Skip to content

Latest commit

 

History

History
4883 lines (4429 loc) · 336 KB

RELEASE.md

File metadata and controls

4883 lines (4429 loc) · 336 KB
Raw

Release 2.6.0

Breaking Changes

* *<THIS SECTION SHOULD CONTAIN API, ABI AND BEHAVIORAL BREAKING CHANGES>

Known Caveats

*<CAVEATS REGARDING THE RELEASE (BUT NOT BREAKING CHANGES).> *<ADDING/BUMPING DEPENDENCIES SHOULD GO HERE> *<KNOWN LACK OF SUPPORT ON SOME PLATFORM, SHOULD GO HERE>

Major Features and Improvements

*<INSERT MAJOR FEATURE HERE, USING MARKDOWN SYNTAX> *<IF RELEASE CONTAINS MULTIPLE FEATURES FROM SAME AREA, GROUP THEM TOGETHER>

  • tf.keras:
    • tf.keras.utils.experimental.DatasetCreator now takes an optional tf.distribute.InputOptions for specific options when used with distribution.

Bug Fixes and Other Changes

*<SIMILAR TO ABOVE SECTION, BUT FOR OTHER IMPORTANT CHANGES / BUG FIXES> *<IF A CHANGE CLOSES A GITHUB ISSUE, IT SHOULD BE DOCUMENTED HERE> *

  • TF Core:
    • Added tf.saved_model.experimental.TrackableResource, which allows the creation of custom wrapper objects for resource tensors.

Thanks to our Contributors

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

, , , , ,

Release 2.5.0

Breaking Changes

  • The TF_CPP_MIN_VLOG_LEVEL environment variable has been renamed to to TF_CPP_MAX_VLOG_LEVEL which correctly describes its effect.

Known Caveats

  • <CAVEATS REGARDING THE RELEASE (BUT NOT BREAKING CHANGES).>
  • <ADDING/BUMPING DEPENDENCIES SHOULD GO HERE>
  • <KNWON LACK OF SUPPORT ON SOME PLATFORM, SHOULD GO HERE>

Major Features and Improvements

  • <INSERT MAJOR FEATURE HERE, USING MARKDOWN SYNTAX>

  • <IF RELEASE CONTAINS MULTIPLE FEATURES FROM SAME AREA, GROUP THEM TOGETHER>

  • TPU embedding support

    • Added profile_data_directory to EmbeddingConfigSpec in _tpu_estimator_embedding.py. This allows embedding lookup statistics gathered at runtime to be used in embedding layer partitioning decisions.

  • tf.keras.metrics.AUC now support logit predictions.

  • Creating tf.random.Generator under tf.distribute.Strategy scopes is now allowed (except for tf.distribute.experimental.CentralStorageStrategy and tf.distribute.experimental.ParameterServerStrategy). Different replicas will get different random-number streams.

  • tf.data:

    • tf.data service now supports strict round-robin reads, which is useful for synchronous training workloads where example sizes vary. With strict round robin reads, users can guarantee that consumers get similar-sized examples in the same step.
    • tf.data service now supports optional compression. Previously data would always be compressed, but now you can disable compression by passing compression=None to tf.data.experimental.service.distribute(...).
    • tf.data.Dataset.batch() now supports num_parallel_calls and deterministic arguments. num_parallel_calls is used to indicate that multiple input batches should be computed in parallel. With num_parallel_calls set, deterministic is used to indicate that outputs can be obtained in the non-deterministic order.
    • Options returned by tf.data.Dataset.options() are no longer mutable.
    • tf.data input pipelines can now be executed in debug mode, which disables any asynchrony, parallelism, or non-determinism and forces Python execution (as opposed to trace-compiled graph execution) of user-defined functions passed into transformations such as map. The debug mode can be enabled through tf.data.experimental.enable_debug_mode().

  • tf.lite

    • Enabled the new MLIR-based quantization backend by default
      • The new backend is used for 8 bits full integer post-training quantization
      • The new backend removes the redundant rescales and fixes some bugs (shared weight/bias, extremely small scales, etc)
      • Set experimental_new_quantizer in tf.lite.TFLiteConverter to False to disable this change

  • tf.keras

    • Enabled a new supported input type in Model.fit, tf.keras.utils.experimental.DatasetCreator, which takes a callable, dataset_fn. DatasetCreator is intended to work across all tf.distribute strategies, and is the only input type supported for Parameter Server strategy.

  • tf.distribute

    • tf.distribute.experimental.ParameterServerStrategy now supports training with Keras Model.fit when used with DatasetCreator.

  • PluggableDevice

  • oneAPI Deep Neural Network Library (oneDNN) CPU performance optimizations from Intel-optimized TensorFlow are now available in the official x86-64 Linux and Windows builds.

    • They are off by default. Enable them by setting the environment variable TF_ENABLE_ONEDNN_OPTS=1.
    • We do not recommend using them in GPU systems, as they have not been sufficiently tested with GPUs yet.

Bug Fixes and Other Changes

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

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

  • tf.keras:

    • Preprocessing layers API consistency changes:
      • StringLookup added output_mode, sparse, and pad_to_max_tokens arguments with same semantics as TextVectorization.
      • IntegerLookup added output_mode, sparse, and pad_to_max_tokens arguments with same semantics as TextVectorization. Renamed max_values, oov_value and mask_value to max_tokens, oov_token and mask_token to align with StringLookup and TextVectorization.
      • TextVectorization default for pad_to_max_tokens switched to False.
      • CategoryEncoding no longer supports adapt, IntegerLookup now supports equivalent functionality. max_tokens argument renamed to num_tokens.
      • Discretization added num_bins argument for learning bins boundaries through calling adapt on a dataset. Renamed bins argument to bin_boundaries for specifying bins without adapt.
    • Improvements to model saving/loading:
      • model.load_weights now accepts paths to saved models.
    • Keras inputs can now be created directly from arbitrary tf.TypeSpecs.
    • Two new learning rate schedules added: tf.keras.optimizers.schedules.CosineDecay and tf.keras.optimizers.schedules.CosineDecayRestarts.

  • tf.data:

    • Exposing tf.data.experimental.ExternalStatePolicy, which can be used to control how external state should be handled during dataset serialization or iterator checkpointing.
    • Changing tf.data.experimental.save to store the type specification of the dataset elements. This avoids the need for explicitly specifying the element_spec argument of tf.data.experimental.load when loading the previously saved dataset.
    • Add .element_spec property to tf.data.DatasetSpec to access the inner spec. This can be used to extract the structure of nested datasets.
    • Add tf.data.experimental.AutoShardingPolicy.HINT which can be used to provide hints to tf.distribute-based auto-sharding as to where in the input pipeline to insert sharding transformations.
    • Make tf.data.Options persistent across tf.function and GraphDef boundaries.
    • If autotuning is ON, tf.data.Dataset.map() will default to be parallelized unless users turn off the optimization option map_parallelization explicitly.

  • XLA compilation:

    • tf.function(experimental_compile=True) has become a stable API, renamed tf.function(jit_compile=True).
    • XLA can now compile MirroredStrategy: the step function passed to strategy.run can now be annoted with jit_compile=True.

  • tf.distribute:

    • Rename experimental_prefetch_to_device in tf.distribute.InputOptions to experimental_fetch_to_device to better reflect the purpose.

  • tf.lite:

    • class tflite::Subgraph:
      • Removed the tensors() method and the non-const overload of the nodes_and_registration() method, both of which were previously documented as temporary and to be removed.
        • Uses of tensors() can be replaced by calling the existing methods tensors_size() and tensor(int).
        • Uses of the non-const overload of nodes_and_registration can be replaced by calling the existing methods nodes_size() and context(), and then calling the GetNodeAndRegistration method in the TfLiteContext returned by context().
    • NNAPI
      • Removed deprecated Interpreter::UseNNAPI(bool) C++ API.
        • Use NnApiDelegate() and related delegate configuration methods directly.
      • Replaced the model cache key for models computation algorithm with one guaranteed to be stable across runs.
    • 16 bits quantization
      • Added int16x8 support for ABS, REDUCE_MAX and REDUCE_MIN operators.
      • Additional tests and fixes for ADD and SUB operators.
    • Added support for saved model's session initializer through TFLiteConverter.from_saved_model.
    • Added DEPTH_TO_SPACE support in Post training quantization.
    • Added dynamic range quantization support for the BatchMatMul op.
      • Both symmetric and asymmetric quantized input tensor are supported.
    • Add RFFT2D as builtin op. (RFFT2D also supports RFFTD.) Currently only supports float32 input.
    • Add 5D support to SLICE op.
    • TFLite Supports SingatureDef:
      • TFLiteConverter exports models with SignatureDef
      • Interpreter supports getting a list of signatures and getting callable function for a given signaturedef.
    • Add int8 support for ReshapeV2.
    • Add experimental support for optimization with sparsity.
    • Add nominal support for unsigned 32-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 static hash tables through TFLiteConverter.from_saved_model.

  • The Python TF Lite Interpreter bindings now have an option experimental_preserve_all_tensors to aid in debugging conversion. * Quantized x86 execution defaults to Ruy GEMM library for platforms with AVX support. * Deprecate tf.compat.v1.lite.experimental.get_potentially_supported_ops. Use tf.lite.TFLiteConverter directly to check whether a model is convertible. * Add support to select one of three different built-in op resolvers to be * Enabled post training with calibrations for models that require user provied TensorFlow Lite custom op libraries via converter.target_spec._experimental_custom_op_registerers. used in Python Interpreter API.

  • TF Core:

    • Several new types have been introduced to describe the return types of tf.function and related APIs:
      • tf.types.experimental.Callable is intended to represent TensorFlow callables in general.
      • tf.types.experimental.ConcreteFunction represents a callable TensofFlow graph function specialized to a particular input.
      • tf.types.experimental.GenericFunction is the output of tf.function and represents a callble that can be backed by multiple ConcreteFunctions (commonly known as traces).
    • Corrected higher-order gradients of control flow constructs (tf.cond, tf.while_loop, and compositions like tf.foldl) computed with tf.GradientTape inside a tf.function.
    • Changed the default step size in gradient_checker_v2.compute_gradients to be exactly representable as a binary floating point numbers. This avoids poluting gradient approximations needlessly, which is some cases leads to false negatives in op gradient tests.
    • Added tf.config.experimental.get_memory_info, returning a dict with the current and peak memory usage. Deprecated tf.config.experimental.get_memory_usage in favor of this new function.
    • Extended tf.config.experimental.enable_tensor_float_32_execution to control Tensor-Float-32 evaluation in RNNs.
    • Added a 'experimental_payloads' field to tf.errors.OpError and its subclasses to support more detailed error reporting. This is inspired from Abseil Status payloads: https://github.com/abseil/abseil-cpp/blob/master/absl/status/status.h
    • Extended tf.matmul to support output_type and mixed input types.

