Hey guys,
I’m looking through the code for the EKF and Dynamic Models and have a few questions (I’m particularly interested in using these libraries for purposes of writing my own dynamic model).
- I’m curious why you guys put use no_grad when writing the jacobians. For example I see the code:
def jacobian(self, dt):
"""
Compute and return cached native state transition Jacobian (F) over
time interval ``dt``.
:param dt: time interval to integrate over.
:return: Read-only Jacobian (F) of integration map (f).
"""
if dt not in self._F_cache:
d = self._dimension
with torch.no_grad():
F = eye_like(self.sa2, d)
F[: d // 2, d // 2 :] = dt * eye_like(self.sa2, d // 2)
self._F_cache[dt] = F
return self._F_cache[dt]
It appears this means that this particular dynamic model will not allow learning to propagate through the matrix F here. Does this mean that learning the transition jacobian is not an intention of the EKF?
- It appears the EKF supports filtering, but not smoothing. Is this correct? For example, I’m seeing this code for the EKFDistribution:
def filter_states(self, value):
"""
Returns the ekf states given measurements
:param value: measurement means of shape `(time_steps, event_shape)`
:type value: torch.Tensor
"""
states = []
state = EKFState(self.dynamic_model, self.x0, self.P0, time=0.0)
assert value.shape[-1] == self.event_shape[-1]
for i, measurement_mean in enumerate(value):
if i:
state = state.predict(self.dt)
measurement = PositionMeasurement(
measurement_mean, self.measurement_cov, time=state.time
)
state, (dz, S) = state.update(measurement)
states.append(state)
return states
- How do you specify the observation matrix? I see that there is measurement_cov, but I don’t see where the observation matrix goes.
It appears this is only forward propagation for message passing. I don’t see any backward message passing. I was curious if this is indeed true, and if it is, whether backward propagation will be supported in the future.