0.5.1

Freeze migrate-ckpt version to v0.2.0.

0.5.0

In this new update, we focused on bringing selection mechanisms to the Global Workspace!
It has seen his API polished and unified (some methods were removed because they were not used).

Version API docs:

https://bdvllrs.github.io/shimmer/v0.5.0/

What changed:

• Replace loss coef buffers by a LossCoef TypedDict.
• Add
  RepeatedDataset
  to shimmer.
• Add docs in docs/, API documentation in https://bdvllrs.github.io/shimmer/, and
  some code examples.
• Replace Black, isort, and flake8 with Ruff (see
  #8).
• Remove GWInterfaces entirely and favor giving encoders and decoders directly to the
  GWModule. See the updated example examples/main_example/train_gw.py to see what
  changes to make (see #9).
• Remove GWModuleBase.translate and GWModuleBase.cycle. Translation and cycles
  can now be done with the utils function translation and cycle.
• Remove GlobalWorkspaceBase.batch_demi_cycles, GlobalWorkspaceBase.batch_cycles,
  and GlobalWorkspaceBase.batch_translations. This can be done with utils
  functions of the same name.
• Rename GWModuleBase.fusion_mechanism to GWModuleBase.fuse,
  GWModuleBase.encode to GWModuleBase.encode_and_fuse, and
  GWModuleBase.encode_pre_fusion to GWModuleBase.encode. Same for the associated
  methods in GlobalWorkspaceBase.
• Remove on_before_gw_encode_{loss} callbacks to allow sharing computation between
  loss functions.
• Remove many _with_uncertainty functions. The GWModuleWithUncertainty now behaves like
  the other GWModules.
• Rename all "with_uncertainty" methods to "bayesian". Note, BayesianGlobalWorkspaces
  are still a work in progress.
• Added selection mechanisms (inheriting from SelectionBase, see
  docs)
  to fuse representations according to different mechanisms (e.g. Attention).
• GlobalWorkspace (and associated GWModule, GWLosses, ...) now uses the
  RandomSelection
  mechanism. For the old behavior, use
  GlobalWorkspace2Domains.

Full Changelog: 0.4.0...v0.5.0

0.4.0

This release brings several breaking changes. Some renaming of classes, and removal of config functions.

• Use ABC for abstract methods.
• Replace DomainDescription with GWInterface.
• Add contrastive_fn attribute in DeterministicGWLosses to compute the contrastive loss.
  It can then be customized.
• Rename every abstract class with ClassNameBase. Rename every "Deterministic" classes
  to remove "Deterministic".
• Remove all config related functions. This is not the role of this repo.

0.3.2

• Revert to using classes for GWs (it's easier when loading from checkpoints.)
• GlobalWorkspace is renamed to GlobalWorkspaceBase and GlobalWorkspace now
  refers to DeterministicGlobalWorkspace.

0.3.1

Fix bugs related to imports and default_decoders.

0.3.0

• Breaking change: remove DeterministGlobaleWorkspace and VariationalGlobalWorkspace
  in favor of the functions: global_workspace and variational_global_workspace.
• Allow setting custom GW encoders and decoders.
• Breaking change: remove self.input_dim, self.encoder_hidden_dim,
  self.encoder_n_layers, self.decoder_hidden_dim, and self.decoder_n_layers
  in GWModules.

0.2.0

• Add callback on_before_gw_encode and individual compute_losses for each loss type.
• fix bugs

0.1.2

fix: wrong module used to compute the cycle losses. Don't do cycle with the same domain as target and source.

0.1.1

Fix: missing individual metrics for translation loss.

0.1.0

Initial release.