v0.2.10

What's Changed

• fix documentation by @LegrandNico in #361
• fix unbounded updates for continuous and volatile nodes in Rust, JAX and vectorised JAX by @LegrandNico in #362
• Bump version from 0.2.9 to 0.2.10 by @LegrandNico in #363

Full Changelog: v0.2.9...v0.2.10

v0.2.9

This release introduces a full Rust backend that is on par with the JAX version on most features, a new vectorised JAX backend to improve support for deep neural networks, and extensive support for deep learning classification tasks in deep networks.

What's Changed

• dissociate fixed and dynamic learning rates for learning in deep networks by @LegrandNico in #320
• check for inf values in the learning updates by @LegrandNico in #321
• Iowa-Gambling Task use-case by @sashapustota in #319
• Add a DeepNetwork class and move the relevant methods by @sashapustota in #318
• add volatile nodes and set them as defaults for hidden layers by @sashapustota in #324
• use prospective configuration in the weight updates by @LegrandNico in #326
• Clip expected mean in binary input for numerical stability by @LegrandNico in #327
• Use a numerically stable version of the uHGF by @LegrandNico in #328
• add Rust backend for continuous nodes by @LegrandNico in #330
• test volatile nodes, fix volatile nodes updates and update learning function by @LegrandNico in #331
• add update_type parameter when using volatile nodes by @LegrandNico in #332
• add volatile node with three possible update types to the Rust backend by @LegrandNico in #333
• fix immutable JAX arrays in plot_nodes by @LegrandNico in #335
• add chaining patterns to Network() in Rust by @LegrandNico in #336
• add add_layers and add_stack_layers to match the JAX version by @LegrandNico in #337
• add a collection of coupling function in Rust and allow the use of coupling_fn in add_nodes by @LegrandNico in #338
• add a learning and fit method to the Rust package by @LegrandNico in #339
• add_nodes can receive additional parameters by @LegrandNico in #340
• add a predict method to the Network class by @LegrandNico in #341
• add support for binary state nodes in Rust by @LegrandNico in #342
• record trajectories is optional when using fit in the Rust backend by @LegrandNico in #343
• add weights initialisation (JAX and Rust) by @LegrandNico in #344
• new learning dynamic for deep networks by @LegrandNico in #345
• various fixes on networks structures and updates dynamics by @LegrandNico in #346
• add constant states and a new weight update function by @LegrandNico in #347
• merge the two learning methods into one parametrisable function by @LegrandNico in #349
• Replace HashMap with flat Structs in the Rust backend by @LegrandNico in #350
• Performance and API improvements in the Rust backend by @LegrandNico in #351
• update the notebook supporting the method paper by @LegrandNico in #352
• add adam optimizer when training with the Rust backend by @LegrandNico in #353
• refactor and simplify the updates for volatile state nodes by @LegrandNico in #354
• Vectorized deep network class by @sashapustota in #348
• use Adam in vectorised deep networks and make it default by @LegrandNico in #355
• add a record_trajectories parameter to the JAX backend by @LegrandNico in #356
• VectorisedDeepNetwork is now replacing DeepNetwork + use three different posterior updates for the volatility parent by @LegrandNico in #357
• change license from GNU to MIT by @LegrandNico in #358
• improve readme, API documentation and the deep learning tutorial by @LegrandNico in #359
• Bump version from 0.2.8 to 0.2.9 by @LegrandNico in #360

New Contributors

• @sashapustota made their first contribution in #319

Full Changelog: v0.2.8...v0.2.9

v0.2.8

What's Changed

• use Python built in types instead of the typing module by @LegrandNico in #307
• use uv for dependencies management by @LegrandNico in #308
• fix JAX version on Colab by @LegrandNico in #310
• Add first components and documentation for deep predictive coding networks by @LegrandNico in #312
• Use better parameters for the update depth example by @LegrandNico in #313
• image link by @LegrandNico in #314
• v0.2.8 by @LegrandNico in #315

