filter.py 文件源码

def compute_forwards(self, reuse=None):
        """Compute the forward step in the Kalman filter.
           The forward pass is intialized with p(z_1)=N(self.mu, self.Sigma).
           We then return the mean and covariances of the predictive distribution p(z_t|z_tm1,u_t), t=2,..T+1
           and the filtering distribution p(z_t|x_1:t,u_1:t), t=1,..T
           We follow the notation of Murphy's book, section 18.3.1
        """

        # To make sure we are not accidentally using the real outputs in the steps with missing values, set them to 0.
        y_masked = tf.multiply(tf.expand_dims(self.mask, 2), self.y)
        inputs = tf.concat([y_masked, self.u, tf.expand_dims(self.mask, 2)], axis=2)

        y_prev = tf.expand_dims(self.y_0, 0)  # (1, dim_y)
        y_prev = tf.tile(y_prev, (tf.shape(self.mu)[0], 1))
        alpha, state, u, buffer = self.alpha(y_prev, self.state, self.u[:, 0], init_buffer=True, reuse= reuse)

        # dummy matrix to initialize B and C in scan
        dummy_init_A = tf.ones([self.Sigma.get_shape()[0], self.dim_z, self.dim_z])
        dummy_init_B = tf.ones([self.Sigma.get_shape()[0], self.dim_z, self.dim_u])
        dummy_init_C = tf.ones([self.Sigma.get_shape()[0], self.dim_y, self.dim_z])
        forward_states = tf.scan(self.forward_step_fn, tf.transpose(inputs, [1, 0, 2]),
                                 initializer=(self.mu, self.Sigma, self.mu, self.Sigma, alpha, u, state, buffer,
                                              dummy_init_A, dummy_init_B, dummy_init_C),
                                 parallel_iterations=1, name='forward')
        return forward_states