TensorFlow Recommenders Addons v0.4.1

Release Notes

This is a patch version for TensorFlow Recommenders-Addons v0.4.x

What's Changed

• Reverse update logic on disabling bp_v2 by @Lifann in #306
  • This optimization can improve the AUC. So we recommend you update it.
  • This optimization only works in session graph mode, and the Keras API would not be affected.

Thanks to @Lifann for finding and fixing it!

TensorFlow Recommenders Addons v0.3.2

Release Notes

This is a patch version for TensorFlow Recommenders-Addons v0.3.x

What's Changed

• Reverse update logic on disabling bp_v2 by @Lifann in #306
  • This optimization can improve the AUC. So we recommend you update it.
  • This optimization only works in session graph mode, and the Keras API would not be affected.

Thanks to @Lifann for finding and fixing it!

TensorFlow Recommenders Addons v0.6.0

Release Notes

This is the 6th release of TensorFlow Recommenders-Addons

Main Features or Fixs:

• [Fix] Reverse update logic within bp_v2 by @Lifann in #306
  • This bug fix is critical and hurts the AUC. So we highly recommend you update it as soon as possible.
  • The related bug only happens in Graph Mode in synchronized training, and the Keras with eager mode would not be affected.
• [Feat] Support unique in FieldWiseEmbedding and SquashedEmbedding by @rhdong in #311
• [Feat] Add unique to keras.layers.BasicEmbedding by @rhdong in #309
• [Feat]Add the features for expanding and shrinking the number of tables in distributed training by independently saving files. by @MoFHeka in #305
• [Fix] fix error that self._saveable_fn would not exist when checkpoint param is False. by @MoFHeka in #310
• [Fix] Fixed Redis table unit test (#308) @LinGeLin
• [Fix]Fix the bug that rehash might not be triggered by @acmore in #299
• [CI] Removed build warning: 'type qualifier is meaningless on cast type' by @LinGeLin in #298
• [Fix] Remove redundant code by @WangShengguang in #294
• [Fix] no register table kernels for int32-float type. by @rhdong in #293

Thanks to our Contributors

@Lifann, @LinGeLin, @MoFHeka and @rhdong

New Contributors

• @WangShengguang made their first contribution in #294
• @LinGeLin made their first contribution in #298

Full Changelog: https://github.com/tensorflow/recommenders-addons/commits/v0.6.0

TensorFlow Recommenders Addons v0.5.1

Release Notes

This is a patch version for TensorFlow Recommenders-Addons v0.5.0

What's Changed

• Reverse update logic on disabling bp_v2 by @Lifann in #306
  • This optimization can improve the AUC. So we recommend you update it.
  • This optimization only works in session graph mode, and the Keras API would not be affected.

Thanks to @Lifann for finding and fixing it!

TensorFlow Recommenders Addons v0.5.0

Release Notes

This is the 5th release of TensorFlow Recommenders-Addons

Main Features:

• tfra.dynamic_embedding

  • Add int64-int64 key-value types on GPU. by @Lifann in #240
  • Add HMACCUM Redis module and enable BPV2 in Redis implement by @thorneliu in #234, Readme
  • Add redis_hash_tags_hypodispersion and using_hash_storage_slice to Redis backend config. Change redis_connection_mode config, now redis_connection_mode = 2 will be standalone mode. by @MoFHeka in #241, Guide
  • Save and load cuckoo hashtable from HDFS by @luliyucoordinate in #257
  • Add embedding table compression script by @xiangzez in #261, Readme
  • [Opt] Avoid reallocating and copying device memory when the inputs of the Remove and ImportValues functions are already device variables. by @MoFHeka in #279
  • Make HDFS feature support S3 and compatible with tensorflow-io and tensorFlow<=2.6. by @MoFHeka in #281
  • Add bfloat16 value type support for CPU by @xiangzez in #268

