Fix broadcasting in calculating emission log probs #22
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Outdated suggestions cannot be applied.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.
Fix reshaping of
emissions
so that broadcasting works withHMM.emission_distribution.log_prob()
works for both scalar and vector observations.The behaviour is now:
This fixes many methods in HMM subclasses with scalar observations (e.g.
CategoricalHMM.filter()
,CategoricalHMM.smoother()
, etc.).The PR also changes a line in
BaseHMM.log_prob()
to only sum over the leading axis, this helps for a failure case whereby extra dimensions in the input can cause broadcasting of the tfp log prob which is subsequently hidden by a.sum()
.With the present change the method will output an array with shape determined by the broadcasting (e.g. if
x
is shape(T,)
inputingx[:,None]
will now output an array of shape(T,)
rather than()
which at least makes it clear that broadcasting has occurred). Going forward it might be useful to have some shape checking steps.