n_sample, rand):
"""
Convert a set of segmentations into a set of labels in [0, n_class - 1]
:param segs: segmentations
:param n_class: number of classes (clusters) for binary splits
:param n_sample: number of samples for clustering structured labels
:param rand: random number generator
"""
w = segs[0].shape[0]
segs = segs.reshape((segs.shape[0], w ** 2))
# compute all possible lookup inds for w x w patches
ids = N.arange(w ** 4, dtype=N.float64)
ids1 = N.floor(ids / w / w)
ids2 = ids - ids1 * w * w
kp = ids2 > ids1
ids1 = ids1[kp]
ids2 = ids2[kp]
# compute n binary codes zs of length nSamples
n_sample = min(n_sample, ids1.shape[0])
kp = rand.permutation(ids1.shape[0])[:n_sample]
n = segs.shape[0]
ids1 = ids1[kp].astype(N.int32)
ids2 = ids2[kp].astype(N.int32)
zs = N.zeros((n, n_sample), dtype=N.float64)
for i in xrange(n):
zs[i] = (segs[i][ids1] == segs[i][ids2])
zs -= N.mean(zs, axis=0)
zs = zs[:, N.any(zs, axis=0)]
if N.count_nonzero(zs) == 0:
lbls = N.ones(n, dtype=N.int32)
segs = segs[0]
else:
# find most representative segs (closest to mean)
ind = N.argmin(N.sum(zs * zs, axis=1))
segs = segs[ind]
# discretize zs by discretizing pca dimensions
d = min(5, n_sample, int(floor(log(n_class, 2))))
zs = robust_pca(zs, d, rand=rand)[0]
lbls = N.zeros(n, dtype=N.int32)
for i in xrange(d):
lbls += (zs[:, i] < 0).astype(N.int32) * 2 ** i
lbls = N.unique(lbls, return_inverse=True)[1].astype(N.int32)
return lbls, segs.reshape((-1, w, w))
def discretize(segs, n_class, n_sample, rand):
"""
Convert a set of segmentations into a set of labels in [0, n_class - 1]
:param segs: segmentations
:param n_class: number of classes (clusters) for binary splits
:param n_sample: number of samples for clustering structured labels
:param rand: random number generator
"""
w = segs[0].shape[0]
segs = segs.reshape((segs.shape[0], w ** 2))
# compute all possible lookup inds for w x w patches
ids = N.arange(w ** 4, dtype=N.float64)
ids1 = N.floor(ids / w / w)
ids2 = ids - ids1 * w * w
kp = ids2 > ids1
ids1 = ids1[kp]
ids2 = ids2[kp]
# compute n binary codes zs of length nSamples
n_sample = min(n_sample, ids1.shape[0])
kp = rand.permutation(ids1.shape[0])[:n_sample]
n = segs.shape[0]
ids1 = ids1[kp].astype(N.int32)
ids2 = ids2[kp].astype(N.int32)
zs = N.zeros((n, n_sample), dtype=N.float64)
for i in xrange(n):
zs[i] = (segs[i][ids1] == segs[i][ids2])
zs -= N.mean(zs, axis=0)
zs = zs[:, N.any(zs, axis=0)]
if N.count_nonzero(zs) == 0:
lbls = N.ones(n, dtype=N.int32)
segs = segs[0]
else:
# find most representative segs (closest to mean)
ind = N.argmin(N.sum(zs * zs, axis=1))
segs = segs[ind]
# discretize zs by discretizing pca dimensions
d = min(5, n_sample, int(floor(log(n_class, 2))))
zs = robust_pca(zs, d, rand=rand)[0]
lbls = N.zeros(n, dtype=N.int32)
for i in xrange(d):
lbls += (zs[:, i] < 0).astype(N.int32) * 2 ** i
lbls = N.unique(lbls, return_inverse=True)[1].astype(N.int32)
return lbls, segs.reshape((-1, w, w))