def LogME(f: tf.Tensor, y: tf.Tensor, regression=False):
f = f.numpy().astype(np.float64)
y = y.numpy()
if regression:
y = y.numpy().astype(np.float64)
fh = f
f = f.transpose()
D, N = f.shape
v, s, vh = np.linalg.svd(f @ fh, full_matrices=True)
evidences = []
if regression:
K = y.shape[1]
for i in range(K):
y_ = y[:, i]
evidence = each_evidence(y_, f, fh, v, s, vh, N, D)
evidences.append(evidence)
else:
K = int(y.max() + 1)
for i in range(K):
y_ = (y == i).astype(np.float64)
evidence = each_evidence(y_, f, fh, v, s, vh, N, D)
evidences.append(evidence)
return np.mean(evidences)
def jsd(p: tf.Tensor, q: tf.Tensor, base=np.e):
"""
Implementation of pairwise Jensen-Shannon Divergence based on
https://en.wikipedia.org/wiki/Jensen%E2%80%93Shannon_divergence
This returns NaNs for all zero probs (unlabeled).
"""
import scipy.stats
p, q = p.numpy(), q.numpy()
# normalize p, q to probabilities
p, q = p / p.sum(axis=-1, keepdims=True), q / q.sum(axis=-1,
keepdims=True)
p, q = p.transpose(), q.transpose()
m = 1.0 / 2 * (p + q)
jsd = (scipy.stats.entropy(p, m, base=base) / 2.0 +
scipy.stats.entropy(q, m, base=base) / 2.0)
jsd = np.clip(jsd, 0.0, 1.0).transpose()
return jsd