What's Changed

• [fix] demo issue for 'horovod_synchronous' -> 'synchronous' by @rhdong in #233
• Add int64-int64 key-value in combination on GPU. by @Lifann in #240
• [Feat] add HMACCUM redis module and enable BPV2 in redis impl by @thorneliu in #234
• [feat] Add redis_hash_tags_hypodispersion and using_hash_storage_slice to redis backend config. Change redis_connection_mode config, now redis_connection_mode = 2 will be standalone mode. by @MoFHeka in #241
• Fix devel container by @rhdong in #248
• [fix] Fixed segment fault due to lambda may capture the thread_context reference with error address by @MoFHeka in #249
• [fix] stopping the Bazel in dev docker to update automatically . by @rhdong in #250
• [refactor] clean some warnings by @rhdong in #252
• fix(comment): Fix error in comment by @Lifann in #255
• [fix] Fixed bug which cause by hiredis updating failed to compile. by @MoFHeka in #260
• fix(optimizer): Bug: Variable set to trainable to false still updating embedding by @Lifann in #263
• [Feat] Save and load cuckoo hashtable from hdfs by @luliyucoordinate in #257
• Add embedding table compression script by @xiangzez in #261
• [Fix]debug CI fail by @rhdong in #267
• Add bfloat16 value type support for CPU by @xiangzez in #268
• [Fix] Avoid double lookup of tables when using ShadowVariable when the Optimizer is updating gradients by @MoFHeka in #262
• [CI] enable unit test on GPU in dev container. by @rhdong in #254
• [fix] Fix bug that user self-defined variable name would be modified to "TrainableWrapper" when Keras eager mode. by @MoFHeka in #275
• [fix] Compatible with compiling problem with TensorFlow version 2.4 and 2.8. by @MoFHeka in #276
• [Fix]Rehash gpu table before importing values and reduce size kernel call. by @acmore in #256
• [feat] Avoid reallocating and copying device memory when the inputs of the Remove and ImportValues functions are already device variables. by @MoFHeka in #279
• [fix] fix shadow variable lookup race condition by @alionkun in #280
• [feat] Make hdfs function compatible with tensorflow-io and TensorFlow version earlier than 2.6 by using FileSystem in TensorFlow Env. Also support S3 and so on now. by @MoFHeka in #281
• [fix] fix device of table shard was lost in savedmodel pb file when save model with keras api. by @MoFHeka in #283
• [release] change version to 0.5.0 without dev by @rhdong in #285

New Contributors

• @xiangzez made their first contribution in #261
• @alionkun made their first contribution in #280

Full Changelog: https://github.com/tensorflow/recommenders-addons/commits/v0.5.0

TensorFlow Recommenders Addons v0.4.0

Release Notes

This is the 4th release of TensorFlow Recommenders-Addons

Main Features:

• tfra.dynamic_embedding

  • Support keras style programming in eager mode and enable tf.function grammar (#196), Demo @Lifann
  • Support warm-start in tfra.dynamic_embedding (#221), Demo @dakabang @Mr-Nineteen
  • Support quasi-synchronous training based on Horovod (#205), Demo @a6802739
    • Synchronous training for dense weights
    • Quasi-synchronous training for sparse weights
  • Support Apple M1(ARM64) on MacOS (#214), Guide @poinwater
  • Add storage_slice_import parameter on Redis for easy rehashing table. (#197) @MoFHeka
  • Add supported key-value type(int32, float32) for GPU @rhdong

What's Changed

• [Doc]Add Apple Silicon installation instruction & fix typo (#228) @poinwater
• [Doc] Update api docs (#229) @Lifann
• [Doc] update readme (#224) @rhdong
• Support keras style programming in eager mode and enable tf.function (#196) @Lifann
• Add patch to tensorflow-docs for pretty on top link (#223) @Lifann
• Support warm-start in tfra.dynamic_embedding (#221) @dakabang
• [fix] Fix an issue where tf_patch.py failed in TF >= 2.6.x. (#222) @MoFHeka
• [Feat]support horovod sync train (#205) @a6802739
• [fix] add mpi & hvd installation to macOS CI and cpu_tests (#219) @rhdong
• [feat]support Mac m1 devices (#214) @poinwater
• [Feat]Remove the lock when choosing the thread context. (#216) @MoFHeka
• [CI] add openmpi and horovod to base_install docker. (#210) @rhdong
• [fix]Avoid Floating point exception (core dumped) when using RedisTable by empty ids inputs. (#206) @MoFHeka
• [fix]version is always "v0.1.0" (#208) @rhdong
• [CI] Install RocksDb and its dependency to CI docker image. (#204) @rhdong
• [fix] embedding variable graph mode example (#189) @Nov11
• [demo] add movielens-100k-estimator-with-warmup (#203) @Mr-Nineteen
• [fix] Updated rules_foreign_cc from 0.0.9 to 0.6.0 (#202) @poinwater
• [CI] remove CI on TF1.15.2 (#201) @rhdong
• [feat] Add storage_slice_import parameter for easy rehashing table. (#197) @MoFHeka
• [CI]after github action updated, we cancel python3.6 for MacOS (#194) @rhdong
• Do not batch reduce as keras optimzer already does. (#183) @acmore
• [fix]Compile compatibility to TF2.6. Add '-lrt' for Redis backend. Fix FindWithExist function in Redis backend. (#191) @MoFHeka
• [feat] add supported key-value type(int32, float32) for GPU (#188) @rhdong
• [fix]Fix misprocessing of "CLUSTER NODES" return (#185) @MoFHeka
• [Fix]change 'KEYS' command in GetKeyBucketsAndOptimizerParamsWithName function to 'SCAN' command (#180) @MoFHeka