  • tf.summary:

  • New tf.summary.graph allows manual write of TensorFlow graph (tf.Graph or tf.compat.v1.GraphDef) as a summary. This is not a replacement for the trace-based API.

  • Set /d2ReducedOptimizeHugeFunctions by default for Windows builds. This provides a big compile-time speedup, and effectively raises the minimum supported MSVC version to 16.4 (current: 16.8).

  • TensorRT

    • Removed the deprecated session_config parameter for the TF1-TRT converter TrtGraphConverter. Previously, we issued a warning when the value of the parameter is not None.
    • The TF2-TRT converter TrtGraphConverterV2 takes an object of class TrtConversionParams as a parameter. Removed three deprecated fields from this class: rewriter_config_template, is_dynamic_op, and max_batch_size. Previously, we issued a warning when the value of rewriter_config_template is not None. We issued an error when the value of is_dynamic_op is not True. We didn't use the value for max_batch_size for building TensorRT engines. Add parameters use_dynamic_shape to enable dynamic shape support. The default is to disable dynamic shape support. Add dynamic_shape_profile_strategy for selecting a dynamic shape profile strategy. The default is profile strategy is Range.
    • Issue a warning when function get_tensorrt_rewriter_config is used.

  • TF XLA

    • Add new enum value MLIR_BRIDGE_ROLLOUT_SAFE_MODE_ENABLED to tf.config.experimental.mlir_bridge_rollout to enable a "safe" mode. This runs the MLIR bridge only when an analysis of the graph only when an analysis of the graph determines that it is safe to run.
    • Add new enum value 'MLIR_BRIDGE_ROLLOUT_SAFE_MODE_FALLBACK_ENABLED' to tf.config.experimental.mlir_bridge_rollout to enable a fallback for the MLIR bridge in a "safe" mode. This runs the MLIR bridge in a FallbackEnabled mode when an analysis of the graph determines that the graph does not have unsupported features.

  • Other

    • Adding show_debug_info to mlir.convert_graph_def and mlir.convert_function.
    • Added Arm Compute Library (ACL) support to --config=mkl_aarch64 build.

Thanks to our Contributors

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

, , , , ,

Release 2.4.1

  • This release removes the AVX2 requirement from TF 2.4.0.

Release 2.3.2

Bug Fixes and Other Changes

  • Fixes an access to unitialized memory in Eigen code (CVE-2020-26266)
  • Fixes a security vulnerability caused by lack of validation in tf.raw_ops.DataFormatVecPermute and tf.raw_ops.DataFormatDimMap (CVE-2020-26267)
  • Fixes a vulnerability caused by attempting to write to immutable memory region in tf.raw_ops.ImmutableConst (CVE-2020-26268
  • Fixes a CHECK-fail in LSTM with zero-length input (CVE-2020-26270)
  • Fixes a security vulnerability caused by accessing heap data outside of bounds when loading a specially crafted SavedModel (CVE-2020-26271)
  • Solves an OOM issue on TPUs when XLA contexts use fused average updates
  • Updates libjpeg-turbo to 2.0.5 to handle CVE-2020-13790.
  • Updates junit to 4.13.1 to handle CVE-2020-15250.
  • Updates PCRE to 8.44 to handle CVE-2019-20838 and CVE-2020-14155.
  • Updates sqlite3 to 3.44.0 to keep in sync with master branch.

Release 2.2.2

Bug Fixes and Other Changes

  • Fixes an access to unitialized memory in Eigen code (CVE-2020-26266)
  • Fixes a security vulnerability caused by lack of validation in tf.raw_ops.DataFormatVecPermute and tf.raw_ops.DataFormatDimMap (CVE-2020-26267)
  • Fixes a vulnerability caused by attempting to write to immutable memory region in tf.raw_ops.ImmutableConst (CVE-2020-26268
  • Fixes a CHECK-fail in LSTM with zero-length input (CVE-2020-26270)
  • Fixes a security vulnerability caused by accessing heap data outside of bounds when loading a specially crafted SavedModel (CVE-2020-26271)
  • Prevents memory leaks in loading SavedModels that import functions
  • Updates libjpeg-turbo to 2.0.5 to handle CVE-2020-13790.
  • Updates junit to 4.13.1 to handle CVE-2020-15250.
  • Updates PCRE to 8.44 to handle CVE-2019-20838 and CVE-2020-14155.
  • Updates sqlite3 to 3.44.0 to keep in sync with master branch.

Release 2.1.3

Bug Fixes and Other Changes

  • Fixes an access to unitialized memory in Eigen code (CVE-2020-26266)
  • Fixes a security vulnerability caused by lack of validation in tf.raw_ops.DataFormatVecPermute and tf.raw_ops.DataFormatDimMap (CVE-2020-26267)
  • Fixes a vulnerability caused by attempting to write to immutable memory region in tf.raw_ops.ImmutableConst (CVE-2020-26268
  • Fixes a CHECK-fail in LSTM with zero-length input (CVE-2020-26270)
  • Fixes a security vulnerability caused by accessing heap data outside of bounds when loading a specially crafted SavedModel (CVE-2020-26271)
  • Updates libjpeg-turbo to 2.0.5 to handle CVE-2020-13790.
  • Updates junit to 4.13.1 to handle CVE-2020-15250.
  • Updates PCRE to 8.44 to handle CVE-2019-20838 and CVE-2020-14155.
  • Updates sqlite3 to 3.44.0 to keep in sync with master branch.
  • Newer ROCm versions are supported on the 2.1 branch.

Release 2.0.4

Note that this is the last patch release for the TensorFlow 2.0.x series.

Bug Fixes and Other Changes

  • Fixes an access to unitialized memory in Eigen code (CVE-2020-26266)
  • Fixes a security vulnerability caused by lack of validation in tf.raw_ops.DataFormatVecPermute and tf.raw_ops.DataFormatDimMap (CVE-2020-26267)
  • Fixes a vulnerability caused by attempting to write to immutable memory region in tf.raw_ops.ImmutableConst (CVE-2020-26268
  • Fixes a CHECK-fail in LSTM with zero-length input (CVE-2020-26270)
  • Fixes a security vulnerability caused by accessing heap data outside of bounds when loading a specially crafted SavedModel (CVE-2020-26271)
  • Updates libjpeg-turbo to 2.0.5 to handle CVE-2020-13790.
  • Updates junit to 4.13.1 to handle CVE-2020-15250.
  • Updates PCRE to 8.44 to handle CVE-2019-20838 and CVE-2020-14155.
  • Updates sqlite3 to 3.44.0 to keep in sync with master branch.

Release 1.15.5

Note that this is the last patch release for the TensorFlow 1.x series.

Bug Fixes and Other Changes

  • Fixes an access to unitialized memory in Eigen code (CVE-2020-26266)
  • Fixes a security vulnerability caused by lack of validation in tf.raw_ops.DataFormatVecPermute and tf.raw_ops.DataFormatDimMap (CVE-2020-26267)
  • Fixes a vulnerability caused by attempting to write to immutable memory region in tf.raw_ops.ImmutableConst (CVE-2020-26268
  • Fixes a CHECK-fail in LSTM with zero-length input (CVE-2020-26270)
  • Fixes a security vulnerability caused by accessing heap data outside of bounds when loading a specially crafted SavedModel (CVE-2020-26271)
  • Updates libjpeg-turbo to 2.0.5 to handle CVE-2020-13790.
  • Updates junit to 4.13.1 to handle CVE-2020-15250.
  • Updates PCRE to 8.44 to handle CVE-2019-20838 and CVE-2020-14155.
  • Updates sqlite3 to 3.44.0 to keep in sync with master branch.

Release 2.4.0

Major Features and Improvements

Breaking Changes

  • TF Core:

    • Certain float32 ops run in lower precision on Ampere based GPUs, including matmuls and convolutions, due to the use of [TensorFloat-32] (https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/). Specifically, inputs to such ops are rounded from 23 bits of precision to 10 bits of precision. This is unlikely to cause issues in practice for deep learning models. In some cases, TensorFloat-32 is also used for complex64 ops. TensorFloat-32 can be disabled by running tf.config.experimental.enable_tensor_float_32_execution(False).
    • The byte layout for string tensors across the C-API has been updated to match TF Core/C++; i.e., a contiguous array of tensorflow::tstring/TF_TStrings.
    • C-API functions TF_StringDecode, TF_StringEncode, and TF_StringEncodedSize are no longer relevant and have been removed; see core/platform/ctstring.h for string access/modification in C.
    • tensorflow.python, tensorflow.core and tensorflow.compiler modules are now hidden. These modules are not part of TensorFlow public API.
    • tf.raw_ops.Max and tf.raw_ops.Min no longer accept inputs of type tf.complex64 or tf.complex128, because the behavior of these ops is not well defined for complex types.
    • XLA:CPU and XLA:GPU devices are no longer registered by default. Use TF_XLA_FLAGS=--tf_xla_enable_xla_devices if you really need them, but this flag will eventually be removed in subsequent releases.

