v0.8.0

What’s Changed

⚠ Breaking Changes

• Add loader transforms @edknv (#740)
• Move transformation-layers to sub package @marcromeyn (#710)

🐜 Bug Fixes

• Bug when predicting with list-column @marcromeyn (#758)
• Fix saving + reloading models with ContrastiveOutput @marcromeyn (#757)
• Fixed inconsistency on multi-task learning metrics due to a mismatch or tasks and metrics order @gabrielspmoreira (#751)
• Implement on_epoch_end method on BatchedDataset @oliverholworthy (#724)
• Fixes RegressionTask which was raising an error because compute_output_shape() was not defined @gabrielspmoreira (#741)

🚀 Features

• Return correct type from data augmentation layer depending on the context @oliverholworthy (#703)
• Add test for TwoTowerModel with InputBlockV2 and Embeddings @oliverholworthy (#759)
• XGBoost - Use DaskDMatrix for evals data to ensure metrics in logs match result of evaluate @oliverholworthy (#682)
• Add BroadcastToSequence Layer to broadcast context features to match sequence shapes. @oliverholworthy (#737)
• Fix saving + reloading models with ContrastiveOutput @marcromeyn (#757)
• Add select_by_tag to ParallelBlock @edknv (#701)
• Add loader transforms @edknv (#740)
• Making InputBlockV2 more generic @marcromeyn (#754)
• Example notebook of Wide&Deep model @gabrielspmoreira (#716)
• Introducing Outputs-V2 @marcromeyn (#715)
• Move transformation-layers to sub package @marcromeyn (#710)
• Add EncoderBlock @marcromeyn (#705)

📄 Documentation

• Fix Feature Engineering Dressipi @bschifferer (#723)

🔧 Maintenance

• Improve reliability of test_train_metrics_steps with a custom test metric @oliverholworthy (#766)
• Remove cluster_type from dask_client test fixture @oliverholworthy (#763)
• Update tensorflow version test matrix version specifiers for clarity @oliverholworthy (#726)
• Update ItemRetrievalScorer to use tf_inspect to get call args @oliverholworthy (#756)
• Added InputBlockV2 support to DeepFMModel (refactored and fixed) and DCNModel @gabrielspmoreira (#717)

v0.7.0

What’s Changed

🐜 Bug Fixes

• Fix ecommerce session based test @edknv (#690)
• Add dtype conversion for continuous columns in synthetic data generation @oliverholworthy (#662)
• Breaks ties of top-k metrics @gabrielspmoreira (#653)
• Enable saving of TwoTowerModel @oliverholworthy (#615)
• Update stacking in DLRMBlock to use correct axis for the dot product. @oliverholworthy (#637)
• Fix lightfm evaluate code @benfred (#616)
• Enable predict in XGBoost without targets in provided dataset @oliverholworthy (#612)
• Use labels instead of predictions in XGBoost eval data @karlhigley (#610)
• Stop reordering columns in merlin.models.xgb.dataset_to_xy @radekosmulski (#603)

🚀 Features

• Extend EmbeddingTable and Embeddings with support for shared embeddings @oliverholworthy (#700)
• Filter features in ParallelBlock @edknv (#686)
• Improvements in Wide and Deep architecture @Timmy00 (#638)
• Add CategoryEncoding transformation block @Timmy00 (#614)
• Add Layer freezing @Timmy00 (#635)
• Add save and load methods to XGBoost class @oliverholworthy (#656)
• Restructuring Embeddings @marcromeyn (#649)
• enhance pretrained-embeddings example to set trainable=False for the pre-trained embeddings @rnyak (#654)
• Tf/categorical prediction v2 @sararb (#633)
• Enable customization of embeddings in DLRM Model @oliverholworthy (#619)
• Add trainable and embeddings_initializers args to Embeddings function @rnyak (#634)
• Add Wide Deep model @Timmy00 (#623)
• Enable static weights to be used with EmbeddingTable.from_pretrained @oliverholworthy (#632)
• Enable TabularData to be used with EmbeddingTable @oliverholworthy (#631)
• Add LazyAdam optimizer @Timmy00 (#602)
• Tf/contrastive prediction @marcromeyn (#594)
• Hashed Crosses for all levels @Timmy00 (#590)
• Supports task-specific sample weights, weighed_metrics and fixes InputBlockV2 @gabrielspmoreira (#600)
• Introducing new base-implementations to replace prediction-tasks @marcromeyn (#589)
• Add multi optimizer @Timmy00 (#581)
• Introducing MapValues @marcromeyn (#591)
• Fix CategoricalOneHot with compute_output_shape and check inputs. @Timmy00 (#597)
• Add EmbeddingOptions arg in the DLRM block @rnyak (#592)
• Decoupling of InputBlock(v2) and Embedding Tables, support to Ragged Embeddings Lookup and AverageEmbeddingsByWeightFeature @gabrielspmoreira (#593)

📄 Documentation

• update logo usecase @bschifferer (#693)
• [WIP] Add Dressipi RecSys Util Function to Merlin Models @bschifferer (#664)
• update 01-Getting-started @radekosmulski (#671)
• update logos for example notebooks 02 - 06 @radekosmulski (#672)
• add tracking logo to examples/07 @radekosmulski (#661)
• Add lightfm and implicit training examples @radekosmulski (#629)

🔧 Maintenance

• Separate the dev and docs requirements @karlhigley (#708)
• Fix ecommerce session based test @edknv (#690)
• Add dtype conversion for continuous columns in synthetic data generation @oliverholworthy (#662)
• Add unit test for the trainable embeddings func @rnyak (#648)
• Update versioneer from 0.20 to 0.23 @oliverholworthy (#646)
• Remove -e from pip install commands in github workflows @oliverholworthy (#645)
• Trying to make model unit-tests faster @marcromeyn (#595)

v0.6.0

What’s Changed

⚠ Breaking Changes

• Remove masking for now @marcromeyn (#557)
• Removing Queue-based negative-sampling for now @marcromeyn (#558)

🐜 Bug Fixes

• Add signature checks for various models like two-tower & DLRM @marcromeyn (#588)
• Keep Dataloader lazy when using map operations @oliverholworthy (#572)
• Enable serialization of TensorInitializer @oliverholworthy (#549)
• EmbeddingFeatures - Move embedding table creation from build to init @oliverholworthy (#532)
• Handle ParallelBlock in Model.from_block @oliverholworthy (#517)
• Propagate schema to branches if needed in ParallelBlock @marcromeyn (#478)

🚀 Features

• Hashed cross @Timmy00 (#587)
• Update UniformNegativeSampling to handle targets and add optional control for testing @oliverholworthy (#583)
• Allow for blocks in the Model to transform the targets @marcromeyn (#554)
• Add In-Batch Negative Sampling Block for positive-only batches @oliverholworthy (#560)
• Add functionality to set activation as a list in MLPBlock @rvk007 (#548)
• Add Cond Layer to tensorflow combinators @oliverholworthy (#552)
• Adding AsRaggedTensors transformation @marcromeyn (#556)
• Breaking out BaseModel to support sub-classing @marcromeyn (#518)
• Add new block EmbeddingTable @oliverholworthy (#541)
• Adding pre & post to Model @marcromeyn (#542)
• XGBoost - Add evals argument to fit to support eval on datasets other than training @oliverholworthy (#538)
• XGBoost - Train/eval with features from schema passed in to the constructor @oliverholworthy (#531)
• XGBoost - Use DaskDeviceQuantileDMatrix with GPU Training @oliverholworthy (#528)
• Use standard Keras embedding layers inside EmbeddingFeatures block @karlhigley (#472)
• Refactored top-k metrics and created TopKMetricsAggregator for optimized metrics computation @gabrielspmoreira (#514)

📄 Documentation

• Attribute source of MF image @mikemckiernan (#568)
• add xgboost example @radekosmulski (#522)
• add usecase with pretrained embeddings @radekosmulski (#508)
• Update URLs to Criteo datasets @mikemckiernan (#516)

🔧 Maintenance

• Adding the last 2 major TF versions to our CPU-tests @marcromeyn (#573)
• Add Keras layer_test function to testing_utils @oliverholworthy (#584)
• Attempt to remove docs dependencies @mikemckiernan (#586)
• Integration Tests for Retrieval models @gabrielspmoreira (#537)
• Move data-augmentation transformations into the new data_augmentation package @marcromeyn (#566)
• Moving tf/core/blocks to tf/core @marcromeyn (#565)
• Replace usage of tensorflow.python.keras with tensorflow.keras @oliverholworthy (#555)
• Make blocks part of Model and not of SequentialBlock @marcromeyn (#551)
• Adding schema's to manifest @marcromeyn (#539)
• Add tests for train_metrics_steps ensuring we can control the frequency of computing metrics @oliverholworthy (#519)
• Add tests for calling filter block when output returns empty @oliverholworthy (#481)

v0.5.0

What’s Changed

• Quick fix to simplify DLRM when there are no continuous features @marcromeyn (#479)
• Move Input-blocks @marcromeyn (#471)
• Fixing the inconsistency when exporting embeddings to cudf DataFrame @gabrielspmoreira (#452)
• Remove the mask variable from the ModelContext @karlhigley (#449)
• Move masking into the FeatureContext @karlhigley (#443)
• Create FeatureContext and pass it to model layers with call_layer @karlhigley (#428)
• Rename "mask schema" to "mask" or "feature mask" @karlhigley (#425)
• Rename BlockContext to ModelContext @karlhigley (#417)
• Create FeatureCollection as a way to pass input features to blocks @karlhigley (#418)
• Added TensorInitializer that can be used to initialize variables/embeddings with pretrained weights @gabrielspmoreira (#357)

⚠ Breaking Changes

• Remove Block.to_model @benfred (#501)
• Moving loss- and metrics-calculation to model instead of the predicti… @marcromeyn (#431)
• Make Model creation explicit @karlhigley (#432)

🐜 Bug Fixes

• Fixes computing metrics each N steps during training and adding regularization_loss metric @gabrielspmoreira (#511)
• Fix to pass shuffle argument to dataloader @benfred (#493)
• fix broken links in the 04-Exporting-ranking-models.ipynb nb @rnyak (#490)
• Fixes errors on retrieval models evaluation in graph mode @gabrielspmoreira (#488)
• Fixes and improvements in ranking metrics @gabrielspmoreira (#475)
• Fixes and improves masking (for both Causal LM and Masked LM) and adds SequenceAggregation.MASKED_MEAN (used by YouTubeDNN) @gabrielspmoreira (#464)
• Replace use of removeprefix function which requires Python 3.9+ @oliverholworthy (#465)
• Parametrize base-model-tests to run in both graph- and eager-mode @marcromeyn (#456)
• Add validation for movielens variant and fix docstring @oliverholworthy (#434)
• Use collections.abc for Sequence import @oliverholworthy (#427)

🚀 Features

• XGBoost - Switch to dask API @oliverholworthy (#466)
• Fixes and improves masking (for both Causal LM and Masked LM) and adds SequenceAggregation.MASKED_MEAN (used by YouTubeDNN) @gabrielspmoreira (#464)
• Moving loss- and metrics-calculation to model instead of the predicti… @marcromeyn (#431)
• XGBoost - Change group_columns to qid_column @oliverholworthy (#462)
• First sketch of XGBoost integration with Merlin Dataset @oliverholworthy (#433)
• Getting rid of feature-variables in ModelContext @marcromeyn (#455)
• Adding the missing logQ correction to sampled softmax @gabrielspmoreira (#457)
• Making YoutubeDNNRetrievalModel more configurable @gabrielspmoreira (#447)
• Making EmbeddingOptions.embedding_dims to have priority over .inferred_embedding_dims @gabrielspmoreira (#398)

📄 Documentation

• Update to refer to merlin-tensorflow container @mikemckiernan (#512)
• Fix prediction task documentation @benfred (#505)
• Fix dataloader docstring @benfred (#492)
• Add common release-drafter workflow @mikemckiernan (#439)
• Add common release-drafter configuration @mikemckiernan (#435)
• Adding container info to notebook 01, 02 and 06 @bschifferer (#424)

🔧 Maintenance

• Fix pytest marker for integration tests. @oliverholworthy (#510)
• Add a GA workflow that requires labels on PR's @benfred (#506)
• Use shared implementation of triage workflow @benfred (#504)
• remove unnecessary line @radekosmulski (#494)
• Adding automatic branch pruning to ParallelBlock @marcromeyn (#477)
• Move tabular tests from test_core to test_tabular @marcromeyn (#476)
• Remove tf.test.is_gpu_available call @benfred (#470)
• Update setup.py to include only the merlin namespace @oliverholworthy (#469)
• Add labels to PRs for change log @mikemckiernan (#453)
• Make use of model_test for all ranking- and benchmark-models @marcromeyn (#440)
• Move tests to tests/unit or tests/integration @jperez999 (#445)
• Creating generic model_test @marcromeyn (#430)
• Create separate a Github-workflow for each backend @marcromeyn (#423)
• rm git pull cmd in test unit @jperez999 (#438)
• Adding CI for notebooks 03, 04, 05 and 06 @bschifferer (#414)

v0.4.0

Known Issues

• model.evaluate() gives much higher acc value for small batch size when we set validation_data int the model.fit() #373

