Description of the Issue:
In our variational inference iterator, we currently only support initializing the variational parameters randomly or from the prior. The latter is only possible for mean-field normal and full-rank normal distributions.
Proposed Solution:
We also need the possibility to manually pass the initial variational parameters, which allows a lot more flexibility. This approach could either replace or expand the current initialization strategies.
Action Items:
No response
Related Issues:
Compare the discussion in #251
Interested Parties:
No response
Description of the Issue:
In our variational inference iterator, we currently only support initializing the variational parameters randomly or from the prior. The latter is only possible for mean-field normal and full-rank normal distributions.
Proposed Solution:
We also need the possibility to manually pass the initial variational parameters, which allows a lot more flexibility. This approach could either replace or expand the current initialization strategies.
Action Items:
No response
Related Issues:
Compare the discussion in #251
Interested Parties:
No response