  • tf.keras:

    • The steps_per_execution argument in model.compile() is no longer experimental; if you were passing experimental_steps_per_execution, rename it to steps_per_execution in your code. This argument controls the number of batches to run during each tf.function call when calling model.fit(). Running multiple batches inside a single tf.function call can greatly improve performance on TPUs or small models with a large Python overhead.
    • A major refactoring of the internals of the Keras Functional API may affect code that is relying on certain internal details:
      • Code that uses isinstance(x, tf.Tensor) instead of tf.is_tensor when checking Keras symbolic inputs/outputs should switch to using tf.is_tensor.
      • Code that is overly dependent on the exact names attached to symbolic tensors (e.g. assumes there will be ":0" at the end of the inputs, treats names as unique identifiers instead of using tensor.ref(), etc.) may break.
      • Code that uses full path for get_concrete_function to trace Keras symbolic inputs directly should switch to building matching tf.TensorSpecs directly and tracing the TensorSpec objects.
      • Code that relies on the exact number and names of the op layers that TensorFlow operations were converted into may have changed.
      • Code that uses tf.map_fn/tf.cond/tf.while_loop/control flow as op layers and happens to work before TF 2.4. These will explicitly be unsupported now. Converting these ops to Functional API op layers was unreliable before TF 2.4, and prone to erroring incomprehensibly or being silently buggy.
      • Code that directly asserts on a Keras symbolic value in cases where ops like tf.rank used to return a static or symbolic value depending on if the input had a fully static shape or not. Now these ops always return symbolic values.
      • Code already susceptible to leaking tensors outside of graphs becomes slightly more likely to do so now.
      • Code that tries directly getting gradients with respect to symbolic Keras inputs/outputs. Use GradientTape on the actual Tensors passed to the already-constructed model instead.
      • Code that requires very tricky shape manipulation via converted op layers in order to work, where the Keras symbolic shape inference proves insufficient.
      • Code that tries manually walking a tf.keras.Model layer by layer and assumes layers only ever have one positional argument. This assumption doesn't hold true before TF 2.4 either, but is more likely to cause issues now.
      • Code that manually enters keras.backend.get_graph() before building a functional model is no longer needed.
      • Start enforcing input shape assumptions when calling Functional API Keras models. This may potentially break some users, in case there is a mismatch between the shape used when creating Input objects in a Functional model, and the shape of the data passed to that model. You can fix this mismatch by either calling the model with correctly-shaped data, or by relaxing Input shape assumptions (note that you can pass shapes with None entries for axes that are meant to be dynamic). You can also disable the input checking entirely by setting model.input_spec = None.
    • Several changes have been made to tf.keras.mixed_precision.experimental. Note that it is now recommended to use the non-experimental tf.keras.mixed_precision API.
    • AutoCastVariable.dtype now refers to the actual variable dtype, not the dtype it will be casted to.
    • When mixed precision is enabled, tf.keras.layers.Embedding now outputs a float16 or bfloat16 tensor instead of a float32 tensor.
    • The property tf.keras.mixed_precision.experimental.LossScaleOptimizer.loss_scale is now a tensor, not a LossScale object. This means to get a loss scale of a LossScaleOptimizer as a tensor, you must now call opt.loss_scaleinstead of opt.loss_scale().
    • The property should_cast_variables has been removed from tf.keras.mixed_precision.experimental.Policy
    • When passing a tf.mixed_precision.experimental.DynamicLossScale to tf.keras.mixed_precision.experimental.LossScaleOptimizer, the DynamicLossScale's multiplier must be 2.
    • When passing a tf.mixed_precision.experimental.DynamicLossScale to tf.keras.mixed_precision.experimental.LossScaleOptimizer, the weights of the DynanmicLossScale are copied into the LossScaleOptimizer instead of being reused. This means modifying the weights of the DynamicLossScale will no longer affect the weights of the LossScaleOptimizer, and vice versa.
    • The global policy can no longer be set to a non-floating point policy in tf.keras.mixed_precision.experimental.set_policy
    • In Layer.call, AutoCastVariables will no longer be casted within MirroredStrategy.run or ReplicaContext.merge_call. This is because a thread local variable is used to determine whether AutoCastVariables are casted, and those two functions run with a different thread. Note this only applies if one of these two functions is called within Layer.call; if one of those two functions calls Layer.call, AutoCastVariables will still be casted.

  • tf.data:

    • tf.data.experimental.service.DispatchServer now takes a config tuple instead of individual arguments. Usages should be updated to tf.data.experimental.service.DispatchServer(dispatcher_config).
    • tf.data.experimental.service.WorkerServer now takes a config tuple instead of individual arguments. Usages should be updated to tf.data.experimental.service.WorkerServer(worker_config).

  • tf.distribute:

    • Removes tf.distribute.Strategy.experimental_make_numpy_dataset. Please use tf.data.Dataset.from_tensor_slices instead.
    • Renames experimental_hints in tf.distribute.StrategyExtended.reduce_to, tf.distribute.StrategyExtended.batch_reduce_to, tf.distribute.ReplicaContext.all_reduce to options.
    • Renames tf.distribute.experimental.CollectiveHints to tf.distribute.experimental.CommunicationOptions.
    • Renames tf.distribute.experimental.CollectiveCommunication to tf.distribute.experimental.CommunicationImplementation.
    • Renames tf.distribute.Strategy.experimental_distribute_datasets_from_function to distribute_datasets_from_function as it is no longer experimental.
    • Removes tf.distribute.Strategy.experimental_run_v2 method, which was deprecated in TF 2.2.

  • tf.lite:

    • tf.quantization.quantize_and_dequantize_v2 has been introduced, which updates the gradient definition for quantization which is outside the range to be 0. To simulate the V1 the behavior of tf.quantization.quantize_and_dequantize(...) use tf.grad_pass_through(tf.quantization.quantize_and_dequantize_v2)(...).

  • Building TensorFlow:

    • Windows platform builds: TensorFlow on Windows under MSVC is now built with --copt=/experimental:preprocessor --host_copt=/experimental:preprocessor (see .bazelrc for more details). Builds including TensorFlow may fail with unexpected syntax errors if these flags are absent. See also this thread on SIG Build.

Known Caveats

  • tf.keras.mixed_precision
    • When using mixed precision, calling RMSprop.apply_gradients or Nadam.apply_gradients outside a tf.function does not work and will raise the AttributeError "Tensor.op is meaningless when eager execution is enabled". See this issue for details and a workaround.

Bug Fixes and Other Changes

TF Core:

  • Introduces experimental support for a new module named [tf.experimental.numpy] (https://www.tensorflow.org/api_docs/python/tf/experimental/numpy), which is a NumPy-compatible API for writing TF programs. This module provides class ndarray, which mimics the ndarray class in NumPy, and wraps an immutable tf.Tensor under the hood. A subset of NumPy functions (e.g. numpy.add) are provided. Their inter-operation with TF facilities is seamless in most cases. See tensorflow/python/ops/numpy_ops/README.md for details of what operations are supported and what are the differences from NumPy.
  • tf.types.experimental.TensorLike is a new Union type that can be used as type annotation for variables representing a Tensor or a value that can be converted to Tensor by tf.convert_to_tensor.
  • Calling ops with a python constants or numpy values is now consistent with tf.convert_to_tensor behavior. This avoids operations like tf.reshape truncating inputs such as from int64 to int32.
  • Adds tf.sparse.map_values to apply a function to the .values of SparseTensor arguments.
  • The Python bitwise operators for Tensor (__and__, __or__, __xor__ and __invert__ now support non-bool arguments and apply the corresponding bitwise ops. bool arguments continue to be supported and dispatch to logical ops. This brings them more in line with Python and NumPy behavior.
  • Adds tf.SparseTensor.with_values. This returns a new SparseTensor with the same sparsity pattern, but with new provided values. It is similar to the with_values function of RaggedTensor.
  • Adds StatelessCase op, and uses it if none of case branches has stateful ops.
  • Adds tf.config.experimental.get_memory_usage to return total memory usage of the device.
  • Adds gradients for RaggedTensorToVariant and RaggedTensorFromVariant.
  • Improve shape inference of nested function calls by supporting constant folding across Arg nodes which makes more static values available to shape inference functions.
  • tf.debugging:
    • tf.debugging.assert_shapes() now works on SparseTensors (Fixes #36268).
  • GPU
    • Adds Support for TensorFloat-32 on Ampere based GPUs.TensorFloat-32, or TF32 for short, is a math mode for NVIDIA Ampere based GPUs which causes certain float32 ops, such as matrix multiplications and convolutions, to run much faster on Ampere GPUs but with reduced precision. This reduced precision has not been found to effect convergence quality of deep learning models in practice. TensorFloat-32 is enabled by default, but can be disabled with tf.config.experimental.enable_tensor_float_32_execution.
  • tf.math:
    • Adds tf.math.erfcinv, the inverse to tf.math.erfc.
  • tf.nn:
    • tf.nn.max_pool2d now supports explicit padding.
  • tf.image:
    • Adds deterministic tf.image.stateless_random_* functions for each tf.image.random_* function. Added a new op stateless_sample_distorted_bounding_box which is a deterministic version of sample_distorted_bounding_box op. Given the same seed, these stateless functions/ops produce the same results independent of how many times the function is called, and independent of global seed settings.
    • Adds deterministic tf.image.resize backprop CUDA kernels for method=ResizeMethod.BILINEAR (the default method). Enable by setting the environment variable TF_DETERMINISTIC_OPS to "true" or "1".
  • tf.print:
    • Bug fix in tf.print() with OrderedDict where if an OrderedDict didn't have the keys sorted, the keys and values were not being printed in accordance with their correct mapping.
  • tf.train.Checkpoint:
    • Now accepts a root argument in the initialization, which generates a checkpoint with a root object. This allows users to create a Checkpoint object that is compatible with Keras model.save_weights() and model.load_weights. The checkpoint is also compatible with the checkpoint saved in the variables/ folder in the SavedModel.
    • When restoring, save_path can be a path to a SavedModel. The function will automatically find the checkpoint in the SavedModel.

tf.data:

  • Adds new tf.data.experimental.service.register_dataset and tf.data.experimental.service.from_dataset_id APIs to enable one process to register a dataset with the tf.data service, and another process to consume data from the dataset.
  • Adds support for dispatcher fault tolerance. To enable fault tolerance, configure a work_dir when running your dispatcher server and set dispatcher_fault_tolerance=True. The dispatcher will store its state to work_dir, so that on restart it can continue from its previous state after restart.
  • Adds support for sharing dataset graphs via shared filesystem instead of over RPC. This reduces load on the dispatcher, improving performance of distributing datasets. For this to work, the dispatcher's work_dir must be accessible from workers. If the worker fails to read from the work_dir, it falls back to using RPC for dataset graph transfer.
  • Adds support for a new "distributed_epoch" processing mode. This processing mode distributes a dataset across all tf.data workers, instead of having each worker process the full dataset. See the tf.data service docs to learn more.
  • Adds optional exclude_cols parameter to CsvDataset. This parameter is the complement of select_cols; at most one of these should be specified.
  • We have implemented an optimization which reorders data-discarding transformations such as take and shard to happen earlier in the dataset when it is safe to do so. The optimization can be disabled via the experimental_optimization.reorder_data_discarding_ops dataset option.
  • tf.data.Options were previously immutable and can now be overridden.
  • tf.data.Dataset.from_generator now supports Ragged and Sparse tensors with a new output_signature argument, which allows from_generator to produce any type describable by a tf.TypeSpec.
  • tf.data.experimental.AUTOTUNE is now available in the core API as tf.data.AUTOTUNE.

