This tracks @eb8680's old idea of making Gaussian more amenable to symbolic computation. One advantage of this approach is that a forward-backward algorithm could be performed symbolically, such that each SVI step would require only a substitution, rather than a complete optimization pass.
Tasks
[x] Support more complex slicing operations #555
[ ] Create a literally_eager interpretation that does not normalize. Use it inside Gaussian.eager_() methods.
This tracks @eb8680's old idea of making Gaussian more amenable to symbolic computation. One advantage of this approach is that a forward-backward algorithm could be performed symbolically, such that each SVI step would require only a substitution, rather than a complete optimization pass.
Tasks
literally_eagerinterpretation that does not normalize. Use it insideGaussian.eager_()methods.GaussianHMM.