What's Changed

• Fixing MF by applying L2-norm in the towers individually by @gabrielspmoreira in #353
• Ensure inferred embedding dim from feature cardinality is multiple of 8 by @gabrielspmoreira in #355
• Update advanced_example notebook by @sararb in #360
• Refactoring the evaluation of RetrievalModel by @marcromeyn in #352
• First design of popularity correction by @sararb in #362
• Fixes and improvements in embeddings export for MatrixFactorizationModel and TwoTowerModel by @gabrielspmoreira in #365
• Adding AlphaDropout support, which is recommended by selu activation by @gabrielspmoreira in #369
• Simplified TopKIndexBlock implementation and fixed intermittent inconsistency on retrieval models eval by @gabrielspmoreira in #370
• Evaluation metrics of a top-k recommender by @sararb in #371
• Adding unttests for example notebooks by @bschifferer in #377
• Adding option for embeddings_l2_reg, which is in particular critical for MF by @gabrielspmoreira in #381
• Creating a reg_factor for PopularityLogitsCorrection by @gabrielspmoreira in #386
• Adding unittests for Notebook 04 and 05 by @bschifferer in #385
• fix accessing item-id embeddings table based on IntDomain name by @sararb in #376
• Fix top-k recommender to return correct top-k relevant items. by @sararb in #389
• Update testbook tests to use absolute notebook paths by @karlhigley in #394
• Add missing features to Ali-ccp data generator by @rnyak in #399
• Fixes error with fit() with validation data by @gabrielspmoreira in #402
• Update README.rst by @bschifferer in #401
• Add unit test for 03_exploring_different_models by @rnyak in #405
• remove unneeded packages by @jperez999 in #412
• Use shuffle_df from merlin-core by @benfred in #413
• Adding CI for movielens notebooks by @bschifferer in #411
• Remove .python-version file by @karlhigley in #416
• Update the tags selection from the schema by excluding the target variable by @sararb in #397
• Create a basic conda recipe by @benfred in #420
• docs: Use sphinx-ext-toc to control examples by @mikemckiernan in #404

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

v0.3.0

What's Changed

• fix for tf when creating cpu dataloader by @jperez999 in #222
• Create the necessary blocks/methods for ItemRetrieval evaluation by @sararb in #189
• Adding wrap_as_model to Block by @marcromeyn in #209
• chore: add PR doc previews by @mikemckiernan in #229
• Refactors and optimizes evaluation with ranking metrics by @gabrielspmoreira in #226
• Dataloader/targets by @marcromeyn in #195
• Break up core.py by @marcromeyn in #227
• Improvements in two-tower model and DCN by @gabrielspmoreira in #237
• Adding NCFModel as benchmark by @sararb in #230
• fix: Typo in workflow script by @mikemckiernan in #248
• Updates container testing by @albert17 in #234
• Specify encoding when reading csv file for ml-100k by @benfred in #249
• chore: Trigger on closed pull requests by @mikemckiernan in #251
• docs: Incremental improvements by @mikemckiernan in #193
• Set up automated docstring coverage checks by @karlhigley in #257
• Automatic setting of retrieval model top-k evaluation by @sararb in #238
• Fixing MLPBlock saving with regularizers by @marcromeyn in #264
• Check if valid data is merlin.io.dataset and then convert to BatchedDataset by @rnyak in #269
• Fix mypy issues & make MF return a RetrievalModel by @marcromeyn in #266
• Update Dataset schema when schema is passed in BatchedDataset by @marcromeyn in #265
• Fix small bug when shuffle is passed in .fit by @marcromeyn in #272
• Moving RetrievalModel._ensure_unique to utils by @marcromeyn in #267
• Fix bug in tower saving where all features were required by @marcromeyn in #275
• Move interrogate docstring coverage settings to pyproject.toml by @karlhigley in #276
• Fix MatrixFactorizationModel bug by @sararb in #280
• Implementation of DeepFM model by @sararb in #271
• Add intersphinx mapping for merlin-core by @benfred in #281
• Examples for Getting Started and Schema File by @bschifferer in #253
• Explore different ranking models by @rnyak in #256
• Item retrieval evaluation fixes by @gabrielspmoreira in #279
• [Doc] a hyperlink missing by @rhdong in #286
• docs: Add API documentation by @mikemckiernan in #212
• Exclude additional directories from docstring coverage req't by @karlhigley in #293
• fix evaluation loss equal to zero and add docstrings by @sararb in #294
• docs: Add nightly multi-version build by @mikemckiernan in #298
• fix: Do not rebuild on PR close by @mikemckiernan in #299
• Retrieval models (ALI-CCP)- Two-Tower model example by @rnyak in #262
• docs: Revise API page by @mikemckiernan in #305
• Exporting ranking models by @rnyak in #300
• Small fixes on Matrix Factorization by @gabrielspmoreira in #306
• first draft of advanced example notebook by @sararb in #295
• Fix default int64 dtype in index.py to avoid dtype errors in Retrieval model validation step by @rnyak in #307
• Specify encoding when reading csv file for ml-1m by @benfred in #310
• Example Readme by @bschifferer in #296
• fix categorical-ohe block by @sararb in #314
• Improving handling of data in examples + tests by @marcromeyn in #309
• quick fix in aliccp schema by @sararb in #323
• fix binary tag in Ali-CCP workflow by @sararb in #325
• updates from readme bash by @sararb in #303
• Pairwise losses fixes by @gabrielspmoreira in #317
• Upgrading black + enabling black on notebooks by @marcromeyn in #326
• docs: Add redirect page by @mikemckiernan in #329
• Fix the evaluation of sampled softmax-based tasks by @sararb in #315
• Fixing synthetic data-generation for aliccp-raw by @marcromeyn in #327
• Adding codespell + flake8-nb to pre-commit by @marcromeyn in #331
• remove barchart from example_utils by @rnyak in #334
• Fix 03-03-05 nbs wrt bug bash by @rnyak in #333
• Merlin models examples fixes by @gabrielspmoreira in #313
• Update NVTabular and merlin-core requirements for the release by @benfred in #347
• Report skipped tests in blossom CI by @benfred in #348
• Fixing error on retrieval model evaluation when using mixed_precision by @gabrielspmoreira in #337
• Automate pushing package to pypi by @benfred in #349

New Contributors

• @rhdong made their first contribution in #286

Full Changelog: v0.2.0...v0.3.0

v0.2.0

What's Changed

• Fix TARGET tag typo in dataloader and sampler initialization in ItemRetrievalTask() by @rnyak in #176
• Pointwise, Pairwise and Listwise Losses by @gabrielspmoreira in #166
• Add download, convert and etl util func for movielens dataset by @rnyak in #155
• Copying Github templates from T4Rec by @marcromeyn in #165
• [REVIEW] docs: initial commit for Sphinx docs by @mikemckiernan in #174
• fix references to item-id column by @sararb in #180
• Activates GPU CI System by @albert17 in #186
• quick fix of setting dtypes from sparse inputs by @sararb in #183
• Adding some basic doc-strings by @marcromeyn in #185
• Introducing TupleAggregator & use it for CosineSimilarity + ElementWiseMultiply by @marcromeyn in #164
• Unified Schema by @benfred in #184
• Adding models by @marcromeyn in #187
• Fix SyntheticData.read_schema with proto text files by @benfred in #191
• Adding batch-predict by @marcromeyn in #145
• Move to 'merlin' namespace by @benfred in #198
• Fix ItemRetrieval assertion and output errors by @rnyak in #200
• Adds initializers and regularizers options to MLPBlock by @gabrielspmoreira in #206
• Update merlin core version by @benfred in #213
• Assume 'merlin' is a first party package for isort by @benfred in #216
• Implemented support for TF mixed precision (mixed_float16) by @gabrielspmoreira in #217
• Add shim to adapt the implicit library by @benfred in #218
• Add a helper function to get the movielens dataset by @benfred in #207
• Add shim to adapt the lightfm library by @benfred in #219

New Contributors

• @mikemckiernan made their first contribution in #174

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

v0.1.0

Initial public release