Question

因此，对于高斯过程的参数，我有以下MCMC代码

代码：

    walkers = 250
    pos_min = np.concatenate((np.array([self.initial_scale, self.initial_noise]), np.zeros(self.ndim)))
    pos_max = np.concatenate((np.array([self.initial_scale, self.initial_noise]), 2.0*np.ones(self.ndim)))


    psize = pos_max - pos_min
    pos = [pos_min + psize*np.random.rand(self.ndim+2) for i in range(nwalkers)]
    with Pool(4) as pool:
      sampler = emcee.EnsembleSampler(nwalkers, self.ndim+2, self.lnprob, pool=pool)
      pos, prob, state = sampler.run_mcmc(pos, 200)
      sampler.reset()
      sampler.run_mcmc(pos, 300)
    samples = sampler.flatchain[-500:,:]
    return samples

函数是self.lnprob在以下代码段中定义

代码

   def lnprior(theta):
      l = theta[2:]
      s2_f = theta[0]
      s2_n = theta[1]
      if 0 < s2_f  and 0 < s2_n  and (l > 0).all() and (l < 2.0).all():
        return np.log(np.log(1 + (0.1/s2_n)**2)) -0.5*(np.log(np.sqrt(s2_f))/1.0)**2 
      return -np.inf

    def lnlike(theta):
      l = theta[2:]
      sigma_f = theta[0]
      sigman  = theta[1]
      self.kernel.k1.k1.constant_value = sigma_f
      self.kernel.k1.k2.length_scale = l
      self.kernel.k2.noise_level = sigman

      K = self.kernel(self.Xi)

      K[np.diag_indices_from(K)] += self.alpha

      L = cholesky(K, lower=True)  # Line 2


      # Support multi-dimensional output of self.y_train_
      y_train = self.Yi
      if y_train.ndim == 1:
        y_train = y_train[:, np.newaxis]

      alpha = cho_solve((L, True), y_train)  # Line 3

      # Compute log-likelihood (compare line 7)
      log_likelihood_dims = -0.5 * np.einsum("ik,ik->k", y_train, alpha)
      log_likelihood_dims -= np.log(np.diag(L)).sum()
      log_likelihood_dims -= K.shape[0] / 2 * np.log(2 * np.pi)
      log_likelihood = log_likelihood_dims.sum(-1)
      return log_likelihood


    lp = lnprior(theta)
    if not np.isfinite(lp):
      return -np.inf
    return lp + lnlike(theta)

代码在我的笔记本电脑（这是Intel Core i5第7代）上花费30秒钟。 theta数组的大小为4，Self.Xi的大小为2。

相同的代码在具有64核的Intel xeon phi处理器上花费超过400s。

此外，我注意到随着多处理池中池数量的增加，时间似乎也在增加。

在英特尔至强phi处理器上花费这么多时间是否有任何原因？我该怎么做才能在Intel Xeon phi处理器上加快速度？

使用emcee的MCMC采样在Intel Xeon phi处理器上运行非常慢

0 个答案: