v23.12.00

What's Changed

• Fix sok import error in ut by @EmmaQiaoCh in #1232
• fix dependencies in docs build by @edknv in #1236
• move merlin dependecy install in horovod-cpu tests by @edknv in #1235

New Contributors

• @EmmaQiaoCh made their first contribution in #1232

Full Changelog: v23.08.01...v23.12.00

v23.08.01

🐜 Bug Fixes

• Change model call default arg values to None instead of bool vals by @oliverholworthy in #1218
• Handle ColumnSchema target in serialization of SequenceTransform by @oliverholworthy in #1214
• Ensure TopKEncoder has correct outputs when model is saved by @oliverholworthy in #1225

🔧 Maintenance

• Use copy-pr-bot by @ajschmidt8 in #1209
• use rapids infra to run testing by @jperez999 in #1216
• improve hpo test by @radekosmulski in #1222

New Contributors

• @ajschmidt8 made their first contribution in #1209

Full Changelog: v23.08.00...v23.08.01

v23.08.00: Makes RetrievalModelV2 support item tower with transforms (e.g. pre-t…

…rained embeddings) (#1198)

* Making retrieval model to_top_k_model(), candidate_embeddings() and batch_predict() support Loader with transforms for pre-trained embeddings in item tower

* Fixing test error and ensuring all batch_predict() with the new API support Loader with transforms (which include pre-trained embeddings)

* Fixing retrieval example, which was using wrong schema to export query and item embeddings

* Added missing importorskip on torch and pytorch_lightning for torch integration tests

* Skiping a test if nvtabular is available

v23.06.00

Update merlin dependencies to match 23.06 release (#1155)

v23.05.00

What’s Changed

⚠ Breaking Changes

• Update minimum version of TensorFlow from 2.8 to 2.9 @oliverholworthy (#1084)

🐜 Bug Fixes

• Extend ContrastiveOutput to support sequential encoders @sararb (#1086)

🚀 Features

• Adding ParallelBlock @marcromeyn (#1088)
• Add dataloader pre-trained embeddings support to Merlin Models @gabrielspmoreira (#1083)
• Introducing Block @marcromeyn (#1087)
• Creates LogLossMetric and generalizes WandbLogger and ExamplesPerSecondCallback @gabrielspmoreira (#1085)

📄 Documentation

• update pretrained embs example @radekosmulski (#1100)
• Change tf.keras.optimizers.Adagrad() to tf.keras.optimizers.legacy.Adagrad() @rnyak (#1098)
• implement review suggestions @radekosmulski (#1062)

🔧 Maintenance

• Skip installing dependencies in tox gpu test @edknv (#1105)
• Change tf.keras.optimizers.Adagrad() to tf.keras.optimizers.legacy.Adagrad() @rnyak (#1098)
• Add Workflows to check base branch and set stable branch @oliverholworthy (#1080)
• Update tag pattern in GitHub Workflows @oliverholworthy (#1081)
• Update minimum version of TensorFlow from 2.8 to 2.9 @oliverholworthy (#1084)
• Remove use of deprecated numpy aliases of builtin types @oliverholworthy (#1082)
• Add publish step for conda package @oliverholworthy (#1075)
• Skip gpu tests on PR closed event @edknv (#1067)
• Update requirements for Merlin packages to minimum version of 23.04 @oliverholworthy (#1069)
• remove blossom-ci @nv-alaiacano (#1074)

v23.04.00

What’s Changed

⚠ Breaking Changes

• New design of the transformer API @sararb (#1022)
• Updates Models to support new dataloader format for lists (__values and __offsets in dict) and scalar (1D) @gabrielspmoreira (#999)

🐜 Bug Fixes

• Refactory/fix of sampled softmax to add logQ correction @gabrielspmoreira (#1051)
• Fixes model.batch_predict() which was not working with ModelOuput @gabrielspmoreira (#1052)
• Test Model Reloading - Fixing some layers @oliverholworthy (#1044)
• Adds Support to embedding sequence_combiner for dense tensors @gabrielspmoreira (#1029)
• Fixed DLRModel / DLRMBlock to accept bottom_block with dropout on top of MLPs @gabrielspmoreira (#1018)

