TensorFlow Java 1.0.0-RC.2

This is the second release candidate for TensorFlow Java 1.0.0, with added support for Linux Arm64 systems, improved in-code documentation (JavaDoc) and other minor fixes.

Conformity to semantic versioning will start with the 1.0.0 release, as per TensorFlow API stability guarantees.

Major Changes

• Support of Linux Arm64 systems by @snadampal in #545
• Upgrade to TF2.16.2 by @karllessard in #557

Other Changes

• Graph is changed to public in Session.java by @manojava98 in #535
• Shape proto fix by @karllessard in #539
• Accept partially known shapes in boolean mask/updates by @karllessard in #541
• Various javadoc fixes by @Craigacp and @karllessard in #542, #555
• Add user instructions on how to use new module names by @karllessard in #554

New Contributors

• @manojava98 made their first contribution in #535
• @snadampal made their first contribution in #545

Full Changelog: v1.0.0-rc.1...v1.0.0-rc.2

TensorFlow Java 1.0.0-RC.1

This is the first release candidate for TensorFlow Java 1.0.0, and includes a major refactoring of the source structure and the CI/CD pipelines (see #521). While most the API has not been impacted, there are some breaking changes that might require updating the code of your application if you've been using TensorFlow Java before, please see these guidelines for more details.

Conformity to semantic versioning will start with the 1.0.0 release, as per TensorFlow API stability guarantees.

Major Changes

• Upgrade to TensorFlow 2.16.1
• Major refactoring of the build and CI/CD pipelines, to facilitate next TensorFlow upgrades
  • Using official TensorFlow native builds in published TensorFlow Java artifacts
  • Publish macOS ARM64 binaries as one of the default platforms
  • Enable compatibility with existing TensorFlow plugins (e.g. tensorflow-text)
• Fix non-deterministic gradient descent during training

All Changes

• Fixed the broken link in install.md by @tiruk007 in #491
• Update ci.yml to ubuntu-20.04 by @Craigacp in #503
• [linux-arm64] set "mkl_aarch64" bazel config for linux-arm64 platform by @snadampal in #502
• Update ci.yml by @Craigacp in #504
• Refactoring native builds and Java binding generation by @karllessard in #521
• Interface should be public for external usage by @nfeybesse in #522
• fix #526 by @nfeybesse in #527
• Upgrade to TF 2.16.1 by @karllessard in #531
• Adding casts to the if test so it passes on GPU by @Craigacp in #528
• Pom profile changes by @Craigacp in #525
• Fixing the native build so it can find the TSL headers by @Craigacp in #534
• Make all TF operations visible via the Ops API by @karllessard in #533

New Contributors

• @tiruk007 made their first contribution in #491
• @snadampal made their first contribution in #502
• @nfeybesse made their first contribution in #522

Full Changelog: v0.5.0...v1.0.0-rc.1

TensorFlow Java 0.5.0

TensorFlow Java is still only available as an alpha release and is not covered by the TensorFlow API stability guarantees

Features

• Upgrade to TensorFlow 2.10.1
• Map sparse tensors as a single tensor entity
• Add utility class for releasing all tensors returned on inference in a single batch
• Support Uint16 tensor type
• Support TF Python Config for deserializing from models saved with the TF Engine outside of Java

Bug Fixes and Other Changes

• Bump minimal JDK version to Java 11
• Enable custom builds for Apple Silicon chip
• Fix fail when a variable receives zero gradient
• Simplify graph initialization for better performance
• Upgrade to NdArray 0.4.0
• Various improvements and doc adjustments

Thanks to our contributors

Adam Pocock (@Craigacp)
Karl Lessard (@karllessard)
Ramón García Fernández (@ramon-garcia)
Joe Yearsley (@joeyearsley)
Samuel Audet (@saudet)
Jim Clarke (@JimClarke5)

TensorFlow Java 0.4.2

TensorFlow Java is still only available as an alpha release and is not covered by the TensorFlow API stability guarantees

Bug Fixes and Other Changes

• Upgrade to TensorFlow 2.7.4 with CVE fixes
• Adding a hook for loading TF IO native libraries (#468)

Thanks to our contributors

Adam Pocock (@Craigacp)

