Bayesian_Source_Space_Connectivity_InvWishart_Prior.py 文件源码

def get_neg_log_post_grad_Qu(Phi, G, MMT, q, Sigma_E, L,Sigma_J, nu, V, GL, prior_on = False):
    """
        Just get the gradient for Qu, ignore other variables
    """
    p = Phi.shape[0]
    Qu = Phi.dot(Phi.T)
    G_Sigma_G = np.zeros(MMT.shape)
    for i in range(Sigma_J.size):
        G_Sigma_G += Sigma_J[i]* np.outer(G[:,i], G[:,i])
    cov = Sigma_E + G_Sigma_G + GL.dot(Qu).dot(GL.T) 
    inv_cov = np.linalg.inv(cov)
    GLT_inv_cov = np.dot(GL.T, inv_cov)
    invQ = np.linalg.inv(Qu)
    if prior_on:
        grad0 = (q*  GL.T.dot(inv_cov).dot(GL) - GLT_inv_cov.dot(MMT).dot(GLT_inv_cov.T) \
           + invQ.dot( (nu+p+1) *np.eye(p) - V.dot(invQ)))
    else:
        grad0 = q* GL.T.dot(inv_cov).dot(GL) - GLT_inv_cov.dot(MMT).dot(GLT_inv_cov.T)
    grad1 = 2.0* grad0.dot(Phi)
    # cholesky decomposition is lower triangular
    grad = np.tril(grad1)
    return grad

#==============================================================================
# gradient descent optimization, using back track
# only update Qu