Diagonal EK1 #30
Diagonal EK1 #30
Conversation
| assert isinstance(P, linops.BlockDiagonal) | ||
| assert isinstance(Pinv, linops.BlockDiagonal) | ||
| assert P.array_stack.shape == (d, n, n) | ||
| assert Pinv.array_stack.shape == (d, n, n) |
There was a problem hiding this comment.
there are a million asserts everywhere. I would propose to leave them while the thing is not complete (it made it much easier to detect issues while developing). I can remove them, though, if you prefer me to.
There was a problem hiding this comment.
it might also be nice as a little bit of documentation for the current, early phase :)
| # Due to the block-diagonal structure in H (and in C), S is diagonal! | ||
| # We can compute the correction really cheaply (like in the EK0, actually) |
There was a problem hiding this comment.
might have been clear to you guys, to me it wasnt and I am really surprised by this!
tornado/ek1.py
Outdated
| innov_stds = innov_chol.array_stack[:, 0, 0] # (was shape (d,1,1)) | ||
| s = z / innov_stds | ||
| sigma_squared_increment = s.T @ s / d | ||
| assert isinstance(sigma_squared_increment, jnp.ndarray) | ||
| assert sigma_squared_increment.shape == () |
There was a problem hiding this comment.
to do the calibration-for-error-estimation, I use the "increment" from the global sigma quasi-MLE. Am I bastardising this too much? It seems that for now, it works okay. Stiff van-der-pol is a different PR though anyway ;) (... I actually tried! :D More on this later...)
| @@ -11,8 +11,8 @@ class MultivariateNormal: | |||
| """Multivariate normal distributions.""" | |||
|
|
|||
| mean: jax.numpy.ndarray | |||
| cov_cholesky: jax.numpy.ndarray | |||
| cov_sqrtm: jax.numpy.ndarray | |||
There was a problem hiding this comment.
see #31. If you dont like this, we can discuss there? Or comment here, I am fine with both...
| @@ -6,10 +6,10 @@ isolated_build = True | |||
| description = Executing tests with pytest | |||
| deps = | |||
| pytest | |||
| pytest-randomly | |||
There was a problem hiding this comment.
shuffles the tests each time they are executed. I used this to assert that there was no weird in-place/variable-lifetime behaviour, and kind of liked it :)
|
I will merge now. there are three PRs open, and one of us has to start :) |
This PR establishes a version of the EK1 with diagonal Jacobians (which always -- successfully -- pushes around block-diagonal matrices).
Adaptive steps? Yes.
Calibration? dynamic.