TensorFlow Java 0.4.1

TensorFlow Java is still only available as an alpha release and is not covered by the TensorFlow API stability guarantees

Bug Fixes and Other Changes

• Upgrade to TensorFlow 2.7.1 with CVE fixes (#412)
• Bumping to protobuf 3.19.4, error-prone 2.10.0 (#412)
• Fix NullPointerException issue. (#407)

Thanks to our contributors

Adam Pocock (@Craigacp)
Frank Liu (@frankfliu)
Karl Lessard (@karllessard)

TensorFlow Java 0.4.0

TensorFlow Java is still only available as an alpha release and is not covered by the TensorFlow API stability guarantees

Major Features

• Upgrade to TensorFlow 2.7.0
• Add support for creating and running native function graphs
• Automatic initialization of the variables when running a graph for the first time (no need of tf.init anymore)
• Register custom gradient computations in Java for operations lacking one in the TensorFlow C++ runtime library
• Add new metrics and regularizers to the TensorFlow Java Framework

Other Features

• Operation inputs and attributes accessors
• Allow output targets in Session runner
• Generate operation wrappers in Java, instead of using a custom C++ binary
• Move out Java NdArray to its own repository
• Add graph walking functions to Graph and GraphOperation
• Security updates
• Various bug and memory leak fixes, and other improvements

Thanks to our contributors

Ryan Nett (@rnett)
Karl Lessard (@karllessard)
Samuel Audet (@saudet)
Adam Pocock (@Craigacp)
Jim Clarke (@JimClarke5)
Till Brychcy (@brychcy)
Mark Daoust (@MarkDaoust)
Avi Perl (@aviperl)
P. Oscar Boykin (@johnynek)

TensorFlow Java 0.3.3

Note: TensorFlow Java is still in alpha release and is not covered by TensorFlow API stability guarantees

Bug Fixes and Other Changes

• Fix a memory leak happening when closing string tensors
• Fix an issue that prevented loading in Python a model trained and saved in Java
• Fix an issue that prevented loading in Java saved models with an empty tag set

Thanks to our contributors

Till Brychcy (@brychcy)
Karl Lessard (@karllessard)
Adam Pocock (@Craigacp)

TensorFlow Java 0.3.2

Note: TensorFlow Java is still in alpha release and is not covered by TensorFlow API stability guarantees

Bug Fixes and Other Changes

• Fix memory leak when closing saved model bundles loaded from disk
• Fix NPE when loading saved model bundles with empty tag

TensorFlow Java 0.3.1

Note: TensorFlow Java is still in alpha release and is not covered by TensorFlow API stability guarantees

Bug Fixes and Other Changes

• Fix memory leaks occurring with new TensorFlow 2.4.1 layout of string tensors
• Fix and use JavaCPP deallocator for releasing graph sessions

TensorFlow Java 0.2.0

This is a first alpha release of the new TensorFlow Java ecosystem.
Note that TensorFlow Java is not covered by the TensorFlow API stability guarantees

Major Features

• Build, train and run inference over a neural network graph using an enhanced Java API based on TensorFlow 2.3.1
• Execute eagerly TensorFlow operations for easy debugging
• Save and load graphs to/from disk using the saved model format
• Access directly tensor memory in a n-dimensional space using the new NdArray Java library
• Leverage native libraries support and installation using JavaCPP
• Train graphs using a new high-level Java framework for machine-learning

Thanks to our contributors

Karl Lessard (@karllessard)
Samuel Audet (@saudet)
Adam Pocock (@Craigacp)
Jim Clarke (@JimClarke5)
Dhruv Rajan (@dhruvrajan)
Shajan Dasan (@Shajan)
Alexey Zinoviev (@zaleslaw)
Christian Tzolov (@tzolov)
Han Dong (@handong0123)
Mark Daoust (@MarkDaoust)
James Ring (@sjamesr)
Lai Wei (@roywei)
Matteo Di Giovinazzo (@matteodg)
Haitang (@hthu)
Gunhan Gulsoy (@gunan)
Fred Reiss (@frreiss)
Dean Thompson (@deansher)
Christoph Henkelmann (@chenkelmann)
Eron Wright (@EronWright)