PyPI · #3785 most downloaded on PyPI

tensorflow-cpu

2.21.01.2M downloads/motensorflow/tensorflow ↗New releases in your daily digest, changelog included.

TensorFlow is an open source machine learning framework for everyone.

Last release 12 days ago
22 Sep 2026
Release timing varies
gaps range from 8 days to 7 months
Nearly every release is documented
notes for 59 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
7 years old
97 releases · first in 2019

Release timeline

97 releases since 2019

20192020202120222023202420252026

Releases

2.22.0rc022 Sept 2026pre-releaseBreaking
<CAVEATS REGARDING THE RELEASE (BUT NOT BREAKING CHANGES).>
The TensorBoard dependency is no longer included by default. If you use tf.summary.* APIs or tf.keras.callbacks.TensorBoard, please install the tensorboard package separately (pip install tensorboard). Otherwise, TensorFlow will raise an ImportError.

In tensorflow/c/experimental/filesystem/filesystem_interface.h, removed TF_TransactionToken and associated APIs.

<CAVEATS REGARDING THE RELEASE (BUT NOT BREAKING CHANGES).>

<ADDING/BUMPING DEPENDENCIES SHOULD GO HERE>

<KNOWN LACK OF SUPPORT ON SOME PLATFORM, SHOULD GO HERE>

tf.lite

Adds support for QUI4 (Quantized Unsigned 4-bit) in Dequantize operator.

Adds support for FP16 and BF16 in Unpack operator.

Adds support for FP16 in Transpose and DynamicUpdateSlice operator.

Transpose now supports up to 8D tensors.

Adds support for FLOAT8_E4M3FN and FLOAT8_E5M2 data types.

BatchFunction Operator

Adds the num_warmup_batch_threads op attribute to support a separate thread pool for processing warmup requests.

Adds the per_criticality_batch_timeout_micros op attribute to support different batch timeouts for different criticality levels.

TensorFlow API

Exports __new__ in public API golden files for subclasses of tuple (like tf.io.FixedLenFeature) to fix false positives during static type checking.>

tf.data

Fixes a bug in tf.data.Dataset.scan where the shape of the state returned by scan_func was not strictly validated against the initial state.

tf.image.adjust_contrast

Registers the missing Python gradient for the AdjustContrastv2 op, so
tf.image.adjust_contrast can now be differentiated with
GradientTape. Fixes
#126083.

tf.nn.softsign

Fixes second-order gradients of tf.nn.softsign. Differentiating twice
previously failed with a lookup error because the SoftsignGrad
backward op had no registered Python gradient.

tf.math.reciprocal

Constrains the XLA registration of Reciprocal and Inv to the types
that have a device kernel, so jit_compile=True no longer silently
accepts the integer inputs that eager execution and autoclustering
reject. Fixes
#126414.

tf.experimental.numpy

tf.experimental.numpy.isclose and tf.experimental.numpy.allclose now
apply rtol and atol to integer inputs, matching NumPy, instead of
comparing with pure equality. The comparison stays in integer
arithmetic; combining integer inputs with floating-point tolerances
(including the defaults) now raises an error unless automatic type
promotion is enabled, which keeps the cost of promotion opt-in.

oneDNN (MKL) convolution and transpose kernels

Raises InvalidArgumentError instead of aborting the process for a
rank-mismatched Conv3DBackpropFilterV2 input and for
ConjugateTranspose on a scalar. Fixes
#118340 and
#118345.

tf.image.non_max_suppression

Fixes a segmentation fault under jit_compile=True when boxes and
scores are empty. The XLA lowering now returns an empty selection,
matching the non-XLA kernels. Fixes
#117245.

<SIMILAR TO ABOVE SECTION, BUT FOR OTHER IMPORTANT CHANGES / BUG FIXES>

<IF A CHANGE CLOSES A GITHUB ISSUE, IT SHOULD BE DOCUMENTED HERE>

This release contains contributions from many people at Google, as well as:

, , , , ,

From the same repository

tensorflow/tensorflow

tensorflowPyPITensorFlow is an open source machine learning framework for everyone.

2.18.025 Oct 2024Deprecated

Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline here.

Release 2.18.0

TensorFlow

Breaking Changes

tf.lite
- C API:
  - An optional, fourth parameter was added TfLiteOperatorCreate as a step forward towards a cleaner API for TfLiteOperator. Function TfLiteOperatorCreate was added recently, in TensorFlow Lite version 2.17.0, released on 7/11/2024, and we do not expect there will be much code using this function yet. Any code breakages can be easily resolved by passing nullptr as the new, 4th parameter.
TensorRT support is disabled in CUDA builds for code health improvement.
Hermetic CUDA support is added.

Hermetic CUDA uses a specific downloadable version of CUDA instead of the user’s locally installed CUDA. Bazel will download CUDA, CUDNN and NCCL distributions, and then use CUDA libraries and tools as dependencies in various Bazel targets. This enables more reproducible builds for Google ML projects and supported CUDA versions.

Known Caveats

Major Features and Improvements

TensorFlow now supports and is compiled with NumPy 2.0 by default. Please see the NumPy 2 release notes and the NumPy 2 migration guide.
- Note that NumPy's type promotion rules have been changed(See NEP 50for details). This may change the precision at which computations happen, leading either to type errors or to numerical changes to results.
- Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline here.
tf.lite:
- The LiteRT repo is live (see announcement), which means that in the coming months there will be changes to the development experience for TFLite. The TF Lite Runtime source will be moved later this year, and sometime after that we will start accepting contributions through that repo.
SignatureRunner is now supported for models with no signatures.

Bug Fixes and Other Changes

tf.data
- Add optional synchronous argument to map, to specify that the map should run synchronously, as opposed to be parallelizable when options.experimental_optimization.map_parallelization=True. This saves memory compared to setting num_parallel_calls=1.
- Add optional use_unbounded_threadpool argument to map, to specify that the map should use an unbounded threadpool instead of the default pool that is based on the number of cores on the machine. This can improve throughput for map functions which perform IO or otherwise release the CPU.
- Add tf.data.experimental.get_model_proto to allow users to peek into the analytical model inside of a dataset iterator.
tf.lite
- Dequantize op supports TensorType_INT4.
  - This change includes per-channel dequantization.
- Add support for stablehlo.composite.
- EmbeddingLookup op supports per-channel quantization and TensorType_INT4 values.
- FullyConnected op supports TensorType_INT16 activation and TensorType_Int4 weight per-channel quantization.
tf.tensor_scatter_update, tf.tensor_scatter_add and of other reduce types.
- Support bad_indices_policy.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Akhil Goel, akhilgoe, Alexander Pivovarov, Amir Samani, Andrew Goodbody, Andrey Portnoy, Anthony Platanios, bernardoArcari, Brett Taylor, buptzyb, Chao, Christian Clauss, Cocoa, Daniil Kutz, Darya Parygina, dependabot[bot], Dimitris Vardoulakis, Dragan Mladjenovic, Elfie Guo, eukub, Faijul Amin, flyingcat, Frédéric Bastien, ganyu.08, Georg Stefan Schmid, Grigory Reznikov, Harsha H S, Harshit Monish, Heiner, Ilia Sergachev, Jan, Jane Liu, Jaroslav Sevcik, Kaixi Hou, Kanvi Khanna, Kristof Maar, Kristóf Maár, LakshmiKalaKadali, Lbertho-Gpsw, lingzhi98, MarcoFalke, Masahiro Hiramori, Mmakevic-Amd, mraunak, Nobuo Tsukamoto, Notheisz57, Olli Lupton, Pearu Peterson, pemeliya, Peyara Nando, Philipp Hack, Phuong Nguyen, Pol Dellaiera, Rahul Batra, Ruturaj Vaidya, sachinmuradi, Sergey Kozub, Shanbin Ke, Sheng Yang, shengyu, Shraiysh, Shu Wang, Surya, sushreebarsa, Swatheesh-Mcw, syzygial, Tai Ly, terryysun, tilakrayal, Tj Xu, Trevor Morris, Tzung-Han Juang, wenchenvincent, wondertx, Xuefei Jiang, Ye Huang, Yimei Sun, Yunlong Liu, Zahid Iqbal, Zhan Lu, Zoranjovanovic-Ns, Zuri Obozuwa

Open source →

2.18.0rc216 Oct 2024pre-releaseDeprecated

Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline here.

Release 2.18.0

TensorFlow

Breaking Changes

tf.lite
- C API:
  - An optional, fourth parameter was added TfLiteOperatorCreate as a step forward towards a cleaner API for TfLiteOperator. Function TfLiteOperatorCreate was added recently, in TensorFlow Lite version 2.17.0, released on 7/11/2024, and we do not expect there will be much code using this function yet. Any code breakages can be easily resolved by passing nullptr as the new, 4th parameter.
- SignatureRunner is now supported for models with no signatures.
TensorRT support is disabled in CUDA builds for code health improvement.
Hermetic CUDA support is added.

Hermetic CUDA uses a specific downloadable version of CUDA instead of the user’s locally installed CUDA. Bazel will download CUDA, CUDNN and NCCL distributions, and then use CUDA libraries and tools as dependencies in various Bazel targets. This enables more reproducible builds for Google ML projects and supported CUDA versions.

Known Caveats

Major Features and Improvements

TensorFlow now supports and is compiled with NumPy 2.0 by default. Please see the NumPy 2 release notes and the NumPy 2 migration guide.
- Note that NumPy's type promotion rules have been changed(See NEP 50 for details). This may change the precision at which computations happen, leading either to type errors or to numerical changes to results.
- Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline here.
tf.lite:
- The LiteRT repo is live (see announcement), which means that in the coming months there will be changes to the development experience for TFLite. The TF Lite Runtime source will be moved later this year, and sometime after that we will start accepting contributions through that repo.

Bug Fixes and Other Changes

tf.data
- Add optional synchronous argument to map, to specify that the map should run synchronously, as opposed to be parallelizable when options.experimental_optimization.map_parallelization=True. This saves memory compared to setting num_parallel_calls=1.
- Add optional use_unbounded_threadpool argument to map, to specify that the map should use an unbounded threadpool instead of the default pool that is based on the number of cores on the machine. This can improve throughput for map functions which perform IO or otherwise release the CPU.
- Add tf.data.experimental.get_model_proto to allow users to peek into the analytical model inside of a dataset iterator.
tf.lite
- Dequantize op supports TensorType_INT4.
  - This change includes per-channel dequantization.
- Add support for stablehlo.composite.
- EmbeddingLookup op supports per-channel quantization and TensorType_INT4 values.
- FullyConnected op supports TensorType_INT16 activation and TensorType_Int4 weight per-channel quantization.
tf.tensor_scatter_update, tf.tensor_scatter_add and of other reduce types.
- Support bad_indices_policy.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.18.0rc17 Oct 2024pre-releaseDeprecated

Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline here.

Release 2.18.0

TensorFlow

Breaking Changes

tf.lite
- C API:
  - An optional, fourth parameter was added TfLiteOperatorCreate as a step forward towards a cleaner API for TfLiteOperator. Function TfLiteOperatorCreate was added recently, in TensorFlow Lite version 2.17.0, released on 7/11/2024, and we do not expect there will be much code using this function yet. Any code breakages can be easily resolved by passing nullptr as the new, 4th parameter.
- SignatureRunner is now supported for models with no signatures.
TensorRT support is disabled in CUDA builds for code health improvement.
Hermetic CUDA support is added.

Hermetic CUDA uses a specific downloadable version of CUDA instead of the user’s locally installed CUDA. Bazel will download CUDA, CUDNN and NCCL distributions, and then use CUDA libraries and tools as dependencies in various Bazel targets. This enables more reproducible builds for Google ML projects and supported CUDA versions.

Known Caveats

Major Features and Improvements

TensorFlow now supports and is compiled with NumPy 2.0 by default. Please see the NumPy 2 release notes and the NumPy 2 migration guide.
- Note that NumPy's type promotion rules have been changed(See NEP 50for details). This may change the precision at which computations happen, leading either to type errors or to numerical changes to results.
- Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline here.
tf.lite:
- The LiteRT repo is live (see announcement), which means that in the coming months there will be changes to the development experience for TFLite. The TF Lite Runtime source will be moved later this year, and sometime after that we will start accepting contributions through that repo.

Bug Fixes and Other Changes

tf.data
- Add optional synchronous argument to map, to specify that the map should run synchronously, as opposed to be parallelizable when options.experimental_optimization.map_parallelization=True. This saves memory compared to setting num_parallel_calls=1.
- Add optional use_unbounded_threadpool argument to map, to specify that the map should use an unbounded threadpool instead of the default pool that is based on the number of cores on the machine. This can improve throughput for map functions which perform IO or otherwise release the CPU.
- Add tf.data.experimental.get_model_proto to allow users to peek into the analytical model inside of a dataset iterator.
tf.lite
- Dequantize op supports TensorType_INT4.
  - This change includes per-channel dequantization.
- Add support for stablehlo.composite.
- EmbeddingLookup op supports per-channel quantization and TensorType_INT4 values.
- FullyConnected op supports TensorType_INT16 activation and TensorType_Int4 weight per-channel quantization.
tf.tensor_scatter_update, tf.tensor_scatter_add and of other reduce types.
- Support bad_indices_policy.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.18.0rc030 Sept 2024pre-releaseDeprecated

Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline here.

Release 2.18.0

TensorFlow

Breaking Changes

tf.lite
- C API:
  - An optional, fourth parameter was added TfLiteOperatorCreate as a step forward towards a cleaner API for TfLiteOperator. Function TfLiteOperatorCreate was added recently, in TensorFlow Lite version 2.17.0, released on 7/11/2024, and we do not expect there will be much code using this function yet. Any code breakages can be easily resolved by passing nullptr as the new, 4th parameter.
- SignatureRunner is now supported for models with no signatures.
TensorRT support is disabled in CUDA builds for code health improvement.
Hermetic CUDA support is added.

Hermetic CUDA uses a specific downloadable version of CUDA instead of the user’s locally installed CUDA. Bazel will download CUDA, CUDNN and NCCL distributions, and then use CUDA libraries and tools as dependencies in various Bazel targets. This enables more reproducible builds for Google ML projects and supported CUDA versions.

Known Caveats

Major Features and Improvements

TensorFlow now supports and is compiled with NumPy 2.0 by default. Please see the NumPy 2 release notes and the NumPy 2 migration guide.
- Note that NumPy's type promotion rules have been changed(See NEP 50for details). This may change the precision at which computations happen, leading either to type errors or to numerical changes to results.
- Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline here.
tf.lite:
- The LiteRT repo is live (see announcement), which means that in the coming months there will be changes to the development experience for TFLite. The TF Lite Runtime source will be moved later this year, and sometime after that we will start accepting contributions through that repo.

Bug Fixes and Other Changes

tf.data
- Add optional synchronous argument to map, to specify that the map should run synchronously, as opposed to be parallelizable when options.experimental_optimization.map_parallelization=True. This saves memory compared to setting num_parallel_calls=1.
- Add optional use_unbounded_threadpool argument to map, to specify that the map should use an unbounded threadpool instead of the default pool that is based on the number of cores on the machine. This can improve throughput for map functions which perform IO or otherwise release the CPU.
- Add tf.data.experimental.get_model_proto to allow users to peek into the analytical model inside of a dataset iterator.
tf.lite
- Dequantize op supports TensorType_INT4.
  - This change includes per-channel dequantization.
- Add support for stablehlo.composite.
- EmbeddingLookup op supports per-channel quantization and TensorType_INT4 values.
- FullyConnected op supports TensorType_INT16 activation and TensorType_Int4 weight per-channel quantization.
tf.tensor_scatter_update, tf.tensor_scatter_add and of other reduce types.
- Support bad_indices_policy.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.17.011 Jul 2024

Support for NVIDIA GPUs with compute capability 5.x (Maxwell generation) has been removed from TF binary distributions (Python wheels).

Release 2.17.0

TensorFlow

Breaking Changes

GPU
- Support for NVIDIA GPUs with compute capability 5.x (Maxwell generation) has been removed from TF binary distributions (Python wheels).

Major Features and Improvements

Add is_cpu_target_available, which indicates whether or not TensorFlow was built with support for a given CPU target. This can be useful for skipping target-specific tests if a target is not supported.
tf.data
- Support data.experimental.distribued_save. distribued_save uses tf.data service (https://www.tensorflow.org/api_docs/python/tf/data/experimental/service) to write distributed dataset snapshots. The call is non-blocking and returns without waiting for the snapshot to finish. Setting wait=True to tf.data.Dataset.load allows the snapshots to be read while they are being written.

Bug Fixes and Other Changes

GPU
- Support for NVIDIA GPUs with compute capability 8.9 (e.g. L4 & L40) has been added to TF binary distributions (Python wheels).
Replace DebuggerOptions of TensorFlow Quantizer, and migrate to DebuggerConfig of StableHLO Quantizer.
Add TensorFlow to StableHLO converter to TensorFlow pip package.
TensorRT support: this is the last release supporting TensorRT. It will be removed in the next release.
NumPy 2.0 support: TensorFlow is going to support NumPy 2.0 in the next release. It may break some edge cases of TensorFlow API usage.
tf.lite
- Quantization for FullyConnected layer is switched from per-tensor to per-channel scales for dynamic range quantization use case (float32 inputs / outputs and int8 weights). The change enables new quantization schema globally in the converter and inference engine. The new behaviour can be disabled via experimental flag converter._experimental_disable_per_channel_quantization_for_dense_layers = True.
- C API:
  - The experimental TfLiteRegistrationExternal type has been renamed as TfLiteOperator, and likewise for the corresponding API functions.
- The Python TF Lite Interpreter bindings now have an option experimental_default_delegate_latest_features to enable all default delegate features.
- Flatbuffer version update:
  - GetTemporaryPointer() bug fixed.
tf.data
- Add wait to tf.data.Dataset.load. If True, for snapshots written with distributed_save, it reads the snapshot while it is being written. For snapshots written with regular save, it waits for the snapshot until it's finished. The default is False for backward compatibility. Users of distributed_save are recommended to set it to True.
tf.tpu.experimental.embedding.TPUEmbeddingV2
- Add compute_sparse_core_stats for sparse core users to profile the data with this API to get the max_ids and max_unique_ids. These numbers will be needed to configure the sparse core embedding mid level api.
- Remove the preprocess_features method since that's no longer needed.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Abdulaziz Aloqeely, Ahmad-M-Al-Khateeb, Akhil Goel, akhilgoe, Alexander Pivovarov, Amir Samani, Andrew Goodbody, Andrey Portnoy, Ashiq Imran, Ben Olson, Chao, Chase Riley Roberts, Clemens Giuliani, dependabot[bot], Dimitris Vardoulakis, Dragan Mladjenovic, ekuznetsov139, Elfie Guo, Faijul Amin, Gauri1 Deshpande, Georg Stefan Schmid, guozhong.zhuang, Hao Wu, Haoyu (Daniel), Harsha H S, Harsha Hs, Harshit Monish, Ilia Sergachev, Jane Liu, Jaroslav Sevcik, Jinzhe Zeng, Justin Dhillon, Kaixi Hou, Kanvi Khanna, LakshmiKalaKadali, Learning-To-Play, lingzhi98, Lu Teng, Matt Bahr, Max Ren, Meekail Zain, Mmakevic-Amd, mraunak, neverlva, nhatle, Nicola Ferralis, Olli Lupton, Om Thakkar, orangekame3, ourfor, pateldeev, Pearu Peterson, pemeliya, Peng Sun, Philipp Hack, Pratik Joshi, prrathi, rahulbatra85, Raunak, redwrasse, Robert Kalmar, Robin Zhang, RoboSchmied, Ruturaj Vaidya, sachinmuradi, Shawn Wang, Sheng Yang, Surya, Thibaut Goetghebuer-Planchon, Thomas Preud'Homme, tilakrayal, Tj Xu, Trevor Morris, wenchenvincent, Yimei Sun, zahiqbal, Zhu Jianjiang, Zoranjovanovic-Ns

Open source →

2.17.0rc12 Jul 2024pre-release

Support for NVIDIA GPUs with compute capability 5.x (Maxwell generation) has been removed from TF binary distributions (Python wheels).

Release 2.17.0

TensorFlow

Breaking Changes

GPU
- Support for NVIDIA GPUs with compute capability 5.x (Maxwell generation) has been removed from TF binary distributions (Python wheels).

Major Features and Improvements

Add is_cpu_target_available, which indicates whether or not TensorFlow was built with support for a given CPU target. This can be useful for skipping target-specific tests if a target is not supported.
tf.data
- Support data.experimental.distribued_save. distribued_save uses tf.data service (https://www.tensorflow.org/api_docs/python/tf/data/experimental/service) to write distributed dataset snapshots. The call is non-blocking and returns without waiting for the snapshot to finish. Setting wait=True to tf.data.Dataset.load allows the snapshots to be read while they are being written.

Bug Fixes and Other Changes

GPU
- Support for NVIDIA GPUs with compute capability 8.9 (e.g. L4 & L40) has been added to TF binary distributions (Python wheels).
Replace DebuggerOptions of TensorFlow Quantizer, and migrate to DebuggerConfig of StableHLO Quantizer.
Add TensorFlow to StableHLO converter to TensorFlow pip package.
TensorRT support: this is the last release supporting TensorRT. It will be removed in the next release.
NumPy 2.0 support: TensorFlow is going to support NumPy 2.0 in the next release. It may break some edge cases of TensorFlow API usage.
tf.lite
- Quantization for FullyConnected layer is switched from per-tensor to per-channel scales for dynamic range quantization use case (float32 inputs / outputs and int8 weights). The change enables new quantization schema globally in the converter and inference engine. The new behaviour can be disabled via experimental flag converter._experimental_disable_per_channel_quantization_for_dense_layers = True.
- C API:
  - The experimental TfLiteRegistrationExternal type has been renamed as TfLiteOperator, and likewise for the corresponding API functions.
- The Python TF Lite Interpreter bindings now have an option experimental_default_delegate_latest_features to enable all default delegate features.
- Flatbuffer version update:
  - GetTemporaryPointer() bug fixed.
tf.data
- Add wait to tf.data.Dataset.load. If True, for snapshots written with distributed_save, it reads the snapshot while it is being written. For snapshots written with regular save, it waits for the snapshot until it's finished. The default is False for backward compatibility. Users of distributed_save are recommended to set it to True.
tf.tpu.experimental.embedding.TPUEmbeddingV2
- Add compute_sparse_core_stats for sparse core users to profile the data with this API to get the max_ids and max_unique_ids. These numbers will be needed to configure the sparse core embedding mid level api.
- Remove the preprocess_features method since that's no longer needed.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.17.0rc018 Jun 2024pre-release

Support for NVIDIA GPUs with compute capability 5.x (Maxwell generation) has been removed from TF binary distributions (Python wheels).

Release 2.17.0

TensorFlow

Breaking Changes

GPU
- Support for NVIDIA GPUs with compute capability 5.x (Maxwell generation) has been removed from TF binary distributions (Python wheels).

Major Features and Improvements

Add is_cpu_target_available, which indicates whether or not TensorFlow was built with support for a given CPU target. This can be useful for skipping target-specific tests if a target is not supported.
tf.data
- Support data.experimental.distribued_save. distribued_save uses tf.data service (https://www.tensorflow.org/api_docs/python/tf/data/experimental/service) to write distributed dataset snapshots. The call is non-blocking and returns without waiting for the snapshot to finish. Setting wait=True to tf.data.Dataset.load allows the snapshots to be read while they are being written.

Bug Fixes and Other Changes

GPU
- Support for NVIDIA GPUs with compute capability 8.9 (e.g. L4 & L40) has been added to TF binary distributions (Python wheels).
Replace DebuggerOptions of TensorFlow Quantizer, and migrate to DebuggerConfig of StableHLO Quantizer.
Add TensorFlow to StableHLO converter to TensorFlow pip package.
TensorRT support: this is the last release supporting TensorRT. It will be removed in the next release.
NumPy 2.0 support: TensorFlow is going to support NumPy 2.0 in the next release. It may break some edge cases of TensorFlow API usage.
tf.lite
- Quantization for FullyConnected layer is switched from per-tensor to per-channel scales for dynamic range quantization use case (float32 inputs / outputs and int8 weights). The change enables new quantization schema globally in the converter and inference engine. The new behaviour can be disabled via experimental flag converter._experimental_disable_per_channel_quantization_for_dense_layers = True.
- C API:
  - The experimental TfLiteRegistrationExternal type has been renamed as TfLiteOperator, and likewise for the corresponding API functions.
- The Python TF Lite Interpreter bindings now have an option experimental_default_delegate_latest_features to enable all default delegate features.
- Flatbuffer version update:
  - GetTemporaryPointer() bug fixed.
tf.data
- Add wait to tf.data.Dataset.load. If True, for snapshots written with distributed_save, it reads the snapshot while it is being written. For snapshots written with regular save, it waits for the snapshot until it's finished. The default is False for backward compatibility. Users of distributed_save are recommended to set it to True.
tf.tpu.experimental.embedding.TPUEmbeddingV2
- Add compute_sparse_core_stats for sparse core users to profile the data with this API to get the max_ids and max_unique_ids. These numbers will be needed to configure the sparse core embedding mid level api.
- Remove the preprocess_features method since that's no longer needed.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.16.19 Mar 2024Deprecated

Mac x86 users: Mac x86 builds are being deprecated and will no longer be released as a Pip package from TF 2.17 onwards.

Release 2.16.1

TensorFlow

TensorFlow Windows Build:
- Clang is now the default compiler to build TensorFlow CPU wheels on the Windows Platform starting with this release. The currently supported version is LLVM/clang 17. The official Wheels-published on PyPI will be based on Clang; however, users retain the option to build wheels using the MSVC compiler following the steps mentioned in https://www.tensorflow.org/install/source_windows as has been the case before
TensorFlow 2.16 will be released as TF 2.16.1 (instead of 2.16.0). The patch release will be done as 2.16.2 during the next release cycle.

Breaking Changes

tf.summary.trace_on now takes a profiler_outdir argument. This must be set if profiler arg is set to True.
- tf.summary.trace_export's profiler_outdir arg is now a no-op. Enabling the profiler now requires setting profiler_outdir in trace_on.
tf.estimator
- The tf.estimator API is removed.
- To continue using tf.estimator, you will need to use TF 2.15 or an earlier version.
Keras 3.0 will be the default Keras version. You may need to update your script to use Keras 3.0.
Please refer to the new Keras documentation for Keras 3.0 (https://keras.io/keras_3).
To continue using Keras 2.0, do the following.
1. Install tf-keras via pip install tf-keras~=2.16
2. To switch tf.keras to use Keras 2 (tf-keras), set the environment variable TF_USE_LEGACY_KERAS=1 directly or in your python program with import os;os.environ["TF_USE_LEGACY_KERAS"]="1". Please note that this will set it for all packages in your Python runtime program
3. Change the keras import: replace import tensorflow.keras as keras or import keras with import tf_keras as keras. Update any tf.keras references to keras.
Apple Silicon users: If you previously installed TensorFlow using pip install tensorflow-macos, please update your installation method. Use pip install tensorflow from now on.
Mac x86 users: Mac x86 builds are being deprecated and will no longer be released as a Pip package from TF 2.17 onwards.

Known Caveats

Full aarch64 Linux and Arm64 macOS wheels are now published to the tensorflow pypi repository and no longer redirect to a separate package.

Major Features and Improvements

Support for Python 3.12 has been added.
tensorflow-tpu package is now available for easier TPU based installs.
TensorFlow pip packages are now built with CUDA 12.3 and cuDNN 8.9.7
Added experimental support for float16 auto-mixed precision using the new AMX-FP16 instruction set on X86 CPUs.

Bug Fixes and Other Changes

tf.lite
- Added support for stablehlo.gather.
- Added support for stablehlo.add.
- Added support for stablehlo.multiply.
- Added support for stablehlo.maximum.
- Added support for stablehlo.minimum.
- Added boolean parameter support for tfl.gather_nd.
- C API:
  - New API functions:
    - tensorflow/lite/c/c_api_experimental.h:
      - TfLiteInterpreterGetVariableTensorCount
      - TfLiteInterpreterGetVariableTensor
      - TfLiteInterpreterGetBufferHandle
      - TfLiteInterpreterSetBufferHandle
    - tensorflow/lite/c/c_api_opaque.h:
      - TfLiteOpaqueTensorSetAllocationTypeToDynamic
  - API functions promoted from experimental to stable:
    - tensorflow/lite/c/c_api.h:
      - TfLiteInterpreterOptionsEnableCancellation
      - TfLiteInterpreterCancel
- C++ API:
  - New virtual methods in the tflite::SimpleDelegateInterface class in tensorflow/lite/delegates/utils/simple_delegate.h, and likewise in the tflite::SimpleOpaqueDelegateInterface class in tensorflow/lite/delegates/utils/simple_opaque_delegate.h:
    - CopyFromBufferHandle
    - CopyToBufferHandle
    - FreeBufferHandle
tf.train.CheckpointOptions and tf.saved_model.SaveOptions
- These now take in a new argument called experimental_sharding_callback. This is a callback function wrapper that will be executed to determine how tensors will be split into shards when the saver writes the checkpoint shards to disk. tf.train.experimental.ShardByTaskPolicy is the default sharding behavior, but tf.train.experimental.MaxShardSizePolicy can be used to shard the checkpoint with a maximum shard file size. Users with advanced use cases can also write their own custom tf.train.experimental.ShardingCallbacks.
tf.train.CheckpointOptions
- Added experimental_skip_slot_variables (a boolean option) to skip restoring of optimizer slot variables in a checkpoint.
tf.saved_model.SaveOptions
- SaveOptions now takes a new argument called experimental_debug_stripper. When enabled, this strips the debug nodes from both the node defs and the function defs of the graph. Note that this currently only strips the Assert nodes from the graph and converts them into NoOps instead.

Keras

keras.layers.experimental.DynamicEmbedding
- Added DynamicEmbedding Keras layer
- Added 'UpdateEmbeddingCallback`
- DynamicEmbedding layer allows for the continuous updating of the vocabulary and embeddings during the training process. This layer maintains a hash table to track the most up-to-date vocabulary based on the inputs received by the layer and the eviction policy. When this layer is used with an UpdateEmbeddingCallback, which is a time-based callback, the vocabulary lookup tensor is updated at the time interval set in the UpdateEmbeddingCallback based on the most up-to-date vocabulary hash table maintained by the layer. If this layer is not used in conjunction with UpdateEmbeddingCallback the behavior of the layer would be same as keras.layers.Embedding.
keras.optimizers.Adam
- Added the option to set adaptive epsilon to match implementations with Jax and PyTorch equivalents.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Aakar Dwivedi, Akhil Goel, Alexander Grund, Alexander Pivovarov, Andrew Goodbody, Andrey Portnoy, Aneta Kaczyńska, AnetaKaczynska, ArkadebMisra, Ashiq Imran, Ayan Moitra, Ben Barsdell, Ben Creech, Benedikt Lorch, Bhavani Subramanian, Bianca Van Schaik, Chao, Chase Riley Roberts, Connor Flanagan, David Hall, David Svantesson, David Svantesson-Yeung, dependabot[bot], Dr. Christoph Mittendorf, Dragan Mladjenovic, ekuznetsov139, Eli Kobrin, Eugene Kuznetsov, Faijul Amin, Frédéric Bastien, fsx950223, gaoyiyeah, Gauri1 Deshpande, Gautam, Giulio C.N, guozhong.zhuang, Harshit Monish, James Hilliard, Jane Liu, Jaroslav Sevcik, jeffhataws, Jerome Massot, Jerry Ge, jglaser, jmaksymc, Kaixi Hou, kamaljeeti, Kamil Magierski, Koan-Sin Tan, lingzhi98, looi, Mahmoud Abuzaina, Malik Shahzad Muzaffar, Meekail Zain, mraunak, Neil Girdhar, Olli Lupton, Om Thakkar, Paul Strawder, Pavel Emeliyanenko, Pearu Peterson, pemeliya, Philipp Hack, Pierluigi Urru, Pratik Joshi, radekzc, Rafik Saliev, Ragu, Rahul Batra, rahulbatra85, Raunak, redwrasse, Rodrigo Gomes, ronaghy, Sachin Muradi, Shanbin Ke, shawnwang18, Sheng Yang, Shivam Mishra, Shu Wang, Strawder, Paul, Surya, sushreebarsa, Tai Ly, talyz, Thibaut Goetghebuer-Planchon, Tj Xu, Tom Allsop, Trevor Morris, Varghese, Jojimon, weihanmines, wenchenvincent, Wenjie Zheng, Who Who Who, Yasir Ashfaq, yasiribmcon, Yoshio Soma, Yuanqiang Liu, Yuriy Chernyshov

Open source →

2.16.0rc023 Feb 2024pre-release

Clang is now the default compiler to build TensorFlow CPU wheels on the Windows Platform starting with this release. The currently supported version i

Release 2.16.0

TensorFlow

TensorFlow Windows Build:
- Clang is now the default compiler to build TensorFlow CPU wheels on the Windows Platform starting with this release. The currently supported version is LLVM/clang 17. The official Wheels-published on PyPI will be based on Clang; however, users retain the option to build wheels using the MSVC compiler following the steps mentioned in https://www.tensorflow.org/install/source_windows as has been the case before

Breaking Changes

tf.summary.trace_on now takes a profiler_outdir argument. This must be set if profiler arg is set to True.
- tf.summary.trace_export's profiler_outdir arg is now a no-op. Enabling the profiler now requires setting profiler_outdir in trace_on.
tf.estimator
- The tf.estimator API is removed.
- To continue using tf.estimator, you will need to use TF 2.15 or an earlier version.
Keras 3 will be the default Keras version. You may need to update your script to use Keras 3. Please refer to the new Keras documentation for Keras 3 (https://keras.io/keras_3). To continue using Keras 2, do the following:

Install tf-keras via pip install tf-keras~=2.16
To switch tf.keras to use Keras 2 (tf-keras), set the environment variable TF_USE_LEGACY_KERAS=1 directly or in your Python program by doing import os;os.environ["TF_USE_LEGACY_KERAS"]=1. Please note that this will set it for all packages in your Python runtime program.

