Change log

rodeo 1.1.3

  • Breaking changes:

    • prior_weight, prior_var arguments are combined as prior_pars for all solver and inference methods

    • rodeo.inference.pseudo_marginal replaces rodeo.inference.random_walk_aux. The API is now exactly like BlackJAX, except that the init method requires a PRNG-key. The signature for the log-posterior/likelihood must be parameters, key.

  • Minor change:

    • Blackjax version is no longer fixed for Python 3.10+.

rodeo 1.1.2

  • Setup is done only using pyproject.toml and setup.cfg is no longer needed.

rodeo 1.1.1

  • No changes besides minor documentation edits.

rodeo 1.1.0

  • Breaking changes:

    • kalman_funs argument is changed to kalman_type. rodeo supports standard Kalman and square-root Kalman and these are directly implemented in the library. The kalman_type argument picks the algorithm to use.

  • New features:

    • Added first_order_pad to help users with zero-padding the initial value and ode_weight matrix.

rodeo 1.0.0

  • Breaking changes:

    • rodeo now depends on jax and previous Cython/C++ implementations are completely removed.

    • Standard Kalman algorithm is now under rodeo.kalmantv.standard instead of just rodeo.kalmantv.

    • The ODE solver in rodeo is now under rodeo.solve.

    • The IBM prior only returns wgt_state, and var_state. That is, mu_state is assumed to be 0.

  • New features:

    • Added new parameter inference algorithms in rodeo.inference: basic, dalton, fenrir, random_walk_aux, magi.

    • Added square-root Kalman algorithms: rodeo.kalmantv.square_root

    • Added new interrogation functions from new research.

    • All ODE solver and parameter inference methods support block-wise computation for efficiency. This is the default option if the IBM prior is used from rodeo.prior.ibm.

rodeo 0.4

  • Initial release of rodeo: a probabilistic ODE solver based on the Bayesian filtering paradigm with Python frontend and three backends:

    • C++ using Eigen

    • Cython using BLAS/LAPACK

    • numba using BLAS/LAPACK

  • Two methods for parameterizing the Gaussian Markov prior are included: IBM and CAR.

  • Three interrogation functions are included: Chkrebtii, Schobert, and a mix of the two called rodeo.

  • Kalman filtering/smoothing algorithms are written in C++ using Eigen.

  • This version is still available at rodeo-legacy.