tf.distribute:

  • Introduces experimental support for asynchronous training of models via tf.distribute.experimental.ParameterServerStrategy:
    • Replaces the existing tf.distribute.experimental.ParameterServerStrategy symbol with a new class that is for parameter server training in TF2. Usage of the old symbol, usually with Estimator API, should be replaced with [tf.compat.v1.distribute.experimental.ParameterServerStrategy].
    • Added tf.distribute.experimental.coordinator.* namespace, including the main API ClusterCoordinator for coordinating the training cluster, the related data structure RemoteValue and PerWorkerValue.
  • MultiWorkerMirroredStrategy](https://www.tensorflow.org/api_docs/python/tf/distribute/MultiWorkerMirroredStrategy) is now a stable API and is no longer considered experimental. Some of the major improvements involve handling peer failure and many bug fixes. Please check out the detailed tutorial on Multi-worer training with Keras.
  • Adds tf.distribute.Strategy.gather and tf.distribute.ReplicaContext.all_gather APIs to support gathering dense distributed values.
  • Fixes various issues with saving a distributed model.

tf.keras:

  • Improvements from the Functional API refactoring:
    • Functional model construction does not need to maintain a global workspace graph, removing memory leaks especially when building many models or very large models.
    • Functional model construction should be ~8-10% faster on average.
    • Functional models can now contain non-symbolic values in their call inputs inside of the first positional argument.
    • Several classes of TF ops that were not reliably converted to Keras layers during functional API construction should now work, e.g.tf.image.ssim_multiscale
    • Error messages when Functional API construction goes wrong (and when ops cannot be converted to Keras layers automatically) should be clearer and easier to understand.
  • Optimizer.minimize can now accept a loss Tensor and a GradientTape as an alternative to accepting a callable loss.
  • Adds beta hyperparameter to FTRL optimizer classes (Keras and others) to match FTRL paper.
  • Optimizer.__init__ now accepts a gradient_aggregator to allow for customization of how gradients are aggregated across devices, as well as gradients_transformers to allow for custom gradient transformations (such as gradient clipping).
  • Improvements to Keras preprocessing layers:
    • TextVectorization can now accept a vocabulary list or file as an init arg.
    • Normalization can now accept mean and variance values as init args.
  • In Attention and AdditiveAttention layers, the call() method now accepts a return_attention_scores argument. When set to True, the layer returns the attention scores as an additional output argument.
  • Adds tf.metrics.log_cosh and tf.metrics.logcosh API entrypoints with the same implementation as their tf.losses equivalent.
  • For Keras model, the individual call of Model.evaluate uses no cached data for evaluation, while Model.fit uses cached data when validation_data arg is provided for better performance.
  • Adds a save_traces argument to model.save/ tf.keras.models.save_model which determines whether the SavedModel format stores the Keras model/layer call functions. The traced functions allow Keras to revive custom models and layers without the original class definition, but if this isn't required the tracing can be disabled with the added option.
  • The tf.keras.mixed_precision API is now non-experimental. The non-experimental API differs from the experimental API in several ways.
    • tf.keras.mixed_precision.Policy no longer takes in a tf.mixed_precision. experimental.LossScale in the constructor, and no longer has a LossScale associated with it. Instead, Model.compile will automatically wrap the optimizer with a LossScaleOptimizer using dynamic loss scaling if Policy.name is "mixed_float16".
    • tf.keras.mixed_precision.LossScaleOptimizer's constructor takes in different arguments. In particular, it no longer takes in a LossScale, and there is no longer a LossScale associated with the LossScaleOptimizer. Instead, LossScaleOptimizer directly implements fixed or dynamic loss scaling. See the documentation of [tf.keras.mixed_precision.experimental.LossScaleOptimizer] (https://www.tensorflow.org/api_docs/python/tf/keras/mixed_precision/experimental/LossScaleOptimizer?version=nightly) for details on the differences between the experimental LossScaleOptimizer and the new non-experimental LossScaleOptimizer.
    • tf.mixed_precision.experimental.LossScale and its subclasses are deprecated, as all of its functionality now exists within tf.keras.mixed_precision.LossScaleOptimizer

tf.lite:

  • TFLiteConverter:
    • Support optional flags inference_input_type and inference_output_type for full integer quantized models. This allows users to modify the model input and output type to integer types (tf.int8, tf.uint8) instead of defaulting to float type (tf.float32).
  • NNAPI
    • Adds NNAPI Delegation support for requantization use cases by converting the operation into a dequantize-quantize pair.
    • Removes deprecated Interpreter.setUseNNAPI(boolean) Java API. Use Interpreter.Options.setUseNNAPI instead.
    • Deprecates Interpreter::UseNNAPI(bool) C++ API. Use NnApiDelegate() and related delegate configuration methods directly.
    • Deprecates Interpreter::SetAllowFp16PrecisionForFp32(bool) C++ API. Prefer controlling this via delegate options, e.g. tflite::StatefulNnApiDelegate::Options::allow_fp16' or TfLiteGpuDelegateOptionsV2::is_precision_loss_allowed`.
  • GPU
    • GPU acceleration now supports quantized models by default
  • DynamicBuffer::AddJoinedString() will now add a separator if the first string to be joined is empty.
  • Adds support for cumulative sum (cumsum), both as builtin op and MLIR conversion.

TensorRT

  • Issues a warning when the session_config parameter for the TF1 converter is used or the rewrite_config_template field in the TF2 converter parameter object is used.

TPU Enhancements:

  • Adds support for the beta parameter of the FTRL optimizer for TPU embeddings. Users of other TensorFlow platforms can implement equivalent behavior by adjusting the l2 parameter.

XLA Support:

  • xla.experimental.compile is deprecated, use tf.function(experimental_compile=True) instead.
  • Adds tf.function.experimental_get_compiler_ir which returns compiler IR (currently 'hlo' and 'optimized_hlo') for given input for given function.

Security:

Other:

  • We have replaced uses of "whitelist" and "blacklist" with "allowlist" and "denylist" where possible. Please see this list for more context.
  • Adds tf.config.experimental.mlir_bridge_rollout which will help us rollout the new MLIR TPU bridge.
  • Adds tf.experimental.register_filesystem_plugin to load modular filesystem plugins from Python

Thanks to our Contributors

This release contains contributions from many people at Google as well as the following external contributors:

8bitmp3, aaa.jq, Abhineet Choudhary, Abolfazl Shahbazi, acxz, Adam Hillier, Adrian Garcia Badaracco, Ag Ramesh, ahmedsabie, Alan Anderson, Alexander Grund, Alexandre Lissy, Alexey Ivanov, Amedeo Cavallo, anencore94, Aniket Kumar Singh, Anthony Platanios, Ashwin Phadke, Balint Cristian, Basit Ayantunde, bbbboom, Ben Barsdell, Benjamin Chetioui, Benjamin Peterson, bhack, Bhanu Prakash Bandaru Venkata, Biagio Montaruli, Brent M. Spell, bubblebooy, bzhao, cfRod, Cheng Chen, Cheng(Kit) Chen, Chris Tessum, Christian, chuanqiw, codeadmin_peritiae, COTASPAR, CuiYifeng, danielknobe, danielyou0230, dannyfriar, daria, DarrenZhang01, Denisa Roberts, dependabot[bot], Deven Desai, Dmitry Volodin, Dmitry Zakharov, drebain, Duncan Riach, Eduard Feicho, Ehsan Toosi, Elena Zhelezina, emlaprise2358, Eugene Kuznetsov, Evaderan-Lab, Evgeniy Polyakov, Fausto Morales, Felix Johnny, fo40225, Frederic Bastien, Fredrik Knutsson, fsx950223, Gaurav Singh, Gauri1 Deshpande, George Grzegorz Pawelczak, gerbauz, Gianluca Baratti, Giorgio Arena, Gmc2, Guozhong Zhuang, Hannes Achleitner, Harirai, HarisWang, Harsh188, hedgehog91, Hemal Mamtora, Hideto Ueno, Hugh Ku, Ian Beauregard, Ilya Persky, jacco, Jakub Beránek, Jan Jongboom, Javier Montalt Tordera, Jens Elofsson, Jerry Shih, jerryyin, jgehw, Jinjing Zhou, jma, jmsmdy, Johan Nordström, John Poole, Jonah Kohn, Jonathan Dekhtiar, jpodivin, Jung Daun, Kai Katsumata, Kaixi Hou, Kamil Rakoczy, Kaustubh Maske Patil, Kazuaki Ishizaki, Kedar Sovani, Koan-Sin Tan, Koki Ibukuro, Krzysztof Laskowski, Kushagra Sharma, Kushan Ahmadian, Lakshay Tokas, Leicong Li, levinxo, Lukas Geiger, Maderator, Mahmoud Abuzaina, Mao Yunfei, Marius Brehler, markf, Martin Hwasser, Martin Kubovčík, Matt Conley, Matthias, mazharul, mdfaijul, Michael137, MichelBr, Mikhail Startsev, Milan Straka, Ml-0, Myung-Hyun Kim, Måns Nilsson, Nathan Luehr, ngc92, nikochiko, Niranjan Hasabnis, nyagato_00, Oceania2018, Oleg Guba, Ongun Kanat, OscarVanL, Patrik Laurell, Paul Tanger, Peter Sobot, Phil Pearl, PlusPlusUltra, Poedator, Prasad Nikam, Rahul-Kamat, Rajeshwar Reddy T, redwrasse, Rickard, Robert Szczepanski, Rohan Lekhwani, Sam Holt, Sami Kama, Samuel Holt, Sandeep Giri, sboshin, Sean Settle, settle, Sharada Shiddibhavi, Shawn Presser, ShengYang1, Shi,Guangyong, Shuxiang Gao, Sicong Li, Sidong-Wei, Srihari Humbarwadi, Srinivasan Narayanamoorthy, Steenu Johnson, Steven Clarkson, stjohnso98, Tamas Bela Feher, Tamas Nyiri, Tarandeep Singh, Teng Lu, Thibaut Goetghebuer-Planchon, Tim Bradley, Tomasz Strejczek, Tongzhou Wang, Torsten Rudolf, Trent Lo, Ty Mick, Tzu-Wei Sung, Varghese, Jojimon, Vignesh Kothapalli, Vishakha Agrawal, Vividha, Vladimir Menshakov, Vladimir Silyaev, VoVAllen, Võ Văn Nghĩa, wondertx, xiaohong1031, Xiaoming (Jason) Cui, Xinan Jiang, Yair Ehrenwald, Yasir Modak, Yasuhiro Matsumoto, Yimei Sun, Yiwen Li, Yixing, Yoav Ramon, Yong Tang, Yong Wu, yuanbopeng, Yunmo Koo, Zhangqiang, Zhou Peng, ZhuBaohe, zilinzhu, zmx

Release 2.3.1

Bug Fixes and Other Changes

Release 2.2.1

Bug Fixes and Other Changes

Release 2.1.2

Bug Fixes and Other Changes

Release 2.0.3

Bug Fixes and Other Changes

Release 1.15.4

Bug Fixes and Other Changes

Release 2.3.0

Major Features and Improvements

  • tf.data adds two new mechanisms to solve input pipeline bottlenecks and save resources:

    In addition checkout the detailed guide for analyzing input pipeline performance with TF Profiler.