Apple Silicon users: If you previously installed TensorFlow using pip install tensorflow-macos, please update your installation method. Use pip install tensorflow from now on. Starting with TF 2.17, the tensorflow-macos package will no longer receive updates.

Known Caveats

Full aarch64 Linux and Arm64 macOS wheels are now published to the tensorflow pypi repository and no longer redirect to a separate package.

Major Features and Improvements

Support for Python 3.12 has been added.
tensorflow-tpu package is now available for easier TPU based installs.
TensorFlow pip packages are now built with CUDA 12.3 and cuDNN 8.9.7

Bug Fixes and Other Changes

tf.lite
- Added support for stablehlo.gather.
- Added support for stablehlo.add.
- Added support for stablehlo.multiply.
- Added support for stablehlo.maximum.
- Added support for stablehlo.minimum.
- Added boolean parameter support for tfl.gather_nd.
tf.train.CheckpointOptions and tf.saved_model.SaveOptions
- These now take in a new argument called experimental_sharding_callback. This is a callback function wrapper that will be executed to determine how tensors will be split into shards when the saver writes the checkpoint shards to disk. tf.train.experimental.ShardByTaskPolicy is the default sharding behavior, but tf.train.experimental.MaxShardSizePolicy can be used to shard the checkpoint with a maximum shard file size. Users with advanced use cases can also write their own custom tf.train.experimental.ShardingCallbacks.
tf.train.CheckpointOptions
- Added experimental_skip_slot_variables (a boolean option) to skip restoring of optimizer slot variables in a checkpoint.
tf.saved_model.SaveOptions
- SaveOptions now takes a new argument called experimental_debug_stripper. When enabled, this strips the debug nodes from both the node defs and the function defs of the graph. Note that this currently only strips the Assert nodes from the graph and converts them into NoOps instead.

Keras

keras.layers.experimental.DynamicEmbedding
- Added DynamicEmbedding Keras layer
- Added 'UpdateEmbeddingCallback`
- DynamicEmbedding layer allows for the continuous updating of the vocabulary and embeddings during the training process. This layer maintains a hash table to track the most up-to-date vocabulary based on the inputs received by the layer and the eviction policy. When this layer is used with an UpdateEmbeddingCallback, which is a time-based callback, the vocabulary lookup tensor is updated at the time interval set in the UpdateEmbeddingCallback based on the most up-to-date vocabulary hash table maintained by the layer. If this layer is not used in conjunction with UpdateEmbeddingCallback the behavior of the layer would be same as keras.layers.Embedding.
keras.optimizers.Adam
- Added the option to set adaptive epsilon to match implementations with Jax and PyTorch equivalents.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.15.014 Nov 2023

tf.types.experimental.GenericFunction has been renamed to tf.types.experimental.PolymorphicFunction.

Release 2.15.0

TensorFlow

Breaking Changes

tf.types.experimental.GenericFunction has been renamed to tf.types.experimental.PolymorphicFunction.

Major Features and Improvements

oneDNN CPU performance optimizations Windows x64 & x86.
- Windows x64 & x86 packages:
  - oneDNN optimizations are enabled by default on X86 CPUs
- To explicitly enable or disable oneDNN optimizations, set the environment variable TF_ENABLE_ONEDNN_OPTS to 1 (enable) or 0 (disable) before running TensorFlow. To fall back to default settings, unset the environment variable.
- oneDNN optimizations can yield slightly different numerical results compared to when oneDNN optimizations are disabled due to floating-point round-off errors from different computation approaches and orders.
- To verify if oneDNN optimizations are on, look for a message with "oneDNN custom operations are on" in the log. If the exact phrase is not there, it means they are off.
Making the tf.function type system fully available:
- tf.types.experimental.TraceType now allows custom tf.function inputs to declare Tensor decomposition and type casting support.
- Introducing tf.types.experimental.FunctionType as the comprehensive representation of the signature of tf.function callables. It can be accessed through the function_type property of tf.functions and ConcreteFunctions. See the tf.types.experimental.FunctionType documentation for more details.
Introducing tf.types.experimental.AtomicFunction as the fastest way to perform TF computations in Python.
- Can be accessed through inference_fn property of ConcreteFunctions
- Does not support gradients.
- See tf.types.experimental.AtomicFunction documentation for how to call and use it.
tf.data:
- Moved option warm_start from tf.data.experimental.OptimizationOptions to tf.data.Options.
tf.lite:
- sub_op and mul_op support broadcasting up to 6 dimensions.
- The tflite::SignatureRunner class, which provides support for named parameters and for multiple named computations within a single TF Lite model, is no longer considered experimental. Likewise for the following signature-related methods of tflite::Interpreter:
  - tflite::Interpreter::GetSignatureRunner
  - tflite::Interpreter::signature_keys
  - tflite::Interpreter::signature_inputs
  - tflite::Interpreter::signature_outputs
  - tflite::Interpreter::input_tensor_by_signature
  - tflite::Interpreter::output_tensor_by_signature
- Similarly, the following signature runner functions in the TF Lite C API are no longer considered experimental:
  - TfLiteInterpreterGetSignatureCount
  - TfLiteInterpreterGetSignatureKey
  - TfLiteInterpreterGetSignatureRunner
  - TfLiteSignatureRunnerAllocateTensors
  - TfLiteSignatureRunnerGetInputCount
  - TfLiteSignatureRunnerGetInputName
  - TfLiteSignatureRunnerGetInputTensor
  - TfLiteSignatureRunnerGetOutputCount
  - TfLiteSignatureRunnerGetOutputName
  - TfLiteSignatureRunnerGetOutputTensor
  - TfLiteSignatureRunnerInvoke
  - TfLiteSignatureRunnerResizeInputTensor
- New C API function TfLiteExtensionApisVersion added to tensorflow/lite/c/c_api.h.
- Add int8 and int16x8 support for RSQRT operator
Android NDK r25 is supported.

Bug Fixes and Other Changes

Add TensorFlow Quantizer to TensorFlow pip package.
tf.sparse.segment_sum tf.sparse.segment_mean tf.sparse.segment_sqrt_n SparseSegmentSum/Mean/SqrtN[WithNumSegments]
- Added sparse_gradient option (default=false) that makes the gradient of these functions/ops sparse (IndexedSlices) instead of dense (Tensor), using new SparseSegmentSum/Mean/SqrtNGradV2 ops.
tf.nn.embedding_lookup_sparse
- Optimized this function for some cases by fusing internal operations.
tf.saved_model.SaveOptions
- Provided a new experimental_skip_saver argument which, if specified, will suppress the addition of SavedModel-native save and restore ops to the SavedModel, for cases where users already build custom save/restore ops and checkpoint formats for the model being saved, and the creation of the SavedModel-native save/restore ops simply cause longer model serialization times.
Add ops to tensorflow.raw_ops that were missing.
tf.CheckpointOptions
- It now takes in a new argument called experimental_write_callbacks. These are callbacks that will be executed after a saving event finishes writing the checkpoint file.
Add an option disable_eager_executer_streaming_enqueue to tensorflow.ConfigProto.Experimental to control the eager runtime's behavior around parallel remote function invocations; when set to True, the eager runtime will be allowed to execute multiple function invocations in parallel.
tf.constant_initializer
- It now takes a new argument called support_partition. If True, constant_initializers can create sharded variables. This is disabled by default, similar to existing behavior.
tf.lite
- Added support for stablehlo.scatter.
tf.estimator
- The tf.estimator API removal is in progress and will be targeted for the 2.16 release.

Keras

This will be the final release before the launch of Keras 3.0, when Keras will become multi-backend. For the compatibility page and other info, please see: https://github.com/keras-team/keras-core

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Aiden Grossman, Akash Patel, Akhil Goel, Alexander Pivovarov, Andrew Goodbody, Ayan Moitra, Ben Barsdell, Ben Olson, Bhavani Subramanian, Boian Petkantchin, Bruce Lai, Chao Chen, Christian Steinmeyer, cjflan, David Korczynski, Donghak Park, Dragan Mladjenovic, Eli Kobrin, Fadi Arafeh, Feiyue Chen, Frédéric Bastien, guozhong.zhuang, halseycamilla, Harshavardhan Bellamkonda, James Ward, jameshollyer, Jane Liu, johnnkp, jswag180, justkw, Kanvi Khanna, Keith Smiley, Koan-Sin Tan, Kulin Seth, Kun-Lu, kushanam, Lu Teng, mdfaijul, Mehdi Drissi, mgokulkrish, mraunak, Mustafa Uzun, Namrata Bhave, Pavel Emeliyanenko, pemeliya, Peng Sun, Philipp Hack, Pratik Joshi, Rahul Batra, Raunak, redwrasse, Saoirse Stewart, SaoirseARM, seanshpark, Shanbin Ke, Spenser Bauman, Surya, sushreebarsa, Tai Ly, Thibaut Goetghebuer-Planchon, tilakrayal, Tirumalesh, Tj Xu, Vladislav, weihanmines, Wen Chen, wenchenvincent, wenscarl, William Muir, Zhoulong, Jiang

Open source →

2.15.0rc13 Nov 2023pre-release

tf.types.experimental.GenericFunction has been renamed to tf.types.experimental.PolymorphicFunction.

Release 2.15.0

TensorFlow

Breaking Changes

tf.types.experimental.GenericFunction has been renamed to tf.types.experimental.PolymorphicFunction.

Known Caveats

Major Features and Improvements

oneDNN CPU performance optimizations Windows x64 & x86.
- Windows x64 & x86 packages:
  - oneDNN optimizations are enabled by default on X86 CPUs
- To explicitly enable or disable oneDNN optimizations, set the environment variable TF_ENABLE_ONEDNN_OPTS to 1 (enable) or 0 (disable) before running TensorFlow. To fall back to default settings, unset the environment variable.
- oneDNN optimizations can yield slightly different numerical results compared to when oneDNN optimizations are disabled due to floating-point round-off errors from different computation approaches and orders.
- To verify if oneDNN optimizations are on, look for a message with "oneDNN custom operations are on" in the log. If the exact phrase is not there, it means they are off.
Making the tf.function type system fully available:
- tf.types.experimental.TraceType now allows custom tf.function inputs to declare Tensor decomposition and type casting support.
- Introducing tf.types.experimental.FunctionType as the comprehensive representation of the signature of tf.function callables. It can be accessed through the function_type property of tf.functions and ConcreteFunctions. See the tf.types.experimental.FunctionType documentation for more details.
Introducing tf.types.experimental.AtomicFunction as the fastest way to perform TF computations in Python.
- Can be accessed through inference_fn property of ConcreteFunctions
- Does not support gradients.
- See tf.types.experimental.AtomicFunction documentation for how to call and use it.
tf.data:
- Moved option warm_start from tf.data.experimental.OptimizationOptions to tf.data.Options.
tf.lite:
- sub_op and mul_op support broadcasting up to 6 dimensions.
- The tflite::SignatureRunner class, which provides support for named parameters and for multiple named computations within a single TF Lite model, is no longer considered experimental. Likewise for the following signature-related methods of tflite::Interpreter:
  - tflite::Interpreter::GetSignatureRunner
  - tflite::Interpreter::signature_keys
  - tflite::Interpreter::signature_inputs
  - tflite::Interpreter::signature_outputs
  - tflite::Interpreter::input_tensor_by_signature
  - tflite::Interpreter::output_tensor_by_signature
- Similarly, the following signature runner functions in the TF Lite C API are no longer considered experimental:
  - TfLiteInterpreterGetSignatureCount
  - TfLiteInterpreterGetSignatureKey
  - TfLiteInterpreterGetSignatureRunner
  - TfLiteSignatureRunnerAllocateTensors
  - TfLiteSignatureRunnerGetInputCount
  - TfLiteSignatureRunnerGetInputName
  - TfLiteSignatureRunnerGetInputTensor
  - TfLiteSignatureRunnerGetOutputCount
  - TfLiteSignatureRunnerGetOutputName
  - TfLiteSignatureRunnerGetOutputTensor
  - TfLiteSignatureRunnerInvoke
  - TfLiteSignatureRunnerResizeInputTensor
- New C API function TfLiteExtensionApisVersion added to tensorflow/lite/c/c_api.h.
- Add int8 and int16x8 support for RSQRT operator
Android NDK r25 is supported.

Bug Fixes and Other Changes

Add TensorFlow Quantizer to TensorFlow pip package.
tf.sparse.segment_sum tf.sparse.segment_mean tf.sparse.segment_sqrt_n SparseSegmentSum/Mean/SqrtN[WithNumSegments]
- Added sparse_gradient option (default=false) that makes the gradient of these functions/ops sparse (IndexedSlices) instead of dense (Tensor), using new SparseSegmentSum/Mean/SqrtNGradV2 ops.
tf.nn.embedding_lookup_sparse
- Optimized this function for some cases by fusing internal operations.
tf.saved_model.SaveOptions
- Provided a new experimental_skip_saver argument which, if specified, will suppress the addition of SavedModel-native save and restore ops to the SavedModel, for cases where users already build custom save/restore ops and checkpoint formats for the model being saved, and the creation of the SavedModel-native save/restore ops simply cause longer model serialization times.
Add ops to tensorflow.raw_ops that were missing.
tf.CheckpointOptions
- It now takes in a new argument called experimental_write_callbacks. These are callbacks that will be executed after a saving event finishes writing the checkpoint file.
Add an option disable_eager_executer_streaming_enqueue to tensorflow.ConfigProto.Experimental to control the eager runtime's behavior around parallel remote function invocations; when set to True, the eager runtime will be allowed to execute multiple function invocations in parallel.
tf.constant_initializer
- It now takes a new argument called support_partition. If True, constant_initializers can create sharded variables. This is disabled by default, similar to existing behavior.
tf.lite
- Added support for stablehlo.scatter.
tf.estimator
- The tf.estimator API removal is in progress and will be targeted for the 2.16 release.

Keras

This will be the final release before the launch of Keras 3.0, when Keras will become multi-backend. For the compatibility page and other info, please see: https://github.com/keras-team/keras-core

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.15.0rc025 Oct 2023pre-release

tf.types.experimental.GenericFunction has been renamed to tf.types.experimental.PolymorphicFunction.

Release 2.15.0

TensorFlow

Breaking Changes

tf.types.experimental.GenericFunction has been renamed to tf.types.experimental.PolymorphicFunction.

Major Features and Improvements

oneDNN CPU performance optimizations Windows x64 & x86.
- Windows x64 & x86 packages:
  - oneDNN optimizations are enabled by default on X86 CPUs
- To explicitly enable or disable oneDNN optimizations, set the environment variable TF_ENABLE_ONEDNN_OPTS to 1 (enable) or 0 (disable) before running TensorFlow. To fall back to default settings, unset the environment variable.
- oneDNN optimizations can yield slightly different numerical results compared to when oneDNN optimizations are disabled due to floating-point round-off errors from different computation approaches and orders.
- To verify if oneDNN optimizations are on, look for a message with "oneDNN custom operations are on" in the log. If the exact phrase is not there, it means they are off.
Making the tf.function type system fully available:
- tf.types.experimental.TraceType now allows custom tf.function inputs to declare Tensor decomposition and type casting support.
- Introducing tf.types.experimental.FunctionType as the comprehensive representation of the signature of tf.function callables. It can be accessed through the function_type property of tf.functions and ConcreteFunctions. See the tf.types.experimental.FunctionType documentation for more details.
Introducing tf.types.experimental.AtomicFunction as the fastest way to perform TF computations in Python.
- Can be accessed through inference_fn property of ConcreteFunctions
- Does not support gradients.
- See tf.types.experimental.AtomicFunction documentation for how to call and use it.
tf.data:
- Moved option warm_start from tf.data.experimental.OptimizationOptions to tf.data.Options.
tf.lite:
- sub_op and mul_op support broadcasting up to 6 dimensions.
- The tflite::SignatureRunner class, which provides support for named parameters and for multiple named computations within a single TF Lite model, is no longer considered experimental. Likewise for the following signature-related methods of tflite::Interpreter:
  - tflite::Interpreter::GetSignatureRunner
  - tflite::Interpreter::signature_keys
  - tflite::Interpreter::signature_inputs
  - tflite::Interpreter::signature_outputs
  - tflite::Interpreter::input_tensor_by_signature
  - tflite::Interpreter::output_tensor_by_signature
- Similarly, the following signature runner functions in the TF Lite C API are no longer considered experimental:
  - TfLiteInterpreterGetSignatureCount
  - TfLiteInterpreterGetSignatureKey
  - TfLiteInterpreterGetSignatureRunner
  - TfLiteSignatureRunnerAllocateTensors
  - TfLiteSignatureRunnerGetInputCount
  - TfLiteSignatureRunnerGetInputName
  - TfLiteSignatureRunnerGetInputTensor
  - TfLiteSignatureRunnerGetOutputCount
  - TfLiteSignatureRunnerGetOutputName
  - TfLiteSignatureRunnerGetOutputTensor
  - TfLiteSignatureRunnerInvoke
  - TfLiteSignatureRunnerResizeInputTensor
- New C API function TfLiteExtensionApisVersion added to tensorflow/lite/c/c_api.h.
- Add int8 and int16x8 support for RSQRT operator
Android NDK r25 is supported.

Bug Fixes and Other Changes

Add TensorFlow Quantizer to TensorFlow pip package.
tf.sparse.segment_sum tf.sparse.segment_mean tf.sparse.segment_sqrt_n SparseSegmentSum/Mean/SqrtN[WithNumSegments]
- Added sparse_gradient option (default=false) that makes the gradient of these functions/ops sparse (IndexedSlices) instead of dense (Tensor), using new SparseSegmentSum/Mean/SqrtNGradV2 ops.
tf.nn.embedding_lookup_sparse
- Optimized this function for some cases by fusing internal operations.
tf.saved_model.SaveOptions
- Provided a new experimental_skip_saver argument which, if specified, will suppress the addition of SavedModel-native save and restore ops to the SavedModel, for cases where users already build custom save/restore ops and checkpoint formats for the model being saved, and the creation of the SavedModel-native save/restore ops simply cause longer model serialization times.

Keras

Bug Fixes and Other Changes

Add ops to tensorflow.raw_ops that were missing.
tf.CheckpointOptions
- It now takes in a new argument called experimental_write_callbacks. These are callbacks that will be executed after a saving event finishes writing the checkpoint file.
Add an option disable_eager_executer_streaming_enqueue to tensorflow.ConfigProto.Experimental to control the eager runtime's behavior around parallel remote function invocations; when set to True, the eager runtime will be allowed to execute multiple function invocations in parallel.
tf.constant_initializer
- It now takes a new argument called support_partition. If True, constant_initializers can create sharded variables. This is disabled by default, similar to existing behavior.
tf.lite
- Added support for stablehlo.scatter.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.14.026 Sept 2023Deprecated

tf.compat.v1.Session.partial_run and tf.compat.v1.Session.partial_run_setup will be deprecated in the next release.

Release 2.14.0

Tensorflow

Breaking Changes

Support for Python 3.8 has been removed starting with TF 2.14. The TensorFlow 2.13.1 patch release will still have Python 3.8 support.
tf.Tensor
- The class hierarchy for tf.Tensor has changed, and there are now explicit EagerTensor and SymbolicTensor classes for eager and tf.function respectively. Users who relied on the exact type of Tensor (e.g. type(t) == tf.Tensor) will need to update their code to use isinstance(t, tf.Tensor). The tf.is_symbolic_tensor helper added in 2.13 may be used when it is necessary to determine if a value is specifically a symbolic tensor.
tf.compat.v1.Session
- tf.compat.v1.Session.partial_run and tf.compat.v1.Session.partial_run_setup will be deprecated in the next release.

Known Caveats

tf.lite
- when converter flag "_experimenal_use_buffer_offset" is enabled, additional metadata is automatically excluded from the generated model. The behaviour is the same as "exclude_conversion_metadata" is set
- If the model is larger than 2GB, then we also require "exclude_conversion_metadata" flag to be set

Major Features and Improvements

The tensorflow pip package has a new, optional installation method for Linux that installs necessary Nvidia CUDA libraries through pip. As long as the Nvidia driver is already installed on the system, you may now run pip install tensorflow[and-cuda] to install TensorFlow's Nvidia CUDA library dependencies in the Python environment. Aside from the Nvidia driver, no other pre-existing Nvidia CUDA packages are necessary.
Enable JIT-compiled i64-indexed kernels on GPU for large tensors with more than 2**32 elements.
- Unary GPU kernels: Abs, Atanh, Acos, Acosh, Asin, Asinh, Atan, Cos, Cosh, Sin, Sinh, Tan, Tanh.
- Binary GPU kernels: AddV2, Sub, Div, DivNoNan, Mul, MulNoNan, FloorDiv, Equal, NotEqual, Greater, GreaterEqual, LessEqual, Less.
tf.lite
- Add experimental supports conversion of models that may be larger than 2GB before buffer deduplication

Bug Fixes and Other Changes

tf.py_function and tf.numpy_function can now be used as function decorators for clearer code:

@tf.py_function(Tout=tf.float32)
def my_fun(x):
  print("This always executes eagerly.")
  return x+1

tf.lite
- Strided_Slice now supports UINT32.
tf.config.experimental.enable_tensor_float_32_execution
- Disabling TensorFloat-32 execution now causes TPUs to use float32 precision for float32 matmuls and other ops. TPUs have always used bfloat16 precision for certain ops, like matmul, when such ops had float32 inputs. Now, disabling TensorFloat-32 by calling tf.config.experimental.enable_tensor_float_32_execution(False) will cause TPUs to use float32 precision for such ops instead of bfloat16.
tf.experimental.dtensor
- API changes for Relayout. Added a new API, dtensor.relayout_like, for relayouting a tensor according to the layout of another tensor.
- Added dtensor.get_default_mesh, for retrieving the current default mesh under the dtensor context.
- *fft* ops now support dtensors with any layout. Fixed bug in 'fft2d/fft3d', 'ifft2d/ifft3d', 'rfft2d/rfft3d', and 'irfft2d/irfft3d' for sharde input. Refer to this blog post for details.
tf.experimental.strict_mode
- Added a new API, strict_mode, which converts all deprecation warnings into runtime errors with instructions on switching to a recommended substitute.
TensorFlow Debugger (tfdbg) CLI: ncurses-based CLI for tfdbg v1 was removed.
TensorFlow now supports C++ RTTI on mobile and Android. To enable this feature, pass the flag --define=tf_force_rtti=true to Bazel when building TensorFlow. This may be needed when linking TensorFlow into RTTI-enabled programs since mixing RTTI and non-RTTI code can cause ABI issues.
tf.ones, tf.zeros, tf.fill, tf.ones_like, tf.zeros_like now take an additional Layout argument that controls the output layout of their results.
tf.nest and tf.data now support user defined classes implementing __tf_flatten__ and __tf_unflatten__ methods. See nest_util code examples for an example.
TensorFlow IO support is now available for Apple Silicon packages.
Refactor CpuExecutable to propagate LLVM errors.

Keras

Keras is a framework built on top of the TensorFlow. See more details on the Keras website.

Major Features and Improvements

tf.keras
- Model.compile now support steps_per_execution='auto' as a parameter, allowing automatic tuning of steps per execution during Model.fit, Model.predict, and Model.evaluate for a significant performance boost.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Aakar Dwivedi, Adrian Popescu, ag.ramesh, Akhil Goel, Albert Zeyer, Alex Rosen, Alexey Vishnyakov, Andrew Goodbody, angerson, Ashiq Imran, Ayan Moitra, Ben Barsdell, Bhavani Subramanian, Boian Petkantchin, BrianWieder, Chris Mc, cloudhan, Connor Flanagan, Daniel Lang, Daniel Yudelevich, Darya Parygina, David Korczynski, David Svantesson, dingyuqing05, Dragan Mladjenovic, dskkato, Eli Kobrin, Erick Ochoa, Erik Schultheis, Frédéric Bastien, gaikwadrahul8, Gauri1 Deshpande, guozhong.zhuang, H. Vetinari, Isaac Cilia Attard, Jake Hall, Jason Furmanek, Jerry Ge, Jinzhe Zeng, JJ, johnnkp, Jonathan Albrecht, jongkweh, justkw, Kanvi Khanna, kikoxia, Koan-Sin Tan, Kun-Lu, ltsai1, Lu Teng, luliyucoordinate, Mahmoud Abuzaina, mdfaijul, Milos Puzovic, Nathan Luehr, Om Thakkar, pateldeev, Peng Sun, Philipp Hack, pjpratik, Poliorcetics, rahulbatra85, rangjiaheng, Renato Arantes, Robert Kalmar, roho, Rylan Justice, Sachin Muradi, samypr100, Saoirse Stewart, Shanbin Ke, Shivam Mishra, shuw, Song Ziming, Stephan Hartmann, Sulav, sushreebarsa, T Coxon, Tai Ly, talyz, Thibaut Goetghebuer-Planchon, Thomas Preud'Homme, tilakrayal, Tirumalesh, Tj Xu, Tom Allsop, Trevor Morris, Varghese, Jojimon, Wen Chen, Yaohui Liu, Yimei Sun, Zhoulong Jiang, Zhoulong, Jiang

Open source →

2.14.0rc11 Sept 2023pre-releaseDeprecated

tf.compat.v1.Session.partial_run and tf.compat.v1.Session.partial_run_setup will be deprecated in the next release.

Release 2.14.0

Tensorflow

Breaking Changes

Support for Python 3.8 has been removed starting with TF 2.14. The TensorFlow 2.13.1 patch release will still have Python 3.8 support.
tf.Tensor
- The class hierarchy for tf.Tensor has changed, and there are now explicit EagerTensor and SymbolicTensor classes for eager and tf.function respectively. Users who relied on the exact type of Tensor (e.g. type(t) == tf.Tensor) will need to update their code to use isinstance(t, tf.Tensor). The tf.is_symbolic_tensor helper added in 2.13 may be used when it is necessary to determine if a value is specifically a symbolic tensor.
tf.compat.v1.Session
- tf.compat.v1.Session.partial_run and tf.compat.v1.Session.partial_run_setup will be deprecated in the next release.
tf.estimator
- tf.estimator API will be removed in the next release. TF Estimator Python package will no longer be released.

Known Caveats

tf.lite
- when converter flag "_experimenal_use_buffer_offset" is enabled, additional metadata is automatically excluded from the generated model. The behaviour is the same as "exclude_conversion_metadata" is set
- If the model is larger than 2GB, then we also require "exclude_conversion_metadata" flag to be set

Major Features and Improvements

The tensorflow pip package has a new, optional installation method for Linux that installs necessary Nvidia CUDA libraries through pip. As long as the Nvidia driver is already installed on the system, you may now run pip install tensorflow[and-cuda] to install TensorFlow's Nvidia CUDA library dependencies in the Python environment. Aside from the Nvidia driver, no other pre-existing Nvidia CUDA packages are necessary.
Enable JIT-compiled i64-indexed kernels on GPU for large tensors with more than 2**32 elements.
- Unary GPU kernels: Abs, Atanh, Acos, Acosh, Asin, Asinh, Atan, Cos, Cosh, Sin, Sinh, Tan, Tanh.
- Binary GPU kernels: AddV2, Sub, Div, DivNoNan, Mul, MulNoNan, FloorDiv, Equal, NotEqual, Greater, GreaterEqual, LessEqual, Less.
tf.lite
- Add experimental supports conversion of models that may be larger than 2GB before buffer deduplication

Bug Fixes and Other Changes

tf.py_function and tf.numpy_function can now be used as function decorators for clearer code:

@tf.py_function(Tout=tf.float32)
def my_fun(x):
  print("This always executes eagerly.")
  return x+1

tf.lite
- Strided_Slice now supports UINT32.
tf.config.experimental.enable_tensor_float_32_execution
- Disabling TensorFloat-32 execution now causes TPUs to use float32 precision for float32 matmuls and other ops. TPUs have always used bfloat16 precision for certain ops, like matmul, when such ops had float32 inputs. Now, disabling TensorFloat-32 by calling tf.config.experimental.enable_tensor_float_32_execution(False) will cause TPUs to use float32 precision for such ops instead of bfloat16.
tf.experimental.dtensor
- API changes for Relayout. Added a new API, dtensor.relayout_like, for relayouting a tensor according to the layout of another tensor.
- Added dtensor.get_default_mesh, for retrieving the current default mesh under the dtensor context.
- *fft* ops now support dtensors with any layout. Fixed bug in 'fft2d/ fft3d', 'ifft2d/ifft3d', 'rfft2d/rfft3d', and 'irfft2d/irfft3d' for sharded input.
tf.experimental.strict_mode
- Added a new API, strict_mode, which converts all deprecation warnings into runtime errors with instructions on switching to recommended substitute.
TensorFlow Debugger (tfdbg) CLI: ncurses-based CLI for tfdbg v1 was removed.
TensorFlow now supports C++ RTTI on mobile and Android. To enable this feature, pass the flag --define=tf_force_rtti=true to Bazel when building TensorFlow. This may be needed when linking TensorFlow into RTTI-enabled programs since mixing RTTI and non-RTTI code can cause ABI issues.
tf.ones, tf.zeros, tf.fill, tf.ones_like, tf.zeros_like now take an additional Layout argument that controls the output layout of their results.
tf.nest and tf.data now support user defined classes implementing __tf_flatten__ and __tf_unflatten__ methods. See nest_util code examples for an example.

Keras

Keras is a framework built on top of the TensorFlow. See more details on the Keras website.

Major Features and Improvements

tf.keras
- Model.compile now support steps_per_execution='auto' as a parameter, allowing automatic tuning of steps per execution during Model fit, Model.predict, and Model.evaluate for a significant performance boost.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Aakar Dwivedi, Adrian Popescu, ag.ramesh, Akhil Goel, Albert Zeyer, Alex Rosen, Alexey Vishnyakov, Andrew Goodbody, angerson, Ashiq Imran, Ayan Moitra, Ben Barsdell, Bhavani Subramanian, Boian Petkantchin, BrianWieder, Chris Mc, cloudhan, Connor Flanagan, Daniel Lang, Daniel Yudelevich, Darya Parygina, David Korczynski, David Svantesson, dingyuqing05, Dragan Mladjenovic, dskkato, Eli Kobrin, Erick Ochoa, Erik Schultheis, Frédéric Bastien, gaikwadrahul8, Gauri1 Deshpande, georgiie, guozhong.zhuang, H. Vetinari, Isaac Cilia Attard, Jake Hall, Jason Furmanek, Jerry Ge, Jinzhe Zeng, JJ, johnnkp, Jonathan Albrecht, jongkweh, justkw, Kanvi Khanna, kikoxia, Koan-Sin Tan, Kun-Lu, Learning-To-Play, ltsai1, Lu Teng, luliyucoordinate, Mahmoud Abuzaina, mdfaijul, Milos Puzovic, Nathan Luehr, Om Thakkar, pateldeev, Peng Sun, Philipp Hack, pjpratik, Poliorcetics, rahulbatra85, rangjiaheng, Renato Arantes, Robert Kalmar, roho, Rylan Justice, Sachin Muradi, samypr100, Saoirse Stewart, Shanbin Ke, Shivam Mishra, shuw, Song Ziming, Stephan Hartmann, Sulav, sushreebarsa, T Coxon, Tai Ly, talyz, Tensorflow Jenkins, Thibaut Goetghebuer-Planchon, Thomas Preud'Homme, tilakrayal, Tirumalesh, Tj Xu, Tom Allsop, Trevor Morris, Varghese, Jojimon, Wen Chen, Yaohui Liu, Yimei Sun, Zhoulong Jiang, Zhoulong, Jiang

Open source →

2.14.0rc017 Aug 2023pre-releaseDeprecated

tf.compat.v1.Session.partial_run and tf.compat.v1.Session.partial_run_setup will be deprecated in the next release.

