Kalman
Dense Kalman filter baseline adapted from the CASSM source.
Source
- Registry name:
kalman - Model class:
ladys.models.kalman.Kalman - Config class:
ladys.models.kalman.KalmanConfig - Source file:
src/ladys/models/kalman.py
When to use
Use Kalman as a dense Bayesian filtering baseline alongside sparse CASSM and GPFA. The method uses the full observation update rather than CASSM's sparse projection, so it is useful for comparing accuracy and runtime against the computation-aware approximation.
Assumptions
Observations are modeled with Gaussian noise and Matern temporal dynamics.
Outputs
The training path returns the Kalman marginal-likelihood objective in
extras["loss"]. predict_rates returns nonnegative filtered rate
predictions shaped like the input observations.
Configuration
Config for the bundled dense Kalman filter baseline.
| Field | Type | Default |
|---|---|---|
name |
Literal['kalman'] |
'kalman' |
objective |
str |
'negative_log_marginal_likelihood' |
dt |
float |
0.01 |
dataset_name |
Optional[str] |
None |
save_model |
bool |
False |
nlb_ridge_alpha |
float |
500.0 |
optimization |
OptimizationConfig |
OptimizationConfig(name='gradient', optimizer='Adam', lr=0.01, weight_decay=0.0, gradient_clip=300.0) |
Contracts
forwardinputs use(batch, time, neurons)observations.- Runtime outputs follow the model output contract.
- Optimizer epochs follow the optimizer contract.