🚀 Features

• Refactory/fix of sampled softmax to add logQ correction @gabrielspmoreira (#1051)
• Add Transformers4Rec repo test @edknv (#1049)
• Fix import gpu without support @jperez999 (#1053)
• Tensorflow 2.11 support @edknv (#1016)
• New design of the transformer API @sararb (#1022)
• Introduce SOKEmbedding using Sparse Operation Kit @WonderingWJ (#863)
• Updates Models to support new dataloader format for lists (__values and __offsets in dict) and scalar (1D) @gabrielspmoreira (#999)
• Save and load Implicit model and schemas @edknv (#1014)
• Updates on MTL example text and illustration @gabrielspmoreira (#1003)

📄 Documentation

• MM transformer training and serving example @radekosmulski (#1045)
• Updates on MTL example text and illustration @gabrielspmoreira (#1003)

🔧 Maintenance

• Add Transformers4Rec repo test @edknv (#1049)
• Install wheel package to enable legacy install method of LightFM package @oliverholworthy (#1060)
• Fix import gpu without support @jperez999 (#1053)
• Update dataset test to expect schema.pbtxt in output @edknv (#1057)
• Update test_synthetic_aliccp_raw_data to check subset of outputs @oliverholworthy (#1056)
• add concurrency param to kill Github Action jobs when new commits come in @nv-alaiacano (#1055)
• Set batch sizes in unit tests to powers of 2 at least 16 @edknv (#1054)
• Enable pickle of model with TensorFlow 2.11 @oliverholworthy (#1040)
• set random seed in implicit tests @edknv (#1031)
• Skip SOK tests if SOK is not installed @edknv (#1030)
• Restoring tox.ini and .github/workflows/tensorflow.yml @gabrielspmoreira (#1023)
• Convert Tensor to RaggedTensor in PrepareFeatures for list features @oliverholworthy (#1028)
• Fix dataset test due to metadata location change @edknv (#1021)
• Updates Models to support new dataloader format for lists (__values and __offsets in dict) and scalar (1D) @gabrielspmoreira (#999)
• Add a pytest.ini file to register custom markers (e.g. notebook) @karlhigley (#1017)
• Add multi-gpu github action tests using horovod @edknv (#1009)
• Set random seed in each lightfm test @edknv (#1012)
• Use globally insatlled installed packages for GPU tests @edknv (#1011)

v23.02.00

What’s Changed

🐜 Bug Fixes

• Fix BroadcastToSequence to enable context features in sequential models @gabrielspmoreira (#991)
• Fixes lower train metrics when using Keras Masking (SequenceMaskRandom, SequenceMaskLast) @gabrielspmoreira (#983)
• Makes ColumnBasedSampleWeight serializable @gabrielspmoreira (#979)
• fix training a sequential model with AverageEmbeddingsByWeightFeature @sararb (#973)
• Fixes support of sequential continuous features for sequential and non-sequential models @gabrielspmoreira (#969)
• add get_config() to AverageEmbeddingsByWeightFeature class @rnyak (#968)

🚀 Features

• Fixes support of sequential continuous features for sequential and non-sequential models @gabrielspmoreira (#969)
• add get_config() to AverageEmbeddingsByWeightFeature class @rnyak (#968)

📄 Documentation

• enhancing the next-item prediction notebook documentation @MarkMoTrin (#975)
• Update 05-Retrieval-Model.ipynb @al-yakubovich (#967)

🔧 Maintenance

• enhancing the next-item prediction notebook documentation @MarkMoTrin (#975)
• Remove Notebook Output before running pre-commit lint checks @oliverholworthy (#984)
• fix gpu visibility with env passthrough @jperez999 (#976)

v0.11.0

What’s Changed

🐜 Bug Fixes

• Fix the inference of transformer-based models trained with masked language modeling @sararb (#909)
• Making InputBlockV2(..., embeddings) deprecated and adjusting tests to use categorical arg @gabrielspmoreira (#880)
• Preventing silent error when evaluating V1 retrieval models (TwoTower, MF) @gabrielspmoreira (#892)
• Fix loading of TwoTowerModel with context query variable @oliverholworthy (#887)
• Enable sequence-testing synthetic data to be generated with different sequence lengths @oliverholworthy (#882)
• Enable concat of sequence features with InputBlockV2 @oliverholworthy (#883)
• Make path optional in get_booking @marcromeyn (#874)
• fix dtype error due to date column in ecommerce-session-based-next-item-prediction-for-fashion nb @rnyak (#921)
• Fix the serialization of SequenceSummary block @sararb (#927)
• Use tf.function for list column operations @edknv (#938)

🚀 Features

• Save output schema of model and add save method to Encoder @oliverholworthy (#886)
• Add the support of different thresholds k in the TopkEncoder @sararb (#869)

📄 Documentation

• Update retrieval notebook to new api @rnyak (#900)
• Add Reference to Transformer4Rec for PyTorch support @bschifferer (#903)
• docs: Add calver to semver banner @mikemckiernan (#865)
• docs: Add basic SEO configuration @mikemckiernan (#864)
• update drafter to work on tags only and update cpu ci to target branch @jperez999 (#901)
• Session-based example using dressipi dataset and XLNet architecture @rnyak (#849)
• Adding Multi-GPU Data Parallel Example @bschifferer (#855)
• T4rec use case @radekosmulski (#853)
• Do not copy source to docs output @mikemckiernan (#888)

🔧 Maintenance

• Support tuple return type from model pre and update test to use this @oliverholworthy (#890)
• Use merlin-dataloader package @edknv (#845)
• update drafter to work on tags only and update cpu ci to target branch @jperez999 (#901)
• Add skip if transformers not available for next item prediction notebook test @oliverholworthy (#899)
• Skip retrieval_with_hpo notebook test when optuna is not installed @edknv (#897)
• Update test of lazy adam notebook to only run when GPU available @oliverholworthy (#894)
• Remove tensorflow dependencies from requirements/dev.txt @oliverholworthy (#895)
• Add test for retrieval model with transformer block @oliverholworthy (#833)
• Pin fiddle to 0.2.2 so that tests run with pip install @oliverholworthy (#872)
• Pass task="binary" to binary metrics to handle latest torchmetrics @oliverholworthy (#920)
• Publish Release Draft on release branch builds @oliverholworthy (#925)
• Update branch name extraction for tag builds @oliverholworthy (#972)

Full Changelog: v0.10.0...v0.11.0

v0.10.0

What’s Changed

🐜 Bug Fixes

• Enable concat of sequence features with InputBlockV2 @oliverholworthy (#883)
• Make path optional in get_booking @marcromeyn (#874)

📄 Documentation

• T4rec use case @radekosmulski (#853)
• Do not copy source to docs output @mikemckiernan (#888)

🔧 Maintenance

• Skip retrieval_with_hpo notebook test when optuna is not installed @edknv (#897)
• Update test of lazy adam notebook to only run when GPU available @oliverholworthy (#894)
• Remove tensorflow dependencies from requirements/dev.txt @oliverholworthy (#895)
• Add test for retrieval model with transformer block @oliverholworthy (#833)
• Pin fiddle to 0.2.2 so that tests run with pip install @oliverholworthy (#872)

v0.9.0

What’s Changed

🐜 Bug Fixes

• Quick fix to parse post in transformer when a string is passed in @marcromeyn (#832)
• Trying to add Bokeh to dev-requirements to fix failing xgb-tests @marcromeyn (#839)
• Ensure metrics passed to model.compile() are always reset @sararb (#830)
• Fix EmbeddingFeatures serialization/deserialization when a feature is named 'name' @oliverholworthy (#817)
• Add support for LogitsTemperatureScaler in the new ModelOutput API @sararb (#815)
• Add LogQ correction support to the new ModelOutput API @sararb (#811)

🚀 Features

• Removing mask from PredictionContext, since it's unused @marcromeyn (#838)
• Add seed to TopKMetric for consistent metrics when ties in ranking @oliverholworthy (#827)
• Adding SequenceSummary transforms for transformers @marcromeyn (#828)
• [Test] - fix pretrained embedding unittest @bschifferer (#826)
• Making sure InputBlockV2 works without the precense of categorical-features @marcromeyn (#821)
• fix OOM issue in accelerate-training-of-large-embedding-tables-by-LazyAdam nb unit test @rnyak (#816)
• Add L2-batch regularization to EmbeddingTable @sararb (#812)

🔧 Maintenance

• fix OOM issue in accelerate-training-of-large-embedding-tables-by-LazyAdam nb unit test @rnyak (#816)
• update temp data paths in example notebooks @nv-alaiacano (#813)
• Use pytest-xdist with tensorflow tests to reduce running time of tests @oliverholworthy (#795)
• Add python_requires specifying 3.8 as minimum Python version @oliverholworthy (#797)