def score_samples(self, X):
"""Compute the pseudo-likelihood of X.
X : {array-like, sparse matrix} shape (n_samples, n_features)
Values of the visible layer. Must be all-boolean (not checked).
Returns
-------
pseudo_likelihood : array-like, shape (n_samples,)
Value of the pseudo-likelihood (proxy for likelihood).
Notes
-----
This method is not deterministic: it computes a quantity called the
free energy on X, then on a randomly corrupted version of X, and
returns the log of the logistic function of the difference.
"""
check_is_fitted(self, "components_")
v = check_array(X, accept_sparse='csr')
fe = self._free_energy(v)
v_, state = self.corrupt(v)
# TODO: If I wanted to be really fancy here, I would do one of those "with..." things.
fe_corrupted = self._free_energy(v)
self.uncorrupt(v, state)
# See https://en.wikipedia.org/wiki/Pseudolikelihood
# Let x be some visible vector. x_i is the ith entry. x_-i is the vector except that entry.
# x_iflipped is x with the ith bit flipped. F() is free energy.
# P(x_i | x_-i) = P(x) / P(x_-i) = P(x) / (P(x) + p(x_iflipped))
# expand def'n of P(x), cancel out the partition function on each term, and divide top and bottom by e^{-F(x)} to get...
# 1 / (1 + e^{F(x) - F(x_iflipped)})
# So we're just calculating the log of that. We multiply by the number of
# visible units because we're approximating P(x) as the product of the conditional likelihood
# of each individual unit. But we're too lazy to do each one individually, so we say the unit
# we tested represents an average.
if hasattr(self, 'codec'):
normalizer = self.codec.shape()[0]
else:
normalizer = v.shape[1]
return normalizer * log_logistic(fe_corrupted - fe)
# TODO: No longer used
评论列表
文章目录
