v0.4.0

New features

• keras_rs.losses.ListMLELoss: Added support for the ListMLE loss function.

New example

• DLRM-DCNv2: Added implementation of the DLRM-DCNv2 model architecture with the MLPerf performance benchmark using the Criteo dataset.

Bug fixes & Improvements

• DistributedEmbedding on JAX
  • Updated DistributedEmbedding to work with the latest jax-tpu-embedding and take advantage of the latest performance improvements, in particular with table stacking.
  • Bug fixes and improvements with the synchronization of table statistics across multiple hosts.
  • Made table statistics updates optional.
  • Improved performance by removing redundant ones/ones_like calls when feature weights are not provided.
• JAX Compatibility
  • Make KerasRS compatible with JAX >= 0.8.0, in particular with regards to jax.jit and shard_map.
• Infrastructure & CI/CD:
  • Added TPU testing for JAX and TensorFlow.
  • Added GPU testing for JAX, TensorFlow and Torch.
  • Upgraded PyTorch dependency to version 2.9.0.
  • Enabled CUDA 13.0 support for GPU testing.

New Contributors

• @adityagupta1089 made their first contribution in #156
• @LakshmiKalaKadali made their first contribution in #130
• @wenyi-guo made their first contribution in #170

Full Changelog: v0.3.0...v0.4.0

v0.4.0.dev0

What's Changed

• update dev version to 0.3.1.dev0 by @sachinprasadhs in #153
• Refactor: Use embedding.EmbeddingVariables in Keras-RS JAX embedding tests. by @adityagupta1089 in #156
• Unwrap variables before passing to jax.jit. by @hertschuh in #157
• Refactor return type of DistributedEmbedding methods to use embedding.EmbeddingVariables by @adityagupta1089 in #158
• Refactor DistributedEmbedding and embedding_utils to streamline table… by @adityagupta1089 in #159
• Add TPU tests for JAX and Tensorflow. by @hertschuh in #160
• Fix for jax import in distributed_embedding_test.py. by @hertschuh in #162
• Refactor: Replace embedding_utils with table_stacking in DistributedEmbedding and tests by @adityagupta1089 in #161
• Fix CPU stats aggregation for JAX multihost. by @cantonios in #164
• Update torch version to 2.9.0. by @hertschuh in #165
• Add heuristic for max ids of stacked tables. by @hertschuh in #167
• Add ListMLE Loss by @LakshmiKalaKadali in #130
• Use the JTE API to set stacked table stats... by @cantonios in #169
• support all tests in layers to tpu by @wenyi-guo in #170
• Prevent max ids per partition fluctuating by always taking the max. by @hertschuh in #168
• Use jax.shard_map with JAX >= 0.8.0 by @hertschuh in #172
• Refactor: Remove explicit np.ones/ones_like for optimized numpy feature_weights=None path. by @adityagupta1089 in #171
• Make stat update optional for JAX distributed embedding by @abheesht17 in #173
• [MLPerf] Add DLRM-DCNv2 by @abheesht17 in #144
• Run GPU tests on runners. by @hertschuh in #174
• Use subprocess.run in pip_build.py to escape wheel path. by @hertschuh in #176
• Use CUDA 13.0 on GPU and other GPU tests tweaks. by @hertschuh in #177
• Version bump to 0.4.0.dev0 by @sachinprasadhs in #178

New Contributors

• @adityagupta1089 made their first contribution in #156
• @LakshmiKalaKadali made their first contribution in #130
• @wenyi-guo made their first contribution in #170

