当使用pykalman (Python卡尔曼滤波器)时，对数似然函数采用什么数据类型？

Question

我正在学习如何使用pykalman package。在使用机器人数据的卡尔曼EM示例中，我遇到了一个错误。EM示例代码是here。机器人示例附带了用于生成数据的描述和代码。

当我运行此部分时：

    # Learn good values for parameters named in `em_vars` using the EM algorithm
    loglikelihoods = np.zeros(10)
    for i in range(len(loglikelihoods)):
    kf = kf.em(X=data.observations, n_iter=1)
    loglikelihoods[i] = kf.loglikelihood(data.observations)

对于kf.loglikelihood(data.observations)函数，我得到以下错误，指出不支持掩码数组。我应该使用什么作为kf.loglikelihood的参数呢？

    C:\Users\Jenny\Anaconda3\lib\site-packages\pykalman-0.9.5-py3.5.egg\pykalman\standard.py in loglikelihood(self, X)
       1472         loglikelihoods = _loglikelihoods(
       1473           observation_matrices, observation_offsets, observation_covariance,
    -> 1474           predicted_state_means, predicted_state_covariances, Z
       1475         )
       1476 

    C:\Users\Jenny\Anaconda3\lib\site-packages\pykalman-0.9.5-py3.5.egg\pykalman\standard.py in _loglikelihoods(observation_matrices, observation_offsets, observation_covariance, predicted_state_means, predicted_state_covariances, observations)
        168                 observation[np.newaxis, :],
        169                 predicted_observation_mean[np.newaxis, :],
    --> 170                 predicted_observation_covariance[np.newaxis, :, :]
        171             )
        172     return loglikelihoods

    C:\Users\Jenny\Anaconda3\lib\site-packages\pykalman-0.9.5-py3.5.egg\pykalman\utils.py in log_multivariate_normal_density(X, means, covars, min_covar)
         71                                       lower=True)
         72         cv_log_det = 2 * np.sum(np.log(np.diagonal(cv_chol)))
    ---> 73         cv_sol = solve_triangular(cv_chol, (X - mu).T, lower=True).T
         74         log_prob[:, c] = - .5 * (np.sum(cv_sol ** 2, axis=1) + \
         75                                      n_dim * np.log(2 * np.pi) + cv_log_det)

    C:\Users\Jenny\Anaconda3\lib\site-packages\scipy\linalg\basic.py in solve_triangular(a, b, trans, lower, unit_diagonal, overwrite_b, debug, check_finite)
        156     """
        157     a1 = _asarray_validated(a, check_finite=check_finite)
    --> 158     b1 = _asarray_validated(b, check_finite=check_finite)
        159     if len(a1.shape) != 2 or a1.shape[0] != a1.shape[1]:
        160         raise ValueError('expected square matrix')

    C:\Users\Jenny\Anaconda3\lib\site-packages\scipy\_lib\_util.py in _asarray_validated(a, check_finite, sparse_ok, objects_ok, mask_ok, as_inexact)
        183     if not mask_ok:
        184         if np.ma.isMaskedArray(a):
    --> 185             raise ValueError('masked arrays are not supported')
        186     toarray = np.asarray_chkfinite if check_finite else np.asarray
       187     a = toarray(a)

    ValueError: masked arrays are not supported

Answer 1

我添加了下面这行

observation=observation.filled()    

就在pykalman.standard.py程序包文件的第168行之前，其内容为：

loglikelihoods[t] = log_multivariate_normal_density(
            observation[np.newaxis, :],
            predicted_observation_mean[np.newaxis, :],
            predicted_observation_covariance[np.newaxis, :, :]
)

假设您有完整的观察结果，这个简单的变通方法将使loglikelihood()方法工作。希望这在包的后续迭代中不是必需的。

Answer 2

在Github上提到了pykalman的问题，solution对我很有效。

链接中提到的解决方案基本上是修改standard.py中第112到173行的代码：

def _loglikelihoods(observation_matrices, observation_offsets,
     for t in range(n_timesteps):
         observation = observations[t]
         if not np.any(np.ma.getmask(observation)):
+            observation_data = np.ma.getdata(observation)
             observation_matrix = _last_dims(observation_matrices, t)
             observation_offset = _last_dims(observation_offsets, t, ndims=1)
             predicted_state_mean = _last_dims(


             loglikelihoods[t] = log_multivariate_normal_density(
-                observation[np.newaxis, :],
+                observation_data[np.newaxis, :],
                 predicted_observation_mean[np.newaxis, :],
                 predicted_observation_covariance[np.newaxis, :, :]
             )