Release 2.14.0

Tensorflow

Breaking Changes

tf.Tensor
- The class hierarchy for tf.Tensor has changed, and there are now explicit EagerTensor and SymbolicTensor classes for eager and tf.function respectively. Users who relied on the exact type of Tensor (e.g. type(t) == tf.Tensor) will need to update their code to use isinstance(t, tf.Tensor). The tf.is_symbolic_tensor helper added in 2.13 may be used when it is necessary to determine if a value is specifically a symbolic tensor.
tf.compat.v1.Session
- tf.compat.v1.Session.partial_run and tf.compat.v1.Session.partial_run_setup will be deprecated in the next release.

Known Caveats

tf.lite
- when converter flag "_experimenal_use_buffer_offset" is enabled, additional metadata is automatically excluded from the generated model. The behaviour is the same as "exclude_conversion_metadata" is set
- If the model is larger than 2GB, then we also require "exclude_conversion_metadata" flag to be set

Major Features and Improvements

Enable JIT-compiled i64-indexed kernels on GPU for large tensors with more than 2**32 elements.
- Unary GPU kernels: Abs, Atanh, Acos, Acosh, Asin, Asinh, Atan, Cos, Cosh, Sin, Sinh, Tan, Tanh.
- Binary GPU kernels: AddV2, Sub, Div, DivNoNan, Mul, MulNoNan, FloorDiv, Equal, NotEqual, Greater, GreaterEqual, LessEqual, Less.
tf.lite
- Add experimental supports conversion of models that may be larger than 2GB before buffer deduplication

Bug Fixes and Other Changes

tf.py_function and tf.numpy_function can now be used as function decorators for clearer code:

@tf.py_function(Tout=tf.float32)
def my_fun(x):
  print("This always executes eagerly.")
  return x+1

tf.lite
- Strided_Slice now supports UINT32.
tf.config.experimental.enable_tensor_float_32_execution
- Disabling TensorFloat-32 execution now causes TPUs to use float32 precision for float32 matmuls and other ops. TPUs have always used bfloat16 precision for certain ops, like matmul, when such ops had float32 inputs. Now, disabling TensorFloat-32 by calling tf.config.experimental.enable_tensor_float_32_execution(False) will cause TPUs to use float32 precision for such ops instead of bfloat16.
tf.experimental.dtensor
- API changes for Relayout. Added a new API, dtensor.relayout_like, for relayouting a tensor according to the layout of another tensor.
- Added dtensor.get_default_mesh, for retrieving the current default mesh under the dtensor context.
- *fft* ops now support dtensors with any layout. Fixed bug in 'fft2d/ fft3d', 'ifft2d/ifft3d', 'rfft2d/rfft3d', and 'irfft2d/irfft3d' for sharded input.
tf.experimental.strict_mode
- Added a new API, strict_mode, which converts all deprecation warnings into runtime errors with instructions on switching to a recommended substitute.
TensorFlow Debugger (tfdbg) CLI: ncurses-based CLI for tfdbg v1 was removed.
TensorFlow now supports C++ RTTI on mobile and Android. To enable this feature, pass the flag --define=tf_force_rtti=true to Bazel when building TensorFlow. This may be needed when linking TensorFlow into RTTI-enabled programs since mixing RTTI and non-RTTI code can cause ABI issues.
tf.ones, tf.zeros, tf.fill, tf.ones_like, tf.zeros_like now take an additional Layout argument that controls the output layout of their results.
tf.nest and tf.data now support user defined classes implementing __tf_flatten__ and __tf_unflatten__ methods. See nest_util code examples for an example.

Keras

Keras is a framework built on top of the TensorFlow. See more details on the Keras website.

Major Features and Improvements

tf.keras
- Model.compile now support steps_per_execution='auto' as a parameter, allowing automatic tuning of steps per execution during Model.fit, Model.predict, and Model.evaluate for a significant performance boost.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Aakar Dwivedi, Adrian Popescu, ag.ramesh, Akhil Goel, Albert Zeyer, Alex Rosen, Alexey Vishnyakov, Andrew Goodbody, angerson, Ashiq Imran, Ayan Moitra, Ben Barsdell, Bhavani Subramanian, Boian Petkantchin, BrianWieder, Chris Mc, cloudhan, Connor Flanagan, Daniel Lang, Daniel Yudelevich, Darya Parygina, David Korczynski, David Svantesson, dingyuqing05, Dragan Mladjenovic, dskkato, Eli Kobrin, Erick Ochoa, Erik Schultheis, Frédéric Bastien, gaikwadrahul8, Gauri1 Deshpande, georgiie, guozhong.zhuang, H. Vetinari, Isaac Cilia Attard, Jake Hall, Jason Furmanek, Jerry Ge, Jinzhe Zeng, JJ, johnnkp, Jonathan Albrecht, jongkweh, justkw, Kanvi Khanna, kikoxia, Koan-Sin Tan, Kun-Lu, Learning-To-Play, ltsai1, Lu Teng, luliyucoordinate, Mahmoud Abuzaina, mdfaijul, Milos Puzovic, Nathan Luehr, Om Thakkar, pateldeev, Peng Sun, Philipp Hack, pjpratik, Poliorcetics, rahulbatra85, rangjiaheng, Renato Arantes, Robert Kalmar, roho, Rylan Justice, Sachin Muradi, samypr100, Saoirse Stewart, Shanbin Ke, Shivam Mishra, shuw, Song Ziming, Stephan Hartmann, Sulav, sushreebarsa, T Coxon, Tai Ly, talyz, Tensorflow Jenkins, Thibaut Goetghebuer-Planchon, Thomas Preud'Homme, tilakrayal, Tirumalesh, Tj Xu, Tom Allsop, Trevor Morris, Varghese, Jojimon, Wen Chen, Yaohui Liu, Yimei Sun, Zhoulong Jiang, Zhoulong, Jiang

Open source →

2.13.05 Jul 2023Security

Fixes correct values rank in UpperBound and LowerBound CVE-2023-33976

Release 2.13.0

TensorFlow

Breaking Changes

The LMDB kernels have been changed to return an error. This is in preparation for completely removing them from TensorFlow. The LMDB dependency that these kernels are bringing to TensorFlow has been dropped, thus making the build slightly faster and more secure.

Major Features and Improvements

tf.lite
- Added 16-bit and 64-bit float type support for built-in op cast.
- The Python TF Lite Interpreter bindings now have an option experimental_disable_delegate_clustering to turn-off delegate clustering.
- Added int16x8 support for the built-in op exp
- Added int16x8 support for the built-in op mirror_pad
- Added int16x8 support for the built-in ops space_to_batch_nd and batch_to_space_nd
- Added 16-bit int type support for built-in op less, greater_than, equal
- Added 8-bit and 16-bit support for floor_div and floor_mod.
- Added 16-bit and 32-bit int support for the built-in op bitcast.
- Added 8-bit/16-bit/32-bit int/uint support for the built-in op bitwise_xor
- Added int16 indices support for built-in op gather and gather_nd.
- Added 8-bit/16-bit/32-bit int/uint support for the built-in op right_shift
- Added reference implementation for 16-bit int unquantized add.
- Added reference implementation for 16-bit int and 32-bit unsigned int unquantized mul.
- add_op supports broadcasting up to 6 dimensions.
- Added 16-bit support for top_k.
tf.function
- ConcreteFunction (tf.types.experimental.ConcreteFunction) as generated through get_concrete_function now performs holistic input validation similar to calling tf.function directly. This can cause breakages where existing calls pass Tensors with the wrong shape or omit certain non-Tensor arguments (including default values).
tf.nn
- tf.nn.embedding_lookup_sparse and tf.nn.safe_embedding_lookup_sparse now support ids and weights described by tf.RaggedTensors.
- Added a new boolean argument allow_fast_lookup to tf.nn.embedding_lookup_sparse and tf.nn.safe_embedding_lookup_sparse, which enables a simplified and typically faster lookup procedure.
tf.data
- tf.data.Dataset.zip now supports Python-style zipping, i.e. Dataset.zip(a, b, c).
- tf.data.Dataset.shuffle now supports tf.data.UNKNOWN_CARDINALITY When doing a "full shuffle" using dataset = dataset.shuffle(dataset.cardinality()). But remember, a "full shuffle" will load the full dataset into memory so that it can be shuffled, so make sure to only use this with small datasets or datasets of small objects (like filenames).
tf.math
- tf.nn.top_k now supports specifying the output index type via parameter index_type. Supported types are tf.int16, tf.int32 (default), and tf.int64.
tf.SavedModel
- Introduced class method tf.saved_model.experimental.Fingerprint.from_proto(proto), which can be used to construct a Fingerprint object directly from a protobuf.
- Introduced member method tf.saved_model.experimental.Fingerprint.singleprint(), which provides a convenient way to uniquely identify a SavedModel.

Bug Fixes and Other Changes

tf.Variable
- Changed resource variables to inherit from tf.compat.v2.Variable instead of tf.compat.v1.Variable. Some checks for isinstance(v, tf compat.v1.Variable) that previously returned True may now return False.
tf.distribute
- Opened an experimental API, tf.distribute.experimental.coordinator.get_current_worker_index, for retrieving the worker index from within a worker, when using parameter server training with a custom training loop.
tf.experimental.dtensor
- Deprecated dtensor.run_on in favor of dtensor.default_mesh to correctly indicate that the context does not override the mesh that the ops and functions will run on, it only sets a fallback default mesh.
- List of members of dtensor.Layout and dtensor.Mesh have slightly changed as part of efforts to consolidate the C++ and Python source code with pybind11. Most notably, dtensor.Layout.serialized_string is removed.
- Minor API changes to represent Single Device Layout for non-distributed Tensors inside DTensor functions. Runtime support will be added soon.
tf.experimental.ExtensionType
- tf.experimental.ExtensionType now supports Python tuple as the type annotation of its fields.
tf.nest
- Deprecated API tf.nest.is_sequence has now been deleted. Please use tf.nest.is_nested instead.

Keras

Keras is a framework built on top of the TensorFlow. See more details on the Keras website.

Breaking Changes

Removed the Keras scikit-learn API wrappers (KerasClassifier and KerasRegressor), which had been deprecated in August 2021. We recommend using SciKeras instead.
The default Keras model saving format is now the Keras v3 format: calling model.save("xyz.keras") will no longer create a H5 file, it will create a native Keras model file. This will only be breaking for you if you were manually inspecting or modifying H5 files saved by Keras under a .keras extension. If this breaks you, simply add save_format="h5" to your .save() call to revert back to the prior behavior.
Added keras.utils.TimedThread utility to run a timed thread every x seconds. It can be used to run a threaded function alongside model training or any other snippet of code.
In the keras PyPI package, accessible symbols are now restricted to symbols that are intended to be public. This may affect your code if you were using import keras and you used keras functions that were not public APIs, but were accessible in earlier versions with direct imports. In those cases, please use the following guideline: - The API may be available in the public Keras API under a different name, so make sure to look for it on keras.io or TensorFlow docs and switch to the public version. - It could also be a simple python or TF utility that you could easily copy over to your own codebase. In those case, just make it your own! - If you believe it should definitely be a public Keras API, please open a feature request in keras GitHub repo. - As a workaround, you could import the same private symbol keras keras.src, but keep in mind the src namespace is not stable and those APIs may change or be removed in the future.

Major Features and Improvements

Added F-Score metrics tf.keras.metrics.FBetaScore, tf.keras.metrics.F1Score, and tf.keras.metrics.R2Score.
Added activation function tf.keras.activations.mish.
Added experimental keras.metrics.experimental.PyMetric API for metrics that run Python code on the host CPU (compiled outside of the TensorFlow graph). This can be used for integrating metrics from external Python libraries (like sklearn or pycocotools) into Keras as first-class Keras metrics.
Added tf.keras.optimizers.Lion optimizer.
Added tf.keras.layers.SpectralNormalization layer wrapper to perform spectral normalization on the weights of a target layer.
The SidecarEvaluatorModelExport callback has been added to Keras as keras.callbacks.SidecarEvaluatorModelExport. This callback allows for exporting the model the best-scoring model as evaluated by a SidecarEvaluator evaluator. The evaluator regularly evaluates the model and exports it if the user-defined comparison function determines that it is an improvement.
Added warmup capabilities to tf.keras.optimizers.schedules.CosineDecay learning rate scheduler. You can now specify an initial and target learning rate, and our scheduler will perform a linear interpolation between the two after which it will begin a decay phase.
Added experimental support for an exactly-once visitation guarantee for evaluating Keras models trained with tf.distribute ParameterServerStrategy, via the exact_evaluation_shards argument in Model.fit and Model.evaluate.
Added tf.keras.__internal__.KerasTensor,tf.keras.__internal__.SparseKerasTensor, and tf.keras.__internal__.RaggedKerasTensor classes. You can use these classes to do instance type checking and type annotations for layer/model inputs and outputs.
All the tf.keras.dtensor.experimental.optimizers classes have been merged with tf.keras.optimizers. You can migrate your code to use tf.keras.optimizers directly. The API namespace for tf.keras.dtensor.experimental.optimizers will be removed in future releases.
Added support for class_weight for 3+ dimensional targets (e.g. image segmentation masks) in Model.fit.
Added a new loss, keras.losses.CategoricalFocalCrossentropy.
Remove the tf.keras.dtensor.experimental.layout_map_scope(). You can user the tf.keras.dtensor.experimental.LayoutMap.scope() instead.

Security

Fixes correct values rank in UpperBound and LowerBound CVE-2023-33976

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

103yiran, 8bitmp3, Aakar, Aakar Dwivedi, Abinash Satapathy, Aditya Kane, ag.ramesh, Alexander Grund, Andrei Pikas, andreii, Andrew Goodbody, angerson, Anthony_256, Ashay Rane, Ashiq Imran, Awsaf, Balint Cristian, Banikumar Maiti (Intel Aipg), Ben Barsdell, bhack, cfRod, Chao Chen, chenchongsong, Chris Mc, Daniil Kutz, David Rubinstein, dianjiaogit, dixr, Dongfeng Yu, dongfengy, drah, Eric Kunze, Feiyue Chen, Frederic Bastien, Gauri1 Deshpande, guozhong.zhuang, hDn248, HYChou, ingkarat, James Hilliard, Jason Furmanek, Jaya, Jens Glaser, Jerry Ge, Jiao Dian'S Power Plant, Jie Fu, Jinzhe Zeng, Jukyy, Kaixi Hou, Kanvi Khanna, Karel Ha, karllessard, Koan-Sin Tan, Konstantin Beluchenko, Kulin Seth, Kun Lu, Kyle Gerard Felker, Leopold Cambier, Lianmin Zheng, linlifan, liuyuanqiang, Lukas Geiger, Luke Hutton, Mahmoud Abuzaina, Manas Mohanty, Mateo Fidabel, Maxiwell S. Garcia, Mayank Raunak, mdfaijul, meatybobby, Meenakshi Venkataraman, Michael Holman, Nathan John Sircombe, Nathan Luehr, nitins17, Om Thakkar, Patrice Vignola, Pavani Majety, per1234, Philipp Hack, pollfly, Prianka Liz Kariat, Rahul Batra, rahulbatra85, ratnam.parikh, Rickard Hallerbäck, Roger Iyengar, Rohit Santhanam, Roman Baranchuk, Sachin Muradi, sanadani, Saoirse Stewart, seanshpark, Shawn Wang, shuw, Srinivasan Narayanamoorthy, Stewart Miles, Sunita Nadampalli, SuryanarayanaY, Takahashi Shuuji, Tatwai Chong, Thibaut Goetghebuer-Planchon, tilakrayal, Tirumalesh, TJ, Tony Sung, Trevor Morris, unda, Vertexwahn, venkat2469, William Muir, Xavier Bonaventura, xiang.zhang, Xiao-Yong Jin, yleeeee, Yong Tang, Yuriy Chernyshov, Zhang, Xiangze, zhaozheng09

Open source →

2.13.0rc222 Jun 2023pre-releaseDeprecated

Deprecated dtensor.run_on in favor of dtensor.default_mesh to correctly indicate that the context does not override the mesh that the ops and function…

Release 2.13.0

TensorFlow

Breaking Changes

The LMDB kernels have been changed to return an error. This is in preparation for completely removing them from TensorFlow. The LMDB dependency that these kernels are bringing to TensorFlow has been dropped, thus making the build slightly faster and more secure.

Major Features and Improvements

tf.lite
- Added 16-bit and 64-bit float type support for built-in op cast.
- The Python TF Lite Interpreter bindings now have an option experimental_disable_delegate_clustering to turn-off delegate clustering.
- Added int16x8 support for the built-in op exp
- Added int16x8 support for the built-in op mirror_pad
- Added int16x8 support for the built-in ops space_to_batch_nd and batch_to_space_nd
- Added 16-bit int type support for built-in op less, greater_than, equal
- Added 8-bit and 16-bit support for floor_div and floor_mod.
- Added 16-bit and 32-bit int support for the built-in op bitcast.
- Added 8-bit/16-bit/32-bit int/uint support for the built-in op bitwise_xor
- Added int16 indices support for built-in op gather and gather_nd.
- Added 8-bit/16-bit/32-bit int/uint support for the built-in op right_shift
- Added reference implementation for 16-bit int unquantized add.
- Added reference implementation for 16-bit int and 32-bit unsigned int unquantized mul.
- add_op supports broadcasting up to 6 dimensions.
- Added 16-bit support for top_k.
tf.function
- ConcreteFunction (tf.types.experimental.ConcreteFunction) as generated through get_concrete_function now performs holistic input validation similar to calling tf.function directly. This can cause breakages where existing calls pass Tensors with the wrong shape or omit certain non-Tensor arguments (including default values).
tf.nn
- tf.nn.embedding_lookup_sparse and tf.nn.safe_embedding_lookup_sparse now support ids and weights described by tf.RaggedTensors.
- Added a new boolean argument allow_fast_lookup to tf.nn.embedding_lookup_sparse and tf.nn.safe_embedding_lookup_sparse, which enables a simplified and typically faster lookup procedure.
tf.data
- tf.data.Dataset.zip now supports Python-style zipping, i.e. Dataset.zip(a, b, c).
- tf.data.Dataset.shuffle now supports tf.data.UNKNOWN_CARDINALITY When doing a "full shuffle" using dataset = dataset.shuffle(dataset.cardinality()). But remember, a "full shuffle" will load the full dataset into memory so that it can be shuffled, so make sure to only use this with small datasets or datasets of small objects (like filenames).
tf.math
- tf.nn.top_k now supports specifying the output index type via parameter index_type. Supported types are tf.int16, tf.int32 (default), and tf.int64.
tf.SavedModel
- Introduced class method tf.saved_model.experimental.Fingerprint.from_proto(proto), which can be used to construct a Fingerprint object directly from a protobuf.
- Introduced member method tf.saved_model.experimental.Fingerprint.singleprint(), which provides a convenient way to uniquely identify a SavedModel.

Bug Fixes and Other Changes

tf.Variable
- Changed resource variables to inherit from tf.compat.v2.Variable instead of tf.compat.v1.Variable. Some checks for isinstance(v, tf compat.v1.Variable) that previously returned True may now return False.
tf.distribute
- Opened an experimental API, tf.distribute.experimental.coordinator.get_current_worker_index, for retrieving the worker index from within a worker, when using parameter server training with a custom training loop.
tf.experimental.dtensor
- Deprecated dtensor.run_on in favor of dtensor.default_mesh to correctly indicate that the context does not override the mesh that the ops and functions will run on, it only sets a fallback default mesh.
- List of members of dtensor.Layout and dtensor.Mesh have slightly changed as part of efforts to consolidate the C++ and Python source code with pybind11. Most notably, dtensor.Layout.serialized_string is removed.
- Minor API changes to represent Single Device Layout for non-distributed Tensors inside DTensor functions. Runtime support will be added soon.
tf.experimental.ExtensionType
- tf.experimental.ExtensionType now supports Python tuple as the type annotation of its fields.
tf.nest
- Deprecated API tf.nest.is_sequence has now been deleted. Please use tf.nest.is_nested instead.

Keras

Keras is a framework built on top of the TensorFlow. See more details on the Keras website.

Breaking Changes

Removed the Keras scikit-learn API wrappers (KerasClassifier and KerasRegressor), which had been deprecated in August 2021. We recommend using SciKeras instead.
The default Keras model saving format is now the Keras v3 format: calling model.save("xyz.keras") will no longer create a H5 file, it will create a native Keras model file. This will only be breaking for you if you were manually inspecting or modifying H5 files saved by Keras under a .keras extension. If this breaks you, simply add save_format="h5" to your .save() call to revert back to the prior behavior.
Added keras.utils.TimedThread utility to run a timed thread every x seconds. It can be used to run a threaded function alongside model training or any other snippet of code.
In the keras PyPI package, accessible symbols are now restricted to symbols that are intended to be public. This may affect your code if you were using import keras and you used keras functions that were not public APIs, but were accessible in earlier versions with direct imports. In those cases, please use the following guideline: - The API may be available in the public Keras API under a different name, so make sure to look for it on keras.io or TensorFlow docs and switch to the public version. - It could also be a simple python or TF utility that you could easily copy over to your own codebase. In those case, just make it your own! - If you believe it should definitely be a public Keras API, please open a feature request in keras GitHub repo. - As a workaround, you could import the same private symbol keras keras.src, but keep in mind the src namespace is not stable and those APIs may change or be removed in the future.

Major Features and Improvements

Added F-Score metrics tf.keras.metrics.FBetaScore, tf.keras.metrics.F1Score, and tf.keras.metrics.R2Score.
Added activation function tf.keras.activations.mish.
Added experimental keras.metrics.experimental.PyMetric API for metrics that run Python code on the host CPU (compiled outside of the TensorFlow graph). This can be used for integrating metrics from external Python libraries (like sklearn or pycocotools) into Keras as first-class Keras metrics.
Added tf.keras.optimizers.Lion optimizer.
Added tf.keras.layers.SpectralNormalization layer wrapper to perform spectral normalization on the weights of a target layer.
The SidecarEvaluatorModelExport callback has been added to Keras as keras.callbacks.SidecarEvaluatorModelExport. This callback allows for exporting the model the best-scoring model as evaluated by a SidecarEvaluator evaluator. The evaluator regularly evaluates the model and exports it if the user-defined comparison function determines that it is an improvement.
Added warmup capabilities to tf.keras.optimizers.schedules.CosineDecay learning rate scheduler. You can now specify an initial and target learning rate, and our scheduler will perform a linear interpolation between the two after which it will begin a decay phase.
Added experimental support for an exactly-once visitation guarantee for evaluating Keras models trained with tf.distribute ParameterServerStrategy, via the exact_evaluation_shards argument in Model.fit and Model.evaluate.
Added tf.keras.__internal__.KerasTensor,tf.keras.__internal__.SparseKerasTensor, and tf.keras.__internal__.RaggedKerasTensor classes. You can use these classes to do instance type checking and type annotations for layer/model inputs and outputs.
All the tf.keras.dtensor.experimental.optimizers classes have been merged with tf.keras.optimizers. You can migrate your code to use tf.keras.optimizers directly. The API namespace for tf.keras.dtensor.experimental.optimizers will be removed in future releases.
Added support for class_weight for 3+ dimensional targets (e.g. image segmentation masks) in Model.fit.
Added a new loss, keras.losses.CategoricalFocalCrossentropy.
Remove the tf.keras.dtensor.experimental.layout_map_scope(). You can user the tf.keras.dtensor.experimental.LayoutMap.scope() instead.

Security

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.13.0rc130 May 2023pre-releaseDeprecated

Deprecated dtensor.run_on in favor of dtensor.default_mesh to correctly indicate that the context does not override the mesh that the ops and function…

Release 2.13.0

TensorFlow

Breaking Changes

The LMDB kernels have been changed to return an error. This is in preparation for completely removing them from TensorFlow. The LMDB dependency that these kernels are bringing to TensorFlow has been dropped, thus making the build slightly faster and more secure.

Major Features and Improvements

tf.lite
- Added 16-bit and 64-bit float type support for built-in op cast.
- The Python TF Lite Interpreter bindings now have an option experimental_disable_delegate_clustering to turn-off delegate clustering.
- Added int16x8 support for the built-in op exp
- Added int16x8 support for the built-in op mirror_pad
- Added int16x8 support for the built-in ops space_to_batch_nd and batch_to_space_nd
- Added 16-bit int type support for built-in op less, greater_than, equal
- Added 8-bit and 16-bit support for floor_div and floor_mod.
- Added 16-bit and 32-bit int support for the built-in op bitcast.
- Added 8-bit/16-bit/32-bit int/uint support for the built-in op bitwise_xor
- Added int16 indices support for built-in op gather and gather_nd.
- Added 8-bit/16-bit/32-bit int/uint support for the built-in op right_shift
- Added reference implementation for 16-bit int unquantized add.
- Added reference implementation for 16-bit int and 32-bit unsigned int unquantized mul.
- add_op supports broadcasting up to 6 dimensions.
- Added 16-bit support for top_k.
tf.function
- ConcreteFunction (tf.types.experimental.ConcreteFunction) as generated through get_concrete_function now performs holistic input validation similar to calling tf.function directly. This can cause breakages where existing calls pass Tensors with the wrong shape or omit certain non-Tensor arguments (including default values).
tf.nn
- tf.nn.embedding_lookup_sparse and tf.nn.safe_embedding_lookup_sparse now support ids and weights described by tf.RaggedTensors.
- Added a new boolean argument allow_fast_lookup to tf.nn.embedding_lookup_sparse and tf.nn.safe_embedding_lookup_sparse, which enables a simplified and typically faster lookup procedure.
tf.data
- tf.data.Dataset.zip now supports Python-style zipping, i.e. Dataset.zip(a, b, c).
- tf.data.Dataset.shuffle now supports tf.data.UNKNOWN_CARDINALITY When doing a "full shuffle" using dataset = dataset.shuffle(dataset.cardinality()). But remember, a "full shuffle" will load the full dataset into memory so that it can be shuffled, so make sure to only use this with small datasets or datasets of small objects (like filenames).
tf.math
- tf.nn.top_k now supports specifying the output index type via parameter index_type. Supported types are tf.int16, tf.int32 (default), and tf.int64.
tf.SavedModel
- Introduced class method tf.saved_model.experimental.Fingerprint.from_proto(proto), which can be used to construct a Fingerprint object directly from a protobuf.
- Introduced member method tf.saved_model.experimental.Fingerprint.singleprint(), which provides a convenient way to uniquely identify a SavedModel.

Bug Fixes and Other Changes

tf.Variable
- Changed resource variables to inherit from tf.compat.v2.Variable instead of tf.compat.v1.Variable. Some checks for isinstance(v, tf compat.v1.Variable) that previously returned True may now return False.
tf.distribute
- Opened an experimental API, tf.distribute.experimental.coordinator.get_current_worker_index, for retrieving the worker index from within a worker, when using parameter server training with a custom training loop.
tf.experimental.dtensor
- Deprecated dtensor.run_on in favor of dtensor.default_mesh to correctly indicate that the context does not override the mesh that the ops and functions will run on, it only sets a fallback default mesh.
- List of members of dtensor.Layout and dtensor.Mesh have slightly changed as part of efforts to consolidate the C++ and Python source code with pybind11. Most notably, dtensor.Layout.serialized_string is removed.
- Minor API changes to represent Single Device Layout for non-distributed Tensors inside DTensor functions. Runtime support will be added soon.
tf.experimental.ExtensionType
- tf.experimental.ExtensionType now supports Python tuple as the type annotation of its fields.
tf.nest
- Deprecated API tf.nest.is_sequence has now been deleted. Please use tf.nest.is_nested instead.

Keras

Keras is a framework built on top of the TensorFlow. See more details on the Keras website.

Breaking Changes

Removed the Keras scikit-learn API wrappers (KerasClassifier and KerasRegressor), which had been deprecated in August 2021. We recommend using SciKeras instead.
The default Keras model saving format is now the Keras v3 format: calling model.save("xyz.keras") will no longer create a H5 file, it will create a native Keras model file. This will only be breaking for you if you were manually inspecting or modifying H5 files saved by Keras under a .keras extension. If this breaks you, simply add save_format="h5" to your .save() call to revert back to the prior behavior.
Added keras.utils.TimedThread utility to run a timed thread every x seconds. It can be used to run a threaded function alongside model training or any other snippet of code.
In the keras PyPI package, accessible symbols are now restricted to symbols that are intended to be public. This may affect your code if you were using import keras and you used keras functions that were not public APIs, but were accessible in earlier versions with direct imports. In those cases, please use the following guideline: - The API may be available in the public Keras API under a different name, so make sure to look for it on keras.io or TensorFlow docs and switch to the public version. - It could also be a simple python or TF utility that you could easily copy over to your own codebase. In those case, just make it your own! - If you believe it should definitely be a public Keras API, please open a feature request in keras GitHub repo. - As a workaround, you could import the same private symbol keras keras.src, but keep in mind the src namespace is not stable and those APIs may change or be removed in the future.

Major Features and Improvements

Added F-Score metrics tf.keras.metrics.FBetaScore, tf.keras.metrics.F1Score, and tf.keras.metrics.R2Score.
Added activation function tf.keras.activations.mish.
Added experimental keras.metrics.experimental.PyMetric API for metrics that run Python code on the host CPU (compiled outside of the TensorFlow graph). This can be used for integrating metrics from external Python libraries (like sklearn or pycocotools) into Keras as first-class Keras metrics.
Added tf.keras.optimizers.Lion optimizer.
Added tf.keras.layers.SpectralNormalization layer wrapper to perform spectral normalization on the weights of a target layer.
The SidecarEvaluatorModelExport callback has been added to Keras as keras.callbacks.SidecarEvaluatorModelExport. This callback allows for exporting the model the best-scoring model as evaluated by a SidecarEvaluator evaluator. The evaluator regularly evaluates the model and exports it if the user-defined comparison function determines that it is an improvement.
Added warmup capabilities to tf.keras.optimizers.schedules.CosineDecay learning rate scheduler. You can now specify an initial and target learning rate, and our scheduler will perform a linear interpolation between the two after which it will begin a decay phase.
Added experimental support for an exactly-once visitation guarantee for evaluating Keras models trained with tf.distribute ParameterServerStrategy, via the exact_evaluation_shards argument in Model.fit and Model.evaluate.
Added tf.keras.__internal__.KerasTensor,tf.keras.__internal__.SparseKerasTensor, and tf.keras.__internal__.RaggedKerasTensor classes. You can use these classes to do instance type checking and type annotations for layer/model inputs and outputs.
All the tf.keras.dtensor.experimental.optimizers classes have been merged with tf.keras.optimizers. You can migrate your code to use tf.keras.optimizers directly. The API namespace for tf.keras.dtensor.experimental.optimizers will be removed in future releases.
Added support for class_weight for 3+ dimensional targets (e.g. image segmentation masks) in Model.fit.
Added a new loss, keras.losses.CategoricalFocalCrossentropy.
Remove the tf.keras.dtensor.experimental.layout_map_scope(). You can user the tf.keras.dtensor.experimental.LayoutMap.scope() instead.

Security

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.13.0rc010 May 2023pre-releaseDeprecated

Deprecated dtensor.run_on in favor of dtensor.default_mesh to correctly indicate that the context does not override the mesh that the ops and function…

Release 2.13.0

TensorFlow

Breaking Changes

The LMDB kernels have been changed to return an error. This is in preparation for completely removing them from TensorFlow. The LMDB dependency that these kernels are bringing to TensorFlow has been dropped, thus making the build slightly faster and more secure.

Major Features and Improvements

tf.lite
- Add 16-bit and 64-bit float type support for built-in op cast.
- The Python TF Lite Interpreter bindings now have an option experimental_disable_delegate_clustering to turn-off delegate clustering.
- Add int16x8 support for the built-in op exp
- Add int16x8 support for the built-in op mirror_pad
- Add int16x8 support for the built-in ops space_to_batch_nd and batch_to_space_nd
- Add 16-bit int type support for built-in op less, greater_than, equal
- Add 8-bit and 16-bit support for floor_div and floor_mod.
- Add 16-bit and 32-bit int support for the built-in op bitcast.
- Add 8-bit/16-bit/32-bit int/uint support for the built-in op bitwise_xor
- Add int16 indices support for built-in op gather and gather_nd.
- Add 8-bit/16-bit/32-bit int/uint support for the built-in op right_shift
- Add reference implementation for 16-bit int unquantized add.
- Add reference implementation for 16-bit int and 32-bit unsigned int unquantized mul.
- add_op supports broadcasting up to 6 dimensions.
- Add 16-bit support for top_k.
tf.function
- ConcreteFunction (tf.types.experimental.ConcreteFunction) as generated through get_concrete_function now performs holistic input validation similar to calling tf.function directly. This can cause breakages where existing calls pass Tensors with the wrong shape or omit certain non-Tensor arguments (including default values).
tf.nn
- tf.nn.embedding_lookup_sparse and tf.nn.safe_embedding_lookup_sparse now support ids and weights described by tf.RaggedTensors.
- Added a new boolean argument allow_fast_lookup to tf.nn.embedding_lookup_sparse and tf.nn.safe_embedding_lookup_sparse, which enables a simplified and typically faster lookup procedure.
tf.data
- tf.data.Dataset.zip now supports Python-style zipping, i.e. Dataset.zip(a, b, c).
- tf.data.Dataset.shuffle now supports full shuffling. To specify that data should be fully shuffled, use dataset = dataset.shuffle(dataset.cardinality()). This will load the full dataset into memory so that it can be shuffled, so make sure to only use this with datasets of filenames or other small datasets.
tf.math
- tf.nn.top_k now supports specifying the output index type via parameter index_type. Supported types are tf.int16, tf.int32 (default), and tf.int64.
tf.SavedModel
- Introduce class method tf.saved_model.experimental.Fingerprint.from_proto(proto), which can be used to construct a Fingerprint object directly from a protobuf.
- Introduce member method tf.saved_model.experimental.Fingerprint.singleprint(), which provides a convenient way to uniquely identify a SavedModel.

