Release 2.10.0-rc2

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(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:

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