Full Changelog: v0.3.0...v0.4.0.dev0

v0.3.0

What's Changed

• Ignore shard_map attr error in mypy. by @cantonios in #97
• Added TF specific documentation to DistributedEmbedding. by @hertschuh in #94
• Fix symbolic calls for EmbedReduce. by @hertschuh in #98
• Move DistributedEmbedding declaration to its own file. by @hertschuh in #99
• Remove dependency on tree and use keras.tree. by @hertschuh in #100
• Only enable JAX on linux_x86_64. by @cantonios in #101
• Add out_sharding argument WrappedKerasInitializer. by @cantonios in #102
• Use Python 3.10 style type annotations. by @hertschuh in #104
• Do not bundle test utils in wheel. by @hertschuh in #105
• Update version number to 0.2.1 by @hertschuh in #106
• Fix invalid escape sequence in unit test. by @hertschuh in #108
• Replace leftover unflatten_as to pack_sequence_as. by @hertschuh in #109
• Make the declaration of Nested compatible with pytype. by @hertschuh in #110
• Add ragged support for default_device placement on JAX. by @cantonios in #107
• Add documentation for using DistributedEmbedding with JAX. by @cantonios in #111
• api_gen now excludes backend specific code. by @hertschuh in #103
• Enable preprocess calls with symbolic input tensors. by @cantonios in #113
• Check for jax_tpu_embedding on JAX backend. by @cantonios in #114
• Version bump to 0.2.2.dev0 by @laxmareddyp in #117
• Fix docstring for DistributedEmbedding. by @hertschuh in #118
• Minor docstring fix in DistributedEmbedding. by @hertschuh in #120
• Fix nightly builds. by @hertschuh in #121
• Bump torch from 2.6.0 to 2.7.1 by @dependabot[bot] in #119
• Checkpoints by @silkyarora in #122
• Add FTRL and Adam optimizers. by @cantonios in #123
• Update references of jax_layout.Layout to jax_layout.Format. by @cantonios in #124
• Update explicit layouts to use JAX 0.6.2 layout.Format. by @cantonios in #125
• Change for DeviceLocalLayout in JAX 0.6.3. by @hertschuh in #126
• Fix for removal of JAX DeviceLocalLayout. by @hertschuh in #127
• Add note about multi-host dataset sharding. by @cantonios in #131
• Fix JAX Layout tiling argument for JAX 0.7.1 by @hertschuh in #133
• Updated documentation regarding DistributedEmbedding batch size. by @hertschuh in #138
• Fix for batch size of FeatureConfig.output_shape on TensorFlow. by @hertschuh in #139
• Fix handling of mixed placement in DistributedEmbedding. by @hertschuh in #136
• Fix batch sizes in distributed_embedding_test.py. by @hertschuh in #142
• Fix and clean up multihost preprocessing. by @cantonios in #146
• Bump Python, JAX and Tensorflow versions by @hertschuh in #148
• DistributedEmbedding: do not hardcode the number of SparseCores with TF. by @hertschuh in #147
• Make TableConfig and FeatureConfig unhashable. by @hertschuh in #150
• version bump to 0.3.0.dev0 by @sachinprasadhs in #151
• Update release version to 0.3.0 by @sachinprasadhs in #152

New Contributors

• @silkyarora made their first contribution in #122
• @sachinprasadhs made their first contribution in #151

Full Changelog: v0.2.1...v0.3.0

v0.3.0.dev0

What's Changed

• Ignore shard_map attr error in mypy. by @cantonios in #97
• Added TF specific documentation to DistributedEmbedding. by @hertschuh in #94
• Fix symbolic calls for EmbedReduce. by @hertschuh in #98
• Move DistributedEmbedding declaration to its own file. by @hertschuh in #99
• Remove dependency on tree and use keras.tree. by @hertschuh in #100
• Only enable JAX on linux_x86_64. by @cantonios in #101
• Add out_sharding argument WrappedKerasInitializer. by @cantonios in #102
• Use Python 3.10 style type annotations. by @hertschuh in #104
• Do not bundle test utils in wheel. by @hertschuh in #105
• Update version number to 0.2.1 by @hertschuh in #106
• Fix invalid escape sequence in unit test. by @hertschuh in #108
• Replace leftover unflatten_as to pack_sequence_as. by @hertschuh in #109
• Make the declaration of Nested compatible with pytype. by @hertschuh in #110
• Add ragged support for default_device placement on JAX. by @cantonios in #107
• Add documentation for using DistributedEmbedding with JAX. by @cantonios in #111
• api_gen now excludes backend specific code. by @hertschuh in #103
• Enable preprocess calls with symbolic input tensors. by @cantonios in #113
• Check for jax_tpu_embedding on JAX backend. by @cantonios in #114
• Version bump to 0.2.2.dev0 by @laxmareddyp in #117
• Fix docstring for DistributedEmbedding. by @hertschuh in #118
• Minor docstring fix in DistributedEmbedding. by @hertschuh in #120
• Fix nightly builds. by @hertschuh in #121
• Bump torch from 2.6.0 to 2.7.1 by @dependabot[bot] in #119
• Checkpoints by @silkyarora in #122
• Add FTRL and Adam optimizers. by @cantonios in #123
• Update references of jax_layout.Layout to jax_layout.Format. by @cantonios in #124
• Update explicit layouts to use JAX 0.6.2 layout.Format. by @cantonios in #125
• Change for DeviceLocalLayout in JAX 0.6.3. by @hertschuh in #126
• Fix for removal of JAX DeviceLocalLayout. by @hertschuh in #127
• Add note about multi-host dataset sharding. by @cantonios in #131
• Fix JAX Layout tiling argument for JAX 0.7.1 by @hertschuh in #133
• Updated documentation regarding DistributedEmbedding batch size. by @hertschuh in #138
• Fix for batch size of FeatureConfig.output_shape on TensorFlow. by @hertschuh in #139
• Fix handling of mixed placement in DistributedEmbedding. by @hertschuh in #136
• Fix batch sizes in distributed_embedding_test.py. by @hertschuh in #142
• Fix and clean up multihost preprocessing. by @cantonios in #146
• Bump Python, JAX and Tensorflow versions by @hertschuh in #148
• DistributedEmbedding: do not hardcode the number of SparseCores with TF. by @hertschuh in #147
• Make TableConfig and FeatureConfig unhashable. by @hertschuh in #150
• version bump to 0.3.0.dev0 by @sachinprasadhs in #151