Bug Fixes and Other Changes

tf.Variable
- Changed resource variables to inherit from tf.compat.v2.Variable instead of tf.compat.v1.Variable. Some checks for isinstance(v, tf compat.v1.Variable) that previously returned True may now return False.
tf.distribute
- Opened an experimental API, tf.distribute.experimental.coordinator.get_current_worker_index, for retrieving the worker index from within a worker, when using parameter server training with a custom training loop.
tf.experimental.dtensor
- Deprecated dtensor.run_on in favor of dtensor.default_mesh to correctly indicate that the context does not override the mesh that the ops and functions will run on, it only sets a fallback default mesh.
- List of members of dtensor.Layout and dtensor.Mesh have slightly changed as part of efforts to consolidate the C++ and Python source code with pybind11. Most notably, Layout.serialized_string is removed.
- Minor API changes to represent Single Device Layout for non-distributed Tensors inside DTensor functions. Runtime support will be added soon.
tf.experimental.ExtensionType
- tf.experimental.ExtensionType now supports Python tuple as the type annotation of its fields.
tf.nest
- Deprecated API tf.nest.is_sequence has now been deleted. Please use tf.nest.is_nested instead.

Keras

Keras is a framework built on top of the TensorFlow. See more details on the Keras website.

Breaking Changes

tf.keras
- Removed the Keras scikit-learn API wrappers (KerasClassifier and KerasRegressor), which had been deprecated in August 2021. We recommend using SciKeras instead.
- The default Keras model saving format is now the Keras v3 format: calling model.save("xyz.keras") will no longer create a H5 file, it will create a native Keras model file. This will only be breaking for you if you were manually inspecting or modifying H5 files saved by Keras under a .keras extension. If this breaks you, simply add save_format="h5" to your .save() call to revert back to the prior behavior.
- Added keras.utils.TimedThread utility to run a timed thread every x seconds. It can be used to run a threaded function alongside model training or any other snippet of code.
- In the keras PyPI package, accessible symbols are now restricted to symbols that are intended to be public. This may affect your code if you were using import keras and you used keras functions that were not public APIs, but were accessible in earlier versions with direct imports. In those cases, please use the following guideline:
  - The API may be available in the public Keras API under a different name, so make sure to look for it on keras.io or TensorFlow docs and switch to the public version.
  - It could also be a simple python or TF utility that you could easily copy over to your own codebase. In those case, just make it your own!
  - If you believe it should definitely be a public Keras API, please open a feature request in keras GitHub repo.
  - As a workaround, you could import the same private symbol keras keras.src, but keep in mind the src namespace is not stable and those APIs may change or be removed in the future.

Major Features and Improvements

tf.keras
- Added F-Score metrics tf.keras.metrics.FBetaScore, tf.keras.metrics.F1Score, and tf.keras.metrics.R2Score.
- Added activation function tf.keras.activations.mish.
- Added experimental keras.metrics.experimental.PyMetric API for metrics that run Python code on the host CPU (compiled outside of the TensorFlow graph). This can be used for integrating metrics from external Python libraries (like sklearn or pycocotools) into Keras as first-class Keras metrics.
- Added tf.keras.optimizers.Lion optimizer.
- Added tf.keras.layers.SpectralNormalization layer wrapper to perform spectral normalization on the weights of a target layer.
- The SidecarEvaluatorModelExport callback has been added to Keras as keras.callbacks.SidecarEvaluatorModelExport. This callback allows for exporting the model the best-scoring model as evaluated by a SidecarEvaluator evaluator. The evaluator regularly evaluates the model and exports it if the user-defined comparison function determines that it is an improvement.
- Added warmup capabilities to tf.keras.optimizers.schedules.CosineDecay learning rate scheduler. You can now specify an initial and target learning rate, and our scheduler will perform a linear interpolation between the two after which it will begin a decay phase.
- Added experimental support for an exactly-once visitation guarantee for evaluating Keras models trained with tf.distribute ParameterServerStrategy, via the exact_evaluation_shards argument in Model.fit and Model.evaluate.
- Added tf.keras.__internal__.KerasTensor,tf.keras.__internal__.SparseKerasTensor, and tf.keras.__internal__.RaggedKerasTensor classes. You can use these classes to do instance type checking and type annotations for layer/model inputs and outputs.
- All the tf.keras.dtensor.experimental.optimizers classes have been merged with tf.keras.optimizers. You can migrate your code to use tf.keras.optimizers directly. The API namespace for tf.keras.dtensor.experimental.optimizers will be removed in future releases.
- Added support for class_weight for 3+ dimensional targets (e.g. image segmentation masks) in Model.fit.
- Added a new loss, keras.losses.CategoricalFocalCrossentropy.
- Remove the tf.keras.dtensor.experimental.layout_map_scope(). You can user the tf.keras.dtensor.experimental.LayoutMap.scope() instead.

Security

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.12.022 Mar 2023Security

Fixes an FPE in TFLite in conv kernel CVE-2023-27579

Release 2.12.0

TensorFlow

Breaking Changes

Build, Compilation and Packaging
- Removed redundant packages tensorflow-gpu and tf-nightly-gpu. These packages were removed and replaced with packages that direct users to switch to tensorflow or tf-nightly respectively. Since TensorFlow 2.1, the only difference between these two sets of packages was their names, so there is no loss of functionality or GPU support. See https://pypi.org/project/tensorflow-gpu for more details.
tf.function:
- tf.function now uses the Python inspect library directly for parsing the signature of the Python function it is decorated on. This change may break code where the function signature is malformed, but was ignored previously, such as:
  - Using functools.wraps on a function with different signature
  - Using functools.partial with an invalid tf.function input
- tf.function now enforces input parameter names to be valid Python identifiers. Incompatible names are automatically sanitized similarly to existing SavedModel signature behavior.
- Parameterless tf.functions are assumed to have an empty input_signature instead of an undefined one even if the input_signature is unspecified.
- tf.types.experimental.TraceType now requires an additional placeholder_value method to be defined.
- tf.function now traces with placeholder values generated by TraceType instead of the value itself.
Experimental APIs tf.config.experimental.enable_mlir_graph_optimization and tf.config.experimental.disable_mlir_graph_optimization were removed.

Major Features and Improvements

Support for Python 3.11 has been added.
Support for Python 3.7 has been removed. We are not releasing any more patches for Python 3.7.
tf.lite:
- Add 16-bit float type support for built-in op fill.
- Transpose now supports 6D tensors.
- Float LSTM now supports diagonal recurrent tensors: https://arxiv.org/abs/1903.08023
tf.experimental.dtensor:
- Coordination service now works with dtensor.initialize_accelerator_system, and enabled by default.
- Add tf.experimental.dtensor.is_dtensor to check if a tensor is a DTensor instance.
tf.data:
- Added support for alternative checkpointing protocol which makes it possible to checkpoint the state of the input pipeline without having to store the contents of internal buffers. The new functionality can be enabled through the experimental_symbolic_checkpoint option of tf.data.Options().
- Added a new rerandomize_each_iteration argument for the tf.data.Dataset.random() operation, which controls whether the sequence of generated random numbers should be re-randomized every epoch or not (the default behavior). If seed is set and rerandomize_each_iteration=True, the random() operation will produce a different (deterministic) sequence of numbers every epoch.
- Added a new rerandomize_each_iteration argument for the tf.data.Dataset.sample_from_datasets() operation, which controls whether the sequence of generated random numbers used for sampling should be re-randomized every epoch or not. If seed is set and rerandomize_each_iteration=True, the sample_from_datasets() operation will use a different (deterministic) sequence of numbers every epoch.
tf.test:
- Added tf.test.experimental.sync_devices, which is useful for accurately measuring performance in benchmarks.
tf.experimental.dtensor:
- Added experimental support to ReduceScatter fuse on GPU (NCCL).

Bug Fixes and Other Changes

tf.SavedModel:
- Introduced new class tf.saved_model.experimental.Fingerprint that contains the fingerprint of the SavedModel. See the SavedModel Fingerprinting RFC for details.
- Introduced API tf.saved_model.experimental.read_fingerprint(export_dir) for reading the fingerprint of a SavedModel.
tf.random
- Added non-experimental aliases for tf.random.split and tf.random.fold_in, the experimental endpoints are still available so no code changes are necessary.
tf.experimental.ExtensionType
- Added function experimental.extension_type.as_dict(), which converts an instance of tf.experimental.ExtensionType to a dict representation.
stream_executor
- Top level stream_executor directory has been deleted, users should use equivalent headers and targets under compiler/xla/stream_executor.
tf.nn
- Added tf.nn.experimental.general_dropout, which is similar to tf.random.experimental.stateless_dropout but accepts a custom sampler function.
tf.types.experimental.GenericFunction
- The experimental_get_compiler_ir method supports tf.TensorSpec compilation arguments.
tf.config.experimental.mlir_bridge_rollout
- Removed enums MLIR_BRIDGE_ROLLOUT_SAFE_MODE_ENABLED and MLIR_BRIDGE_ROLLOUT_SAFE_MODE_FALLBACK_ENABLED which are no longer used by the tf2xla bridge

Keras

Keras is a framework built on top of the TensorFlow. See more details on the Keras website.

Breaking Changes

tf.keras:

Moved all saving-related utilities to a new namespace, keras.saving, for example: keras.saving.load_model, keras.saving.save_model, keras.saving.custom_object_scope, keras.saving.get_custom_objects, keras.saving.register_keras_serializable,keras.saving.get_registered_name and keras.saving.get_registered_object. The previous API locations (in keras.utils and keras.models) will be available indefinitely, but we recommend you update your code to point to the new API locations.
Improvements and fixes in Keras loss masking:
- Whether you represent a ragged tensor as a tf.RaggedTensor or using keras masking, the returned loss values should be the identical to each other. In previous versions Keras may have silently ignored the mask.
If you use masked losses with Keras the loss values may be different in TensorFlow 2.12 compared to previous versions.
In cases where the mask was previously ignored, you will now get an error if you pass a mask with an incompatible shape.

Major Features and Improvements

tf.keras:

The new Keras model saving format (.keras) is available. You can start using it via model.save(f"{fname}.keras", save_format="keras_v3"). In the future it will become the default for all files with the .keras extension. This file format targets the Python runtime only and makes it possible to reload Python objects identical to the saved originals. The format supports non-numerical state such as vocabulary files and lookup tables, and it is easy to customize in the case of custom layers with exotic elements of state (e.g. a FIFOQueue). The format does not rely on bytecode or pickling, and is safe by default. Note that as a result, Python lambdas are disallowed at loading time. If you want to use lambdas, you can pass safe_mode=False to the loading method (only do this if you trust the source of the model).
Added a model.export(filepath) API to create a lightweight SavedModel artifact that can be used for inference (e.g. with TF-Serving).
Added keras.export.ExportArchive class for low-level customization of the process of exporting SavedModel artifacts for inference. Both ways of exporting models are based on tf.function tracing and produce a TF program composed of TF ops. They are meant primarily for environments where the TF runtime is available, but not the Python interpreter, as is typical for production with TF Serving.
Added utility tf.keras.utils.FeatureSpace, a one-stop shop for structured data preprocessing and encoding.
Added tf.SparseTensor input support to tf.keras.layers.Embedding layer. The layer now accepts a new boolean argument sparse. If sparse is set to True, the layer returns a SparseTensor instead of a dense Tensor. Defaults to False.
Added jit_compile as a settable property to tf.keras.Model.
Added synchronized optional parameter to layers.BatchNormalization.
Added deprecation warning to layers.experimental.SyncBatchNormalization and suggested to use layers.BatchNormalization with synchronized=True instead.
Updated tf.keras.layers.BatchNormalization to support masking of the inputs (mask argument) when computing the mean and variance.
Add tf.keras.layers.Identity, a placeholder pass-through layer.
Add show_trainable option to tf.keras.utils.model_to_dot to display layer trainable status in model plots.
Add ability to save a tf.keras.utils.FeatureSpace object, via feature_space.save("myfeaturespace.keras"), and reload it via feature_space = tf.keras.models.load_model("myfeaturespace.keras").
Added utility tf.keras.utils.to_ordinal to convert class vector to ordinal regression / classification matrix.

Bug Fixes and Other Changes

Security

Fixes an FPE in TFLite in conv kernel CVE-2023-27579
Fixes a double free in Fractional(Max/Avg)Pool CVE-2023-25801
Fixes a null dereference on ParallelConcat with XLA CVE-2023-25676
Fixes a segfault in Bincount with XLA CVE-2023-25675
Fixes an NPE in RandomShuffle with XLA enable CVE-2023-25674
Fixes an FPE in TensorListSplit with XLA CVE-2023-25673
Fixes segmentation fault in tfg-translate CVE-2023-25671
Fixes an NPE in QuantizedMatMulWithBiasAndDequantize CVE-2023-25670
Fixes an FPE in AvgPoolGrad with XLA CVE-2023-25669
Fixes a heap out-of-buffer read vulnerability in the QuantizeAndDequantize operation CVE-2023-25668
Fixes a segfault when opening multiframe gif CVE-2023-25667
Fixes an NPE in SparseSparseMaximum CVE-2023-25665
Fixes an FPE in AudioSpectrogram CVE-2023-25666
Fixes a heap-buffer-overflow in AvgPoolGrad CVE-2023-25664
Fixes a NPE in TensorArrayConcatV2 CVE-2023-25663
Fixes a Integer overflow in EditDistance CVE-2023-25662
Fixes a Seg fault in tf.raw_ops.Print CVE-2023-25660
Fixes a OOB read in DynamicStitch CVE-2023-25659
Fixes a OOB Read in GRUBlockCellGrad CVE-2023-25658

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

103yiran, 8bitmp3, Aakar, Aakar Dwivedi, Abinash Satapathy, Aditya Kane, ag.ramesh, Alexander Grund, Andrei Pikas, andreii, Andrew Goodbody, angerson, Anthony_256, Ashay Rane, Ashiq Imran, Awsaf, Balint Cristian, Banikumar Maiti (Intel Aipg), Ben Barsdell, bhack, cfRod, Chao Chen, chenchongsong, Chris Mc, Daniil Kutz, David Rubinstein, dianjiaogit, dixr, Dongfeng Yu, dongfengy, drah, Eric Kunze, Feiyue Chen, Frederic Bastien, Gauri1 Deshpande, guozhong.zhuang, hDn248, HYChou, ingkarat, James Hilliard, Jason Furmanek, Jaya, Jens Glaser, Jerry Ge, Jiao Dian'S Power Plant, Jie Fu, Jinzhe Zeng, Jukyy, Kaixi Hou, Kanvi Khanna, Karel Ha, karllessard, Koan-Sin Tan, Konstantin Beluchenko, Kulin Seth, Kun Lu, Kyle Gerard Felker, Leopold Cambier, Lianmin Zheng, linlifan, liuyuanqiang, Lukas Geiger, Luke Hutton, Mahmoud Abuzaina, Manas Mohanty, Mateo Fidabel, Maxiwell S. Garcia, Mayank Raunak, mdfaijul, meatybobby, Meenakshi Venkataraman, Michael Holman, Nathan John Sircombe, Nathan Luehr, nitins17, Om Thakkar, Patrice Vignola, Pavani Majety, per1234, Philipp Hack, pollfly, Prianka Liz Kariat, Rahul Batra, rahulbatra85, ratnam.parikh, Rickard Hallerbäck, Roger Iyengar, Rohit Santhanam, Roman Baranchuk, Sachin Muradi, sanadani, Saoirse Stewart, seanshpark, Shawn Wang, shuw, Srinivasan Narayanamoorthy, Stewart Miles, Sunita Nadampalli, SuryanarayanaY, Takahashi Shuuji, Tatwai Chong, Thibaut Goetghebuer-Planchon, tilakrayal, Tirumalesh, TJ, Tony Sung, Trevor Morris, unda, Vertexwahn, Vinila S, William Muir, Xavier Bonaventura, xiang.zhang, Xiao-Yong Jin, yleeeee, Yong Tang, Yuriy Chernyshov, Zhang, Xiangze, zhaozheng09

Open source →

2.12.0rc17 Mar 2023pre-releaseDeprecated

Added deprecation warning to layers.experimental.SyncBatchNormalization and suggested to use layers.BatchNormalization with synchronized=True instead.

Release 2.12.0

Breaking Changes

Build, Compilation and Packaging
- Removal of redundant packages: the tensorflow-gpu and tf-nightly-gpu packages have been effectively removed and replaced with packages that direct users to switch to tensorflow or tf-nightly respectively. The naming difference was the only difference between the two sets of packages ever since TensorFlow 2.1, so there is no loss of functionality or GPU support. See https://pypi.org/project/tensorflow-gpu for more details.
tf.function:
- tf.function now uses the Python inspect library directly for parsing the signature of the Python function it is decorated on.
- This can break certain cases that were previously ignored where the signature is malformed, such as:
  - Using functools.wraps on a function with different signature
  - Using functools.partial with an invalid tf.function input
- tf.function now enforces input parameter names to be valid Python identifiers. Incompatible names are automatically sanitized similarly to existing SavedModel signature behavior.
- Parameterless tf.functions are assumed to have an empty input_signature instead of an undefined one even if the input_signature is unspecified.
- tf.types.experimental.TraceType now requires an additional placeholder_value method to be defined.
- tf.function now traces with placeholder values generated by TraceType instead of the value itself.
Experimental APIs tf.config.experimental.enable_mlir_graph_optimization and tf.config.experimental.disable_mlir_graph_optimization were removed.
tf.keras:
- Moved all saving-related utilities to a new namespace, keras.saving, i.e. keras.saving.load_model, keras.saving.save_model, keras.saving.custom_object_scope, keras.saving.get_custom_objects, keras.saving.register_keras_serializable,keras.saving.get_registered_name and keras.saving.get_registered_object. The previous API locations (in keras.utils and keras.models) will stay available indefinitely, but we recommend that you update your code to point to the new API locations.
- Improvements and fixes in Keras loss masking:
  - Whether you represent a ragged tensor as a tf.RaggedTensor or using keras masking, the returned loss values should be the identical to each other. In previous versions Keras may have silently ignored the mask.
  - If you use masked losses with Keras the loss values may be different in TensorFlow 2.12 compared to previous versions.
  - In cases where the mask was previously ignored, you will now get an error if you pass a mask with an incompatible shape.

Major Features and Improvements

tf.lite:
- Add 16-bit float type support for built-in op fill.
- Transpose now supports 6D tensors.
- Float LSTM now supports diagonal recurrent tensors: https://arxiv.org/abs/1903.08023
tf.keras:
- The new Keras model saving format (.keras) is available. You can start using it via model.save(f"{fname}.keras", save_format="keras_v3"). In the future it will become the default for all files with the .keras extension. This file format targets the Python runtime only and makes it possible to reload Python objects identical to the saved originals. The format supports non-numerical state such as vocabulary files and lookup tables, and it is easy to customize in the case of custom layers with exotic elements of state (e.g. a FIFOQueue). The format does not rely on bytecode or pickling, and is safe by default. Note that as a result, Python lambdas are disallowed at loading time. If you want to use lambdas, you can pass safe_mode=False to the loading method (only do this if you trust the source of the model).
- Added a model.export(filepath) API to create a lightweight SavedModel artifact that can be used for inference (e.g. with TF-Serving).
- Added keras.export.ExportArchive class for low-level customization of the process of exporting SavedModel artifacts for inference. Both ways of exporting models are based on tf.function tracing and produce a TF program composed of TF ops. They are meant primarily for environments where the TF runtime is available, but not the Python interpreter, as is typical for production with TF Serving.
- Added utility tf.keras.utils.FeatureSpace, a one-stop shop for structured data preprocessing and encoding.
- Added tf.SparseTensor input support to tf.keras.layers.Embedding layer. The layer now accepts a new boolean argument sparse. If sparse is set to True, the layer returns a SparseTensor instead of a dense Tensor. Defaults to False.
- Added jit_compile as a settable property to tf.keras.Model.
- Added synchronized optional parameter to layers.BatchNormalization.
- Added deprecation warning to layers.experimental.SyncBatchNormalization and suggested to use layers.BatchNormalization with synchronized=True instead.
- Updated tf.keras.layers.BatchNormalization to support masking of the inputs (mask argument) when computing the mean and variance.
- Add tf.keras.layers.Identity, a placeholder pass-through layer.
- Add show_trainable option to tf.keras.utils.model_to_dot to display layer trainable status in model plots.
- Add ability to save a tf.keras.utils.FeatureSpace object, via feature_space.save("myfeaturespace.keras"), and reload it via feature_space = tf.keras.models.load_model("myfeaturespace.keras").
- Added utility tf.keras.utils.to_ordinal to convert class vector to ordinal regression / classification matrix.
tf.experimental.dtensor:
- Coordination service now works with dtensor.initialize_accelerator_system, and enabled by default.
- Add tf.experimental.dtensor.is_dtensor to check if a tensor is a DTensor instance.
tf.data:
- Added support for alternative checkpointing protocol which makes it possible to checkpoint the state of the input pipeline without having to store the contents of internal buffers. The new functionality can be enabled through the experimental_symbolic_checkpoint option of tf.data.Options().
- Added a new rerandomize_each_iteration argument for the tf.data.Dataset.random() operation, which controls whether the sequence of generated random numbers should be re-randomized every epoch or not (the default behavior). If seed is set and rerandomize_each_iteration=True, the random() operation will produce a different (deterministic) sequence of numbers every epoch.
- Added a new rerandomize_each_iteration argument for the tf.data.Dataset.sample_from_datasets() operation, which controls whether the sequence of generated random numbers used for sampling should be re-randomized every epoch or not. If seed is set and rerandomize_each_iteration=True, the sample_from_datasets() operation will use a different (deterministic) sequence of numbers every epoch.
tf.test:
- Added tf.test.experimental.sync_devices, which is useful for accurately measuring performance in benchmarks.
tf.experimental.dtensor:
- Added experimental support to ReduceScatter fuse on GPU (NCCL).

Bug Fixes and Other Changes

tf.SavedModel:
- Introduced new class tf.saved_model.experimental.Fingerprint that contains the fingerprint of the SavedModel. See the SavedModel Fingerprinting RFC for details.
- Introduced API tf.saved_model.experimental.read_fingerprint(export_dir) for reading the fingerprint of a SavedModel.
tf.random
- Added non-experimental aliases for tf.random.split and tf.random.fold_in, the experimental endpoints are still available so no code changes are necessary.
tf.experimental.ExtensionType
- Added function experimental.extension_type.as_dict(), which converts an instance of tf.experimental.ExtensionType to a dict representation.
stream_executor
- Top level stream_executor directory has been deleted, users should use equivalent headers and targets under compiler/xla/stream_executor.
tf.nn
- Added tf.nn.experimental.general_dropout, which is similar to tf.random.experimental.stateless_dropout but accepts a custom sampler function.
tf.types.experimental.GenericFunction
- The experimental_get_compiler_ir method supports tf.TensorSpec compilation arguments.
tf.config.experimental.mlir_bridge_rollout
- Removed enums MLIR_BRIDGE_ROLLOUT_SAFE_MODE_ENABLED and MLIR_BRIDGE_ROLLOUT_SAFE_MODE_FALLBACK_ENABLED which are no longer used by the tf2xla bridge

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

103yiran, 8bitmp3, Aakar, Aakar Dwivedi, Abinash Satapathy, Aditya Kane, ag.ramesh, Alexander Grund, Andrei Pikas, andreii, Andrew Goodbody, angerson, Anthony_256, Ashay Rane, Ashiq Imran, Awsaf, Balint Cristian, Banikumar Maiti (Intel Aipg), Ben Barsdell, bhack, cfRod, Chao Chen, chenchongsong, Chris Mc, Daniil Kutz, David Rubinstein, dianjiaogit, dixr, Dongfeng Yu, dongfengy, drah, Eric Kunze, Feiyue Chen, Frederic Bastien, Gauri1 Deshpande, guozhong.zhuang, hDn248, HYChou, ingkarat, James Hilliard, Jason Furmanek, Jaya, Jens Glaser, Jerry Ge, Jiao Dian'S Power Plant, Jie Fu, Jinzhe Zeng, Jukyy, Kaixi Hou, Kanvi Khanna, Karel Ha, karllessard, Koan-Sin Tan, Konstantin Beluchenko, Kulin Seth, Kun Lu, Kyle Gerard Felker, Leopold Cambier, Lianmin Zheng, linlifan, liuyuanqiang, Lukas Geiger, Luke Hutton, Mahmoud Abuzaina, Manas Mohanty, Mateo Fidabel, Maxiwell S. Garcia, Mayank Raunak, mdfaijul, meatybobby, Meenakshi Venkataraman, Michael Holman, Nathan John Sircombe, Nathan Luehr, nitins17, Om Thakkar, Patrice Vignola, Pavani Majety, per1234, Philipp Hack, pollfly, Prianka Liz Kariat, Rahul Batra, rahulbatra85, ratnam.parikh, Rickard Hallerbäck, Roger Iyengar, Rohit Santhanam, Roman Baranchuk, Sachin Muradi, sanadani, Saoirse Stewart, seanshpark, Shawn Wang, shuw, Srinivasan Narayanamoorthy, Stewart Miles, Sunita Nadampalli, SuryanarayanaY, Takahashi Shuuji, Tatwai Chong, Thibaut Goetghebuer-Planchon, tilakrayal, Tirumalesh, TJ, Tony Sung, Trevor Morris, unda, Vertexwahn, Vinila S, William Muir, Xavier Bonaventura, xiang.zhang, Xiao-Yong Jin, yleeeee, Yong Tang, Yuriy Chernyshov, Zhang, Xiangze, zhaozheng09

Open source →

2.12.0rc015 Feb 2023pre-releaseDeprecated

Added deprecation warning to layers.experimental.SyncBatchNormalization and suggested to use layers.BatchNormalization with synchronized=True instead.

Release 2.12.0

Breaking Changes

Build, Compilation and Packaging
- Removal of redundant packages: the tensorflow-gpu and tf-nightly-gpu packages have been effectively removed and replaced with packages that direct users to switch to tensorflow or tf-nightly respectively. The naming difference was the only difference between the two sets of packages ever since TensorFlow 2.1, so there is no loss of functionality or GPU support. See https://pypi.org/project/tensorflow-gpu for more details.
tf.function:
- tf.function now uses the Python inspect library directly for parsing the signature of the Python function it is decorated on.
- This can break certain cases that were previously ignored where the signature is malformed, such as:
  - Using functools.wraps on a function with different signature
  - Using functools.partial with an invalid tf.function input
- tf.function now enforces input parameter names to be valid Python identifiers. Incompatible names are automatically sanitized similarly to existing SavedModel signature behavior.
- Parameterless tf.functions are assumed to have an empty input_signature instead of an undefined one even if the input_signature is unspecified.
- tf.types.experimental.TraceType now requires an additional placeholder_value method to be defined.
- tf.function now traces with placeholder values generated by TraceType instead of the value itself.
Experimental APIs tf.config.experimental.enable_mlir_graph_optimization and tf.config.experimental.disable_mlir_graph_optimization were removed.
tf.keras:
- Moved all saving-related utilities to a new namespace, keras.saving, i.e. keras.saving.load_model, keras.saving.save_model, keras.saving.custom_object_scope, keras.saving.get_custom_objects, keras.saving.register_keras_serializable,keras.saving.get_registered_name and keras.saving.get_registered_object. The previous API locations (in keras.utils and keras.models) will stay available indefinitely, but we recommend that you update your code to point to the new API locations.
- Improvements and fixes in Keras loss masking:
  - Whether you represent a ragged tensor as a tf.RaggedTensor or using keras masking, the returned loss values should be the identical to each other. In previous versions Keras may have silently ignored the mask.
  - If you use masked losses with Keras the loss values may be different in TensorFlow 2.12 compared to previous versions.
  - In cases where the mask was previously ignored, you will now get an error if you pass a mask with an incompatible shape.
tf.SavedModel:
- Introduced new class tf.saved_model.experimental.Fingerprint that contains the fingerprint of the SavedModel. See the SavedModel Fingerprinting RFC for details.
- Introduced API tf.saved_model.experimental.read_fingerprint(export_dir) for reading the fingerprint of a SavedModel.

Major Features and Improvements

tf.lite:
- Add 16-bit float type support for built-in op fill.
- Transpose now supports 6D tensors.
- Float LSTM now supports diagonal recurrent tensors: https://arxiv.org/abs/1903.08023
tf.keras:
- The new Keras model saving format (.keras) is available. You can start using it via model.save(f"{fname}.keras", save_format="keras_v3"). In the future it will become the default for all files with the .keras extension. This file format targets the Python runtime only and makes it possible to reload Python objects identical to the saved originals. The format supports non-numerical state such as vocabulary files and lookup tables, and it is easy to customize in the case of custom layers with exotic elements of state (e.g. a FIFOQueue). The format does not rely on bytecode or pickling, and is safe by default. Note that as a result, Python lambdas are disallowed at loading time. If you want to use lambdas, you can pass safe_mode=False to the loading method (only do this if you trust the source of the model).
- Added a model.export(filepath) API to create a lightweight SavedModel artifact that can be used for inference (e.g. with TF-Serving).
- Added keras.export.ExportArchive class for low-level customization of the process of exporting SavedModel artifacts for inference. Both ways of exporting models are based on tf.function tracing and produce a TF program composed of TF ops. They are meant primarily for environments where the TF runtime is available, but not the Python interpreter, as is typical for production with TF Serving.
- Added utility tf.keras.utils.FeatureSpace, a one-stop shop for structured data preprocessing and encoding.
- Added tf.SparseTensor input support to tf.keras.layers.Embedding layer. The layer now accepts a new boolean argument sparse. If sparse is set to True, the layer returns a SparseTensor instead of a dense Tensor. Defaults to False.
- Added jit_compile as a settable property to tf.keras.Model.
- Added synchronized optional parameter to layers.BatchNormalization.
- Added deprecation warning to layers.experimental.SyncBatchNormalization and suggested to use layers.BatchNormalization with synchronized=True instead.
- Updated tf.keras.layers.BatchNormalization to support masking of the inputs (mask argument) when computing the mean and variance.
- Add tf.keras.layers.Identity, a placeholder pass-through layer.
- Add show_trainable option to tf.keras.utils.model_to_dot to display layer trainable status in model plots.
- Add ability to save a tf.keras.utils.FeatureSpace object, via feature_space.save("myfeaturespace.keras"), and reload it via feature_space = tf.keras.models.load_model("myfeaturespace.keras").
- Added utility tf.keras.utils.to_ordinal to convert class vector to ordinal regression / classification matrix.
tf.experimental.dtensor:
- Coordination service now works with dtensor.initialize_accelerator_system, and enabled by default.
- Add tf.experimental.dtensor.is_dtensor to check if a tensor is a DTensor instance.
tf.data:
- Added support for alternative checkpointing protocol which makes it possible to checkpoint the state of the input pipeline without having to store the contents of internal buffers. The new functionality can be enabled through the experimental_symbolic_checkpoint option of tf.data.Options().
- Added a new rerandomize_each_iteration argument for the tf.data.Dataset.random() operation, which controls whether the sequence of generated random numbers should be re-randomized every epoch or not (the default behavior). If seed is set and rerandomize_each_iteration=True, the random() operation will produce a different (deterministic) sequence of numbers every epoch.
- Added a new rerandomize_each_iteration argument for the tf.data.Dataset.sample_from_datasets() operation, which controls whether the sequence of generated random numbers used for sampling should be re-randomized every epoch or not. If seed is set and rerandomize_each_iteration=True, the sample_from_datasets() operation will use a different (deterministic) sequence of numbers every epoch.
tf.test:
- Added tf.test.experimental.sync_devices, which is useful for accurately measuring performance in benchmarks.
tf.experimental.dtensor:
- Added experimental support to ReduceScatter fuse on GPU (NCCL).

