def score_samples(X,model_weights,model_covars,use_scipy_misc=False):
inv_covars = 1.0/model_covars
n_features = X.shape[1]
lpr= - 0.5 * ( n_features * numpy.log(2*numpy.pi) +
numpy.sum(numpy.log(model_covars),1)
+ numpy.dot(numpy.abs(X)**2, inv_covars.T))
if numpy.any(numpy.isnan(lpr)):
import pdb; pdb.set_trace()
lpr += numpy.log(model_weights)
logprob = logsumexp(lpr)
responsibilities = numpy.exp(lpr - logprob[:,numpy.newaxis])
return logprob, responsibilities
def score_samples(X,model_weights,model_covars,use_scipy_misc=False):
n_samples, n_dim = X.shape
nmix = len(model_covars)
lpr = numpy.empty((n_samples, nmix))
for c, cv in enumerate(model_covars):
cv_chol = linalg.cholesky(cv, lower=True)
cv_log_det = 2 * numpy.sum(numpy.log(numpy.diagonal(cv_chol)))
cv_sol = linalg.solve_triangular(cv_chol, X.T, lower=True).T
lpr[:, c] = - .5 * (numpy.sum(cv_sol ** 2, axis=1) +
n_dim * numpy.log(2 * numpy.pi) + cv_log_det)
if numpy.any(numpy.isnan(lpr)):
import pdb; pdb.set_trace()
lpr += numpy.log(model_weights)
logprob = logsumexp(lpr)
responsibilities = numpy.exp(lpr - logprob[:,numpy.newaxis])
return logprob, responsibilities