  • tf.distribute.TPUStrategy is now a stable API and no longer considered experimental for TensorFlow. (earlier tf.distribute.experimental.TPUStrategy).

  • TF Profiler introduces two new tools: a memory profiler to visualize your model’s memory usage over time and a python tracer which allows you to trace python function calls in your model. Usability improvements include better diagnostic messages and profile options to customize the host and device trace verbosity level.

  • Introduces experimental support for Keras Preprocessing Layers API (tf.keras.layers.experimental.preprocessing.*) to handle data preprocessing operations, with support for composite tensor inputs. Please see below for additional details on these layers.

  • TFLite now properly supports dynamic shapes during conversion and inference. We’ve also added opt-in support on Android and iOS for XNNPACK, a highly optimized set of CPU kernels, as well as opt-in support for executing quantized models on the GPU.

  • Libtensorflow packages are available in GCS starting this release. We have also started to release a nightly version of these packages.

  • The experimental Python API tf.debugging.experimental.enable_dump_debug_info() now allows you to instrument a TensorFlow program and dump debugging information to a directory on the file system. The directory can be read and visualized by a new interactive dashboard in TensorBoard 2.3 called Debugger V2, which reveals the details of the TensorFlow program including graph structures, history of op executions at the Python (eager) and intra-graph levels, the runtime dtype, shape, and numerical composition of tensors, as well as their code locations.

Breaking Changes

  • Increases the minimum bazel version required to build TF to 3.1.0.
  • tf.data
    • Makes the following (breaking) changes to the tf.data.
    • C++ API: - IteratorBase::RestoreInternal, IteratorBase::SaveInternal, and DatasetBase::CheckExternalState become pure-virtual and subclasses are now expected to provide an implementation.
    • The deprecated DatasetBase::IsStateful method is removed in favor of DatasetBase::CheckExternalState.
    • Deprecated overrides of DatasetBase::MakeIterator and MakeIteratorFromInputElement are removed.
    • The signature of tensorflow::data::IteratorBase::SaveInternal and tensorflow::data::IteratorBase::SaveInput has been extended with SerializationContext argument to enable overriding the default policy for the handling external state during iterator checkpointing. This is not a backwards compatible change and all subclasses of IteratorBase need to be updated accordingly.
  • tf.keras
    • Add a new BackupAndRestore callback for handling distributed training failures & restarts. Please take a look at this tutorial for details on how to use the callback.
  • tf.image.extract_glimpse has been updated to correctly process the case where centered=False and normalized=False. This is a breaking change as the output is different from (incorrect) previous versions. Note this breaking change only impacts tf.image.extract_glimpse and tf.compat.v2.image.extract_glimpse API endpoints. The behavior of tf.compat.v1.image.extract_glimpse does not change. The behavior of existing C++ kernel ExtractGlimpse does not change either, so saved models using tf.raw_ops.ExtractGlimpse will not be impacted.

Known Caveats

  • tf.lite
    • Keras-based LSTM models must be converted with an explicit batch size in the input layer.

Bug Fixes and Other Changes

TF Core:

  • Set tf2_behavior to 1 to enable V2 for early loading cases.
  • Add execute_fn_for_device function to dynamically choose the implementation based on underlying device placement.
  • Eager:
    • Add reduce_logsumexp benchmark with experiment compile.
    • Give EagerTensors a meaningful __array__ implementation.
    • Add another version of defun matmul for performance analysis.
  • tf.function/AutoGraph:
    • AutoGraph now includes into TensorFlow loops any variables that are closed over by local functions. Previously, such variables were sometimes incorrectly ignored.
    • functions returned by the get_concrete_function method of tf.function objects can now be called with arguments consistent with the original arguments or type specs passed to get_concrete_function. This calling convention is now the preferred way to use concrete functions with nested values and composite tensors. Please check the guide for more details on concrete_ function.
    • Update tf.function's experimental_relax_shapes to handle composite tensors appropriately.
    • Optimize tf.function invocation, by removing redundant list converter.
    • tf.function will retrace when called with a different variable instead of simply using the dtype & shape.
    • Improve support for dynamically-sized TensorArray inside tf.function.
  • tf.math:
    • Narrow down argmin/argmax contract to always return the smallest index for ties.
    • tf.math.reduce_variance and tf.math.reduce_std return correct computation for complex types and no longer support integer types.
    • Add Bessel functions of order 0,1 to tf.math.special.
    • tf.divide now always returns a tensor to be consistent with documentation and other APIs.
  • tf.image:
    • Replaced tf.image.non_max_suppression_padded with a new implementation that supports batched inputs, which is considerably faster on TPUs and GPUs. Boxes with area=0 will be ignored. Existing usage with single inputs should still work as before.
  • tf.linalg
    • Add tf.linalg.banded_triangular_solve.
  • tf.random:
    • Add tf.random.stateless_parameterized_truncated_normal.
  • tf.ragged:
    • Add tf.ragged.cross and tf.ragged.cross_hashed operations.
  • tf.RaggedTensor:
    • RaggedTensor.to_tensor() now preserves static shape.
    • Add tf.strings.format() and tf.print() to support RaggedTensors.
  • tf.saved_model:
    • @tf.function from SavedModel no longer ignores args after a RaggedTensor when selecting the concrete function to run.
    • Fix save model issue for ops with a list of functions.
    • Add tf.saved_model.LoadOptions with experimental_io_device as arg with default value None to choose the I/O device for loading models and weights.
    • Update tf.saved_model.SaveOptions with experimental_io_device as arg with default value None to choose the I/O device for saving models and weights.
    • Mutable tables now restore checkpointed values when loaded from SavedModel.
    • The user object metadata field in the SavedModel proto has been deprecated as part of the updates to Keras SavedModel. Keras was the only consumer of this field prior to the update.
  • GPU
    • TF 2.3 includes PTX kernels only for compute capability 7.0 to reduce the TF pip binary size. Earlier releases included PTX for a variety of older compute capabilities.
    • Remove environmental variable TF_USE_CUDNN.
  • Others
    • Retain parent namescope for ops added inside tf.while_loop/tf.cond/tf.switch_case.
    • Update tf.vectorized_map to support vectorizing tf.while_loop and TensorList operations.
    • tf.custom_gradient can now be applied to functions that accept nested structures of tensors as inputs (instead of just a list of tensors). Note that Python structures such as tuples and lists now won't be treated as tensors, so if you still want them to be treated that way, you need to wrap them with tf.convert_to_tensor.
    • No lowering on gradient case op when input is DeviceIndex op.
    • Extend the ragged version of tf.gather to support batch_dims and axis args.
    • Update tf.map_fn to support RaggedTensors and SparseTensors.
    • Deprecate tf.group. It is not useful in eager mode.
    • Add CPU and GPU implementation of modified variation of FTRL/FTRLV2 that can triggerred by multiply_linear_by_lr allowing a learning rate of zero.

tf.data:

  • tf.data.experimental.dense_to_ragged_batch works correctly with tuples.
  • tf.data.experimental.dense_to_ragged_batch to output variable ragged rank.
  • tf.data.experimental.cardinality is now a method on tf.data.Dataset.
  • tf.data.Dataset now supports len(Dataset) when the cardinality is finite.

tf.distribute:

  • Expose experimental tf.distribute.DistributedDataset and tf.distribute.DistributedIterator to distribute input data when using tf.distribute to scale training on multiple devices.
  • Allow var.assign on MirroredVariables with aggregation=NONE in replica context. Previously this would raise an error. We now allow this because many users and library writers find using .assign in replica context to be more convenient, instead of having to use Strategy.extended.update which was the previous way of updating variables in this situation.
  • tf.distribute.experimental.MultiWorkerMirroredStrategy adds support for partial batches. Workers running out of data now continue to participate in the training with empty inputs, instead of raising an error. Learn more about partial batches here.
  • Improve the performance of reading metrics eagerly under tf.distribute.experimental.MultiWorkerMirroredStrategy.
  • Fix the issue that strategy.reduce() inside tf.function may raise exceptions when the values to reduce are from loops or if-clauses.
  • Fix the issue that tf.distribute.MirroredStrategy cannot be used together with tf.distribute.experimental.MultiWorkerMirroredStrategy.
  • Add a tf.distribute.cluster_resolver.TPUClusterResolver.connect API to simplify TPU initialization.
  • Add tf.distribute.Strategy.gather and tf.distribute.ReplicaContext.all_gather methods to gather and concatenate tf.distribute.DistributedValues across workers and devices.

tf.keras:

