v0.6.1

What has changed

• Fix installing setup requirements
• Remove use of deprecated jnp.NINF
• Remove use of deprecated np.product

v0.6.0

What has changed

⚠️ This release changes the high-level API ⚠️

Deprecation notice

• bp.run_bp is deprecated in favor of bp.run. bp.run_bp will be removed in future releases.
• Dropped support for python 3.7.

New features

• Added a SDLP solver that solves the smoothed dual of the LP relaxation of the MAP problem by accelerated gradient descent.
  • An alternative to Belief Propagation with convergence guarantees.
  • Based on the Smooth Dual LP-MAP problem introduced in this paper.
  • Leverages existing PGMax infrastructure to implement the underlying message-passing algorithm.
• Added a solver for the primal of the LP relaxation based on the LP solver from cvxpy.

Changes

• Implemented a unified interface for the BP and the SDLP solvers.
  • Solvers can be created via

     inferer = infer.build_inferer(fg.bp_state, backend=BACKEND)

    where BACKEND can be "bp" or "sdlp".
• Added a run_with_diffs function to the BP solver to monitor BP convergence.
• Improved the stability of message updates at low temperatures for logical factors and pool factors.
• Improved the numerical stability of the messages / potentials clipping.
• Improved the handling of various corner cases, including variables with single state, empty variable groups, energy with infinite potentials, etc..

Examples

• Updated all example colabs to use the new unified interface.
• Added a colab notebook which uses the SDLP solver to
  1. Run inference on an Ising Model and compare its results with BP
  2. Extract sparse feature activations from visually complex binary scenes.

v0.5.1

What has changed

• Added a new module to compute the energy of a MAP decoding
• Defined inference arguments for each factor type which split all the 2D wiring arrays of the different factor types into contiguous 1D arrays; leading to 5x compile time speedups
• Stored new information in var_states_for_edges which makes inference faster
• Added PGMax documentation to ReadTheDocs
• Added a nicer representation of NDVarArray

v0.5.0

PGMax is now part of DeepMind 🎉!

Following the acquisition of Vicarious by Alphabet in 2022, PGMax developments have now moved from Vicarious to DeepMind.

What has changed

• Added PoolFactors with optimized inference procedure.
• License changed from MIT License to Apache 2.0 License.
• Adopted pyink as the new auto-formatter.