Bug Fixes and Other Changes

tf.random
- Added non-experimental aliases for tf.random.split and tf.random.fold_in, the experimental endpoints are still available so no code changes are necessary.
tf.experimental.ExtensionType
- Added function experimental.extension_type.as_dict(), which converts an instance of tf.experimental.ExtensionType to a dict representation.
stream_executor
- Top level stream_executor directory has been deleted, users should use equivalent headers and targets under compiler/xla/stream_executor.
tf.nn
- Added tf.nn.experimental.general_dropout, which is similar to tf.random.experimental.stateless_dropout but accepts a custom sampler function.
tf.types.experimental.GenericFunction
- The experimental_get_compiler_ir method supports tf.TensorSpec compilation arguments.
tf.config.experimental.mlir_bridge_rollout
- Removed enums MLIR_BRIDGE_ROLLOUT_SAFE_MODE_ENABLED and MLIR_BRIDGE_ROLLOUT_SAFE_MODE_FALLBACK_ENABLED which are no longer used by the tf2xla bridge

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

103yiran, 8bitmp3, Aakar, Aakar Dwivedi, Abinash Satapathy, Aditya Kane, ag.ramesh, Alexander Grund, Andrei Pikas, andreii, Andrew Goodbody, angerson, Anthony_256, Ashay Rane, Ashiq Imran, Awsaf, Balint Cristian, Banikumar Maiti (Intel Aipg), Ben Barsdell, bhack, cfRod, Chao Chen, chenchongsong, Chris Mc, Daniil Kutz, David Rubinstein, dianjiaogit, dixr, Dongfeng Yu, dongfengy, drah, Eric Kunze, Feiyue Chen, Frederic Bastien, Gauri1 Deshpande, guozhong.zhuang, hDn248, HYChou, ingkarat, James Hilliard, Jason Furmanek, Jaya, Jens Glaser, Jerry Ge, Jiao Dian'S Power Plant, Jie Fu, Jinzhe Zeng, Jukyy, Kaixi Hou, Kanvi Khanna, Karel Ha, karllessard, Koan-Sin Tan, Konstantin Beluchenko, Kulin Seth, Kun Lu, Kyle Gerard Felker, Leopold Cambier, Lianmin Zheng, linlifan, liuyuanqiang, Lukas Geiger, Luke Hutton, Mahmoud Abuzaina, Manas Mohanty, Mateo Fidabel, Maxiwell S. Garcia, Mayank Raunak, mdfaijul, meatybobby, Meenakshi Venkataraman, Michael Holman, Nathan John Sircombe, Nathan Luehr, nitins17, Om Thakkar, Patrice Vignola, Pavani Majety, per1234, Philipp Hack, pollfly, Prianka Liz Kariat, Rahul Batra, rahulbatra85, ratnam.parikh, Rickard Hallerbäck, Roger Iyengar, Rohit Santhanam, Roman Baranchuk, Sachin Muradi, sanadani, Saoirse Stewart, seanshpark, Shawn Wang, shuw, Srinivasan Narayanamoorthy, Stewart Miles, Sunita Nadampalli, SuryanarayanaY, Takahashi Shuuji, Tatwai Chong, Thibaut Goetghebuer-Planchon, tilakrayal, Tirumalesh, TJ, Tony Sung, Trevor Morris, unda, Vertexwahn, Vinila S, William Muir, Xavier Bonaventura, xiang.zhang, Xiao-Yong Jin, yleeeee, Yong Tang, Yuriy Chernyshov, Zhang, Xiangze, zhaozheng09

Open source →

2.11.018 Nov 2022Security

TF is currently using giflib 5.2.1 which has CVE-2022-28506. TF is not affected by the CVE as it does not use DumpScreen2RGB at all.

Release 2.11.0

Breaking Changes

The tf.keras.optimizers.Optimizer base class now points to the new Keras optimizer, while the old optimizers have been moved to the tf.keras.optimizers.legacy namespace.

If you find your workflow failing due to this change, you may be facing one of the following issues:
- Checkpoint loading failure. The new optimizer handles optimizer state differently from the old optimizer, which simplifies the logic of checkpoint saving/loading, but at the cost of breaking checkpoint backward compatibility in some cases. If you want to keep using an old checkpoint, please change your optimizer to tf.keras.optimizer.legacy.XXX (e.g. tf.keras.optimizer.legacy.Adam).
- TF1 compatibility. The new optimizer, tf.keras.optimizers.Optimizer, does not support TF1 any more, so please use the legacy optimizer tf.keras.optimizer.legacy.XXX. We highly recommend migrating your workflow to TF2 for stable support and new features.
- Old optimizer API not found. The new optimizer, tf.keras.optimizers.Optimizer, has a different set of public APIs from the old optimizer. These API changes are mostly related to getting rid of slot variables and TF1 support. Please check the API documentation to find alternatives to the missing API. If you must call the deprecated API, please change your optimizer to the legacy optimizer.
- Learning rate schedule access. When using a tf.keras.optimizers.schedules.LearningRateSchedule, the new optimizer's learning_rate property returns the current learning rate value instead of a LearningRateSchedule object as before. If you need to access the LearningRateSchedule object, please use optimizer._learning_rate.
- If you implemented a custom optimizer based on the old optimizer. Please set your optimizer to subclass tf.keras.optimizer.legacy.XXX. If you want to migrate to the new optimizer and find it does not support your optimizer, please file an issue in the Keras GitHub repo.
- Errors, such as Cannot recognize variable.... The new optimizer requires all optimizer variables to be created at the first apply_gradients() or minimize() call. If your workflow calls the optimizer to update different parts of the model in multiple stages, please call optimizer.build(model.trainable_variables) before the training loop.
- Timeout or performance loss. We don't anticipate this to happen, but if you see such issues, please use the legacy optimizer, and file an issue in the Keras GitHub repo.
The old Keras optimizer will never be deleted, but will not see any new feature additions. New optimizers (for example, tf.keras.optimizers.Adafactor) will only be implemented based on the new tf.keras.optimizers.Optimizer base class.
tensorflow/python/keras code is a legacy copy of Keras since the TensorFlow v2.7 release, and will be deleted in the v2.12 release. Please remove any import of tensorflow.python.keras and use the public API with from tensorflow import keras or import tensorflow as tf; tf.keras.

Major Features and Improvements

tf.lite:
- New operations supported: tf.math.unsorted_segment_sum, tf.atan2 and tf.sign.
- Updates to existing operations:
  - tfl.mul now supports complex32 inputs.
tf.experimental.StructuredTensor:
- Introduced tf.experimental.StructuredTensor, which provides a flexible and TensorFlow-native way to encode structured data such as protocol buffers or pandas dataframes.
tf.keras:
- Added a new get_metrics_result() method to tf.keras.models.Model.
  - Returns the current metrics values of the model as a dict.
- Added a new group normalization layer - tf.keras.layers.GroupNormalization.
- Added weight decay support for all Keras optimizers via the weight_decay argument.
- Added the Adafactor optimizer - tf.keras.optimizers.Adafactor.
- Added warmstart_embedding_matrix to tf.keras.utils.
  - This utility can be used to warmstart an embedding matrix, so you reuse previously-learned word embeddings when working with a new set of words which may include previously unseen words (the embedding vectors for unseen words will be randomly initialized).
tf.Variable:
- Added CompositeTensor as a base class to ResourceVariable.
  - This allows tf.Variables to be nested in tf.experimental.ExtensionTypes.
- Added a new constructor argument experimental_enable_variable_lifting to tf.Variable, defaulting to True.
  - When it's set to False, the variable won't be lifted out of tf.function; thus it can be used as a tf.function-local variable: during each execution of the tf.function, the variable will be created and then disposed, similar to a local (that is, stack-allocated) variable in C/C++. Currently, experimental_enable_variable_lifting=False only works on non-XLA devices (for example, under @tf.function(jit_compile=False)).
TF SavedModel:
- Added fingerprint.pb to the SavedModel directory. The fingerprint.pb file is a protobuf containing the "fingerprint" of the SavedModel. See the RFC for more details regarding its design and properties.
TF pip:
- Windows CPU-builds for x86/x64 processors are now built, maintained, tested and released by a third party: Intel. Installing the Windows-native pip packages for tensorflow or tensorflow-cpu would install Intel's tensorflow-intel package. These packages are provided on an as-is basis. TensorFlow will use reasonable efforts to maintain the availability and integrity of this pip package. There may be delays if the third party fails to release the pip package. For using TensorFlow GPU on Windows, you will need to install TensorFlow in WSL2.

Bug Fixes and Other Changes

tf.image:
- Added an optional parameter return_index_map to tf.image.ssim, which causes the returned value to be the local SSIM map instead of the global mean.
TF Core:
- tf.custom_gradient can now be applied to functions that accept "composite" tensors, such as tf.RaggedTensor, as inputs.
- Fix device placement issues related to datasets with ragged tensors of strings (i.e. variant encoded data with types not supported on GPU).
- experimental_follow_type_hints for tf.function has been deprecated. Please use input_signature or reduce_retracing to minimize retracing.
tf.SparseTensor:
- Introduced set_shape, which sets the static dense shape of the sparse tensor and has the same semantics as tf.Tensor.set_shape.

Security

TF is currently using giflib 5.2.1 which has CVE-2022-28506. TF is not affected by the CVE as it does not use DumpScreen2RGB at all.
Fixes an OOB seg fault in DynamicStitch due to missing validation (CVE-2022-41883)
Fixes an overflow in tf.keras.losses.poisson (CVE-2022-41887)
Fixes a heap OOB failure in ThreadUnsafeUnigramCandidateSampler caused by missing validation (CVE-2022-41880)
Fixes a segfault in ndarray_tensor_bridge (CVE-2022-41884)
Fixes an overflow in FusedResizeAndPadConv2D (CVE-2022-41885)
Fixes a overflow in ImageProjectiveTransformV2 (CVE-2022-41886)
Fixes an FPE in tf.image.generate_bounding_box_proposals on GPU (CVE-2022-41888)
Fixes a segfault in pywrap_tfe_src caused by invalid attributes (CVE-2022-41889)
Fixes a CHECK fail in BCast (CVE-2022-41890)
Fixes a segfault in TensorListConcat (CVE-2022-41891)
Fixes a CHECK_EQ fail in TensorListResize (CVE-2022-41893)
Fixes an overflow in CONV_3D_TRANSPOSE on TFLite (CVE-2022-41894)
Fixes a heap OOB in MirrorPadGrad (CVE-2022-41895)
Fixes a crash in Mfcc (CVE-2022-41896)
Fixes a heap OOB in FractionalMaxPoolGrad (CVE-2022-41897)
Fixes a CHECK fail in SparseFillEmptyRowsGrad (CVE-2022-41898)
Fixes a CHECK fail in SdcaOptimizer (CVE-2022-41899)
Fixes a heap OOB in FractionalAvgPool and FractionalMaxPool(CVE-2022-41900)
Fixes a CHECK_EQ in SparseMatrixNNZ (CVE-2022-41901)
Fixes an OOB write in grappler (CVE-2022-41902)
Fixes a overflow in ResizeNearestNeighborGrad (CVE-2022-41907)
Fixes a CHECK fail in PyFunc (CVE-2022-41908)
Fixes a segfault in CompositeTensorVariantToComponents (CVE-2022-41909)
Fixes a invalid char to bool conversion in printing a tensor (CVE-2022-41911)
Fixes a heap overflow in QuantizeAndDequantizeV2 (CVE-2022-41910)
Fixes a CHECK failure in SobolSample via missing validation (CVE-2022-35935)
Fixes a CHECK fail in TensorListScatter and TensorListScatterV2 in eager mode (CVE-2022-35935)

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

103yiran, 8bitmp3, Aakar Dwivedi, Alexander Grund, alif_elham, Aman Agarwal, amoitra, Andrei Ivanov, andreii, Andrew Goodbody, angerson, Ashay Rane, Azeem Shaikh, Ben Barsdell, bhack, Bhavani Subramanian, Cedric Nugteren, Chandra Kumar Ramasamy, Christopher Bate, CohenAriel, Cotarou, cramasam, Enrico Minack, Francisco Unda, Frederic Bastien, gadagashwini, Gauri1 Deshpande, george, Jake, Jeff, Jerry Ge, Jingxuan He, Jojimon Varghese, Jonathan Dekhtiar, Kaixi Hou, Kanvi Khanna, kcoul, Keith Smiley, Kevin Hu, Kun Lu, kushanam, Lianmin Zheng, liuyuanqiang, Louis Sugy, Mahmoud Abuzaina, Marius Brehler, mdfaijul, Meenakshi Venkataraman, Milos Puzovic, mohantym, Namrata-Ibm, Nathan John Sircombe, Nathan Luehr, Olaf Lipinski, Om Thakkar, Osman F Bayram, Patrice Vignola, Pavani Majety, Philipp Hack, Prianka Liz Kariat, Rahul Batra, RajeshT, Renato Golin, riestere, Roger Iyengar, Rohit Santhanam, Rsanthanam-Amd, Sadeed Pv, Samuel Marks, Shimokawa, Naoaki, Siddhesh Kothadi, Simengliu-Nv, Sindre Seppola, snadampal, Srinivasan Narayanamoorthy, sushreebarsa, syedshahbaaz, Tamas Bela Feher, Tatwai Chong, Thibaut Goetghebuer-Planchon, tilakrayal, Tom Anderson, Tomohiro Endo, Trevor Morris, vibhutisawant, Victor Zhang, Vremold, Xavier Bonaventura, Yanming Wang, Yasir Modak, Yimei Sun, Yong Tang, Yulv-Git, zhuoran.liu, zotanika

Open source →

2.11.0rc22 Nov 2022pre-releaseDeprecated

Please check the API documentation to find alternatives to the missing API. If you must call the deprecated API, please change your optimizer to the l…

Release 2.11.0

Breaking Changes

tf.keras.optimizers.Optimizer now points to the new Keras optimizer, and old optimizers have moved to the tf.keras.optimizers.legacy namespace.
If you find your workflow failing due to this change, you may be facing one of the following issues:
- Checkpoint loading failure. The new optimizer handles optimizer state differently from the old optimizer, which simplifies the logic of checkpoint saving/loading, but at the cost of breaking checkpoint backward compatibility in some cases. If you want to keep using an old checkpoint, please change your optimizer to tf.keras.optimizer.legacy.XXX (e.g. tf.keras.optimizer.legacy.Adam).
- TF1 compatibility. The new optimizer, tf.keras.optimizers.Optimizer, does not support TF1 any more, so please use the legacy optimizer tf.keras.optimizer.legacy.XXX. We highly recommend to migrate your workflow to TF2 for stable support and new features.
- Old optimizer API not found. The new optimizer, tf.keras.optimizers.Optimizer, has a different set of public APIs from the old optimizer. These API changes are mostly related to getting rid of slot variables and TF1 support. Please check the API documentation to find alternatives to the missing API. If you must call the deprecated API, please change your optimizer to the legacy optimizer.
- Learning rate schedule access. When using a LearningRateSchedule, The new optimizer's learning_rate property returns the current learning rate value instead of a LearningRateSchedule object as before. If you need to access the LearningRateSchedule object, please use optimizer._learning_rate.
- If you implemented a custom optimizer based on the old optimizer. Please set your optimizer to subclass tf.keras.optimizer.legacy.XXX. If you want to migrate to the new optimizer and find it does not support your optimizer, please file an issue in the Keras GitHub repo.
- Errors, such as Cannot recognize variable.... The new optimizer requires all optimizer variables to be created at the first apply_gradients() or minimize() call. If your workflow calls the optimizer to update different parts of the model in multiple stages, please call optimizer.build(model.trainable_variables) before the training loop.
- Timeout or performance loss. We don't anticipate this to happen, but if you see such issues, please use the legacy optimizer, and file an issue in the Keras GitHub repo.
The old Keras optimizer will never be deleted, but will not see any new feature additions. New optimizers (for example, tf.keras.optimizers.Adafactor) will only be implemented based on tf.keras.optimizers.Optimizer, the new base class.
tensorflow/python/keras code is a legacy copy of Keras since 2.7 release, and will be deleted in 2.12 release. Please remove any import of tensorflow.python.keras and use public API with from tensorflow import keras or import tensorflow as tf; tf.keras.

Major Features and Improvements

tf.lite:
- New operations supported: tf.unsortedsegmentmin, tf.atan2 and tf.sign.
- Updates to existing operations:
  - tfl.mul now supports complex32 inputs.
tf.experimental.StructuredTensor
- Introduced tf.experimental.StructuredTensor, which provides a flexible and TensorFlow-native way to encode structured data such as protocol buffers or pandas dataframes.
tf.keras:
- Added a new get_metrics_result() method to tf.keras.models.Model.
  - Returns the current metrics values of the model as a dict.
- Added a new group normalization layer - tf.keras.layers.GroupNormalization.
- Added weight decay support for all Keras optimizers.
- Added Adafactor optimizer tf.keras.optimizers.Adafactor.
- Added warmstart_embedding_matrix to tf.keras.utils.
  - This utility can be used to warmstart an embeddings matrix, so you reuse previously-learned word embeddings when working with a new set of words which may include previously unseen words (the embedding vectors for unseen words will be randomly initialized).
tf.Variable:
- Added CompositeTensor as a baseclass to ResourceVariable.
  - This allows tf.Variables to be nested in tf.experimental.ExtensionTypes.
- Added a new constructor argument experimental_enable_variable_lifting to tf.Variable, defaulting to True.
  - When it's False, the variable won't be lifted out of tf.function, thus it can be used as a tf.function-local variable: during each execution of the tf.function, the variable will be created and then disposed, similar to a local (that is, stack-allocated) variable in C/C++. Currently, experimental_enable_variable_lifting=False only works on non-XLA devices (for example, under @tf.function(jit_compile=False)).
TF SavedModel:
- Added fingerprint.pb to the SavedModel directory. The fingerprint.pb file is a protobuf containing the "fingerprint" of the SavedModel. See the RFC for more details regarding its design and properties.
TF pip:
- Windows CPU-builds for x86/x64 processors are now built, maintained, tested and released by a third party: Intel. Installing the windows-native pip packages for tensorflow or tensorflow-cpu would install Intel's tensorflow-intel package. These packages are provided as-is. Tensorflow will use reasonable efforts to maintain the availability and integrity of this pip package. There may be delays if the third party fails to release the pip package. For using TensorFlow GPU on Windows, you will need to install TensorFlow in WSL2.

Bug Fixes and Other Changes

tf.image
- Added an optional parameter return_index_map to tf.image.ssim which causes the returned value to be the local SSIM map instead of the global mean.
TF Core:
- tf.custom_gradient can now be applied to functions that accept "composite" tensors, such as tf.RaggedTensor, as inputs.
- Fix device placement issues related to datasets with ragged tensors of strings (i.e. variant encoded data with types not supported on GPU).
- experimental_follow_type_hints for tf.function has been deprecated. Please use input_signature or reduce_retracing to minimize retracing.
tf.SparseTensor:
- Introduced set_shape, which sets the static dense shape of the sparse tensor and has the same semantics as tf.Tensor.set_shape.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.11.0rc119 Oct 2022pre-releaseDeprecated

Please check the API documentation to find alternatives to the missing API. If you must call the deprecated API, please change your optimizer to the l…

Release 2.11.0

Breaking Changes

tf.keras.optimizers.Optimizer now points to the new Keras optimizer, and old optimizers have moved to the tf.keras.optimizers.legacy namespace. If you find your workflow failing due to this change, you may be facing one of the following issues:
- Checkpoint loading failure. The new optimizer handles optimizer state differently from the old optimizer, which simplies the logic of checkpoint saving/loading, but at the cost of breaking checkpoint backward compatibility in some cases. If you want to keep using an old checkpoint, please change your optimizer to tf.keras.optimizer.legacy.XXX (e.g. tf.keras.optimizer.legacy.Adam).
- TF1 compatibility. The new optimizer, tf.keras.optimizers.Optimizer, does not support TF1 any more, so please use the legacy optimizer tf.keras.optimizer.legacy.XXX. We highly recommend to migrate your workflow to TF2 for stable support and new features.
- Old optimizer API not found. The new optimizer, tf.keras.optimizers.Optimizer, has a different set of public APIs from the old optimizer. These API changes are mostly related to getting rid of slot variables and TF1 support. Please check the API documentation to find alternatives to the missing API. If you must call the deprecated API, please change your optimizer to the legacy optimizer.
- Learning rate schedule access. When using a LearningRateSchedule, The new optimizer's learning_rate property returns the current learning rate value instead of a LearningRateSchedule object as before. If you need to access the LearningRateSchedule object, please use optimizer._learning_rate.
- If you implemented a custom optimizer based on the old optimizer. Please set your optimizer to subclass tf.keras.optimizer.legacy.XXX. If you want to migrate to the new optimizer and find it does not support your optimizer, please file an issue in the Keras GitHub repo.
- Errors, such as Cannot recognize variable.... The new optimizer requires all optimizer variables to be created at the first apply_gradients() or minimize() call. If your workflow calls optimizer to update different parts of model in multiple stages, please call optimizer.build(model.trainable_variables) before the training loop.
- Timeout or performance loss. We don't anticipate this to happen, but if you see such issues, please use the legacy optimizer, and file an issue in the Keras GitHub repo.
The old Keras optimizer will never be deleted, but will not see any new feature additions. New optimizers (for example, tf.keras.optimizers.Adafactor) will only be implemented based on tf.keras.optimizers.Optimizer, the new base class.

Major Features and Improvements

tf.lite:
- New operations supported: tf.unsortedsegmentmin, tf.atan2 and tf.sign.
- Updates to existing operations:
  - tfl.mul now supports complex32 inputs.
tf.experimental.StructuredTensor
- Introduced tf.experimental.StructuredTensor, which provides a flexible and TensorFlow-native way to encode structured data such as protocol buffers or pandas dataframes.
tf.keras:
- Added a new get_metrics_result() method to tf.keras.models.Model.
  - Returns the current metrics values of the model as a dict.
- Added a new group normalization layer - tf.keras.layers.GroupNormalization.
- Added weight decay support for all Keras optimizers.
- Added Adafactor optimizer tf.keras.optimizers.Adafactor.
- Added warmstart_embedding_matrix to tf.keras.utils.
  - This utility can be used to warmstart an embeddings matrix, so you reuse previously-learned word embeddings when working with a new set of words which may include previously unseen words (the embedding vectors for unseen words will be randomly initialized).
tf.Variable:
- Added CompositeTensor as a baseclass to ResourceVariable.
  - This allows tf.Variables to be nested in tf.experimental.ExtensionTypes.
- Added a new constructor argument experimental_enable_variable_lifting to tf.Variable, defaulting to True.
  - When it's False, the variable won't be lifted out of tf.function, thus it can be used as a tf.function-local variable: during each execution of the tf.function, the variable will be created and then disposed, similar to a local (that is, stack-allocated) variable in C/C++. Currently, experimental_enable_variable_lifting=False only works on non-XLA devices (for example, under @tf.function(jit_compile=False)).
TF SavedModel:
- Added fingerprint.pb to the SavedModel directory. The fingerprint.pb file is a protobuf containing the "fingerprint" of the SavedModel. See the RFC for more details regarding its design and properties.
TF pip:
- Windows CPU-builds for x86/x64 processors are now built, maintained, tested and released by a third party: Intel. Installing the windows-native pip packages for tensorflow or tensorflow-cpu would install Intel's tensorflow-intel package. These packages are provided as-is. Tensorflow will use reasonable efforts to maintain the availability and integrity of this pip package. There may be delays if the third party fails to release the pip package. For using TensorFlow GPU on Windows, you will need to install TensorFlow in WSL2.

Bug Fixes and Other Changes

tf.image
- Added an optional parameter return_index_map to tf.image.ssim which causes the returned value to be the local SSIM map instead of the global mean.
TF Core:
- tf.custom_gradient can now be applied to functions that accept "composite" tensors, such as tf.RaggedTensor, as inputs.
- Fix device placement issues related to datasets with ragged tensors of strings (i.e. variant encoded data with types not supported on GPU).
- experimental_follow_type_hints for tf.function has been deprecated. Please use input_signature or reduce_retracing to minimize retracing.
tf.SparseTensor:
- Introduced set_shape, which sets the static dense shape of the sparse tensor and has the same semantics as tf.Tensor.set_shape.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.11.0rc018 Oct 2022pre-releaseDeprecated

Please check the API documentation to find alternatives to the missing API. If you must call the deprecated API, please change your optimizer to the l…

Release 2.11.0

Breaking Changes

tf.keras.optimizers.Optimizer now points to the new Keras optimizer, and old optimizers have moved to the tf.keras.optimizers.legacy namespace. If you find your workflow failing due to this change, you may be facing one of the following issues:
- Checkpoint loading failure. The new optimizer handles optimizer state differently from the old optimizer, which simplies the logic of checkpoint saving/loading, but at the cost of breaking checkpoint backward compatibility in some cases. If you want to keep using an old checkpoint, please change your optimizer to tf.keras.optimizer.legacy.XXX (e.g. tf.keras.optimizer.legacy.Adam).
- TF1 compatibility. The new optimizer, tf.keras.optimizers.Optimizer, does not support TF1 any more, so please use the legacy optimizer tf.keras.optimizer.legacy.XXX. We highly recommend to migrate your workflow to TF2 for stable support and new features.
- Old optimizer API not found. The new optimizer, tf.keras.optimizers.Optimizer, has a different set of public APIs from the old optimizer. These API changes are mostly related to getting rid of slot variables and TF1 support. Please check the API documentation to find alternatives to the missing API. If you must call the deprecated API, please change your optimizer to the legacy optimizer.
- Learning rate schedule access. When using a LearningRateSchedule, The new optimizer's learning_rate property returns the current learning rate value instead of a LearningRateSchedule object as before. If you need to access the LearningRateSchedule object, please use optimizer._learning_rate.
- If you implemented a custom optimizer based on the old optimizer. Please set your optimizer to subclass tf.keras.optimizer.legacy.XXX. If you want to migrate to the new optimizer and find it does not support your optimizer, please file an issue in the Keras GitHub repo.
- Errors, such as Cannot recognize variable.... The new optimizer requires all optimizer variables to be created at the first apply_gradients() or minimize() call. If your workflow calls optimizer to update different parts of model in multiple stages, please call optimizer.build(model.trainable_variables) before the training loop.
- Timeout or performance loss. We don't anticipate this to happen, but if you see such issues, please use the legacy optimizer, and file an issue in the Keras GitHub repo.
The old Keras optimizer will never be deleted, but will not see any new feature additions. New optimizers (for example, tf.keras.optimizers.Adafactor) will only be implemented based on tf.keras.optimizers.Optimizer, the new base class.

Major Features and Improvements

tf.lite:
- New operations supported: tf.unsortedsegmentmin, tf.atan2 and tf.sign.
- Updates to existing operations:
  - tfl.mul now supports complex32 inputs.
tf.experimental.StructuredTensor
- Introduced tf.experimental.StructuredTensor, which provides a flexible and TensorFlow-native way to encode structured data such as protocol buffers or pandas dataframes.
tf.keras:
- Added a new get_metrics_result() method to tf.keras.models.Model.
  - Returns the current metrics values of the model as a dict.
- Added a new group normalization layer - tf.keras.layers.GroupNormalization.
- Added weight decay support for all Keras optimizers.
- Added Adafactor optimizer tf.keras.optimizers.Adafactor.
- Added warmstart_embedding_matrix to tf.keras.utils.
  - This utility can be used to warmstart an embeddings matrix, so you reuse previously-learned word embeddings when working with a new set of words which may include previously unseen words (the embedding vectors for unseen words will be randomly initialized).
tf.Variable:
- Added CompositeTensor as a baseclass to ResourceVariable.
  - This allows tf.Variables to be nested in tf.experimental.ExtensionTypes.
- Added a new constructor argument experimental_enable_variable_lifting to tf.Variable, defaulting to True.
  - When it's False, the variable won't be lifted out of tf.function, thus it can be used as a tf.function-local variable: during each execution of the tf.function, the variable will be created and then disposed, similar to a local (that is, stack-allocated) variable in C/C++. Currently, experimental_enable_variable_lifting=False only works on non-XLA devices (for example, under @tf.function(jit_compile=False)).
TF SavedModel:
- Added fingerprint.pb to the SavedModel directory. The fingerprint.pb file is a protobuf containing the "fingerprint" of the SavedModel. See the RFC for more details regarding its design and properties.
TF pip:
- Windows CPU-builds for x86/x64 processors are now built, maintained, tested and released by a third party: Intel. Installing the windows-native pip packages for tensorflow or tensorflow-cpu would install Intel's tensorflow-intel package. These packages are provided as-is. Tensorflow will use reasonable efforts to maintain the availability and integrity of this pip package. There may be delays if the third party fails to release the pip package. For using TensorFlow GPU on Windows, you will need to install TensorFlow in WSL2.

Bug Fixes and Other Changes

tf.image
- Added an optional parameter return_index_map to tf.image.ssim which causes the returned value to be the local SSIM map instead of the global mean.
TF Core:
- tf.custom_gradient can now be applied to functions that accept "composite" tensors, such as tf.RaggedTensor, as inputs.
- Fix device placement issues related to datasets with ragged tensors of strings (i.e. variant encoded data with types not supported on GPU).
- experimental_follow_type_hints for tf.function has been deprecated. Please use input_signature or reduce_retracing to minimize retracing.
tf.SparseTensor:
- Introduced set_shape, which sets the static dense shape of the sparse tensor and has the same semantics as tf.Tensor.set_shape.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.10.06 Sept 2022Security

Fixes a CHECK failure in tf.reshape caused by overflows (CVE-2022-35934)

Release 2.10.0

Breaking Changes

Causal attention in keras.layers.Attention and keras.layers.AdditiveAttention is now specified in the call() method via the use_causal_mask argument (rather than in the constructor), for consistency with other layers.
Some files in tensorflow/python/training have been moved to tensorflow/python/tracking and tensorflow/python/checkpoint. Please update your imports accordingly, the old files will be removed in Release 2.11.
tf.keras.optimizers.experimental.Optimizer will graduate in Release 2.11, which means tf.keras.optimizers.Optimizer will be an alias of tf.keras.optimizers.experimental.Optimizer. The current tf.keras.optimizers.Optimizer will continue to be supported as tf.keras.optimizers.legacy.Optimizer, e.g.,tf.keras.optimizers.legacy.Adam. Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaptions. If you decide to keep using the old optimizer, please explicitly change your optimizer to tf.keras.optimizers.legacy.Optimizer.
RNG behavior change for tf.keras.initializers. Keras initializers will now use stateless random ops to generate random numbers.
- Both seeded and unseeded initializers will always generate the same values every time they are called (for a given variable shape). For unseeded initializers (seed=None), a random seed will be created and assigned at initializer creation (different initializer instances get different seeds).
- An unseeded initializer will raise a warning if it is reused (called) multiple times. This is because it would produce the same values each time, which may not be intended.

Deprecations

The C++ tensorflow::Code and tensorflow::Status will become aliases of respectively absl::StatusCode and absl::Status in some future release.
- Use tensorflow::OkStatus() instead of tensorflow::Status::OK().
- Stop constructing Status objects from tensorflow::error::Code.
- One MUST NOT access tensorflow::errors::Code fields. Accessing tensorflow::error::Code fields is fine.
  - Use the constructors such as tensorflow::errors:InvalidArgument to create status using an error code without accessing it.
  - Use the free functions such as tensorflow::errors::IsInvalidArgument if needed.
  - In the last resort, use e.g.static_cast<tensorflow::errors::Code>(error::Code::INVALID_ARGUMENT) or static_cast<int>(code) for comparisons.
tensorflow::StatusOr will also become in the future alias to absl::StatusOr, so use StatusOr::value instead of StatusOr::ConsumeValueOrDie.