  • Introduces experimental preprocessing layers API (tf.keras.layers.experimental.preprocessing) to handle data preprocessing operations such as categorical feature encoding, text vectorization, data normalization, and data discretization (binning). The newly added layers provide a replacement for the legacy feature column API, and support composite tensor inputs.
  • Added categorical data processing layers:
    • IntegerLookup & StringLookup: build an index of categorical feature values
    • CategoryEncoding: turn integer-encoded categories into one-hot, multi-hot, or tf-idf encoded representations
    • CategoryCrossing: create new categorical features representing co-occurrences of previous categorical feature values
    • Hashing: the hashing trick, for large-vocabulary categorical features
    • Discretization: turn continuous numerical features into categorical features by binning their values
  • Improved image preprocessing layers: CenterCrop, Rescaling
  • Improved image augmentation layers: RandomCrop, RandomFlip, RandomTranslation, RandomRotation, RandomHeight, RandomWidth, RandomZoom, RandomContrast
  • Improved TextVectorization layer, which handles string tokenization, n-gram generation, and token encoding
    • The TextVectorization layer now accounts for the mask_token as part of the vocabulary size when output_mode='int'. This means that, if you have a max_tokens value of 5000, your output will have 5000 unique values (not 5001 as before).
    • Change the return value of TextVectorization.get_vocabulary() from byte to string. Users who previously were calling 'decode' on the output of this method should no longer need to do so.
  • Introduce new Keras dataset generation utilities :
    • image_dataset_from_directory is a utility based on tf.data.Dataset, meant to replace the legacy ImageDataGenerator. It takes you from a structured directory of images to a labeled dataset, in one function call. Note that it doesn't perform image data augmentation (which is meant to be done using preprocessing layers).
    • text_dataset_from_directory takes you from a structured directory of text files to a labeled dataset, in one function call.
    • timeseries_dataset_from_array is a tf.data.Dataset-based replacement of the legacy TimeseriesGenerator. It takes you from an array of timeseries data to a dataset of shifting windows with their targets.
  • Added experimental_steps_per_execution arg to model.compile to indicate the number of batches to run per tf.function call. This can speed up Keras Models on TPUs up to 3x.
  • Extends tf.keras.layers.Lambda layers to support multi-argument lambdas, and keyword arguments when calling the layer.
  • Functional models now get constructed if any tensor in a layer call's arguments/keyword arguments comes from a keras input. Previously the functional api would only work if all of the elements in the first argument to the layer came from a keras input.
  • Clean up BatchNormalization layer's trainable property to act like standard python state when it's used inside tf.functions (frozen at tracing time), instead of acting like a pseudo-variable whose updates kind of sometimes get reflected in already-traced tf.function traces.
  • Add the Conv1DTranspose layer.
  • Refine the semantics of SensitivitySpecificityBase derived metrics. See the updated API docstrings for tf.keras.metrics.SensitivityAtSpecificity and tf.keras.metrics.SpecificityAtSensitivty.

tf.lite:

  • Converter
    • Restored inference_input_type and inference_output_type flags in TF 2.x TFLiteConverter (backward compatible with TF 1.x) to support integer (tf.int8, tf.uint8) input and output types in post training full integer quantized models.
    • Added support for converting and resizing models with dynamic (placeholder) dimensions. Previously, there was only limited support for dynamic batch size, and even that did not guarantee that the model could be properly resized at runtime.
    • Enabled experimental support for a new quantization mode with 16-bit activations and 8-bit weights. See lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8.
  • CPU
    • Fix an issue w/ dynamic weights and Conv2D on x86.
    • Add a runtime Android flag for enabling XNNPACK for optimized CPU performance.
    • Add a runtime iOS flag for enabling XNNPACK for optimized CPU performance.
    • Add a compiler flag to enable building a TFLite library that applies XNNPACK delegate automatically when the model has a fp32 operation.
  • GPU
    • Allow GPU acceleration starting with internal graph nodes
    • Experimental support for quantized models with the Android GPU delegate
    • Add GPU delegate whitelist.
    • Rename GPU whitelist -> compatibility (list).
    • Improve GPU compatibility list entries from crash reports.
  • NNAPI
    • Set default value for StatefulNnApiDelegate::Options::max_number_delegated_partitions to 3.
    • Add capability to disable NNAPI CPU and check NNAPI Errno.
    • Fix crashes when using NNAPI with target accelerator specified with model containing Conv2d or FullyConnected or LSTM nodes with quantized weights.
    • Fix ANEURALNETWORKS_BAD_DATA execution failures with sum/max/min/reduce operations with scalar inputs.
  • Hexagon
    • TFLite Hexagon Delegate out of experimental.
    • Experimental int8 support for most hexagon ops.
    • Experimental per-channel quant support for conv in Hexagon delegate.
    • Support dynamic batch size in C++ API.
  • CoreML
    • Opensource CoreML delegate
  • Misc
    • Enable building Android TFLite targets on Windows
    • Add support for BatchMatMul.
    • Add support for half_pixel_centers with ResizeNearestNeighbor.
    • Add 3D support for BatchToSpaceND.
    • Add 5D support for BroadcastSub, Maximum, Minimum, Transpose and BroadcastDiv.
    • Rename kTfLiteActRelu1 to kTfLiteActReluN1To1.
    • Enable flex delegate on tensorflow.lite.Interpreter Python package.
    • Add Buckettize, SparseCross and BoostedTreesBucketize to the flex whitelist.
    • Add support for selective registration of flex ops.
    • Add missing kernels for flex delegate whitelisted ops.
    • Fix issue when using direct ByteBuffer inputs with graphs that have dynamic shapes.
    • Fix error checking supported operations in a model containing HardSwish.

Packaging Support

  • Added tf.sysconfig.get_build_info(). Returns a dict that describes the build environment of the currently installed TensorFlow package, e.g. the NVIDIA CUDA and NVIDIA CuDNN versions used when TensorFlow was built.

Profiler

  • Fix a subtle use-after-free issue in XStatVisitor::RefValue().

TPU Enhancements

  • Adds 3D mesh support in TPU configurations ops.
  • Added TPU code for FTRL with multiply_linear_by_lr.
  • Silently adds a new file system registry at gstpu.
  • Support restartType in cloud tpu client.
  • Depend on a specific version of google-api-python-client.
  • Fixes apiclient import.

Tracing and Debugging

  • Add a TFE_Py_Execute traceme.

XLA Support

  • Implement stable argmin and argmax

Thanks to our Contributors

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

902449@58880@bigcat_chen@ASIC, Abdul Baseer Khan, Abhineet Choudhary, Abolfazl Shahbazi, Adam Hillier, ag.ramesh, Agoniii, Ajay P, Alex Hoffman, Alexander Bayandin, Alexander Grund, Alexandre Abadie, Alexey Rogachevskiy, amoitra, Andrew Stevens, Angus-Luo, Anshuman Tripathy, Anush Elangovan, Artem Mavrin, Ashutosh Hathidara, autoih, Ayushman Kumar, ayushmankumar7, Bairen Yi, Bas Aarts, Bastian Eichenberger, Ben Barsdell, bhack, Bharat Raghunathan, Biagio Montaruli, Bigcat-Himax, blueyi, Bryan Cutler, Byambaa, Carlos Hernandez-Vaquero, Chen Lei, Chris Knorowski, Christian Clauss, chuanqiw, CuiYifeng, Daniel Situnayake, Daria Zhuravleva, Dayananda-V, Deven Desai, Devi Sandeep Endluri, Dmitry Zakharov, Dominic Jack, Duncan Riach, Edgar Liberis, Ehsan Toosi, ekuznetsov139, Elena Zhelezina, Eugene Kuznetsov, Eugene Mikhantiev, Evgenii Zheltonozhskii, Fabio Di Domenico, Fausto Morales, Fei Sun, feihugis, Felix E. Klee, flyingcat, Frederic Bastien, Fredrik Knutsson, frreiss, fsx950223, ganler, Gaurav Singh, Georgios Pinitas, Gian Marco Iodice, Giorgio Arena, Giuseppe Rossini, Gregory Keith, Guozhong Zhuang, gurushantj, Hahn Anselm, Harald Husum, Harjyot Bagga, Hristo Vrigazov, Ilya Persky, Ir1d, Itamar Turner-Trauring, jacco, Jake Tae, Janosh Riebesell, Jason Zaman, jayanth, Jeff Daily, Jens Elofsson, Jinzhe Zeng, JLZ, Jonas Skog, Jonathan Dekhtiar, Josh Meyer, Joshua Chia, Judd, justkw, Kaixi Hou, Kam D Kasravi, Kamil Rakoczy, Karol Gugala, Kayou, Kazuaki Ishizaki, Keith Smiley, Khaled Besrour, Kilaru Yasaswi Sri Chandra Gandhi, Kim, Young Soo, Kristian Hartikainen, Kwabena W. Agyeman, Leslie-Fang, Leslie-Fang-Intel, Li, Guizi, Lukas Geiger, Lutz Roeder, M\U00E5Ns Nilsson, Mahmoud Abuzaina, Manish, Marcel Koester, Marcin Sielski, marload, Martin Jul, Matt Conley, mdfaijul, Meng, Peng, Meteorix, Michael Käufl, Michael137, Milan Straka, Mitchell Vitez, Ml-0, Mokke Meguru, Mshr-H, nammbash, Nathan Luehr, naumkin, Neeraj Bhadani, ngc92, Nick Morgan, nihui, Niranjan Hasabnis, Niranjan Yadla, Nishidha Panpaliya, Oceania2018, oclyke, Ouyang Jin, OverLordGoldDragon, Owen Lyke, Patrick Hemmer, Paul Andrey, Peng Sun, periannath, Phil Pearl, Prashant Dandriyal, Prashant Kumar, Rahul Huilgol, Rajan Singh, Rajeshwar Reddy T, rangjiaheng, Rishit Dagli, Rohan Reddy, rpalakkal, rposts, Ruan Kunliang, Rushabh Vasani, Ryohei Ikegami, Semun Lee, Seo-Inyoung, Sergey Mironov, Sharada Shiddibhavi, ShengYang1, Shraiysh Vaishay, Shunya Ueta, shwetaoj, Siyavash Najafzade, Srinivasan Narayanamoorthy, Stephan Uphoff, storypku, sunchenggen, sunway513, Sven-Hendrik Haase, Swapnil Parekh, Tamas Bela Feher, Teng Lu, tigertang, tomas, Tomohiro Ubukata, tongxuan.ltx, Tony Tonev, Tzu-Wei Huang, Téo Bouvard, Uday Bondhugula, Vaibhav Jade, Vijay Tadikamalla, Vikram Dattu, Vincent Abriou, Vishnuvardhan Janapati, Vo Van Nghia, VoVAllen, Will Battel, William D. Irons, wyzhao, Xiaoming (Jason) Cui, Xiaoquan Kong, Xinan Jiang, xutianming, Yair Ehrenwald, Yasir Modak, Yasuhiro Matsumoto, Yixing Fu, Yong Tang, Yuan Tang, zhaozheng09, Zilin Zhu, zilinzhu, 张志豪

Release 2.1.1

Bug Fixes and Other Changes

Release 2.0.2

Bug Fixes and Other Changes

Release 1.15.3

Bug Fixes and Other Changes

Release 2.2.0

TensorFlow 2.2 discontinues support for Python 2, previously announced as following Python 2's EOL on January 1, 2020.

Coinciding with this change, new releases of TensorFlow's Docker images provide Python 3 exclusively. Because all images now use Python 3, Docker tags containing -py3 will no longer be provided and existing -py3 tags like latest-py3 will not be updated.

Major Features and Improvements

  • Replaced the scalar type for string tensors from std::string to tensorflow::tstring which is now ABI stable.

