- Fixes some issues with threading.
- Improves speed of SMC^2 when we require an increase in state particles by checking acceptance rate threshold in kernel instead.
- Check average acceptance rate for MH kernel instead.
- Bug fix for matrix shaped parameters.
- Use
torch.Tensor.copy_
instead offill_
.
- Uses stoch-proc version 0.3.0, which means that whenever the context is updated in an inference algorithm we rebuild the model.
- Speed improvement for SISR.
- Major change in backend by interchaning batch and sample shape in inference algorithms. This enables much more natural handling of parameters.
- Adds support for finding mode of distribution by means of `functorch
- Adds the Gaussian Particle Filter
- Reworks correct/predict logic by enforcing that correct only takes a prediction as input
- Simplifies proposal logic
- Uses latest version of stoch-proc
- Reworks how we define parameters by initializing all parameters to their "correct" shapes from the get-go by introducing a shape object stored on
InferenceContext
.
- Reworks
ParameterContext
logic by removing the requirements for the context being on the stack in order to register parameters. - Adds fixed-lag smoothing.
- Renames
ParameterContext
toInferenceContext
- Adds support for using QMC points in
SMC2
- Utilizes
int
instead oflong
- Adds a proposal for locally linearized observation dynamics.
- Adds support for performing variational inference using particle filter.
- Improves plotting functionality by handling multi-dimensional parameters.
- Adds support for applying function on context and return a copy of that context.