there are two small issues with the documentation of experimental.sequential.extended_kalman_filter.
1) The code in the example does not run, giving
InvalidArgumentError: cannot compute Sub as input #1(zero-based) was expected to be a double tensor but is a float tensor [Op:Sub] name.
Changing
x = [np.zeros((2,), dtype=np.float32)]
to
x = [tf.zeros((2,), dtype=tf.float32)]
solves this.
2) The documentation states
observation_jacobian_fn: a Python `callable` that accepts a (batched) vector
of length `state_size` and returns a (batched) matrix of size
`[state_size, event_size]`, representing the Jacobian of `observation_fn`.
but the correct matrix size to be returned should be [event_size, state_size].
In the example, making observation_jacobian_fn return a Tensor of size [state_size, event_size] will cause the Kalman Filter to fail, while [event_size, state_size] works.
Hi all,
there are two small issues with the documentation of experimental.sequential.extended_kalman_filter.
1) The code in the example does not run, giving
InvalidArgumentError: cannot compute Sub as input #1(zero-based) was expected to be a double tensor but is a float tensor [Op:Sub] name.Changing
x = [np.zeros((2,), dtype=np.float32)]tox = [tf.zeros((2,), dtype=tf.float32)]solves this.2) The documentation states
but the correct matrix size to be returned should be
[event_size, state_size].In the example, making
observation_jacobian_fnreturn a Tensor of size[state_size, event_size]will cause the Kalman Filter to fail, while[event_size, state_size]works.See also this colab https://colab.research.google.com/drive/1vUNMHQ1Fc3CwdpOMnKXqzqmQ2zM4A8Pp?usp=sharing
Thank you for the great package!