v2.3.0

What's Changed

• Multinode training: timing profiles, seeding, distributed asynchronous training (including selective policy averaging) by @josephdviviano, @chirayuharyan, @younik in #260, #335, #418, #341, #395, #401, #403, #411, #417, #420
• Move to PolicyMixin to factorize Sampler and Estimator logic by @josephdviviano in #413 and #410
• Diffusion Sampling by @hyeok9855 and @josephdviviano in #419
• Simplification and enhancement of node classes in graph sampling by @younik in #386, #387
• Expanded off policy sampling methods by @josephdviviano in #388
• Efficient TerminatingStateBuffer by @hyeok9855 in #392
• hypergrid refactor by @josephdviviano in #402
• Misc fixes by @hyeok9855, @josephdviviano in #391, #397, #399, #398

Full Changelog: v2.2.2...v2.3.0

v2.2.2

What's Changed

• Fix conditional GFlowNet training by @saleml in #385

Full Changelog: v2.2.1...v2.2.2

v2.2.1

What's Changed

• [HOTFIX] Fix mode discovery metrics by @saleml in #384

Full Changelog: v2.2.0...v2.2.1

v2.2.0

High Level

• Added Generative Augmented Flow Networks example (https://arxiv.org/abs/2210.03308), which specifies intermediate rewards with intrinsic motivation to help with the exploration problem in sparse reward environments.
• Extended graph functionality (e.g., add nodes as an action).
• More efficient autoregressive GFN training.

What's Changed

• Add GAFN for hypergrid with sparse reward by @hyeok9855 in #375
• Add graph example with add node action by @younik in #378
• Misc efficiency improvements by @hyeok9855 in #376
• Simplify GraphActionGNN by @younik in #382
• Don't allow removal of nodes with edges by @younik in #381
• minor change to allow for pb to be None when the gflownet's DAG is a tree by @josephdviviano in #371

Full Changelog: v2.1.1...v2.2.0

v2.1.1

What's Changed

• Add custom mRNA codon design environment by @ayalaajil in #350
• Fix broken conditional training script by @saleml in #372

New Contributors

• @ayalaajil made their first contribution in #350

Full Changelog: v2.1.0...v2.1.1

v2.1.0

What's Changed

• Documentation updates by @hyeok9855 in #352
• Rename GFNModule to Estimator. (removing the notion of GFNModules from the library entirely) by @hyeok9855 in #353
• consistent dtype handling library-wide by @josephdviviano in #358
• typing consistency fixes library-wide
• misc minor fixes

Full Changelog: v2.0.1...v2.1.0

v2.0.1

What's Changed

• Multiple improvements to replay buffers.
• Support for conditional GFNs.
• Support for graph-based states (using torch_geometric).
• Major improvements supporting reproducible experiments in distributed settings.
• Various quality of life improvements, small features, optimizations, and refactorings.
• Complete type safety - library-wide adherence to pyright, while dropping torchtyping.
• Improved documentation.
• Addition of many new environments in gfn/gym.
• Addition of new tutorials and examples demonstrating advanced usage of torchgfn.
• Reproduction of DAG-GFlowNet (Bayesian Structure learning, Deleu et al., 2022).
• Local search sampler (Local Search GFlowNets, Kim et al., 2024).

New Contributors

• @inelgnu made their first contribution in #192
• @younik made their first contribution in #202
• @hyeok9855 made their first contribution in #203
• @alexandrelarouche made their first contribution in #232
• @Idriss-Malek made their first contribution in #236
• @emirceyani made their first contribution in #240
• @alip67 made their first contribution in #210

Full Changelog: v1.2.1...v2.0.1

New Replay Buffer, Computation Caching, Helper Functions, and Tutorials

• License / Readme updates.
• Updates to package requirements.
• Addition of a Prioritised Replay Buffer.
• GFlowNets now optionally save_logprobs or estimator_outputs -- this is to prevent unnecessary re-computation (depending on whether you are performing on-policy or off-policy learning).
• Added self.logF*_parameters() methods to help when passing to a dedicated optimizer (differently from say pf and pb.
• Helper functions (e.g., stack_states).
• Improved tutorials - new use-cases and improved notebooks.

v1.2 Substantial Updates to Environment Definition and Sampling

• Sampling now saves estimator outputs to avoid recomputation.
• The user no longer has to define a class factory when defining environments.
• New examples added.
• Other small quality of life improvements to prevent silent bugs (often these require the user to more explicitly define expected behaviours when sampling etc).

Version 1.1.1

Bug fix : #134

• From now on, the published version (on pypi) and the release should correspond to the stable branch