Thanks to our Contributors

@Lifann, @MoFHeka, @Mr-Nineteen, @Nov11, @a6802739, @acmore, @dakabang, @poinwater, @rhdong, @zhangyafeikimi

Full Changelog: https://github.com/tensorflow/recommenders-addons/commits/v0.4.0

TensorFlow Recommenders Addons v0.3.1

Release Notes

This is a patch version for TensorFlow Recommenders-Addons v0.3.0

What's Changed

• [fix] Illegal Instruction error on AMD CPUs by @rhdong in #176

TensorFlow Recommenders Addons v0.3.0

Release Notes

This is the 3rd release of TensorFlow Recommenders-Addons
Main Features:

• tfra.dynamic_embedding
  • Support Redis as Key-Value Backend by @MoFHeka , demo, doc
  • Support bpv2 (update de by adding delta instead of setting) by @hxfxjun @rhdong , demo
  • Add Clear API for dynamic_embedding.Variable by @yuanqingsunny , PR
• tfra.embedding_variable
  • Support tf.fixed_size_partitioner by @qqsun8819

What's Changed

• add dynamic_embedding.Variable clear by @yuanqingsunny in #100
• [doc] add DynamicEmbeddingOptimizer to rfc by @rhdong in #101
• [doc-fix]add TF_VERSION to env. by @rhdong in #106
• [fix] Import of GPUHashTable should clear firstly. by @rhdong in #108
• [fix] style guide spell error by @rhdong in #120
• [feat] add Accum and FindWithExists APIs for hashtable by @rhdong in #116
• [fix] fix syntax error in EmbeddingVariable tutorial by @qqsun8819 in #127
• [fix] fix runtime error in EmbeddingVariable tutorial for graph mode by @qqsun8819 in #129
• [fix] embedding variable is_initialized not executed in graph mode by @qqsun8819 in #130
• Fix ktype not set in embedding variable tutorial by @qqsun8819 in #131
• Fix embedding variable tutorial notebook format error by @qqsun8819 in #132
• [feat] bpv2 (update de by adding delta instead of setting) by @nrailgun in #133
• Support Keras APIs and fix save-restore issue in TF2 by @Lifann in #119
• EmbeddingVariable add local and distributed examples by @qqsun8819 in #141
• Doc: add instruction for sign git commit for PR by @qqsun8819 in #142
• [demo] move ev example to demo. by @rhdong in #144
• Fix bug: incompatible jobs with explicit placement. by @Lifann in #145
• EmbeddingVarialbe support partitioner by @qqsun8819 in #140
• Support DynamicEmbeddingOptimizer bp_v2 parameter Settings, solve the… by @hxfxjun in #148
• embedding_variable.get_variable support int32 key type by @qqsun8819 in #149
• Revert "Add Redis backend feature to TFRA" by @rhdong in #154
• correct ev doc link by @qqsun8819 in #152
• [feat]Add Redis backend feature to TFRA by @MoFHeka in #157
• add qqsun8819 as code owner by @qqsun8819 in #161
• fix cleate_slot_var import xla_sharding error in tf >=2.5 by @qqsun8819 in #160
• [fix] Added more type and value check for Redis backend config file to prevent unconditional setting. by @MoFHeka in #159
• [fix]Github auto merge hide the email address, revert and push the right address by @MoFHeka in #163
• [feat] Update dependent & CI & readme for TF version 2.5.1 by @rhdong in #165
• [fix] CI fail on TF2.4.1 by @rhdong in #168
• [hotfix]I forget to modify RestoreFromDisk and HscanGetKeysValsInBucket in cluster mode code by @MoFHeka in #166

