mmd.py 文件源码

def _mmd2_and_ratio(K_XX, K_XY, K_YY, const_diagonal=False, biased=False):
    mmd2, var_est = _mmd2_and_variance(K_XX, K_XY, K_YY, const_diagonal=const_diagonal, biased=biased)
    loss = mmd2 / torch.sqrt(torch.clamp(var_est, min=min_var_est))
    return loss, mmd2, var_est