  • A new Profiler for TF 2 for CPU/GPU/TPU. It offers both device and host performance analysis, including input pipeline and TF Ops. Optimization advisory is provided whenever possible. Please see this tutorial and guide for usage guidelines.

  • Export C++ functions to Python using pybind11 as opposed to SWIG as a part of our deprecation of swig efforts.

  • tf.distribute:

    • Support added for global sync BatchNormalization by using the newly added tf.keras.layers.experimental.SyncBatchNormalization layer. This layer will sync BatchNormalization statistics every step across all replicas taking part in sync training.
    • Performance improvements for GPU multi-worker distributed training using tf.distribute.experimental.MultiWorkerMirroredStrategy
      • Update NVIDIA NCCL to 2.5.7-1 for better performance and performance tuning. Please see nccl developer guide for more information on this.
      • Support gradient allreduce in float16. See this example usage.
      • Experimental support of all reduce gradient packing to allow overlapping gradient aggregation with backward path computation.
      • Deprecated experimental_run_v2 method for distribution strategies and renamed the method run as it is no longer experimental.
      • Add CompositeTensor support for DistributedIterators. This should help prevent unnecessary function retracing and memory leaks.

  • tf.keras:

    • Model.fit major improvements:
      • You can now use custom training logic with Model.fit by overriding Model.train_step.
      • Easily write state-of-the-art training loops without worrying about all of the features Model.fit handles for you (distribution strategies, callbacks, data formats, looping logic, etc)
      • See the default Model.train_step for an example of what this function should look like. Same applies for validation and inference via Model.test_step and Model.predict_step.
      • SavedModel uses its own Model._saved_model_inputs_spec attr now instead of relying on Model.inputs and Model.input_names, which are no longer set for subclass Models. This attr is set in eager, tf.function, and graph modes. This gets rid of the need for users to manually call Model._set_inputs when using Custom Training Loops(CTLs).
      • Dynamic shapes are supported for generators by calling the Model on the first batch we "peek" from the generator. This used to happen implicitly in Model._standardize_user_data. Long-term, a solution where the DataAdapter doesn't need to call the Model is probably preferable.
    • The SavedModel format now supports all Keras built-in layers (including metrics, preprocessing layers, and stateful RNN layers)
    • Update Keras batch normalization layer to use the running mean and average computation in the fused_batch_norm. You should see significant performance improvements when using fused_batch_norm in Eager mode.

  • tf.lite:

    • Enable TFLite experimental new converter by default.

  • XLA

    • XLA now builds and works on windows. All prebuilt packages come with XLA available.
    • XLA can be enabled for a tf.function with “compile or throw exception” semantics on CPU and GPU.

Breaking Changes

  • tf.keras:
    • In tf.keras.applications the name of the "top" layer has been standardized to "predictions". This is only a problem if your code relies on the exact name of the layer.
    • Huber loss function has been updated to be consistent with other Keras losses. It now computes mean over the last axis of per-sample losses before applying the reduction function.
  • AutoGraph no longer converts functions passed to tf.py_function, tf.py_func and tf.numpy_function.
  • Deprecating XLA_CPU and XLA_GPU devices with this release.
  • Increasing the minimum bazel version to build TF to 2.0.0 to use Bazel's cc_experimental_shared_library.
  • Keras compile/fit behavior for functional and subclassed models have been unified. Model properties such as metrics, metrics_names will now be available only after training/evaluating the model on actual data for functional models. metrics will now include model loss and output losses.loss_functions property has been removed from the model. This was an undocumented property that was accidentally public and has now been removed.

Known Caveats

  • The current TensorFlow release now requires gast version 0.3.3.

Bug Fixes and Other Changes

  • tf.data:
    • Removed autotune_algorithm from experimental optimization options.
  • TF Core:
    • tf.constant always creates CPU tensors irrespective of the current device context.
    • Eager TensorHandles maintain a list of mirrors for any copies to local or remote devices. This avoids any redundant copies due to op execution.
    • For tf.Tensor & tf.Variable, .experimental_ref() is no longer experimental and is available as simply .ref().
    • pfor/vectorized_map: Added support for vectorizing 56 more ops. Vectorizing tf.cond is also supported now.
    • Set as much partial shape as we can infer statically within the gradient impl of the gather op.
    • Gradient of tf.while_loop emits StatelessWhile op if cond and body functions are stateless. This allows multiple gradients while ops to run in parallel under distribution strategy.
    • Speed up GradientTape in eager mode by auto-generating list of op inputs/outputs which are unused and hence not cached for gradient functions.
    • Support back_prop=False in while_v2 but mark it as deprecated.
    • Improve error message when attempting to use None in data-dependent control flow.
    • Add RaggedTensor.numpy().
    • Update RaggedTensor.__getitem__ to preserve uniform dimensions & allow indexing into uniform dimensions.
    • Update tf.expand_dims to always insert the new dimension as a non-ragged dimension.
    • Update tf.embedding_lookup to use partition_strategy and max_norm when ids is ragged.
    • Allow batch_dims==rank(indices) in tf.gather.
    • Add support for bfloat16 in tf.print.
  • tf.distribute:
    • Support embedding_column with variable-length input features for MultiWorkerMirroredStrategy.
  • tf.keras:
    • Added experimental_aggregate_gradients argument to tf.keras.optimizer.Optimizer.apply_gradients. This allows custom gradient aggregation and processing aggregated gradients in custom training loop.
    • Allow pathlib.Path paths for loading models via Keras API.
  • tf.function/AutoGraph:
    • AutoGraph is now available in ReplicaContext.merge_call, Strategy.extended.update and Strategy.extended.update_non_slot.
    • Experimental support for shape invariants has been enabled in tf.function. See the API docs for tf.autograph.experimental.set_loop_options for additional info.
    • AutoGraph error messages now exclude frames corresponding to APIs internal to AutoGraph.
    • Improve shape inference for tf.function input arguments to unlock more Grappler optimizations in TensorFlow 2.x.
    • Improve automatic control dependency management of resources by allowing resource reads to occur in parallel and synchronizing only on writes.
    • Fix execution order of multiple stateful calls to experimental_run_v2 in tf.function.
    • You can now iterate over RaggedTensors using a for loop inside tf.function.
  • tf.lite:
    • Migrated the tf.lite C inference API out of experimental into lite/c.
    • Add an option to disallow NNAPI CPU / partial acceleration on Android 10
    • TFLite Android AARs now include the C headers and APIs are required to use TFLite from native code.
    • Refactors the delegate and delegate kernel sources to allow usage in the linter.
    • Limit delegated ops to actually supported ones if a device name is specified or NNAPI CPU Fallback is disabled.
    • TFLite now supports tf.math.reciprocal1 op by lowering to tf.div op.
    • TFLite's unpack op now supports boolean tensor inputs.
    • Microcontroller and embedded code moved from experimental to main TensorFlow Lite folder
    • Check for large TFLite tensors.
    • Fix GPU delegate crash with C++17.
    • Add 5D support to TFLite strided_slice.
    • Fix error in delegation of DEPTH_TO_SPACE to NNAPI causing op not to be accelerated.
    • Fix segmentation fault when running a model with LSTM nodes using NNAPI Delegate
    • Fix NNAPI delegate failure when an operand for Maximum/Minimum operation is a scalar.
    • Fix NNAPI delegate failure when Axis input for reduce operation is a scalar.
    • Expose option to limit the number of partitions that will be delegated to NNAPI.
    • If a target accelerator is specified, use its feature level to determine operations to delegate instead of SDK version.
  • tf.random:
    • Various random number generation improvements:
    • Add a fast path for default random_uniform
    • random_seed documentation improvement.
    • RandomBinomial broadcasts and appends the sample shape to the left rather than the right.
    • Added tf.random.stateless_binomial, tf.random.stateless_gamma, tf.random.stateless_poisson
    • tf.random.stateless_uniform now supports unbounded sampling of int types.
  • Math and Linear Algebra:
    • Add tf.linalg.LinearOperatorTridiag.
    • Add LinearOperatorBlockLowerTriangular
    • Add broadcasting support to tf.linalg.triangular_solve#26204, tf.math.invert_permutation.
    • Add tf.math.sobol_sample op.
    • Add tf.math.xlog1py.
    • Add tf.math.special.{dawsn,expi,fresnel_cos,fresnel_sin,spence}.
    • Add a Modified Discrete Cosine Transform (MDCT) and its inverse to tf.signal.
  • TPU Enhancements:
    • Refactor TpuClusterResolver to move shared logic to a separate pip package.
    • Support configuring TPU software version from cloud tpu client.
    • Allowed TPU embedding weight decay factor to be multiplied by learning rate.
  • XLA Support:
    • Add standalone XLA AOT runtime target + relevant .cc sources to pip package.
    • Add check for memory alignment to MemoryAllocation::MemoryAllocation() on 32-bit ARM. This ensures a deterministic early exit instead of a hard to debug bus error later.
    • saved_model_cli aot_compile_cpu allows you to compile saved models to XLA header+object files and include them in your C++ programs.
    • Enable Igamma, Igammac for XLA.
  • Deterministic Op Functionality:
    • XLA reduction emitter is deterministic when the environment variable TF_DETERMINISTIC_OPS is set to "true" or "1". This extends deterministic tf.nn.bias_add back-prop functionality (and therefore also deterministic back-prop of bias-addition in Keras layers) to include when XLA JIT compilation is enabled.
    • Fix problem, when running on a CUDA GPU and when either environment variable TF_DETERMINISTIC_OPS or environment variable TF_CUDNN_DETERMINISTIC is set to "true" or "1", in which some layer configurations led to an exception with the message "No algorithm worked!"
  • Tracing and Debugging:
    • Add source, destination name to _send traceme to allow easier debugging.
    • Add traceme event to fastpathexecute.
  • Other:
    • Fix an issue with AUC.reset_states for multi-label AUC #35852
    • Fix the TF upgrade script to not delete files when there is a parsing error and the output mode is in-place.
    • Move tensorflow/core:framework/*_pyclif rules to tensorflow/core/framework:*_pyclif.