Thanks to our Contributors

• @Lifann
• @yuanqingsunny
• @qqsun8819
• @nrailgun
• @hxfxjun
• @MoFHeka
• @rhdong

Full Changelog: https://github.com/tensorflow/recommenders-addons/commits/v0.3.0

TensorFlow Recommenders Addons v0.2.0

Release Notes

This is the 2nd release of TensorFlow Recommenders-Addons
Main Features:

• tfra.dynamic_embedding
  • Support Hashtable on GPU and some GPU OPs.
  • Support TensorFlow 1.15.2
  • Add tfra.dynamic_embedding.embedding_lookup_unique API
  • Improved the documentation and demo
• tfra.embedding_variable
  • Support save/restore

Release Notes

• [CI]change version to 0.2.0 (#91) @rhdong
• [Feat] compatible with TF1.15.2 (#90) @rhdong
• [Feat] Add GPU operator: sparse_fill_empty_rows and sparse_reshape (#87) @funsimple
• [CI]release-drafter fail for shortened SHA (#89) @rhdong
• [CI]Support TF Serving compilation (#85) @Mr-Nineteen
• [Doc]Update api_docs for dynamic_embedding (#82) @Lifann
• [Doc]update readme, added the way to compile GPU Serving. (#83) @Mr-Nineteen
• [Feat]GPU supports Unified Memory (#79) @Mr-Nineteen
• [Feat]Add GPU hashtable and GPU workflow (#78) @rhdong
• [Feat]Support sparse_segment_sum ops on GPU. (#55) @Lifann
• [Fix] optimizer slots have 1/N init_size of primary var. (#75) @rhdong
• [feat] add Hashtable On GPU (#74) @rhdong
• [demo]add estimator demo (#72) @tracyxzh001
• [feat] add dynamic_embedding.embedding_lookup_unique API. (#62) @funsimple
• [Demo] add a distributed training demo on movielens for dynamic_embedding (#61) @rhdong
• [Fix] optimized hashtable not work (#66) @rhdong
• [CI]Add style guide for bazel BUILD files. (#58) @Lifann
• [Opt]Remove ‘LaunchTensorsInsert num_worker_threads’ log printing. (#63) @Mr-Nineteen
• [Fix]Fix bazel build error with cuda (#53) @Lifann
• [opt] reduce memory consuming of de.Variable and support string type Keys and Values. (#57) @rhdong
• [fix] wrong TF version warning (#54) @rhdong
• For LaunchTensorsInsert, num_worker_threads is used only when TFRA_NUM_WORKER_THREADS_FOR_LOOKUP_TABLE_INSERT env var is set to k, where k>0 and k<tf current number of cpu worker threads. Otherwise, nothing will change. (#51) @Mr-Nineteen
• [doc-fix] optimizer apply updates on TrainableWrapper. (#46) @wuxianxingkong
• [doc]add API doc for dynamic_embedding.DynamicEmbeddingOptimizer (#50) @rhdong
• [fix]batch port bot error (#40) @rhdong
• [doc-fix] wrap the optimizer in tfra.dynamic_embedding.DynamicEmbeddingOptimizer (#43) @funsimple
• [Feat]EmbeddingVariable support save/restore (#39) @candyzone

Thanks to our Contributors

@Lifann, @Mr-Nineteen, @candyzone, @funsimple, @rhdong, @tracyxzh001, @wuxianxingkong
We also want to extend a thank to @郁凡 from NVIDIA Co. who have helped with HashTable, and Nvidia China team that helped implement SparseFillEmptyRowsOP and SparseReshapeOP On GPU!

TensorFlow Recommenders-Addons v0.1.0

Release Notes

This is the first release of TensorFlow Recommenders-Addons!

Change Log

• Add tfra.dynamic_embedding
• Add tfra.embedding_variable

Thank You to All Our Contributors

• Cheng Chuan
• Ding Chen
• Fan Guifeng
• Li Fan
• Rong Haidong
• Wang Dongxin

We also want to extend a thank you to the Google team members who have helped with CI setup and reviews!