New Contributors

• @silkyarora made their first contribution in #122
• @sachinprasadhs made their first contribution in #151

Full Changelog: v0.2.1...v0.3.0.dev0

v0.2.1

Summary:

• Distributed Embeddings & JAX Support:

  • Improvements to the DistributedEmbedding layer (documentation, bug fixes, code structure) and the JAX backend (platform-specific enabling, new API arguments, ragged tensor support).

What's Changed

• Version bump 0.2.1.dev0 by @laxmareddyp in #115
• Version bump to 0.2.1 by @laxmareddyp in #116

New Contributors

• @laxmareddyp made their first contribution in #115

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

v0.2.1.dev0

What's Changed

• Version bump 0.2.1.dev0 by @laxmareddyp in #115

New Contributors

• @laxmareddyp made their first contribution in #115

Full Changelog: v0.2.0...v0.2.1.dev0

v0.2.0

What's Changed

• Update version.py to 0.1.1 by @rtg0795 in #78
• Change main version to 0.1.1.dev0 by @abheesht17 in #79
• Add assertNotAllClose to TestCase. by @hertschuh in #80
• Add Nested generic type for type hints. by @hertschuh in #81
• Add no_automatic_dependency_tracking utility. by @hertschuh in #82
• Add workflows for stale PR reminder, issue template, auto-assigning issues by @abheesht17 in #84
• Add EmbedReduce layer. by @hertschuh in #83
• Added TensorShape alias for the recently renamed Shape. by @hertschuh in #85
• Add jax-tpu-embedding dependency for large embeddings. by @cantonios in #86
• Fix embed_reduce for sparse inputs. by @cantonios in #87
• Add DistributedEmbedding cross-backend default implementation. by @hertschuh in #88
• Add backend/hardware-independent doc for DistributedEmbedding. by @hertschuh in #90
• Add JAX DistributedEmbedding implementation. by @cantonios in #89
• Remove Python 3.9 support. by @cantonios in #91
• Add TensorFlow implementation of DistributedEmbedding. by @hertschuh in #92
• Move preprocess to base distributed embedding class. by @cantonios in #93
• Version bump to 0.2.0.dev0 by @rtg0795 in #95
• Version bump to 0.2.0 by @rtg0795 in #96

New Contributors

• @cantonios made their first contribution in #86

Full Changelog: v0.1.0...v0.2.0

v0.2.0.dev0

What's Changed

• Update version.py to 0.1.1 by @rtg0795 in #78
• Change main version to 0.1.1.dev0 by @abheesht17 in #79
• Add assertNotAllClose to TestCase. by @hertschuh in #80
• Add Nested generic type for type hints. by @hertschuh in #81
• Add no_automatic_dependency_tracking utility. by @hertschuh in #82
• Add workflows for stale PR reminder, issue template, auto-assigning issues by @abheesht17 in #84
• Add EmbedReduce layer. by @hertschuh in #83
• Added TensorShape alias for the recently renamed Shape. by @hertschuh in #85
• Add jax-tpu-embedding dependency for large embeddings. by @cantonios in #86
• Fix embed_reduce for sparse inputs. by @cantonios in #87
• Add DistributedEmbedding cross-backend default implementation. by @hertschuh in #88
• Add backend/hardware-independent doc for DistributedEmbedding. by @hertschuh in #90
• Add JAX DistributedEmbedding implementation. by @cantonios in #89
• Remove Python 3.9 support. by @cantonios in #91
• Add TensorFlow implementation of DistributedEmbedding. by @hertschuh in #92
• Move preprocess to base distributed embedding class. by @cantonios in #93

