You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
We originally had Kernel. But this is very overloaded across the ecosystem. Also, what we have isn't quite a Markov kernel, since it maps to more general measures. We had changed it to Kleisli, which is accurate, but the term is not in wide wide use, especially among those doing stats work.
This PR tries to find a middle ground with TransitionKernel. It's descriptive and more intuitive, while still being accurate. The smart constructor has been changed back to kernel, but we're no longer exporting it in order to avoid conflicts with other packages.
Conditional meausres
Soss and now Tilde use | for conditional models. But this should really be at the level of measures. So you can not express measure | constraint in the case where measure is over NamedTuples, and constraint is a particular named tuple.
One thing that may be surprising is that | in this PR does not change the dimensionality, but instead transforms the constrained dimension to a Dirac measure. I think this will be easier to later generalize to constraints that are not axis-aligned.
Likelihoods
Some updates here, as well as introduction of log_likelihood_ratio.
Pointwise Products
Slot names are changed to prior and likelihood for easier interpretation. I've added an iterate method, allowing μ, ℓ = (p::PointwiseProductMeasure).
Power measures
logdensity_def is simplified to use sum instead of an explicit for loop. Hopefully this will later make it easier to port to GPU methods.
PowerWeighted
In for example GLMs or Bayesian bootstrap, we want to consider each observation to have some weight. This is an exponent on the likelihood factor, or equivalently a multiplicative weight on the log-likelihood. Since ^ is already taken (and much more common than this) I've added
Merging #49 (417b2d0) into master (1d1ad53) will increase coverage by 1.28%.
The diff coverage is 23.65%.
❗ Current head 417b2d0 differs from pull request most recent head 11adfcf. Consider uploading reports for the commit 11adfcf to get more accurate results
The reason will be displayed to describe this comment to others. Learn more.
I did not look at everything in detail (this is a massive PR). But generally, it looks like many tests are missing, I think I caught some wrong implementations which would have been caught with some unit tests. At some point, I would be happy to help with the reorganization of the tests for the package.
I would not delete completely kleisli, it's still a nice thing to know and learn, and it can be exported. Maybe just leave it as an alias to kernel ? It's just my opinion though.
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Kernel/Kleisli contention
We originally had
Kernel. But this is very overloaded across the ecosystem. Also, what we have isn't quite a Markov kernel, since it maps to more general measures. We had changed it toKleisli, which is accurate, but the term is not in wide wide use, especially among those doing stats work.This PR tries to find a middle ground with
TransitionKernel. It's descriptive and more intuitive, while still being accurate. The smart constructor has been changed back tokernel, but we're no longer exporting it in order to avoid conflicts with other packages.Conditional meausres
Soss and now Tilde use
|for conditional models. But this should really be at the level of measures. So you can not expressmeasure | constraintin the case wheremeasureis over NamedTuples, andconstraintis a particular named tuple.One thing that may be surprising is that
|in this PR does not change the dimensionality, but instead transforms the constrained dimension to a Dirac measure. I think this will be easier to later generalize to constraints that are not axis-aligned.Likelihoods
Some updates here, as well as introduction of
log_likelihood_ratio.Pointwise Products
Slot names are changed to
priorandlikelihoodfor easier interpretation. I've added aniteratemethod, allowingμ, ℓ = (p::PointwiseProductMeasure).Power measures
logdensity_defis simplified to usesuminstead of an explicitforloop. Hopefully this will later make it easier to port to GPU methods.PowerWeighted
In for example GLMs or Bayesian bootstrap, we want to consider each observation to have some weight. This is an exponent on the likelihood factor, or equivalently a multiplicative weight on the log-likelihood. Since
^is already taken (and much more common than this) I've added