Major Features and Improvements

tf.lite:
- New operations supported:
  - tflite SelectV2 now supports 5D.
  - tf.einsum is supported with multiple unknown shapes.
  - tf.unsortedsegmentprod op is supported.
  - tf.unsortedsegmentmax op is supported.
  - tf.unsortedsegmentsum op is supported.
- Updates to existing operations:
  - tfl.scatter_nd now supports I1 for update arg.
- Upgrade Flatbuffers v2.0.5 from v1.12.0
tf.keras:
- EinsumDense layer is moved from experimental to core. Its import path is moved from tf.keras.layers.experimental.EinsumDense to tf.keras.layers.EinsumDense.
- Added tf.keras.utils.audio_dataset_from_directory utility to easily generate audio classification datasets from directories of .wav files.
- Added subset="both" support in tf.keras.utils.image_dataset_from_directory,tf.keras.utils.text_dataset_from_directory, and audio_dataset_from_directory, to be used with the validation_split argument, for returning both dataset splits at once, as a tuple.
- Added tf.keras.utils.split_dataset utility to split a Dataset object or a list/tuple of arrays into two Dataset objects (e.g. train/test).
- Added step granularity to BackupAndRestore callback for handling distributed training failures & restarts. The training state can now be restored at the exact epoch and step at which it was previously saved before failing.
- Added tf.keras.dtensor.experimental.optimizers.AdamW. This optimizer is similar as the existing keras.optimizers.experimental.AdamW, and works in the DTensor training use case.
- Improved masking support for tf.keras.layers.MultiHeadAttention.
  - Implicit masks for query, key and value inputs will automatically be used to compute a correct attention mask for the layer. These padding masks will be combined with any attention_mask passed in directly when calling the layer. This can be used with tf.keras.layers.Embedding with mask_zero=True to automatically infer a correct padding mask.
  - Added a use_causal_mask call time arugment to the layer. Passing use_causal_mask=True will compute a causal attention mask, and optionally combine it with any attention_mask passed in directly when calling the layer.
- Added ignore_class argument in the loss SparseCategoricalCrossentropy and metrics IoU and MeanIoU, to specify a class index to be ignored during loss/metric computation (e.g. a background/void class).
- Added tf.keras.models.experimental.SharpnessAwareMinimization. This class implements the sharpness-aware minimization technique, which boosts model performance on various tasks, e.g., ResNet on image classification.
tf.data:
- Added support for cross-trainer data caching in tf.data service. This saves computation resources when concurrent training jobs train from the same dataset. See (https://www.tensorflow.org/api_docs/python/tf/data/experimental/service#sharing_tfdata_service_with_concurrent_trainers) for more details.
- Added dataset_id to tf.data.experimental.service.register_dataset. If provided, tf.data service will use the provided ID for the dataset. If the dataset ID already exists, no new dataset will be registered. This is useful if multiple training jobs need to use the same dataset for training. In this case, users should call register_dataset with the same dataset_id.
- Added a new field, inject_prefetch, to tf.data.experimental.OptimizationOptions. If it is set to True,tf.data will now automatically add a prefetch transformation to datasets that end in synchronous transformations. This enables data generation to be overlapped with data consumption. This may cause a small increase in memory usage due to buffering. To enable this behavior, set inject_prefetch=True in tf.data.experimental.OptimizationOptions.
- Added a new value to tf.data.Options.autotune.autotune_algorithm: STAGE_BASED. If the autotune algorithm is set to STAGE_BASED, then it runs a new algorithm that can get the same performance with lower CPU/memory usage.
- Added tf.data.experimental.from_list, a new API for creating Datasets from lists of elements.
tf.distribute:
- Added tf.distribute.experimental.PreemptionCheckpointHandler to handle worker preemption/maintenance and cluster-wise consistent error reporting for tf.distribute.MultiWorkerMirroredStrategy. Specifically, for the type of interruption with advance notice, it automatically saves a checkpoint, exits the program without raising an unrecoverable error, and restores the progress when training restarts.
tf.math:
- Added tf.math.approx_max_k and tf.math.approx_min_k which are the optimized alternatives to tf.math.top_k on TPU. The performance difference range from 8 to 100 times depending on the size of k. When running on CPU and GPU, a non-optimized XLA kernel is used.
tf.train:
- Added tf.train.TrackableView which allows users to inspect the TensorFlow Trackable object (e.g. tf.Module, Keras Layers and models).
tf.vectorized_map:
- Added an optional parameter: warn. This parameter controls whether or not warnings will be printed when operations in the provided fn fall back to a while loop.
XLA:
- MWMS is now compilable with XLA.
- Compute Library for the Arm® Architecture (ACL) is supported for aarch64 CPU XLA runtime
CPU performance optimizations:
- x86 CPUs: oneDNN bfloat16 auto-mixed precision grappler graph optimization pass has been renamed from auto_mixed_precision_mkl to auto_mixed_precision_onednn_bfloat16. See example usage here.
- aarch64 CPUs: Experimental performance optimizations from Compute Library for the Arm® Architecture (ACL) are available through oneDNN in the default Linux aarch64 package (pip install tensorflow).
  - The optimizations are disabled by default.
  - Set the environment variable TF_ENABLE_ONEDNN_OPTS=1 to enable the optimizations. Setting the variable to 0 or unsetting it will disable the optimizations.
  - These optimizations can yield slightly different numerical results from when they are off due to floating-point round-off errors from different computation approaches and orders.
  - To verify that the optimizations are on, look for a message with "oneDNN custom operations are on" in the log. If the exact phrase is not there, it means they are off.

Bug Fixes and Other Changes

New argument experimental_device_ordinal in LogicalDeviceConfiguration to control the order of logical devices. (GPU only)
tf.keras:
- Changed the TensorBoard tag names produced by the tf.keras.callbacks.TensorBoard callback, so that summaries logged automatically for model weights now include either a /histogram or /image suffix in their tag names, in order to prevent tag name collisions across summary types.
When running on GPU (with cuDNN version 7.6.3 or later),tf.nn.depthwise_conv2d backprop to filter (and therefore also tf.keras.layers.DepthwiseConv2D) now operate deterministically (and tf.errors.UnimplementedError is no longer thrown) when op-determinism has been enabled via tf.config.experimental.enable_op_determinism. This closes issue 47174.
tf.random
- Added tf.random.experimental.stateless_shuffle, a stateless version of tf.random.shuffle.

Security

Fixes a CHECK failure in tf.reshape caused by overflows (CVE-2022-35934)
Fixes a CHECK failure in SobolSample caused by missing validation (CVE-2022-35935)
Fixes an OOB read in Gather_nd op in TF Lite (CVE-2022-35937)
Fixes a CHECK failure in TensorListReserve caused by missing validation (CVE-2022-35960)
Fixes an OOB write in Scatter_nd op in TF Lite (CVE-2022-35939)
Fixes an integer overflow in RaggedRangeOp (CVE-2022-35940)
Fixes a CHECK failure in AvgPoolOp (CVE-2022-35941)
Fixes a CHECK failures in UnbatchGradOp (CVE-2022-35952)
Fixes a segfault TFLite converter on per-channel quantized transposed convolutions (CVE-2022-36027)
Fixes a CHECK failures in AvgPool3DGrad (CVE-2022-35959)
Fixes a CHECK failures in FractionalAvgPoolGrad (CVE-2022-35963)
Fixes a segfault in BlockLSTMGradV2 (CVE-2022-35964)
Fixes a segfault in LowerBound and UpperBound (CVE-2022-35965)
Fixes a segfault in QuantizedAvgPool (CVE-2022-35966)
Fixes a segfault in QuantizedAdd (CVE-2022-35967)
Fixes a CHECK fail in AvgPoolGrad (CVE-2022-35968)
Fixes a CHECK fail in Conv2DBackpropInput (CVE-2022-35969)
Fixes a segfault in QuantizedInstanceNorm (CVE-2022-35970)
Fixes a CHECK fail in FakeQuantWithMinMaxVars (CVE-2022-35971)
Fixes a segfault in Requantize (CVE-2022-36017)
Fixes a segfault in QuantizedBiasAdd (CVE-2022-35972)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsPerChannel (CVE-2022-36019)
Fixes a segfault in QuantizedMatMul (CVE-2022-35973)
Fixes a segfault in QuantizeDownAndShrinkRange (CVE-2022-35974)
Fixes segfaults in QuantizedRelu and QuantizedRelu6 (CVE-2022-35979)
Fixes a CHECK fail in FractionalMaxPoolGrad (CVE-2022-35981)
Fixes a CHECK fail in RaggedTensorToVariant (CVE-2022-36018)
Fixes a CHECK fail in QuantizeAndDequantizeV3 (CVE-2022-36026)
Fixes a segfault in SparseBincount (CVE-2022-35982)
Fixes a CHECK fail in Save and SaveSlices (CVE-2022-35983)
Fixes a CHECK fail in ParameterizedTruncatedNormal (CVE-2022-35984)
Fixes a CHECK fail in LRNGrad (CVE-2022-35985)
Fixes a segfault in RaggedBincount (CVE-2022-35986)
Fixes a CHECK fail in DenseBincount (CVE-2022-35987)
Fixes a CHECK fail in tf.linalg.matrix_rank (CVE-2022-35988)
Fixes a CHECK fail in MaxPool (CVE-2022-35989)
Fixes a CHECK fail in Conv2DBackpropInput (CVE-2022-35999)
Fixes a CHECK fail in EmptyTensorList (CVE-2022-35998)
Fixes a CHECK fail in tf.sparse.cross (CVE-2022-35997)
Fixes a floating point exception in Conv2D (CVE-2022-35996)
Fixes a CHECK fail in AudioSummaryV2 (CVE-2022-35995)
Fixes a CHECK fail in CollectiveGather (CVE-2022-35994)
Fixes a CHECK fail in SetSize (CVE-2022-35993)
Fixes a CHECK fail in TensorListFromTensor (CVE-2022-35992)
Fixes a CHECK fail in TensorListScatter and TensorListScatterV2 (CVE-2022-35991)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsPerChannelGradient (CVE-2022-35990)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsGradient (CVE-2022-36005)
Fixes a CHECK fail in tf.random.gamma (CVE-2022-36004)
Fixes a CHECK fail in RandomPoissonV2 (CVE-2022-36003)
Fixes a CHECK fail in Unbatch (CVE-2022-36002)
Fixes a CHECK fail in DrawBoundingBoxes (CVE-2022-36001)
Fixes a CHECK fail in Eig (CVE-2022-36000)
Fixes a null dereference on MLIR on empty function attributes (CVE-2022-36011)
Fixes an assertion failure on MLIR empty edge names (CVE-2022-36012)
Fixes a null-dereference in mlir::tfg::GraphDefImporter::ConvertNodeDef (CVE-2022-36013)
Fixes a null-dereference in mlir::tfg::TFOp::nameAttr (CVE-2022-36014)
Fixes an integer overflow in math ops (CVE-2022-36015)
Fixes a CHECK-fail in tensorflow::full_type::SubstituteFromAttrs (CVE-2022-36016)
Fixes an OOB read in Gather_nd op in TF Lite Micro (CVE-2022-35938)

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Abolfazl Shahbazi, Adam Lanicek, Amin Benarieb, andreii, Andrew Fitzgibbon, Andrew Goodbody, angerson, Ashiq Imran, Aurélien Geron, Banikumar Maiti (Intel Aipg), Ben Barsdell, Ben Mares, bhack, Bhavani Subramanian, Bill Schnurr, Byungsoo Oh, Chandra Sr Potula, Chengji Yao, Chris Carpita, Christopher Bate, chunduriv, Cliff Woolley, Cliffs Dover, Cloud Han, Code-Review-Doctor, DEKHTIARJonathan, Deven Desai, Djacon, Duncan Riach, fedotoff, fo40225, Frederic Bastien, gadagashwini, Gauri1 Deshpande, guozhong.zhuang, Hui Peng, James Gerity, Jason Furmanek, Jonathan Dekhtiar, Jueon Park, Kaixi Hou, Kanvi Khanna, Keith Smiley, Koan-Sin Tan, Kulin Seth, kushanam, Learning-To-Play, Li-Wen Chang, lipracer, liuyuanqiang, Louis Sugy, Lucas David, Lukas Geiger, Mahmoud Abuzaina, Marius Brehler, Maxiwell S. Garcia, mdfaijul, Meenakshi Venkataraman, Michal Szutenberg, Michele Di Giorgio, Mickaël Salamin, Nathan John Sircombe, Nathan Luehr, Neil Girdhar, Nils Reichardt, Nishidha Panpaliya, Nobuo Tsukamoto, Om Thakkar, Patrice Vignola, Philipp Hack, Pooya Jannaty, Prianka Liz Kariat, pshiko, Rajeshwar Reddy T, rdl4199, Rohit Santhanam, Rsanthanam-Amd, Sachin Muradi, Saoirse Stewart, Serge Panev, Shu Wang, Srinivasan Narayanamoorthy, Stella Stamenova, Stephan Hartmann, Sunita Nadampalli, synandi, Tamas Bela Feher, Tao Xu, Thibaut Goetghebuer-Planchon, Trevor Morris, Xiaoming (Jason) Cui, Yimei Sun, Yong Tang, Yuanqiang Liu, Yulv-Git, Zhoulong Jiang, ZihengJiang

Open source →

2.10.0rc329 Aug 2022pre-releaseDeprecated

Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaption…

Release 2.10.0

Breaking Changes

Causal attention in keras.layers.Attention and keras.layers.AdditiveAttention is now specified in the call() method via the use_causal_mask argument (rather than in the constructor), for consistency with other layers.
Some files in tensorflow/python/training have been moved to tensorflow/python/tracking and tensorflow/python/checkpoint. Please update your imports accordingly, the old files will be removed in Release 2.11.
tf.keras.optimizers.experimental.Optimizer will graduate in Release 2.11, which means tf.keras.optimizers.Optimizer will be an alias of tf.keras.optimizers.experimental.Optimizer. The current tf.keras.optimizers.Optimizer will continue to be supported as tf.keras.optimizers.legacy.Optimizer, e.g.,tf.keras.optimizers.legacy.Adam. Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaptions. If you decide to keep using the old optimizer, please explicitly change your optimizer to tf.keras.optimizers.legacy.Optimizer.
RNG behavior change for tf.keras.initializers. Keras initializers will now use stateless random ops to generate random numbers.
- Both seeded and unseeded initializers will always generate the same values every time they are called (for a given variable shape). For unseeded initializers (seed=None), a random seed will be created and assigned at initializer creation (different initializer instances get different seeds).
- An unseeded initializer will raise a warning if it is reused (called) multiple times. This is because it would produce the same values each time, which may not be intended.

Deprecations

The C++ tensorflow::Code and tensorflow::Status will become aliases of respectively absl::StatusCode and absl::Status in some future release.
- Use tensorflow::OkStatus() instead of tensorflow::Status::OK().
- Stop constructing Status objects from tensorflow::error::Code.
- One MUST NOT access tensorflow::errors::Code fields. Accessing tensorflow::error::Code fields is fine.
  - Use the constructors such as tensorflow::errors:InvalidArgument to create status using an error code without accessing it.
  - Use the free functions such as tensorflow::errors::IsInvalidArgument if needed.
  - In the last resort, use e.g.static_cast<tensorflow::errors::Code>(error::Code::INVALID_ARGUMENT) or static_cast<int>(code) for comparisons.
tensorflow::StatusOr will also become in the future alias to absl::StatusOr, so use StatusOr::value instead of StatusOr::ConsumeValueOrDie.

Major Features and Improvements

tf.lite:
- New operations supported:
  - tflite SelectV2 now supports 5D.
  - tf.einsum is supported with multiple unknown shapes.
  - tf.unsortedsegmentprod op is supported.
  - tf.unsortedsegmentmax op is supported.
  - tf.unsortedsegmentsum op is supported.
- Updates to existing operations:
  - tfl.scatter_nd now supports I1 for update arg.
- Upgrade Flatbuffers v2.0.5 from v1.12.0
tf.keras:
- EinsumDense layer is moved from experimental to core. Its import path is moved from tf.keras.layers.experimental.EinsumDense to tf.keras.layers.EinsumDense.
- Added tf.keras.utils.audio_dataset_from_directory utility to easily generate audio classification datasets from directories of .wav files.
- Added subset="both" support in tf.keras.utils.image_dataset_from_directory,tf.keras.utils.text_dataset_from_directory, and audio_dataset_from_directory, to be used with the validation_split argument, for returning both dataset splits at once, as a tuple.
- Added tf.keras.utils.split_dataset utility to split a Dataset object or a list/tuple of arrays into two Dataset objects (e.g. train/test).
- Added step granularity to BackupAndRestore callback for handling distributed training failures & restarts. The training state can now be restored at the exact epoch and step at which it was previously saved before failing.
- Added tf.keras.dtensor.experimental.optimizers.AdamW. This optimizer is similar as the existing keras.optimizers.experimental.AdamW, and works in the DTensor training use case.
- Improved masking support for tf.keras.layers.MultiHeadAttention.
  - Implicit masks for query, key and value inputs will automatically be used to compute a correct attention mask for the layer. These padding masks will be combined with any attention_mask passed in directly when calling the layer. This can be used with tf.keras.layers.Embedding with mask_zero=True to automatically infer a correct padding mask.
  - Added a use_causal_mask call time arugment to the layer. Passing use_causal_mask=True will compute a causal attention mask, and optionally combine it with any attention_mask passed in directly when calling the layer.
- Added ignore_class argument in the loss SparseCategoricalCrossentropy and metrics IoU and MeanIoU, to specify a class index to be ignored during loss/metric computation (e.g. a background/void class).
- Added tf.keras.models.experimental.SharpnessAwareMinimization. This class implements the sharpness-aware minimization technique, which boosts model performance on various tasks, e.g., ResNet on image classification.
tf.data:
- Added support for cross-trainer data caching in tf.data service. This saves computation resources when concurrent training jobs train from the same dataset. See (https://www.tensorflow.org/api_docs/python/tf/data/experimental/service#sharing_tfdata_service_with_concurrent_trainers) for more details.
- Added dataset_id to tf.data.experimental.service.register_dataset. If provided, tf.data service will use the provided ID for the dataset. If the dataset ID already exists, no new dataset will be registered. This is useful if multiple training jobs need to use the same dataset for training. In this case, users should call register_dataset with the same dataset_id.
- Added a new field, inject_prefetch, to tf.data.experimental.OptimizationOptions. If it is set to True,tf.data will now automatically add a prefetch transformation to datasets that end in synchronous transformations. This enables data generation to be overlapped with data consumption. This may cause a small increase in memory usage due to buffering. To enable this behavior, set inject_prefetch=True in tf.data.experimental.OptimizationOptions.
- Added a new value to tf.data.Options.autotune.autotune_algorithm: STAGE_BASED. If the autotune algorithm is set to STAGE_BASED, then it runs a new algorithm that can get the same performance with lower CPU/memory usage.
- Added tf.data.experimental.from_list, a new API for creating Datasets from lists of elements.
tf.distribute:
- Added tf.distribute.experimental.PreemptionCheckpointHandler to handle worker preemption/maintenance and cluster-wise consistent error reporting for tf.distribute.MultiWorkerMirroredStrategy. Specifically, for the type of interruption with advance notice, it automatically saves a checkpoint, exits the program without raising an unrecoverable error, and restores the progress when training restarts.
tf.math:
- Added tf.math.approx_max_k and tf.math.approx_min_k which are the optimized alternatives to tf.math.top_k on TPU. The performance difference range from 8 to 100 times depending on the size of k. When running on CPU and GPU, a non-optimized XLA kernel is used.
tf.train:
- Added tf.train.TrackableView which allows users to inspect the TensorFlow Trackable object (e.g. tf.Module, Keras Layers and models).
tf.vectorized_map:
- Added an optional parameter: warn. This parameter controls whether or not warnings will be printed when operations in the provided fn fall back to a while loop.
XLA:
- MWMS is now compilable with XLA.
- Compute Library for the Arm® Architecture (ACL) is supported for aarch64 CPU XLA runtime
CPU performance optimizations:
- x86 CPUs: oneDNN bfloat16 auto-mixed precision grappler graph optimization pass has been renamed from auto_mixed_precision_mkl to auto_mixed_precision_onednn_bfloat16. See example usage here.
- aarch64 CPUs: Experimental performance optimizations from Compute Library for the Arm® Architecture (ACL) are available through oneDNN in the default Linux aarch64 package (pip install tensorflow).
  - The optimizations are disabled by default.
  - Set the environment variable TF_ENABLE_ONEDNN_OPTS=1 to enable the optimizations. Setting the variable to 0 or unsetting it will disable the optimizations.
  - These optimizations can yield slightly different numerical results from when they are off due to floating-point round-off errors from different computation approaches and orders.
  - To verify that the optimizations are on, look for a message with "oneDNN custom operations are on" in the log. If the exact phrase is not there, it means they are off.

Bug Fixes and Other Changes

New argument experimental_device_ordinal in LogicalDeviceConfiguration to control the order of logical devices. (GPU only)
tf.keras:
- Changed the TensorBoard tag names produced by the tf.keras.callbacks.TensorBoard callback, so that summaries logged automatically for model weights now include either a /histogram or /image suffix in their tag names, in order to prevent tag name collisions across summary types.
When running on GPU (with cuDNN version 7.6.3 or later),tf.nn.depthwise_conv2d backprop to filter (and therefore also tf.keras.layers.DepthwiseConv2D) now operate deterministically (and tf.errors.UnimplementedError is no longer thrown) when op-determinism has been enabled via tf.config.experimental.enable_op_determinism. This closes issue 47174.
tf.random
- Added tf.random.experimental.stateless_shuffle, a stateless version of tf.random.shuffle.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.10.0rc223 Aug 2022pre-releaseDeprecated

Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaption…

Release 2.10.0

Breaking Changes

Causal attention in keras.layers.Attention and keras.layers.AdditiveAttention is now specified in the call() method via the use_causal_mask argument (rather than in the constructor), for consistency with other layers.
Some files in tensorflow/python/training have been moved to tensorflow/python/tracking and tensorflow/python/checkpoint. Please update your imports accordingly, the old files will be removed in Release 2.11.
tf.keras.optimizers.experimental.Optimizer will graduate in Release 2.11, which means tf.keras.optimizers.Optimizer will be an alias of tf.keras.optimizers.experimental.Optimizer. The current tf.keras.optimizers.Optimizer will continue to be supported as tf.keras.optimizers.legacy.Optimizer, e.g.,tf.keras.optimizers.legacy.Adam. Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaptions. If you decide to keep using the old optimizer, please explicitly change your optimizer to tf.keras.optimizers.legacy.Optimizer.
RNG behavior change for tf.keras.initializers. Keras initializers will now use stateless random ops to generate random numbers.
- Both seeded and unseeded initializers will always generate the same values every time they are called (for a given variable shape). For unseeded initializers (seed=None), a random seed will be created and assigned at initializer creation (different initializer instances get different seeds).
- An unseeded initializer will raise a warning if it is reused (called) multiple times. This is because it would produce the same values each time, which may not be intended.

Major Features and Improvements

tf.lite:
- New operations supported:
  - tflite SelectV2 now supports 5D.
  - tf.einsum is supported with multiple unknown shapes.
  - tf.unsortedsegmentprod op is supported.
  - tf.unsortedsegmentmax op is supported.
  - tf.unsortedsegmentsum op is supported.
- Updates to existing operations:
  - tfl.scatter_nd now supports I1 for update arg.
- Upgrade Flatbuffers v2.0.5 from v1.12.0
tf.keras:
- EinsumDense layer is moved from experimental to core. Its import path is moved from tf.keras.layers.experimental.EinsumDense to tf.keras.layers.EinsumDense.
- Added tf.keras.utils.audio_dataset_from_directory utility to easily generate audio classification datasets from directories of .wav files.
- Added subset="both" support in tf.keras.utils.image_dataset_from_directory,tf.keras.utils.text_dataset_from_directory, and audio_dataset_from_directory, to be used with the validation_split argument, for returning both dataset splits at once, as a tuple.
- Added tf.keras.utils.split_dataset utility to split a Dataset object or a list/tuple of arrays into two Dataset objects (e.g. train/test).
- Added step granularity to BackupAndRestore callback for handling distributed training failures & restarts. The training state can now be restored at the exact epoch and step at which it was previously saved before failing.
- Added tf.keras.dtensor.experimental.optimizers.AdamW. This optimizer is similar as the existing keras.optimizers.experimental.AdamW, and works in the DTensor training use case.
- Improved masking support for tf.keras.layers.MultiHeadAttention.
  - Implicit masks for query, key and value inputs will automatically be used to compute a correct attention mask for the layer. These padding masks will be combined with any attention_mask passed in directly when calling the layer. This can be used with tf.keras.layers.Embedding with mask_zero=True to automatically infer a correct padding mask.
  - Added a use_causal_mask call time arugment to the layer. Passing use_causal_mask=True will compute a causal attention mask, and optionally combine it with any attention_mask passed in directly when calling the layer.
- Added ignore_class argument in the loss SparseCategoricalCrossentropy and metrics IoU and MeanIoU, to specify a class index to be ignored during loss/metric computation (e.g. a background/void class).
- Added tf.keras.models.experimental.SharpnessAwareMinimization. This class implements the sharpness-aware minimization technique, which boosts model performance on various tasks, e.g., ResNet on image classification.
tf.data:
- Added support for cross-trainer data caching in tf.data service. This saves computation resources when concurrent training jobs train from the same dataset. See (https://www.tensorflow.org/api_docs/python/tf/data/experimental/service#sharing_tfdata_service_with_concurrent_trainers) for more details.
- Added dataset_id to tf.data.experimental.service.register_dataset. If provided, tf.data service will use the provided ID for the dataset. If the dataset ID already exists, no new dataset will be registered. This is useful if multiple training jobs need to use the same dataset for training. In this case, users should call register_dataset with the same dataset_id.
- Added a new field, inject_prefetch, to tf.data.experimental.OptimizationOptions. If it is set to True,tf.data will now automatically add a prefetch transformation to datasets that end in synchronous transformations. This enables data generation to be overlapped with data consumption. This may cause a small increase in memory usage due to buffering. To enable this behavior, set inject_prefetch=True in tf.data.experimental.OptimizationOptions.
- Added a new value to tf.data.Options.autotune.autotune_algorithm: STAGE_BASED. If the autotune algorithm is set to STAGE_BASED, then it runs a new algorithm that can get the same performance with lower CPU/memory usage.
- Added tf.data.experimental.from_list, a new API for creating Datasets from lists of elements.
tf.distribute:
- Added tf.distribute.experimental.PreemptionCheckpointHandler to handle worker preemption/maintenance and cluster-wise consistent error reporting for tf.distribute.MultiWorkerMirroredStrategy. Specifically, for the type of interruption with advance notice, it automatically saves a checkpoint, exits the program without raising an unrecoverable error, and restores the progress when training restarts.
tf.math:
- Added tf.math.approx_max_k and tf.math.approx_min_k which are the optimized alternatives to tf.math.top_k on TPU. The performance difference range from 8 to 100 times depending on the size of k. When running on CPU and GPU, a non-optimized XLA kernel is used.
tf.train:
- Added tf.train.TrackableView which allows users to inspect the TensorFlow Trackable object (e.g. tf.Module, Keras Layers and models).
tf.vectorized_map:
- Added an optional parameter: warn. This parameter controls whether or not warnings will be printed when operations in the provided fn fall back to a while loop.
XLA:
- MWMS is now compilable with XLA.
oneDNN CPU performance optimizations:
- x86 CPUs: oneDNN bfloat16 auto-mixed precision grappler graph optimization pass has been renamed from auto_mixed_precision_mkl to auto_mixed_precision_onednn_bfloat16. See example usage here.
- aarch64 CPUs: Experimental Arm Compute Library (ACL) CPU performance optimizations through oneDNN are available in the default Linux aarch64 package (pip install tensorflow).
  - The optimizations are disabled by default.
  - Set the environment variable TF_ENABLE_ONEDNN_OPTS=1 to enable the optimizations. Setting the variable to 0 or unsetting it will disable the optimizations.
  - These optimizations can yield slightly different numerical results from when they are off due to floating-point round-off errors from different computation approaches and orders.
  - To verify that the optimizations are on, look for a message with "oneDNN custom operations are on" in the log. If the exact phrase is not there, it means they are off.

Bug Fixes and Other Changes

New argument experimental_device_ordinal in LogicalDeviceConfiguration to control the order of logical devices. (GPU only)
tf.keras:
- Changed the TensorBoard tag names produced by the tf.keras.callbacks.TensorBoard callback, so that summaries logged automatically for model weights now include either a /histogram or /image suffix in their tag names, in order to prevent tag name collisions across summary types.
When running on GPU (with cuDNN version 7.6.3 or later),tf.nn.depthwise_conv2d backprop to filter (and therefore also tf.keras.layers.DepthwiseConv2D) now operate deterministically (and tf.errors.UnimplementedError is no longer thrown) when op-determinism has been enabled via tf.config.experimental.enable_op_determinism. This closes issue 47174.
tf.random
- Added tf.random.experimental.stateless_shuffle, a stateless version of tf.random.shuffle.

Deprecations

The C++ tensorflow::Code and tensorflow::Status will become aliases of respectively absl::StatusCode and absl::Status in some future release.
- Use tensorflow::OkStatus() instead of tensorflow::Status::OK().
- Stop constructing Status objects from tensorflow::error::Code.
- One MUST NOT access tensorflow::errors::Code fields. Accessing tensorflow::error::Code fields is fine.
  - Use the constructors such as tensorflow::errors:InvalidArgument to create status using an error code without accessing it.
  - Use the free functions such as tensorflow::errors::IsInvalidArgument if needed.
  - In the last resort, use e.g.static_cast<tensorflow::errors::Code>(error::Code::INVALID_ARGUMENT) or static_cast<int>(code) for comparisons.
tensorflow::StatusOr will also become in the future alias to absl::StatusOr, so use StatusOr::value instead of StatusOr::ConsumeValueOrDie.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.10.0rc115 Aug 2022pre-releaseDeprecated

Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaption…

Release 2.10.0

Breaking Changes

Causal attention in keras.layers.Attention and keras.layers.AdditiveAttention is now specified in the call() method via the use_causal_mask argument (rather than in the constructor), for consistency with other layers.
Some files in tensorflow/python/training have been moved to tensorflow/python/tracking and tensorflow/python/checkpoint. Please update your imports accordingly, the old files will be removed in Release 2.11.
tf.keras.optimizers.experimental.Optimizer will graduate in Release 2.11, which means tf.keras.optimizers.Optimizer will be an alias of tf.keras.optimizers.experimental.Optimizer. The current tf.keras.optimizers.Optimizer will continue to be supported as tf.keras.optimizers.legacy.Optimizer, e.g.,tf.keras.optimizers.legacy.Adam. Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaptions. If you decide to keep using the old optimizer, please explicitly change your optimizer to tf.keras.optimizers.legacy.Optimizer.
RNG behavior change for tf.keras.initializers. Keras initializers will now use stateless random ops to generate random numbers.
- Both seeded and unseeded initializers will always generate the same values every time they are called (for a given variable shape). For unseeded initializers (seed=None), a random seed will be created and assigned at initializer creation (different initializer instances get different seeds).
- An unseeded initializer will raise a warning if it is reused (called) multiple times. This is because it would produce the same values each time, which may not be intended.

Major Features and Improvements

tf.lite:
- New operations supported:
  - tflite SelectV2 now supports 5D.
  - tf.einsum is supported with multiple unknown shapes.
  - tf.unsortedsegmentprod op is supported.
  - tf.unsortedsegmentmax op is supported.
  - tf.unsortedsegmentsum op is supported.
- Updates to existing operations:
  - tfl.scatter_nd now supports I1 for update arg.
- Upgrade Flatbuffers v2.0.5 from v1.12.0
tf.keras:
- EinsumDense layer is moved from experimental to core. Its import path is moved from tf.keras.layers.experimental.EinsumDense to tf.keras.layers.EinsumDense.
- Added tf.keras.utils.audio_dataset_from_directory utility to easily generate audio classification datasets from directories of .wav files.
- Added subset="both" support in tf.keras.utils.image_dataset_from_directory,tf.keras.utils.text_dataset_from_directory, and audio_dataset_from_directory, to be used with the validation_split argument, for returning both dataset splits at once, as a tuple.
- Added tf.keras.utils.split_dataset utility to split a Dataset object or a list/tuple of arrays into two Dataset objects (e.g. train/test).
- Added step granularity to BackupAndRestore callback for handling distributed training failures & restarts. The training state can now be restored at the exact epoch and step at which it was previously saved before failing.
- Added tf.keras.dtensor.experimental.optimizers.AdamW. This optimizer is similar as the existing keras.optimizers.experimental.AdamW, and works in the DTensor training use case.
- Improved masking support for tf.keras.layers.MultiHeadAttention.
  - Implicit masks for query, key and value inputs will automatically be used to compute a correct attention mask for the layer. These padding masks will be combined with any attention_mask passed in directly when calling the layer. This can be used with tf.keras.layers.Embedding with mask_zero=True to automatically infer a correct padding mask.
  - Added a use_causal_mask call time arugment to the layer. Passing use_causal_mask=True will compute a causal attention mask, and optionally combine it with any attention_mask passed in directly when calling the layer.
- Added ignore_class argument in the loss SparseCategoricalCrossentropy and metrics IoU and MeanIoU, to specify a class index to be ignored during loss/metric computation (e.g. a background/void class).
- Added tf.keras.models.experimental.SharpnessAwareMinimization. This class implements the sharpness-aware minimization technique, which boosts model performance on various tasks, e.g., ResNet on image classification.
tf.data:
- Added support for cross-trainer data caching in tf.data service. This saves computation resources when concurrent training jobs train from the same dataset. See (https://www.tensorflow.org/api_docs/python/tf/data/experimental/service#sharing_tfdata_service_with_concurrent_trainers) for more details.
- Added dataset_id to tf.data.experimental.service.register_dataset. If provided, tf.data service will use the provided ID for the dataset. If the dataset ID already exists, no new dataset will be registered. This is useful if multiple training jobs need to use the same dataset for training. In this case, users should call register_dataset with the same dataset_id.
- Added a new field, inject_prefetch, to tf.data.experimental.OptimizationOptions. If it is set to True,tf.data will now automatically add a prefetch transformation to datasets that end in synchronous transformations. This enables data generation to be overlapped with data consumption. This may cause a small increase in memory usage due to buffering. To enable this behavior, set inject_prefetch=True in tf.data.experimental.OptimizationOptions.
- Added a new value to tf.data.Options.autotune.autotune_algorithm: STAGE_BASED. If the autotune algorithm is set to STAGE_BASED, then it runs a new algorithm that can get the same performance with lower CPU/memory usage.
- Added tf.data.experimental.from_list, a new API for creating Datasets from lists of elements.
tf.distribute:
- Added tf.distribute.experimental.PreemptionCheckpointHandler to handle worker preemption/maintenance and cluster-wise consistent error reporting for tf.distribute.MultiWorkerMirroredStrategy. Specifically, for the type of interruption with advance notice, it automatically saves a checkpoint, exits the program without raising an unrecoverable error, and restores the progress when training restarts.
tf.math:
- Added tf.math.approx_max_k and tf.math.approx_min_k which are the optimized alternatives to tf.math.top_k on TPU. The performance difference range from 8 to 100 times depending on the size of k. When running on CPU and GPU, a non-optimized XLA kernel is used.
tf.train:
- Added tf.train.TrackableView which allows users to inspect the TensorFlow Trackable object (e.g. tf.Module, Keras Layers and models).
tf.vectorized_map:
- Added an optional parameter: warn. This parameter controls whether or not warnings will be printed when operations in the provided fn fall back to a while loop.
XLA:
- MWMS is now compilable with XLA.
oneDNN CPU performance optimizations:
- x86 CPUs: oneDNN bfloat16 auto-mixed precision grappler graph optimization pass has been renamed from auto_mixed_precision_mkl to auto_mixed_precision_onednn_bfloat16. See example usage here.
- aarch64 CPUs: Experimental oneDNN optimizations are available in the default Linux aarch64 package (pip install tensorflow).
  - The optimizations are disabled by default.
  - Set the environment variable TF_ENABLE_ONEDNN_OPTS=1 to enable the optimizations. Setting the variable to 0 or unsetting it will disable the optimizations.
  - These optimizations can yield slightly different numerical results from when they are off due to floating-point round-off errors from different computation approaches and orders.
  - To verify that the optimizations are on, look for a message with "oneDNN custom operations are on" in the log. If the exact phrase is not there, it means they are off.

