GaussianProcess.py 文件源码

def param_grad(self, k_param):
        """
        Returns gradient over hyperparameters. It is recommended to use `self._grad` instead.

        Parameters
        ----------
        k_param: dict
            Dictionary with keys being hyperparameters and values their queried values.

        Returns
        -------
        np.ndarray
            Gradient corresponding to each hyperparameters. Order given by `k_param.keys()`
        """
        k_param_key = list(k_param.keys())
        covfunc = self.covfunc.__class__(**k_param)
        n = self.X.shape[0]
        K = covfunc.K(self.X, self.X)
        L = cholesky(K).T
        alpha = solve(L.T, solve(L, self.y))
        inner = np.dot(np.atleast_2d(alpha).T, np.atleast_2d(alpha)) - np.linalg.inv(K)
        grads = []
        for param in k_param_key:
            gradK = covfunc.gradK(self.X, self.X, param=param)
            gradK = .5 * np.trace(np.dot(inner, gradK))
            grads.append(gradK)
        return np.array(grads)