estimator.py 文件源码

def _calc_gamma_psi(log_d, alpha, log_beta, gamma, log_psi0):
    log_psi = log_psi0
    count = 0
    #print ((np.exp(log_psi1 - log_psi) ** 2).sum())

    while count == 0 \
        or ((np.exp(log_psi1) - np.exp(log_psi)) ** 2).sum() > 0.001 \
        or ((gamma1 - gamma) ** 2).sum() > 0.001:
        #print ('gamma psi:', count, ((np.exp(log_psi1) - np.exp(log_psi)) ** 2).sum())
        log_psi1 = log_psi
        gamma1 = gamma

        psi_offset = (digamma(gamma))[:, np.newaxis, np.newaxis, :]

        log_psi = log_beta[np.newaxis, :, :, :] + psi_offset
        log_psi = log_normalize(log_psi, axis=3)
        gamma = np.exp(logsumexp(logsumexp(log_d[:, :, :, np.newaxis] + log_psi, axis=1), axis=1)) + alpha[np.newaxis, :]
        count += 1

    #log_psi = np.average([log_psi0, log_psi], axis=0, weights=[0.9, 0.1])   # weak learning
    return (gamma, log_psi)