Sampler.py 文件源码

def sample_at_prob(self, prob, mean, var, rstate=None):
        """
        """
        shape = mean.shape[0]
        # Get a sample from a distribution N(0,I)
        scale = spstat.norm.ppf(prob, loc=0, scale=1)
        v = spstat.multivariate_normal.rvs(
            mean=None,
            cov=1,
            size=shape,
            random_state=rstate)

        v *= np.fabs(scale) / np.sqrt((v**2).sum())

        # Spectral decomposition of target dist covariance
        eigs, vects = np.linalg.eigh(var)

        assert eigs.shape[0] == shape, 'Too few eigenvalues'
        assert np.all(eigs >= 0.0), 'Negative eigenvalues'
        assert np.all(np.isreal(eigs)), 'Imaginary eigenvalues'
        assert np.all(np.fabs(
            np.dot(vects.T, vects) - np.eye(shape)) < 1E-14),\
            'Eigenvectors are not orthogonal'


        # Calculate map from N(0,I) to N(0,E)
        a_mat = np.dot(vects.T, np.dot(np.diag(np.sqrt(eigs)), vects))

        # Add the mean to get N(u,E)
        #return mean + np.copysign(np.dot(a_mat, v), scale)

        return mean + scale * np.diag(np.sqrt(var))