Bug Fixes and Other Changes

New argument experimental_device_ordinal in LogicalDeviceConfiguration to control the order of logical devices. (GPU only)
tf.keras:
- Changed the TensorBoard tag names produced by the tf.keras.callbacks.TensorBoard callback, so that summaries logged automatically for model weights now include either a /histogram or /image suffix in their tag names, in order to prevent tag name collisions across summary types.
When running on GPU (with cuDNN version 7.6.3 or later),tf.nn.depthwise_conv2d backprop to filter (and therefore also tf.keras.layers.DepthwiseConv2D) now operate deterministically (and tf.errors.UnimplementedError is no longer thrown) when op-determinism has been enabled via tf.config.experimental.enable_op_determinism. This closes issue 47174.
tf.random
- Added tf.random.experimental.stateless_shuffle, a stateless version of tf.random.shuffle.

Deprecations

The C++ tensorflow::Code and tensorflow::Status will become aliases of respectively absl::StatusCode and absl::Status in some future release.
- Use tensorflow::OkStatus() instead of tensorflow::Status::OK().
- Stop constructing Status objects from tensorflow::error::Code.
- One MUST NOT access tensorflow::errors::Code fields. Accessing tensorflow::error::Code fields is fine.
  - Use the constructors such as tensorflow::errors:InvalidArgument to create status using an error code without accessing it.
  - Use the free functions such as tensorflow::errors::IsInvalidArgument if needed.
  - In the last resort, use e.g.static_cast<tensorflow::errors::Code>(error::Code::INVALID_ARGUMENT) or static_cast<int>(code) for comparisons.
tensorflow::StatusOr will also become in the future alias to absl::StatusOr, so use StatusOr::value instead of StatusOr::ConsumeValueOrDie.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.10.0rc03 Aug 2022pre-releaseDeprecated

Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaption…

Release 2.10.0

Breaking Changes

Causal attention in keras.layers.Attention and keras.layers.AdditiveAttention is now specified in the call() method via the use_causal_mask argument (rather than in the constructor), for consistency with other layers.
Some files in tensorflow/python/training have been moved to tensorflow/python/tracking and tensorflow/python/checkpoint. Please update your imports accordingly, the old files will be removed in Release 2.11.
tf.keras.optimizers.experimental.Optimizer will graduate in Release 2.11, which means tf.keras.optimizers.Optimizer will be an alias of tf.keras.optimizers.experimental.Optimizer. The current tf.keras.optimizers.Optimizer will continue to be supported as tf.keras.optimizers.legacy.Optimizer, e.g., tf.keras.optimizers.legacy.Adam. Most users won't be affected by this change, but please check the API doc if any API used in your workflow is changed or deprecated, and make adaptions. If you decide to keep using the old optimizer, please explicitly change your optimizer to tf.keras.optimizers.legacy.Optimizer.
RNG behavior change for tf.keras.initializers. Keras initializers will now use stateless random ops to generate random numbers.
- Both seeded and unseeded initializers will always generate the same values every time they are called (for a given variable shape). For unseeded initializers (seed=None), a random seed will be created and assigned at initializer creation (different initializer instances get different seeds).
- An unseeded initializer will raise a warning if it is reused (called) multiple times. This is because it would produce the same values each time, which may not be intended.

Major Features and Improvements

tf.lite:
- New operations supported:
  - tflite SelectV2 now supports 5D.
  - tf.einsum is supported with multiple unknown shapes.
  - tf.unsortedsegmentprod op is supported.
  - tf.unsortedsegmentmax op is supported.
  - tf.unsortedsegmentsum op is supported.
- Updates to existing operations:
  - tfl.scatter_nd now supports I1 for update arg.
- Upgrade Flatbuffers v2.0.5 from v1.12.0
tf.keras:
- EinsumDense layer moved from experimental to core. Its import path moved from tf.keras.layers.experimental.EinsumDense to tf.keras.layers.EinsumDense.
- Added tf.keras.utils.audio_dataset_from_directory utility to easily generate audio classification datasets from directories of .wav files.
- Added subset="both" support in tf.keras.utils.image_dataset_from_directory, tf.keras.utils.text_dataset_from_directory, and audio_dataset_from_directory, to be used with the validation_split argument, for returning both dataset splits at once, as a tuple.
- Added tf.keras.utils.split_dataset utility to split a Dataset object or a list/tuple of arrays into two Dataset objects (e.g. train/test).
- Added step granularity to BackupAndRestore callback for handling distributed training failures & restarts. The training state can now be restored at the exact epoch and step at which it was previously saved before failing.
- Added tf.keras.dtensor.experimental.optimizers.AdamW. This optimizer is similar as the existing keras.optimizers.experimental.AdamW, and works in the DTensor training use case.
- Improved masking support for tf.keras.layers.MultiHeadAttention.
  - Implicit masks for query, key and value inputs will automatically be used to compute a correct attention mask for the layer. These padding masks will be combined with any attention_mask passed in directly when calling the layer. This can be used with tf.keras.layers.Embedding with mask_zero=True to automatically infer a correct padding mask.
  - Added a use_causal_mask call time arugment to the layer. Passing use_causal_mask=True will compute a causal attention mask, and optionally combine it with any attention_mask passed in directly when calling the layer.
- Added ignore_class argument in the loss SparseCategoricalCrossentropy and metrics IoU and MeanIoU, to specify a class index to be ignored during loss/metric computation (e.g. a background/void class).
- Added tf.keras.models.experimental.SharpnessAwareMinimization. This class implements the sharpness-aware minimization technique, which boosts model performance on various tasks, e.g., ResNet on image classification.
tf.data:
- Added support for cross-trainer data caching in tf.data service. This saves computation resources when concurrent training jobs train from the same dataset. See https://www.tensorflow.org/api_docs/python/tf/data/experimental/service#sharing_tfdata_service_with_concurrent_trainers for more details.
- Added dataset_id to tf.data.experimental.service.register_dataset. If provided, tf.data service will use the provided ID for the dataset. If the dataset ID already exists, no new dataset will be registered. This is useful if multiple training jobs need to use the same dataset for training. In this case, users should call register_dataset with the same dataset_id.
- Added a new field, inject_prefetch, to tf.data.experimental.OptimizationOptions. If it is set to True, tf.data will now automatically add a prefetch transformation to datasets that end in synchronous transformations. This enables data generation to be overlapped with data consumption. This may cause a small increase in memory usage due to buffering. To enable this behavior, set inject_prefetch=True in tf.data.experimental.OptimizationOptions.
- Added a new value to tf.data.Options.autotune.autotune_algorithm: STAGE_BASED. If the autotune algorithm is set to STAGE_BASED, then it runs a new algorithm that can get the same performance with lower CPU/memory usage.
- Added tf.data.experimental.from_list, a new API for creating Datasets from lists of elements.
tf.distribute:
- Added tf.distribute.experimental.PreemptionCheckpointHandler to handle worker preemption/maintenance and cluster-wise consistent error reporting for tf.distribute.MultiWorkerMirroredStrategy. Specifically, for the type of interruption with advance notice, it automatically saves a checkpoint, exits the program without raising an unrecoverable error, and restores the progress when training restarts.
tf.math:
- Added tf.math.approx_max_k and tf.math.approx_min_k which are the optimized alternatives to tf.math.top_k on TPU. The performance difference range from 8 to 100 times depending on the size of k. When running on CPU and GPU, a non-optimized XLA kernel is used.
tf.train:
- Added tf.train.TrackableView which allows users to inspect the TensorFlow Trackable object (e.g. tf.Module, Keras Layers and models).
tf.vectorized_map:
- Added an optional parameter: warn. This parameter controls whether or not warnings will be printed when operations in the provided fn fall back to a while loop.
XLA:
- MWMS is now compilable with XLA.

Bug Fixes and Other Changes

New argument experimental_device_ordinal in LogicalDeviceConfiguration to control the order of logical devices. (GPU only)
tf.keras:
- Changed the TensorBoard tag names produced by the tf.keras.callbacks.TensorBoard callback, so that summaries logged automatically for model weights now include either a /histogram or /image suffix in their tag names, in order to prevent tag name collisions across summary types.
When running on GPU (with cuDNN version 7.6.3 or later), tf.nn.depthwise_conv2d backprop to filter (and therefore also tf.keras.layers.DepthwiseConv2D) now operate deterministically (and tf.errors.UnimplementedError is no longer thrown) when op-determinism has been enabled via tf.config.experimental.enable_op_determinism. This closes issue 47174.
tf.random
- Added tf.random.experimental.stateless_shuffle, a stateless version of tf.random.shuffle.

Deprecations

The C++ tensorflow::Code and tensorflow::Status will become aliases of respectively absl::StatusCode and absl::Status in some future release.
- Use tensorflow::OkStatus() instead of tensorflow::Status::OK().
- Stop constructing Status objects from tensorflow::error::Code.
- One MUST NOT access tensorflow::errors::Code fields. Accessing tensorflow::error::Code fields is fine.
  - Use the constructors such as tensorflow::errors:InvalidArgument to create status using an error code without accessing it.
  - Use the free functions such as tensorflow::errors::IsInvalidArgument if needed.
  - In the last resort, use e.g. static_cast<tensorflow::errors::Code>(error::Code::INVALID_ARGUMENT) or static_cast<int>(code) for comparisons.
tensorflow::StatusOr will also become in the future alias to absl::StatusOr, so use StatusOr::value instead of StatusOr::ConsumeValueOrDie.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.9.23 Sept 2022Security

This releases introduces several vulnerability fixes:

Release 2.9.2

This releases introduces several vulnerability fixes:

Fixes a CHECK failure in tf.reshape caused by overflows (CVE-2022-35934)
Fixes a CHECK failure in SobolSample caused by missing validation (CVE-2022-35935)
Fixes an OOB read in Gather_nd op in TF Lite (CVE-2022-35937)
Fixes a CHECK failure in TensorListReserve caused by missing validation (CVE-2022-35960)
Fixes an OOB write in Scatter_nd op in TF Lite (CVE-2022-35939)
Fixes an integer overflow in RaggedRangeOp (CVE-2022-35940)
Fixes a CHECK failure in AvgPoolOp (CVE-2022-35941)
Fixes a CHECK failures in UnbatchGradOp (CVE-2022-35952)
Fixes a segfault TFLite converter on per-channel quantized transposed convolutions (CVE-2022-36027)
Fixes a CHECK failures in AvgPool3DGrad (CVE-2022-35959)
Fixes a CHECK failures in FractionalAvgPoolGrad (CVE-2022-35963)
Fixes a segfault in BlockLSTMGradV2 (CVE-2022-35964)
Fixes a segfault in LowerBound and UpperBound (CVE-2022-35965)
Fixes a segfault in QuantizedAvgPool (CVE-2022-35966)
Fixes a segfault in QuantizedAdd (CVE-2022-35967)
Fixes a CHECK fail in AvgPoolGrad (CVE-2022-35968)
Fixes a CHECK fail in Conv2DBackpropInput (CVE-2022-35969)
Fixes a segfault in QuantizedInstanceNorm (CVE-2022-35970)
Fixes a CHECK fail in FakeQuantWithMinMaxVars (CVE-2022-35971)
Fixes a segfault in Requantize (CVE-2022-36017)
Fixes a segfault in QuantizedBiasAdd (CVE-2022-35972)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsPerChannel (CVE-2022-36019)
Fixes a segfault in QuantizedMatMul (CVE-2022-35973)
Fixes a segfault in QuantizeDownAndShrinkRange (CVE-2022-35974)
Fixes segfaults in QuantizedRelu and QuantizedRelu6 (CVE-2022-35979)
Fixes a CHECK fail in FractionalMaxPoolGrad (CVE-2022-35981)
Fixes a CHECK fail in RaggedTensorToVariant (CVE-2022-36018)
Fixes a CHECK fail in QuantizeAndDequantizeV3 (CVE-2022-36026)
Fixes a segfault in SparseBincount (CVE-2022-35982)
Fixes a CHECK fail in Save and SaveSlices (CVE-2022-35983)
Fixes a CHECK fail in ParameterizedTruncatedNormal (CVE-2022-35984)
Fixes a CHECK fail in LRNGrad (CVE-2022-35985)
Fixes a segfault in RaggedBincount (CVE-2022-35986)
Fixes a CHECK fail in DenseBincount (CVE-2022-35987)
Fixes a CHECK fail in tf.linalg.matrix_rank (CVE-2022-35988)
Fixes a CHECK fail in MaxPool (CVE-2022-35989)
Fixes a CHECK fail in Conv2DBackpropInput (CVE-2022-35999)
Fixes a CHECK fail in EmptyTensorList (CVE-2022-35998)
Fixes a CHECK fail in tf.sparse.cross (CVE-2022-35997)
Fixes a floating point exception in Conv2D (CVE-2022-35996)
Fixes a CHECK fail in AudioSummaryV2 (CVE-2022-35995)
Fixes a CHECK fail in CollectiveGather (CVE-2022-35994)
Fixes a CHECK fail in SetSize (CVE-2022-35993)
Fixes a CHECK fail in TensorListFromTensor (CVE-2022-35992)
Fixes a CHECK fail in TensorListScatter and TensorListScatterV2 (CVE-2022-35991)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsPerChannelGradient (CVE-2022-35990)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsGradient (CVE-2022-36005)
Fixes a CHECK fail in tf.random.gamma (CVE-2022-36004)
Fixes a CHECK fail in RandomPoissonV2 (CVE-2022-36003)
Fixes a CHECK fail in Unbatch (CVE-2022-36002)
Fixes a CHECK fail in DrawBoundingBoxes (CVE-2022-36001)
Fixes a CHECK fail in Eig (CVE-2022-36000)
Fixes a null dereference on MLIR on empty function attributes (CVE-2022-36011)
Fixes an assertion failure on MLIR empty edge names (CVE-2022-36012)
Fixes a null-dereference in mlir::tfg::GraphDefImporter::ConvertNodeDef (CVE-2022-36013)
Fixes a null-dereference in mlir::tfg::TFOp::nameAttr (CVE-2022-36014)
Fixes an integer overflow in math ops (CVE-2022-36015)
Fixes a CHECK-fail in tensorflow::full_type::SubstituteFromAttrs (CVE-2022-36016)
Fixes an OOB read in Gather_nd op in TF Lite Micro (CVE-2022-35938)

Open source →

2.9.016 May 2022Security

Fixes a code injection in saved_model_cli (CVE-2022-29216)

Release 2.9.0

Breaking Changes

Due to security issues in TF 2.8, all boosted trees code has now been removed (after being deprecated in TF 2.8). Users should switch to TensorFlow Decision Forests.
Build, Compilation and Packaging
- TensorFlow is now compiled with _GLIBCXX_USE_CXX11_ABI=1. Downstream projects that encounter std::__cxx11 or [abi:cxx11] linker errors will need to adopt this compiler option. See the GNU C++ Library docs on Dual ABI.
- TensorFlow Python wheels now specifically conform to manylinux2014, an upgrade from manylinux2010. The minimum Pip version supporting manylinux2014 is Pip 19.3 (see pypa/manylinux. This change may affect you if you have been using TensorFlow on a very old platform equivalent to CentOS 6, as manylinux2014 targets CentOS 7 as a compatibility base. Note that TensorFlow does not officially support either platform.
- Discussion for these changes can be found on SIG Build's TensorFlow Community Forum thread
The tf.keras.mixed_precision.experimental API has been removed. The non-experimental symbols under tf.keras.mixed_precision have been available since TensorFlow 2.4 and should be used instead.
- The non-experimental API has some minor differences from the experimental API. In most cases, you only need to make three minor changes:
  - Remove the word "experimental" from tf.keras.mixed_precision symbols. E.g., replace tf.keras.mixed_precision.experimental.global_policy with tf.keras.mixed_precision.global_policy.
  - Replace tf.keras.mixed_precision.experimental.set_policy with tf.keras.mixed_precision.set_global_policy. The experimental symbol set_policy was renamed to set_global_policy in the non-experimental API.
  - Replace LossScaleOptimizer(opt, "dynamic") with LossScaleOptimizer(opt). If you pass anything other than "dynamic" to the second argument, see (1) of the next section.
- In the following rare cases, you need to make more changes when switching to the non-experimental API:
  - If you passed anything other than "dynamic" to the loss_scale argument (the second argument) of LossScaleOptimizer:
    - The LossScaleOptimizer constructor takes in different arguments. See the TF 2.7 documentation of tf.keras.mixed_precision.experimental.LossScaleOptimizer for details on the differences, which has examples on how to convert to the non-experimental LossScaleOptimizer.
  - If you passed a value to the loss_scale argument (the second argument) of Policy:
    - The experimental version of Policy optionally took in a tf.compat.v1.mixed_precision.LossScale in the constructor, which defaulted to a dynamic loss scale for the "mixed_float16" policy and no loss scale for other policies. In Model.compile, if the model's policy had a loss scale, the optimizer would be wrapped with a LossScaleOptimizer. With the non-experimental Policy, there is no loss scale associated with the Policy, and Model.compile wraps the optimizer with a LossScaleOptimizer if and only if the policy is a "mixed_float16" policy. If you previously passed a LossScale to the experimental Policy, consider just removing it, as the default loss scaling behavior is usually what you want. If you really want to customize the loss scaling behavior, you can wrap your optimizer with a LossScaleOptimizer before passing it to Model.compile.
  - If you use the very rarely-used function tf.keras.mixed_precision.experimental.get_layer_policy:
    - Replace tf.keras.mixed_precision.experimental.get_layer_policy(layer) with layer.dtype_policy.
tf.mixed_precision.experimental.LossScale and its subclasses have been removed from the TF2 namespace. This symbols were very rarely used and were only useful in TF2 for use in the now-removed tf.keras.mixed_precision.experimental API. The symbols are still available under tf.compat.v1.mixed_precision.
The experimental_relax_shapes heuristic for tf.function has been deprecated and replaced with reduce_retracing which encompasses broader heuristics to reduce the number of retraces (see below)

Major Features and Improvements

tf.keras:
- Added tf.keras.applications.resnet_rs models. This includes the ResNetRS50, ResNetRS101, ResNetRS152, ResNetRS200, ResNetRS270, ResNetRS350 and ResNetRS420 model architectures. The ResNetRS models are based on the architecture described in Revisiting ResNets: Improved Training and Scaling Strategies
- Added tf.keras.optimizers.experimental.Optimizer. The reworked optimizer gives more control over different phases of optimizer calls, and is easier to customize. We provide Adam, SGD, Adadelta, AdaGrad and RMSprop optimizers based on tf.keras.optimizers.experimental.Optimizer. Generally the new optimizers work in the same way as the old ones, but support new constructor arguments. In the future, the symbols tf.keras.optimizers.Optimizer/Adam/etc will point to the new optimizers, and the previous generation of optimizers will be moved to tf.keras.optimizers.legacy.Optimizer/Adam/etc.
- Added L2 unit normalization layer tf.keras.layers.UnitNormalization.
- Added tf.keras.regularizers.OrthogonalRegularizer, a new regularizer that encourages orthogonality between the rows (or columns) or a weight matrix.
- Added tf.keras.layers.RandomBrightness layer for image preprocessing.
- Added APIs for switching between interactive logging and absl logging. By default, Keras always writes the logs to stdout. However, this is not optimal in a non-interactive environment, where you don't have access to stdout, but can only view the logs. You can use tf.keras.utils.disable_interactive_logging() to write the logs to ABSL logging. You can also use tf.keras.utils.enable_interactive_logging() to change it back to stdout, or tf.keras.utils.is_interactive_logging_enabled() to check if interactive logging is enabled.
- Changed default value for the verbose argument of Model.evaluate() and Model.predict() to "auto", which defaults to verbose=1 for most cases and defaults to verbose=2 when used with ParameterServerStrategy or with interactive logging disabled.
- Argument jit_compile in Model.compile() now applies to Model.evaluate() and Model.predict(). Setting jit_compile=True in compile() compiles the model's training, evaluation, and inference steps to XLA. Note that jit_compile=True may not necessarily work for all models.
- Added DTensor-related Keras APIs under tf.keras.dtensor namespace. The APIs are still classified as experimental. You are welcome to try it out. Please check the tutoral and guide on https://www.tensorflow.org/ for more details about DTensor.
tf.lite:
- Added TFLite builtin op support for the following TF ops:
  - tf.math.argmin/tf.math.argmax for input data type tf.bool on CPU.
  - tf.nn.gelu op for output data type tf.float32 and quantization on CPU.
- Add nominal support for unsigned 16-bit integer tensor types. Note that very few TFLite kernels support this type natively, so its use in mobile ML authoring is generally discouraged.
- Add support for unsigned 16-bit integer tensor types in cast op.
- Experimental support for lowering list_ops.tensor_list_set_item with DynamicUpdateSlice.
- Enabled a new MLIR-based dynamic range quantization backend by default
  - The new backend is used for post-training int8 dynamic range quantization and post-training float16 quantization.
  - Set experimental_new_dynamic_range_quantizer in tf.lite.TFLiteConverter to False to disable this change
- Native TF Lite variables are now enabled during conversion by default on all v2 TfLiteConverter entry points. experimental_enable_resource_variables on tf.lite.TFLiteConverter is now True by default and will be removed in the future.
tf.function:
- Custom classes used as arguments for tf.function can now specify rules regarding when retracing needs to occur by implementing the Tracing Protocol available through tf.types.experimental.SupportsTracingProtocol.
- TypeSpec classes (as associated with ExtensionTypes) also implement the Tracing Protocol which can be overriden if necessary.
- The newly introduced reduce_retracing option also uses the Tracing Protocol to proactively generate generalized traces similar to experimental_relax_shapes (which has now been deprecated).
Unified eager and tf.function execution:
- Eager mode can now execute each op as a tf.function, allowing for more consistent feature support in future releases.
- It is available for immediate use.
  - See the TF_RUN_EAGER_OP_AS_FUNCTION environment variable in eager context.
  - Eager performance should be similar with this feature enabled.
    - A roughly 5us per-op overhead may be observed when running many small functions.
    - Note a known issue with GPU performance.
  - The behavior of tf.function itself is unaffected.
- Note: This feature will be enabled by default in an upcoming version of TensorFlow.
tf.experimental.dtensor: Added DTensor, an extension to TensorFlow for large-scale modeling with minimal changes to user code. You are welcome to try it out, though be aware that the DTensor API is experimental and up-to backward-incompatible changes. DTensor and Keras integration is published under tf.keras.dtensor in this release (refer to the tf.keras entry). The tutoral and guide for DTensor will be published on https://www.tensorflow.org/. Please stay tuned.
oneDNN CPU performance optimizations are available in Linux x86, Windows x86, and Linux aarch64 packages.
- Linux x86 packages:
  - oneDNN optimizations are enabled by default on CPUs with neural-network-focused hardware features such as AVX512_VNNI, AVX512_BF16, AMX, etc. (Intel Cascade Lake and newer CPUs.)
    - Example performance speedups.
  - For older CPUs, oneDNN optimizations are disabled by default.
- Windows x86 package: oneDNN optimizations are disabled by default.
- Linux aach64 (--config=mkl_aarch64) package:
  - Experimental oneDNN optimizations are disabled by default.
  - If you experience issues with oneDNN optimizations on, we recommend turning them off.
- To explicitly enable or disable oneDNN optimizations, set the environment variable TF_ENABLE_ONEDNN_OPTS to 1 (enable) or 0 (disable) before running TensorFlow. (The variable is checked during import tensorflow.) To fall back to default settings, unset the environment variable.
- These optimizations can yield slightly different numerical results from when they are off due to floating-point round-off errors from different computation approaches and orders.
- To verify that the optimizations are on, look for a message with "oneDNN custom operations are on" in the log. If the exact phrase is not there, it means they are off.

Bug Fixes and Other Changes

tf.data:
- Fixed bug in tf.data.experimental.parse_example_dataset when tf.io.RaggedFeatures would specify value_key but no partitions. Before the fix, setting value_key but no partitions would result in the feature key being replaced by the value key, e.g. {'value_key': <RaggedTensor>} instead of {'key': <RaggedTensor>}. Now the correct feature key will be used. This aligns the behavior of tf.data.experimental.parse_example_dataset to match the behavior of tf.io.parse_example.
- Added a new field, filter_parallelization, to tf.data.experimental.OptimizationOptions. If it is set to True, tf.data will run Filter transformation with multiple threads. Its default value is False if not specified.
tf.keras:
- Fixed bug in optimizers that prevented them from properly checkpointing slot variables when they are ShardedVariables (used for training with tf.distribute.experimental.ParameterServerStrategy).
tf.random:
- Added tf.random.experimental.index_shuffle, for shuffling a sequence without materializing the sequence in memory.
tf.RaggedTensor:
- Introduced tf.experimental.RowPartition, which encodes how one dimension in a RaggedTensor relates to another, into the public API.
- Introduced tf.experimental.DynamicRaggedShape, which represents the shape of a RaggedTensor.

Security

Fixes a code injection in saved_model_cli (CVE-2022-29216)
Fixes a missing validation which causes TensorSummaryV2 to crash (CVE-2022-29193)
Fixes a missing validation which crashes QuantizeAndDequantizeV4Grad (CVE-2022-29192)
Fixes a missing validation which causes denial of service via DeleteSessionTensor (CVE-2022-29194)
Fixes a missing validation which causes denial of service via GetSessionTensor (CVE-2022-29191)
Fixes a missing validation which causes denial of service via StagePeek (CVE-2022-29195)
Fixes a missing validation which causes denial of service via UnsortedSegmentJoin (CVE-2022-29197)
Fixes a missing validation which causes denial of service via LoadAndRemapMatrix (CVE-2022-29199)
Fixes a missing validation which causes denial of service via SparseTensorToCSRSparseMatrix (CVE-2022-29198)
Fixes a missing validation which causes denial of service via LSTMBlockCell (CVE-2022-29200)
Fixes a missing validation which causes denial of service via Conv3DBackpropFilterV2 (CVE-2022-29196)
Fixes a CHECK failure in depthwise ops via overflows (CVE-2021-41197)
Fixes issues arising from undefined behavior stemming from users supplying invalid resource handles (CVE-2022-29207)
Fixes a segfault due to missing support for quantized types (CVE-2022-29205)
Fixes a missing validation which results in undefined behavior in SparseTensorDenseAdd (CVE-2022-29206)
Fixes a missing validation which results in undefined behavior in QuantizedConv2D (CVE-2022-29201)
Fixes an integer overflow in SpaceToBatchND (CVE-2022-29203)
Fixes a segfault and OOB write due to incomplete validation in EditDistance (CVE-2022-29208)
Fixes a missing validation which causes denial of service via Conv3DBackpropFilterV2 (CVE-2022-29204)
Fixes a denial of service in tf.ragged.constant due to lack of validation (CVE-2022-29202)
Fixes a segfault when tf.histogram_fixed_width is called with NaN values (CVE-2022-29211)
Fixes a core dump when loading TFLite models with quantization (CVE-2022-29212)
Fixes crashes stemming from incomplete validation in signal ops (CVE-2022-29213)
Fixes a type confusion leading to CHECK-failure based denial of service (CVE-2022-29209)
Fixes a heap buffer overflow due to incorrect hash function (CVE-2022-29210)
Updates curl to 7.83.1 to handle (CVE-2022-22576, (CVE-2022-27774, (CVE-2022-27775, (CVE-2022-27776, (CVE-2022-27778, (CVE-2022-27779, (CVE-2022-27780, (CVE-2022-27781, (CVE-2022-27782 and (CVE-2022-30115
Updates zlib to 1.2.12 after 1.2.11 was pulled due to security issue

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Aaron Debattista, Abel Soares Siqueira, Abhishek Varma, Andrei Ivanov, andreii, Andrew Goodbody, apeltop, Arnab Dutta, Ashiq Imran, Banikumar Maiti (Intel Aipg), Ben Greiner, Benjamin Peterson, bhack, Christopher Bate, chunduriv, Copybara-Service, DEKHTIARJonathan, Deven Desai, Duncan Riach, Eric Kunze, Everton Constantino, Faruk D, Fredrik Knutsson, gadagashwini, Gauri1 Deshpande, gtiHibGele, Guozhong Zhuang, Islem-Esi, Ivanov Viktor, Jason Furmanek, Jason Zaman, Jim, Jinzhe Zeng, John Laxson, Jonas Eschle, Jonas Eschle 'Mayou36, Jonathan Dekhtiar, Kaixi Hou, Kanvi Khanna, KaurkerDevourer, Koan-Sin Tan, kushanam, Laramie Leavitt, Li-Wen Chang, lipracer, Louis Sugy, Lu Teng, Mahmoud Abuzaina, Malcolm Slaney, Malik Shahzad Muzaffar, Marek Šuppa, Matt Conley, Michael Melesse, Milos Puzovic, mohantym, Nathan John Sircombe, Nathan Luehr, Nilesh Agarwalla, Patrice Vignola, peterjc123, Philip Turner, Rajeshwar Reddy T, Robert Kalmar, Rodrigo Formigone, Rohit Santhanam, rui, Sachin Muradi, Saduf2019, sandip, Scott Leishman, Serge Panev, Shi,Guangyong, Srinivasan Narayanamoorthy, stanley, Steven I Reeves, stevenireeves, sushreebarsa, Tamas Bela Feher, Tao He, Thomas Schmeyer, Tiago Almeida, Trevor Morris, Uday Bondhugula, Uwe L. Korn, Varghese, Jojimon, Vishnuvardhan Janapati, William Muir, William Raveane, xutianming, Yasuhiro Matsumoto, Yimei Sun, Yong Tang, Yu Feng, Yuriy Chernyshov, zhaozheng09

Open source →

2.9.0rc24 May 2022pre-releaseDeprecated

Due to security issues in TF 2.8, all boosted trees code has now been removed (after being deprecated in TF 2.8). Users should switch to TensorFlow De…

