Seth Axen
Seth Axen
> 2. Make `InferenceData` a supported `chain_type` for `AbstractMCMC.sample` (https://beta.turing.ml/AbstractMCMC.jl/dev/api/#Chains), which would bypass `Chains`'s flattening entirely. I'm not sure this should live here, but it should not live in InferenceObjects....
Since there have been no objections to these steps, I'm going to move forward with opening a PR for Step 1.
There are a few more edge cases to handle, and docstrings and tests are still needed, but the PR should otherwise be function-complete.
The implementation of the generic shooting method is now quite general. It differs in 2 ways from the one in the algorithm in the paper: 1. the initial guess is...
> Cool! For the first point we could introduce a `ScaledProjectionRetraction` (which is a retraction as long as in the limit for smaller and smaller vectors the scaling tends to...
Should `shooting` and `ShootingInverseRetraction` be immediately upstreamed to ManifoldsBase?
We should probably also use this to implement `log` for `EuclideanMetric`.
> I agree, the shooting method is generic enough that it can be in ManifoldsBase.jl. Okay! I'll split it out into a PR directly to ManifoldsBase.
I opted to create a factorization type and implement only the necessary operations on this type. This is better documented, and it also allows us to still use generic algorithms....
> I just noticed that the retraction is only implemented for this new metric. Since retractions are defined just using the differential, they are actually retraction independent of which metric...