New Contributors

• @cantonios made their first contribution in #86

Full Changelog: v0.1.0...v0.2.0.dev0

v0.1.0

What's Changed

• Add MRR by @abheesht17 in #46
• Update data parallel example to use TPUs by @abheesht17 in #56
• Add Mean Average Precision by @abheesht17 in #54
• Add Recall@k by @abheesht17 in #59
• Bump torch from 2.4.1+cu121 to 2.6.0 by @dependabot in #50
• Update link to Keras.io. by @hertschuh in #61
• Remove torchvision from requirements files. by @hertschuh in #62
• Add DCG by @abheesht17 in #55
• Add temperature arg to pairwise losses by @abheesht17 in #60
• Add Precision@k by @abheesht17 in #58
• Add nDCG by @abheesht17 in #57
• Regenerate APIs with the new version of namex. by @hertschuh in #63
• Add SASRec example by @abheesht17 in #64
• Add list-wise ranking example by @abheesht17 in #65
• Add doc-string examples by @abheesht17 in #66
• Fix shape issue in deep rec example by @abheesht17 in #68
• Update README with links and quickstart by @abheesht17 in #67
• Add example in doc-string in retrieval layers. by @hertschuh in #69
• Fix data parallel inference by @abheesht17 in #70
• Small changes in example by @abheesht17 in #71
• Fixes related to documentation rendering. by @hertschuh in #72
• Fix pesky annoying keras.io rendering issues by @abheesht17 in #73
• Cast tensors to the correct dtype for metrics by @abheesht17 in #74
• Example fixes by @abheesht17 in #75
• Add logo to README. by @hertschuh in #76
• Version bump to 0.1.0 by @rtg0795 in #77

New Contributors

• @dependabot made their first contribution in #50

Full Changelog: v0.1.0.dev0...v0.1.0

v0.1.0.dev0

What's Changed

• Added initial README and updated LICENSE. by @hertschuh in #1
• Rename keras-recommenders to keras-rs. by @hertschuh in #2
• Added SECURITY.md file. by @hertschuh in #3
• Initial commit with one class and one example. by @hertschuh in #4
• Add requirements.txt, .gitignore, setup.py and github actions. by @hertschuh in #5
• Add pip_build.py and required __init__.py files, update README.md. by @hertschuh in #6
• Add basic ranking example. by @hertschuh in #7
• Add Github action to publish nightly to PyPi by @hertschuh in #8
• Switch from black, isort, flake8 to ruff by @hertschuh in #9
• Fix requirements.txt for MacOS, use Tensorflow 2.18. by @hertschuh in #10
• Add pre-commit hooks by @abheesht17 in #11
• Move project metadata from setup.py to pyproject.toml. by @hertschuh in #12
• Add FeatureCross Layer by @abheesht17 in #13
• Add retrieval with data parallel training example by @abheesht17 in #14
• Add get_config to BruteForceRetrieval layer. by @hertschuh in #15
• Fix Keras RS object serialization. by @hertschuh in #17
• Make clone_initializer work better with type checking. by @hertschuh in #18
• Change LR in examples by @abheesht17 in #19
• Add GRU4Rec example by @abheesht17 in #20
• Use tuple instead of Tuple in type annotation. by @hertschuh in #26
• Add DotInteraction layer by @abheesht17 in #22
• Fix docstring arguments in basic retrieval example. by @hertschuh in #27
• Add DCN example by @abheesht17 in #21
• Add multi-task (retrieval + ranking) example by @abheesht17 in #29
• Add ScANN example by @abheesht17 in #30
• Add Abheesht to README citation by @abheesht17 in #31
• Fix dot interaction by @abheesht17 in #28
• Add deep recommender example. by @hertschuh in #23
• Add Hard Negative Mining Layer. by @hertschuh in #24
• Add Sampling Probability Correction layer. by @hertschuh in #35
• Add pairwise losses (MSE, Hinge, Logistic, SoftZeroOne) by @abheesht17 in #32
• Move layers.modeling to layers.feature_interaction by @abheesht17 in #40
• Small edit in basic ranking example by @abheesht17 in #41
• Add Remove Accidental Hits layer. by @hertschuh in #37
• Use ml_dtypes instead of numpy for finfo. by @hertschuh in #42
• Change order of call args in RemoveAccidentalHits layer. by @hertschuh in #45
• Fix disabling of traceback filtering in unit test. by @hertschuh in #47
• Fix invalid type annotation. by @hertschuh in #48
• Add base retrieval class by @abheesht17 in #49
• Add support for any dimensions to HardNegativeMining layer. by @hertschuh in #44
• Test SamplingProbabilityCorrection layer with different dimensions. by @hertschuh in #43
• Add publish_to_pypi.yml and RELEASE_PROCESS.md by @rtg0795 in #51
• Rename to PYPI_API_TOKEN by @rtg0795 in #52
• Version bump to 0.1.0.dev0 by @rtg0795 in #53

New Contributors

• @hertschuh made their first contribution in #1
• @abheesht17 made their first contribution in #11
• @rtg0795 made their first contribution in #51

Full Changelog: https://github.com/keras-team/keras-rs/commits/v0.1.0.dev0