Release 2.9.0

Breaking Changes

Due to security issues in TF 2.8, all boosted trees code has now been removed (after being deprecated in TF 2.8). Users should switch to TensorFlow Decision Forests.
Build, Compilation and Packaging
- TensorFlow is now compiled with _GLIBCXX_USE_CXX11_ABI=1. Downstream projects that encounter std::__cxx11 or [abi:cxx11] linker errors will need to adopt this compiler option. See the GNU C++ Library docs on Dual ABI.
- TensorFlow Python wheels now specifically conform to manylinux2014, an upgrade from manylinux2010. The minimum Pip version supporting manylinux2014 is Pip 19.3 (see pypa/manylinux. This change may affect you if you have been using TensorFlow on a very old platform equivalent to CentOS 6, as manylinux2014 targets CentOS 7 as a compatibility base. Note that TensorFlow does not officially support either platform.
- Discussion for these changes can be found on SIG Build's TensorFlow Community Forum thread
The tf.keras.mixed_precision.experimental API has been removed. The non-experimental symbols under tf.keras.mixed_precision have been available since TensorFlow 2.4 and should be used instead.
- The non-experimental API has some minor differences from the experimental API. In most cases, you only need to make three minor changes:
  - Remove the word "experimental" from tf.keras.mixed_precision symbols. E.g., replace tf.keras.mixed_precision.experimental.global_policy with tf.keras.mixed_precision.global_policy.
  - Replace tf.keras.mixed_precision.experimental.set_policy with tf.keras.mixed_precision.set_global_policy. The experimental symbol set_policy was renamed to set_global_policy in the non-experimental API.
  - Replace LossScaleOptimizer(opt, "dynamic") with LossScaleOptimizer(opt). If you pass anything other than "dynamic" to the second argument, see (1) of the next section.
- In the following rare cases, you need to make more changes when switching to the non-experimental API:
  - If you passed anything other than "dynamic" to the loss_scale argument (the second argument) of LossScaleOptimizer:
    - The LossScaleOptimizer constructor takes in different arguments. See the TF 2.7 documentation of tf.keras.mixed_precision.experimental.LossScaleOptimizer for details on the differences, which has examples on how to convert to the non-experimental LossScaleOptimizer.
  - If you passed a value to the loss_scale argument (the second argument) of Policy:
    - The experimental version of Policy optionally took in a tf.compat.v1.mixed_precision.LossScale in the constructor, which defaulted to a dynamic loss scale for the "mixed_float16" policy and no loss scale for other policies. In Model.compile, if the model's policy had a loss scale, the optimizer would be wrapped with a LossScaleOptimizer. With the non-experimental Policy, there is no loss scale associated with the Policy, and Model.compile wraps the optimizer with a LossScaleOptimizer if and only if the policy is a "mixed_float16" policy. If you previously passed a LossScale to the experimental Policy, consider just removing it, as the default loss scaling behavior is usually what you want. If you really want to customize the loss scaling behavior, you can wrap your optimizer with a LossScaleOptimizer before passing it to Model.compile.
  - If you use the very rarely-used function tf.keras.mixed_precision.experimental.get_layer_policy:
    - Replace tf.keras.mixed_precision.experimental.get_layer_policy(layer) with layer.dtype_policy.
tf.mixed_precision.experimental.LossScale and its subclasses have been removed from the TF2 namespace. This symbols were very rarely used and were only useful in TF2 for use in the now-removed tf.keras.mixed_precision.experimental API. The symbols are still available under tf.compat.v1.mixed_precision.
The experimental_relax_shapes heuristic for tf.function has been deprecated and replaced with reduce_retracing which encompasses broader heuristics to reduce the number of retraces (see below)

Major Features and Improvements

tf.keras:
- Added tf.keras.applications.resnet_rs models. This includes the ResNetRS50, ResNetRS101, ResNetRS152, ResNetRS200, ResNetRS270, ResNetRS350 and ResNetRS420 model architectures. The ResNetRS models are based on the architecture described in Revisiting ResNets: Improved Training and Scaling Strategies
- Added tf.keras.optimizers.experimental.Optimizer. The reworked optimizer gives more control over different phases of optimizer calls, and is easier to customize. We provide Adam, SGD, Adadelta, AdaGrad and RMSprop optimizers based on tf.keras.optimizers.experimental.Optimizer. Generally the new optimizers work in the same way as the old ones, but support new constructor arguments. In the future, the symbols tf.keras.optimizers.Optimizer/Adam/etc will point to the new optimizers, and the previous generation of optimizers will be moved to tf.keras.optimizers.legacy.Optimizer/Adam/etc.
- Added L2 unit normalization layer tf.keras.layers.UnitNormalization.
- Added tf.keras.regularizers.OrthogonalRegularizer, a new regularizer that encourages orthogonality between the rows (or columns) or a weight matrix.
- Added tf.keras.layers.RandomBrightness layer for image preprocessing.
- Added APIs for switching between interactive logging and absl logging. By default, Keras always writes the logs to stdout. However, this is not optimal in a non-interactive environment, where you don't have access to stdout, but can only view the logs. You can use tf.keras.utils.disable_interactive_logging() to write the logs to ABSL logging. You can also use tf.keras.utils.enable_interactive_logging() to change it back to stdout, or tf.keras.utils.is_interactive_logging_enabled() to check if interactive logging is enabled.
- Changed default value for the verbose argument of Model.evaluate() and Model.predict() to "auto", which defaults to verbose=1 for most cases and defaults to verbose=2 when used with ParameterServerStrategy or with interactive logging disabled.
- Argument jit_compile in Model.compile() now applies to Model.evaluate() and Model.predict(). Setting jit_compile=True in compile() compiles the model's training, evaluation, and inference steps to XLA. Note that jit_compile=True may not necessarily work for all models.
- Added DTensor-related Keras APIs under tf.keras.dtensor namespace. The APIs are still classified as experimental. You are welcome to try it out. Please check the tutoral and guide on https://www.tensorflow.org/ for more details about DTensor.
tf.lite:
- Added TFLite builtin op support for the following TF ops:
  - tf.math.argmin/tf.math.argmax for input data type tf.bool on CPU.
  - tf.nn.gelu op for output data type tf.float32 and quantization on CPU.
- Add nominal support for unsigned 16-bit integer tensor types. Note that very few TFLite kernels support this type natively, so its use in mobile ML authoring is generally discouraged.
- Add support for unsigned 16-bit integer tensor types in cast op.
- Experimental support for lowering list_ops.tensor_list_set_item with DynamicUpdateSlice.
- Enabled a new MLIR-based dynamic range quantization backend by default
  - The new backend is used for post-training int8 dynamic range quantization and post-training float16 quantization.
  - Set experimental_new_dynamic_range_quantizer in tf.lite.TFLiteConverter to False to disable this change
- Native TF Lite variables are now enabled during conversion by default on all v2 TfLiteConverter entry points. experimental_enable_resource_variables on tf.lite.TFLiteConverter is now True by default and will be removed in the future.
tf.function:
- Custom classes used as arguments for tf.function can now specify rules regarding when retracing needs to occur by implementing the Tracing Protocol available through tf.types.experimental.SupportsTracingProtocol.
- TypeSpec classes (as associated with ExtensionTypes) also implement the Tracing Protocol which can be overriden if necessary.
- The newly introduced reduce_retracing option also uses the Tracing Protocol to proactively generate generalized traces similar to experimental_relax_shapes (which has now been deprecated).
Unified eager and tf.function execution:
- Eager mode can now execute each op as a tf.function, allowing for more consistent feature support in future releases.
- It is available for immediate use.
  - See the TF_RUN_EAGER_OP_AS_FUNCTION environment variable in eager context.
  - Eager performance should be similar with this feature enabled.
    - A roughly 5us per-op overhead may be observed when running many small functions.
    - Note a known issue with GPU performance.
  - The behavior of tf.function itself is unaffected.
- Note: This feature will be enabled by default in an upcoming version of TensorFlow.
tf.experimental.dtensor: Added DTensor, an extension to TensorFlow for large-scale modeling with minimal changes to user code. You are welcome to try it out, though be aware that the DTensor API is experimental and up-to backward-incompatible changes. DTensor and Keras integration is published under tf.keras.dtensor in this release (refer to the tf.keras entry). The tutoral and guide for DTensor will be published on https://www.tensorflow.org/. Please stay tuned.

Bug Fixes and Other Changes

tf.data:
- Fixed bug in tf.data.experimental.parse_example_dataset when tf.io.RaggedFeatures would specify value_key but no partitions. Before the fix, setting value_key but no partitions would result in the feature key being replaced by the value key, e.g. {'value_key': <RaggedTensor>} instead of {'key': <RaggedTensor>}. Now the correct feature key will be used. This aligns the behavior of tf.data.experimental.parse_example_dataset to match the behavior of tf.io.parse_example.
- Added a new field, filter_parallelization, to tf.data.experimental.OptimizationOptions. If it is set to True, tf.data will run Filter transformation with multiple threads. Its default value is False if not specified.
tf.keras:
- Fixed bug in optimizers that prevented them from properly checkpointing slot variables when they are ShardedVariables (used for training with tf.distribute.experimental.ParameterServerStrategy).
tf.random:
- Added tf.random.experimental.index_shuffle, for shuffling a sequence without materializing the sequence in memory.
tf.RaggedTensor:
- Introduced tf.experimental.RowPartition, which encodes how one dimension in a RaggedTensor relates to another, into the public API.
- Introduced tf.experimental.DynamicRaggedShape, which represents the shape of a RaggedTensor.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.9.0rc121 Apr 2022pre-releaseDeprecated

Due to security issues in TF 2.8, all boosted trees code has now been removed (after being deprecated in TF 2.8). Users should switch to TensorFlow De…

Release 2.9.0

Breaking Changes

Due to security issues in TF 2.8, all boosted trees code has now been removed (after being deprecated in TF 2.8). Users should switch to TensorFlow Decision Forests.
Build, Compilation and Packaging
- TensorFlow is now compiled with _GLIBCXX_USE_CXX11_ABI=1. Downstream projects that encounter std::__cxx11 or [abi:cxx11] linker errors will need to adopt this compiler option. See the GNU C++ Library docs on Dual ABI.
- TensorFlow Python wheels now specifically conform to manylinux2014, an upgrade from manylinux2010. The minimum Pip version supporting manylinux2014 is Pip 19.3 (see pypa/manylinux. This change may affect you if you have been using TensorFlow on a very old platform equivalent to CentOS 6, as manylinux2014 targets CentOS 7 as a compatibility base. Note that TensorFlow does not officially support either platform.
- Discussion for these changes can be found on SIG Build's TensorFlow Community Forum thread
The tf.keras.mixed_precision.experimental API has been removed. The non-experimental symbols under tf.keras.mixed_precision have been available since TensorFlow 2.4 and should be used instead.
- The non-experimental API has some minor differences from the experimental API. In most cases, you only need to make three minor changes:
  - Remove the word "experimental" from tf.keras.mixed_precision symbols. E.g., replace tf.keras.mixed_precision.experimental.global_policy with tf.keras.mixed_precision.global_policy.
  - Replace tf.keras.mixed_precision.experimental.set_policy with tf.keras.mixed_precision.set_global_policy. The experimental symbol set_policy was renamed to set_global_policy in the non-experimental API.
  - Replace LossScaleOptimizer(opt, "dynamic") with LossScaleOptimizer(opt). If you pass anything other than "dynamic" to the second argument, see (1) of the next section.
- In the following rare cases, you need to make more changes when switching to the non-experimental API:
  - If you passed anything other than "dynamic" to the loss_scale argument (the second argument) of LossScaleOptimizer:
    - The LossScaleOptimizer constructor takes in different arguments. See the TF 2.7 documentation of tf.keras.mixed_precision.experimental.LossScaleOptimizer for details on the differences, which has examples on how to convert to the non-experimental LossScaleOptimizer.
  - If you passed a value to the loss_scale argument (the second argument) of Policy:
    - The experimental version of Policy optionally took in a tf.compat.v1.mixed_precision.LossScale in the constructor, which defaulted to a dynamic loss scale for the "mixed_float16" policy and no loss scale for other policies. In Model.compile, if the model's policy had a loss scale, the optimizer would be wrapped with a LossScaleOptimizer. With the non-experimental Policy, there is no loss scale associated with the Policy, and Model.compile wraps the optimizer with a LossScaleOptimizer if and only if the policy is a "mixed_float16" policy. If you previously passed a LossScale to the experimental Policy, consider just removing it, as the default loss scaling behavior is usually what you want. If you really want to customize the loss scaling behavior, you can wrap your optimizer with a LossScaleOptimizer before passing it to Model.compile.
  - If you use the very rarely-used function tf.keras.mixed_precision.experimental.get_layer_policy:
    - Replace tf.keras.mixed_precision.experimental.get_layer_policy(layer) with layer.dtype_policy.
tf.mixed_precision.experimental.LossScale and its subclasses have been removed from the TF2 namespace. This symbols were very rarely used and were only useful in TF2 for use in the now-removed tf.keras.mixed_precision.experimental API. The symbols are still available under tf.compat.v1.mixed_precision.
The experimental_relax_shapes heuristic for tf.function has been deprecated and replaced with reduce_retracing which encompasses broader heuristics to reduce the number of retraces (see below)

Major Features and Improvements

tf.keras:
- Added tf.keras.applications.resnet_rs models. This includes the ResNetRS50, ResNetRS101, ResNetRS152, ResNetRS200, ResNetRS270, ResNetRS350 and ResNetRS420 model architectures. The ResNetRS models are based on the architecture described in Revisiting ResNets: Improved Training and Scaling Strategies
- Added tf.keras.optimizers.experimental.Optimizer. The reworked optimizer gives more control over different phases of optimizer calls, and is easier to customize. We provide Adam, SGD, Adadelta, AdaGrad and RMSprop optimizers based on tf.keras.optimizers.experimental.Optimizer. Generally the new optimizers work in the same way as the old ones, but support new constructor arguments. In the future, the symbols tf.keras.optimizers.Optimizer/Adam/etc will point to the new optimizers, and the previous generation of optimizers will be moved to tf.keras.optimizers.legacy.Optimizer/Adam/etc.
- Added L2 unit normalization layer tf.keras.layers.UnitNormalization.
- Added tf.keras.regularizers.OrthogonalRegularizer, a new regularizer that encourages orthogonality between the rows (or columns) or a weight matrix.
- Added tf.keras.layers.RandomBrightness layer for image preprocessing.
- Added APIs for switching between interactive logging and absl logging. By default, Keras always writes the logs to stdout. However, this is not optimal in a non-interactive environment, where you don't have access to stdout, but can only view the logs. You can use tf.keras.utils.disable_interactive_logging() to write the logs to ABSL logging. You can also use tf.keras.utils.enable_interactive_logging() to change it back to stdout, or tf.keras.utils.is_interactive_logging_enabled() to check if interactive logging is enabled.
- Changed default value for the verbose argument of Model.evaluate() and Model.predict() to "auto", which defaults to verbose=1 for most cases and defaults to verbose=2 when used with ParameterServerStrategy or with interactive logging disabled.
- Argument jit_compile in Model.compile() now applies to Model.evaluate() and Model.predict(). Setting jit_compile=True in compile() compiles the model's training, evaluation, and inference steps to XLA. Note that jit_compile=True may not necessarily work for all models.
- Added DTensor-related Keras APIs under tf.keras.dtensor namespace. The APIs are still classified as experimental. You are welcome to try it out. Please check the tutoral and guide on https://www.tensorflow.org/ for more details about DTensor.
tf.lite:
- Added TFLite builtin op support for the following TF ops:
  - tf.math.argmin/tf.math.argmax for input data type tf.bool on CPU.
  - tf.nn.gelu op for output data type tf.float32 and quantization on CPU.
- Add nominal support for unsigned 16-bit integer tensor types. Note that very few TFLite kernels support this type natively, so its use in mobile ML authoring is generally discouraged.
- Add support for unsigned 16-bit integer tensor types in cast op.
- Experimental support for lowering list_ops.tensor_list_set_item with DynamicUpdateSlice.
- Enabled a new MLIR-based dynamic range quantization backend by default
  - The new backend is used for post-training int8 dynamic range quantization and post-training float16 quantization.
  - Set experimental_new_dynamic_range_quantizer in tf.lite.TFLiteConverter to False to disable this change
- Native TF Lite variables are now enabled during conversion by default on all v2 TfLiteConverter entry points. experimental_enable_resource_variables on tf.lite.TFLiteConverter is now True by default and will be removed in the future.
tf.function:
- Custom classes used as arguments for tf.function can now specify rules regarding when retracing needs to occur by implementing the Tracing Protocol available through tf.types.experimental.SupportsTracingProtocol.
- TypeSpec classes (as associated with ExtensionTypes) also implement the Tracing Protocol which can be overriden if necessary.
- The newly introduced reduce_retracing option also uses the Tracing Protocol to proactively generate generalized traces similar to experimental_relax_shapes (which has now been deprecated).
Unified eager and tf.function execution:
- Eager mode can now execute each op as a tf.function, allowing for more consistent feature support in future releases.
- It is available for immediate use.
  - See the TF_RUN_EAGER_OP_AS_FUNCTION environment variable in eager context.
  - Eager performance should be similar with this feature enabled.
    - A roughly 5us per-op overhead may be observed when running many small functions.
    - Note a known issue with GPU performance.
  - The behavior of tf.function itself is unaffected.
- Note: This feature will be enabled by default in an upcoming version of TensorFlow.
tf.experimental.dtensor: Added DTensor, an extension to TensorFlow for large-scale modeling with minimal changes to user code. You are welcome to try it out, though be aware that the DTensor API is experimental and up-to backward-incompatible changes. DTensor and Keras integration is published under tf.keras.dtensor in this release (refer to the tf.keras entry). The tutoral and guide for DTensor will be published on https://www.tensorflow.org/. Please stay tuned.

Bug Fixes and Other Changes

tf.data:
- Fixed bug in tf.data.experimental.parse_example_dataset when tf.io.RaggedFeatures would specify value_key but no partitions. Before the fix, setting value_key but no partitions would result in the feature key being replaced by the value key, e.g. {'value_key': <RaggedTensor>} instead of {'key': <RaggedTensor>}. Now the correct feature key will be used. This aligns the behavior of tf.data.experimental.parse_example_dataset to match the behavior of tf.io.parse_example.
- Added a new field, filter_parallelization, to tf.data.experimental.OptimizationOptions. If it is set to True, tf.data will run Filter transformation with multiple threads. Its default value is False if not specified.
tf.keras:
- Fixed bug in optimizers that prevented them from properly checkpointing slot variables when they are ShardedVariables (used for training with tf.distribute.experimental.ParameterServerStrategy).
tf.random:
- Added tf.random.experimental.index_shuffle, for shuffling a sequence without materializing the sequence in memory.
tf.RaggedTensor:
- Introduced tf.experimental.RowPartition, which encodes how one dimension in a RaggedTensor relates to another, into the public API.
- Introduced tf.experimental.DynamicRaggedShape, which represents the shape of a RaggedTensor.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.9.0rc012 Apr 2022pre-releaseDeprecated

Due to security issues in TF 2.8, all boosted trees code has now been removed (after being deprecated in TF 2.8). Users should switch to TensorFlow De…

Release 2.9.0

Breaking Changes

Due to security issues in TF 2.8, all boosted trees code has now been removed (after being deprecated in TF 2.8). Users should switch to TensorFlow Decision Forests.
Build, Compilation and Packaging
- TensorFlow is now compiled with _GLIBCXX_USE_CXX11_ABI=1. Downstream projects that encounter std::__cxx11 or [abi:cxx11] linker errors will need to adopt this compiler option. See the GNU C++ Library docs on Dual ABI.
- TensorFlow Python wheels now specifically conform to manylinux2014, an upgrade from manylinux2010. The minimum Pip version supporting manylinux2014 is Pip 19.3 (see pypa/manylinux. This change may affect you if you have been using TensorFlow on a very old platform equivalent to CentOS 6, as manylinux2014 targets CentOS 7 as a compatibility base. Note that TensorFlow does not officially support either platform.
- Discussion for these changes can be found on SIG Build's TensorFlow Community Forum thread
The tf.keras.mixed_precision.experimental API has been removed. The non-experimental symbols under tf.keras.mixed_precision have been available since TensorFlow 2.4 and should be used instead.
- The non-experimental API has some minor differences from the experimental API. In most cases, you only need to make three minor changes:
  - Remove the word "experimental" from tf.keras.mixed_precision symbols. E.g., replace tf.keras.mixed_precision.experimental.global_policy with tf.keras.mixed_precision.global_policy.
  - Replace tf.keras.mixed_precision.experimental.set_policy with tf.keras.mixed_precision.set_global_policy. The experimental symbol set_policy was renamed to set_global_policy in the non-experimental API.
  - Replace LossScaleOptimizer(opt, "dynamic") with LossScaleOptimizer(opt). If you pass anything other than "dynamic" to the second argument, see (1) of the next section.
- In the following rare cases, you need to make more changes when switching to the non-experimental API:
  - If you passed anything other than "dynamic" to the loss_scale argument (the second argument) of LossScaleOptimizer:
    - The LossScaleOptimizer constructor takes in different arguments. See the TF 2.7 documentation of tf.keras.mixed_precision.experimental.LossScaleOptimizer for details on the differences, which has examples on how to convert to the non-experimental LossScaleOptimizer.
  - If you passed a value to the loss_scale argument (the second argument) of Policy:
    - The experimental version of Policy optionally took in a tf.compat.v1.mixed_precision.LossScale in the constructor, which defaulted to a dynamic loss scale for the "mixed_float16" policy and no loss scale for other policies. In Model.compile, if the model's policy had a loss scale, the optimizer would be wrapped with a LossScaleOptimizer. With the non-experimental Policy, there is no loss scale associated with the Policy, and Model.compile wraps the optimizer with a LossScaleOptimizer if and only if the policy is a "mixed_float16" policy. If you previously passed a LossScale to the experimental Policy, consider just removing it, as the default loss scaling behavior is usually what you want. If you really want to customize the loss scaling behavior, you can wrap your optimizer with a LossScaleOptimizer before passing it to Model.compile.
  - If you use the very rarely-used function tf.keras.mixed_precision.experimental.get_layer_policy:
    - Replace tf.keras.mixed_precision.experimental.get_layer_policy(layer) with layer.dtype_policy.
tf.mixed_precision.experimental.LossScale and its subclasses have been removed from the TF2 namespace. This symbols were very rarely used and were only useful in TF2 for use in the now-removed tf.keras.mixed_precision.experimental API. The symbols are still available under tf.compat.v1.mixed_precision.
The experimental_relax_shapes heuristic for tf.function has been deprecated and replaced with reduce_retracing which encompasses broader heuristics to reduce the number of retraces (see below)

Major Features and Improvements

tf.keras:
- Added tf.keras.applications.resnet_rs models. This includes the ResNetRS50, ResNetRS101, ResNetRS152, ResNetRS200, ResNetRS270, ResNetRS350 and ResNetRS420 model architectures. The ResNetRS models are based on the architecture described in Revisiting ResNets: Improved Training and Scaling Strategies
- Added tf.keras.optimizers.experimental.Optimizer. The reworked optimizer gives more control over different phases of optimizer calls, and is easier to customize. We provide Adam, SGD, Adadelta, AdaGrad and RMSprop optimizers based on tf.keras.optimizers.experimental.Optimizer. Generally the new optimizers work in the same way as the old ones, but support new constructor arguments. In the future, the symbols tf.keras.optimizers.Optimizer/Adam/etc will point to the new optimizers, and the previous generation of optimizers will be moved to tf.keras.optimizers.legacy.Optimizer/Adam/etc.
- Added L2 unit normalization layer tf.keras.layers.UnitNormalization.
- Added tf.keras.regularizers.OrthogonalRegularizer, a new regularizer that encourages orthogonality between the rows (or columns) or a weight matrix.
- Added tf.keras.layers.RandomBrightness layer for image preprocessing.
- Added APIs for switching between interactive logging and absl logging. By default, Keras always writes the logs to stdout. However, this is not optimal in a non-interactive environment, where you don't have access to stdout, but can only view the logs. You can use tf.keras.utils.disable_interactive_logging() to write the logs to ABSL logging. You can also use tf.keras.utils.enable_interactive_logging() to change it back to stdout, or tf.keras.utils.is_interactive_logging_enabled() to check if interactive logging is enabled.
- Changed default value for the verbose argument of Model.evaluate() and Model.predict() to "auto", which defaults to verbose=1 for most cases and defaults to verbose=2 when used with ParameterServerStrategy or with interactive logging disabled.
- Argument jit_compile in Model.compile() now applies to Model.evaluate() and Model.predict(). Setting jit_compile=True in compile() compiles the model's training, evaluation, and inference steps to XLA. Note that jit_compile=True may not necessarily work for all models.
- Added DTensor-related Keras APIs under tf.keras.dtensor namespace. The APIs are still classified as experimental. You are welcome to try it out. Please check the tutoral and guide on https://www.tensorflow.org/ for more details about DTensor.
tf.lite:
- Added TFLite builtin op support for the following TF ops:
  - tf.math.argmin/tf.math.argmax for input data type tf.bool on CPU.
  - tf.nn.gelu op for output data type tf.float32 and quantization on CPU.
- Add nominal support for unsigned 16-bit integer tensor types. Note that very few TFLite kernels support this type natively, so its use in mobile ML authoring is generally discouraged.
- Add support for unsigned 16-bit integer tensor types in cast op.
- Experimental support for lowering list_ops.tensor_list_set_item with DynamicUpdateSlice.
- Enabled a new MLIR-based dynamic range quantization backend by default
  - The new backend is used for post-training int8 dynamic range quantization and post-training float16 quantization.
  - Set experimental_new_dynamic_range_quantizer in tf.lite.TFLiteConverter to False to disable this change
- Native TF Lite variables are now enabled during conversion by default on all v2 TfLiteConverter entry points. experimental_enable_resource_variables on tf.lite.TFLiteConverter is now True by default and will be removed in the future.
tf.function:
- Custom classes used as arguments for tf.function can now specify rules regarding when retracing needs to occur by implementing the Tracing Protocol available through tf.types.experimental.SupportsTracingProtocol.
- TypeSpec classes (as associated with ExtensionTypes) also implement the Tracing Protocol which can be overriden if necessary.
- The newly introduced reduce_retracing option also uses the Tracing Protocol to proactively generate generalized traces similar to experimental_relax_shapes (which has now been deprecated).
Unified eager and tf.function execution:
- Eager mode can now execute each op as a tf.function, allowing for more consistent feature support in future releases.
- It is available for immediate use.
  - See the TF_RUN_EAGER_OP_AS_FUNCTION environment variable in eager context.
  - Eager performance should be similar with this feature enabled.
    - A roughly 5us per-op overhead may be observed when running many small functions.
    - Note a known issue with GPU performance.
  - The behavior of tf.function itself is unaffected.
- Note: This feature will be enabled by default in an upcoming version of TensorFlow.

Bug Fixes and Other Changes

tf.data:
- Fixed bug in tf.data.experimental.parse_example_dataset when tf.io.RaggedFeatures would specify value_key but no partitions. Before the fix, setting value_key but no partitions would result in the feature key being replaced by the value key, e.g. {'value_key': <RaggedTensor>} instead of {'key': <RaggedTensor>}. Now the correct feature key will be used. This aligns the behavior of tf.data.experimental.parse_example_dataset to match the behavior of tf.io.parse_example.
- Added a new field, filter_parallelization, to tf.data.experimental.OptimizationOptions. If it is set to True, tf.data will run Filter transformation with multiple threads. Its default value is False if not specified.
tf.keras:
- Fixed bug in optimizers that prevented them from properly checkpointing slot variables when they are ShardedVariables (used for training with tf.distribute.experimental.ParameterServerStrategy).
tf.random:
- Added tf.random.experimental.index_shuffle, for shuffling a sequence without materializing the sequence in memory.
tf.RaggedTensor:
- Introduced tf.experimental.RowPartition, which encodes how one dimension in a RaggedTensor relates to another, into the public API.
- Introduced tf.experimental.DynamicRaggedShape, which represents the shape of a RaggedTensor.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Open source →

2.8.32 Sept 2022Security

This releases introduces several vulnerability fixes:

Release 2.8.3

This releases introduces several vulnerability fixes:

Fixes a CHECK failure in tf.reshape caused by overflows (CVE-2022-35934)
Fixes a CHECK failure in SobolSample caused by missing validation (CVE-2022-35935)
Fixes an OOB read in Gather_nd op in TF Lite (CVE-2022-35937)
Fixes a CHECK failure in TensorListReserve caused by missing validation (CVE-2022-35960)
Fixes an OOB write in Scatter_nd op in TF Lite (CVE-2022-35939)
Fixes an integer overflow in RaggedRangeOp (CVE-2022-35940)
Fixes a CHECK failure in AvgPoolOp (CVE-2022-35941)
Fixes a CHECK failures in UnbatchGradOp (CVE-2022-35952)
Fixes a segfault TFLite converter on per-channel quantized transposed convolutions (CVE-2022-36027)
Fixes a CHECK failures in AvgPool3DGrad (CVE-2022-35959)
Fixes a CHECK failures in FractionalAvgPoolGrad (CVE-2022-35963)
Fixes a segfault in BlockLSTMGradV2 (CVE-2022-35964)
Fixes a segfault in LowerBound and UpperBound (CVE-2022-35965)
Fixes a segfault in QuantizedAvgPool (CVE-2022-35966)
Fixes a segfault in QuantizedAdd (CVE-2022-35967)
Fixes a CHECK fail in AvgPoolGrad (CVE-2022-35968)
Fixes a CHECK fail in Conv2DBackpropInput (CVE-2022-35969)
Fixes a segfault in QuantizedInstanceNorm (CVE-2022-35970)
Fixes a CHECK fail in FakeQuantWithMinMaxVars (CVE-2022-35971)
Fixes a segfault in Requantize (CVE-2022-36017)
Fixes a segfault in QuantizedBiasAdd (CVE-2022-35972)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsPerChannel (CVE-2022-36019)
Fixes a segfault in QuantizedMatMul (CVE-2022-35973)
Fixes a segfault in QuantizeDownAndShrinkRange (CVE-2022-35974)
Fixes segfaults in QuantizedRelu and QuantizedRelu6 (CVE-2022-35979)
Fixes a CHECK fail in FractionalMaxPoolGrad (CVE-2022-35981)
Fixes a CHECK fail in RaggedTensorToVariant (CVE-2022-36018)
Fixes a CHECK fail in QuantizeAndDequantizeV3 (CVE-2022-36026)
Fixes a segfault in SparseBincount (CVE-2022-35982)
Fixes a CHECK fail in Save and SaveSlices (CVE-2022-35983)
Fixes a CHECK fail in ParameterizedTruncatedNormal (CVE-2022-35984)
Fixes a CHECK fail in LRNGrad (CVE-2022-35985)
Fixes a segfault in RaggedBincount (CVE-2022-35986)
Fixes a CHECK fail in DenseBincount (CVE-2022-35987)
Fixes a CHECK fail in tf.linalg.matrix_rank (CVE-2022-35988)
Fixes a CHECK fail in MaxPool (CVE-2022-35989)
Fixes a CHECK fail in Conv2DBackpropInput (CVE-2022-35999)
Fixes a CHECK fail in EmptyTensorList (CVE-2022-35998)
Fixes a CHECK fail in tf.sparse.cross (CVE-2022-35997)
Fixes a floating point exception in Conv2D (CVE-2022-35996)
Fixes a CHECK fail in AudioSummaryV2 (CVE-2022-35995)
Fixes a CHECK fail in CollectiveGather (CVE-2022-35994)
Fixes a CHECK fail in SetSize (CVE-2022-35993)
Fixes a CHECK fail in TensorListFromTensor (CVE-2022-35992)
Fixes a CHECK fail in TensorListScatter and TensorListScatterV2 (CVE-2022-35991)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsPerChannelGradient (CVE-2022-35990)
Fixes a CHECK fail in FakeQuantWithMinMaxVarsGradient (CVE-2022-36005)
Fixes a CHECK fail in tf.random.gamma (CVE-2022-36004)
Fixes a CHECK fail in RandomPoissonV2 (CVE-2022-36003)
Fixes a CHECK fail in Unbatch (CVE-2022-36002)
Fixes a CHECK fail in DrawBoundingBoxes (CVE-2022-36001)
Fixes a CHECK fail in Eig (CVE-2022-36000)
Fixes a null dereference on MLIR on empty function attributes (CVE-2022-36011)
Fixes an assertion failure on MLIR empty edge names (CVE-2022-36012)
Fixes a null-dereference in mlir::tfg::GraphDefImporter::ConvertNodeDef (CVE-2022-36013)
Fixes a null-dereference in mlir::tfg::TFOp::nameAttr (CVE-2022-36014)
Fixes an integer overflow in math ops (CVE-2022-36015)
Fixes a CHECK-fail in tensorflow::full_type::SubstituteFromAttrs (CVE-2022-36016)
Fixes an OOB read in Gather_nd op in TF Lite Micro (CVE-2022-35938)

Open source →

tensorflow-cpu

Release timeline

Releases

From the same repository

Release 2.21.0

TensorFlow

Breaking Changes

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors

Release 2.21.0

TensorFlow

Breaking Changes

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors

Release 2.21.0

TensorFlow

Breaking Changes

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors

Release 2.20.0

TensorFlow

Breaking Changes

Major Features and Improvements

Thanks to our Contributors

Release 2.20.0

TensorFlow

Breaking Changes

Major Features and Improvements

Thanks to our Contributors

Release 2.19.1

Bug Fixes and Other Changes

Release 2.19.0

TensorFlow

Breaking Changes

Known Caveats

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors

Release 2.19.0

TensorFlow

Breaking Changes

Known Caveats

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors

Release 2.18.1

Security

Bug Fixes and Other Changes

Breaking Changes

Release 2.18.0

TensorFlow

Breaking Changes

Known Caveats

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors

Release 2.18.0

TensorFlow

Breaking Changes

Known Caveats

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors

Release 2.18.0

TensorFlow

Breaking Changes

Known Caveats

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors

Release 2.18.0

TensorFlow

Breaking Changes

Known Caveats

Major Features and Improvements

Bug Fixes and Other Changes

Thanks to our Contributors