Sequential Hidden Variable Model
TODO:
- Allow models not to condition on actions.
- Allow models to condition on invariant information throughtout a sequence.
- Allow models to have independent transition or observation.
- Allow algorithms to handle non-reparameterized models and the reparameterized ones.
- Make all models and proposals to share the same state dict
- Make all models and proposals be able to condition on some prior knowledges.
- Modify algorithms to accomodate the above changes
- Modify SMC to be TwistedSMC and returns only loss term
- Implement FilteringProposal and SmoothingProposal respectively, and replace the current ones
- Make twisters a separate class.
- Make sure the shape of observations and actions are expand to batch_shape before passing to models and proposals, where the batch_shape includes the particle dimension when there are multiple particles.
- Make sequence elbo suitable for the general case