Full Changelog: v0.2.7...v0.2.8

v0.2.7

What's Changed

• add unbounded posterior updates for volatility parents by @LegrandNico in #262
• support python 3.13 by @LegrandNico in #299
• v0.2.7 by @LegrandNico in #300

Full Changelog: v0.2.6...v0.2.7

v0.2.6

What's Changed

• add graphviz as dependency by @LegrandNico in #293
• update docs on binary node by @LegrandNico in #294
• use jax.nn.sigmoid by default by @LegrandNico in #295
• hide unobserved values when using plot_node() by @LegrandNico in #296
• create wheels for multiple Python version and OS support by @LegrandNico in #297
• update cargo configuration and CI file by @LegrandNico in #298

Full Changelog: v0.2.5...v0.2.6

v0.2.5

What's Changed

• Add observation types to belief_propagation by @SylvainEstebe in #274
• add an action component to the belief propagation step by @LegrandNico in #287
• fix links and proofs rendering in the documentation by @LegrandNico in #288
• add generative sampling functionalities to the Network class by @SylvainEstebe in #290
• use ruff for linting by @LegrandNico in #291
• update maturin and pyo3 by @LegrandNico in #292

Full Changelog: v0.2.4...v0.2.5

v0.2.4

What's Changed

This release fixes an important bug from the add_nodes function that could create networks with inconsistent edges/attributes regarding value and volatility parents:

• fix the insert_nodes function for multiple parents/children by @LegrandNico in #279

The rest are maintenance issues related to publishing mixed Python/Ruste packages:

• v0.2.4 by @LegrandNico in #280
• use upload/download-artifact@v4 by @LegrandNico in #281
• fix pypi action by @LegrandNico in #282
• set rshgf as a sub-module by @LegrandNico in #283
• adapt the github action to the Rust submodule by @LegrandNico in #284
• use modern pyproject.toml structure to avoid conflicts with Maturin by @LegrandNico in #285

Full Changelog: v0.2.3...v0.2.4

v0.2.3

What's Changed

• update Github action for code coverage by @LegrandNico in #272
• add more tests and improve code coverage by @LegrandNico in #273
• set package version in pyproject.toml by @LegrandNico in #276
• use pytest-benchmark in CI by @LegrandNico in #277
• add a plot_network method using NetworkX as backende by @LouieMH in #275
• Add update function and tutorial on causal prediction errors by @LinaWalk13 in #249
• v0.2.3 by @LegrandNico in #278

New Contributors

• @LinaWalk13 made their first contribution in #249

Full Changelog: v0.2.2...v0.2.3

v0.2.2

What's Changed

• Update README.md by @LegrandNico in #265
• add remove_node function by @LouieMH in #267
• Add add_parent function by @LouieMH in #268
• input nodes can receive arrays of observations by @LegrandNico in #270
• Add support for generalised Bayesian filtering with dynamic learning rate in JAX by @LegrandNico in #266
• v0.2.2 by @LegrandNico in #271

New Contributors

• @LouieMH made their first contribution in #267

Full Changelog: v0.2.1...v0.2.2

v0.2.1

What's Changed

• use poetry for packaging by @LegrandNico in #246
• add citation guidelines by @LegrandNico in #248
• Add a Rust implementation for 1d Gaussian exponential family state nodes by @LegrandNico in #229
• Use Hashmaps, vector preallocation and avoid cloning [Rust] by @LegrandNico in #250
• refactor the add nodes methods by @LegrandNico in #251
• update links to the new tapas repository by @LegrandNico in #253
• fix maturin version in CI by @LegrandNico in #254
• refactoring utils module by @SylvainEstebe in #256
• update pyo3 to 0.23.3 by @LegrandNico in #258
• fix pre-commit paths and Rust CI by @LegrandNico in #260
• refactor continuous posterior updates module by @LegrandNico in #261
• Split posterior precision update into two branches by @LegrandNico in #263
• v0.2.1 by @LegrandNico in #264

New Contributors

• @SylvainEstebe made their first contribution in #256

Full Changelog: v0.2.0...v0.2.1