Thanks to our Contributors

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

372046933, 8bitmp3, aaronhma, Abin Shahab, Aditya Patwardhan, Agoniii, Ahti Kitsik, Alan Yee, Albin Joy, Alex Hoffman, Alexander Grund, Alexandre E. Eichenberger, Amit Kumar Jaiswal, amoitra, Andrew Anderson, Angus-Luo, Anthony Barbier, Anton Kachatkou, Anuj Rawat, archis, Arpan-Dhatt, Arvind Sundararajan, Ashutosh Hathidara, autoih, Bairen Yi, Balint Cristian, Bas Aarts, BashirSbaiti, Basit Ayantunde, Ben Barsdell, Benjamin Gaillard, boron, Brett Koonce, Bryan Cutler, Christian Goll, Christian Sachs, Clayne Robison, comet, Daniel Falbel, Daria Zhuravleva, darsh8200, David Truby, Dayananda-V, deepakm, Denis Khalikov, Devansh Singh, Dheeraj R Reddy, Diederik Van Liere, Diego Caballero, Dominic Jack, dothinking, Douman, Drake Gens, Duncan Riach, Ehsan Toosi, ekuznetsov139, Elena Zhelezina, elzino, Ending2015a, Eric Schweitz, Erik Zettel, Ethan Saadia, Eugene Kuznetsov, Evgeniy Zheltonozhskiy, Ewout Ter Hoeven, exfalso, FAIJUL, Fangjun Kuang, Fei Hu, Frank Laub, Frederic Bastien, Fredrik Knutsson, frreiss, Frédéric Rechtenstein, fsx950223, Gaurav Singh, gbaned, George Grzegorz Pawelczak, George Sterpu, Gian Marco Iodice, Giorgio Arena, Hans Gaiser, Hans Pabst, Haoyu Wu, Harry Slatyer, hsahovic, Hugo, Hugo Sjöberg, IrinaM21, jacco, Jake Tae, Jean-Denis Lesage, Jean-Michel Gorius, Jeff Daily, Jens Elofsson, Jerry Shih, jerryyin, Jin Mingjian, Jinjing Zhou, JKIsaacLee, jojimonv, Jonathan Dekhtiar, Jose Ignacio Gomez, Joseph-Rance, Judd, Julian Gross, Kaixi Hou, Kaustubh Maske Patil, Keunwoo Choi, Kevin Hanselman, Khor Chean Wei, Kilaru Yasaswi Sri Chandra Gandhi, Koan-Sin Tan, Koki Ibukuro, Kristian Holsheimer, kurileo, Lakshay Tokas, Lee Netherton, leike666666, Leslie-Fang-Intel, Li, Guizi, LIUJIAN435, Lukas Geiger, Lyo Nguyen, madisetti, Maher Jendoubi, Mahmoud Abuzaina, Manuel Freiberger, Marcel Koester, Marco Jacopo Ferrarotti, Markus Franke, marload, Mbah-Javis, mbhuiyan, Meng Zhang, Michael Liao, MichaelKonobeev, Michal Tarnowski, Milan Straka, minoring, Mohamed Nour Abouelseoud, MoussaMM, Mrinal Jain, mrTsjolder, Måns Nilsson, Namrata Bhave, Nicholas Gao, Niels Ole Salscheider, nikochiko, Niranjan Hasabnis, Nishidha Panpaliya, nmostafa, Noah Trenaman, nuka137, Officium, Owen L - Sfe, Pallavi G, Paul Andrey, Peng Sun, Peng Wu, Phil Pearl, PhilipMay, pingsutw, Pooya Davoodi, PragmaTwice, pshiko, Qwerty71, R Gomathi, Rahul Huilgol, Richard Xiao, Rick Wierenga, Roberto Rosmaninho, ruchit2801, Rushabh Vasani, Sami, Sana Damani, Sarvesh Dubey, Sasan Jafarnejad, Sergii Khomenko, Shane Smiskol, Shaochen Shi, sharkdtu, Shawn Presser, ShengYang1, Shreyash Patodia, Shyam Sundar Dhanabalan, Siju Samuel, Somyajit Chakraborty Sam, Srihari Humbarwadi, srinivasan.narayanamoorthy, Srishti Yadav, Steph-En-M, Stephan Uphoff, Stephen Mugisha, SumanSudhir, Taehun Kim, Tamas Bela Feher, TengLu, Tetragramm, Thierry Herrmann, Tian Jin, tigertang, Tom Carchrae, Tom Forbes, Trent Lo, Victor Peng, vijayphoenix, Vincent Abriou, Vishal Bhola, Vishnuvardhan Janapati, vladbataev, VoVAllen, Wallyss Lima, Wen-Heng (Jack) Chung, wenxizhu, William D. Irons, William Zhang, Xiaoming (Jason) Cui, Xiaoquan Kong, Xinan Jiang, Yasir Modak, Yasuhiro Matsumoto, Yaxun (Sam) Liu, Yong Tang, Ytyt-Yt, yuan, Yuan Mingshuai, Yuan Tang, Yuki Ueda, Yusup, zhangshijin, zhuwenxi

Release 2.0.1

Bug Fixes and Other Changes

Release 1.15.2

Bug Fixes and Other Changes

Release 2.1.0

TensorFlow 2.1 will be the last TF release supporting Python 2. Python 2 support officially ends an January 1, 2020. As announced earlier, TensorFlow will also stop supporting Python 2 starting January 1, 2020, and no more releases are expected in 2019.

Major Features and Improvements

  • The tensorflow pip package now includes GPU support by default (same as tensorflow-gpu) for both Linux and Windows. This runs on machines with and without NVIDIA GPUs. tensorflow-gpu is still available, and CPU-only packages can be downloaded at tensorflow-cpu for users who are concerned about package size.
  • Windows users: Officially-released tensorflow Pip packages are now built with Visual Studio 2019 version 16.4 in order to take advantage of the new /d2ReducedOptimizeHugeFunctions compiler flag. To use these new packages, you must install "Microsoft Visual C++ Redistributable for Visual Studio 2015, 2017 and 2019", available from Microsoft's website here.
    • This does not change the minimum required version for building TensorFlow from source on Windows, but builds enabling EIGEN_STRONG_INLINE can take over 48 hours to compile without this flag. Refer to configure.py for more information about EIGEN_STRONG_INLINE and /d2ReducedOptimizeHugeFunctions.
    • If either of the required DLLs, msvcp140.dll (old) or msvcp140_1.dll (new), are missing on your machine, import tensorflow will print a warning message.
  • The tensorflow pip package is built with CUDA 10.1 and cuDNN 7.6.
  • tf.keras
    • Experimental support for mixed precision is available on GPUs and Cloud TPUs. See usage guide.
    • Introduced the TextVectorization layer, which takes as input raw strings and takes care of text standardization, tokenization, n-gram generation, and vocabulary indexing. See this end-to-end text classification example.
    • Keras .compile .fit .evaluate and .predict are allowed to be outside of the DistributionStrategy scope, as long as the model was constructed inside of a scope.
    • Experimental support for Keras .compile, .fit, .evaluate, and .predict is available for Cloud TPUs, Cloud TPU, for all types of Keras models (sequential, functional and subclassing models).
    • Automatic outside compilation is now enabled for Cloud TPUs. This allows tf.summary to be used more conveniently with Cloud TPUs.
    • Dynamic batch sizes with DistributionStrategy and Keras are supported on Cloud TPUs.
    • Support for .fit, .evaluate, .predict on TPU using numpy data, in addition to tf.data.Dataset.
    • Keras reference implementations for many popular models are available in the TensorFlow Model Garden.
  • tf.data
    • Changes rebatching for tf.data datasets + DistributionStrategy for better performance. Note that the dataset also behaves slightly differently, in that the rebatched dataset cardinality will always be a multiple of the number of replicas.
    • tf.data.Dataset now supports automatic data distribution and sharding in distributed environments, including on TPU pods.
    • Distribution policies for tf.data.Dataset can now be tuned with 1. tf.data.experimental.AutoShardPolicy(OFF, AUTO, FILE, DATA) 2. tf.data.experimental.ExternalStatePolicy(WARN, IGNORE, FAIL)
  • tf.debugging
    • Add tf.debugging.enable_check_numerics() and tf.debugging.disable_check_numerics() to help debugging the root causes of issues involving infinities and NaNs.
  • tf.distribute
    • Custom training loop support on TPUs and TPU pods is available through strategy.experimental_distribute_dataset, strategy.experimental_distribute_datasets_from_function, strategy.experimental_run_v2, strategy.reduce.
    • Support for a global distribution strategy through tf.distribute.experimental_set_strategy(), in addition to strategy.scope().
  • TensorRT
    • TensorRT 6.0 is now supported and enabled by default. This adds support for more TensorFlow ops including Conv3D, Conv3DBackpropInputV2, AvgPool3D, MaxPool3D, ResizeBilinear, and ResizeNearestNeighbor. In addition, the TensorFlow-TensorRT python conversion API is exported as tf.experimental.tensorrt.Converter.
  • Environment variable TF_DETERMINISTIC_OPS has been added. When set to "true" or "1", this environment variable makes tf.nn.bias_add operate deterministically (i.e. reproducibly), but currently only when XLA JIT compilation is not enabled. Setting TF_DETERMINISTIC_OPS to "true" or "1" also makes cuDNN convolution and max-pooling operate deterministically. This makes Keras Conv*D and MaxPool*D layers operate deterministically in both the forward and backward directions when running on a CUDA-enabled GPU.

Breaking Changes

  • Deletes Operation.traceback_with_start_lines for which we know of no usages.
  • Removed id from tf.Tensor.__repr__() as id is not useful other than internal debugging.
  • Some tf.assert_* methods now raise assertions at operation creation time if the input tensors' values are known at that time, not during the session.run(). This only changes behavior when the graph execution would have resulted in an error. When this happens, a noop is returned and the input tensors are marked non-feedable. In other words, if they are used as keys in feed_dict argument to session.run(), an error will be raised. Also, because some assert ops don't make it into the graph, the graph structure changes. A different graph can result in different per-op random seeds when they are not given explicitly (most often).
  • The following APIs are not longer experimental: tf.config.list_logical_devices, tf.config.list_physical_devices, tf.config.get_visible_devices, tf.config.set_visible_devices, tf.config.get_logical_device_configuration, tf.config.set_logical_device_configuration.
  • tf.config.experimentalVirtualDeviceConfiguration has been renamed to tf.config.LogicalDeviceConfiguration.
  • tf.config.experimental_list_devices has been removed, please use tf.config.list_logical_devices.

Bug Fixes and Other Changes

  • tf.data
    • Fixes concurrency issue with tf.data.experimental.parallel_interleave with sloppy=True.
    • Add tf.data.experimental.dense_to_ragged_batch().
    • Extend tf.data parsing ops to support RaggedTensors.
  • tf.distribute
    • Fix issue where GRU would crash or give incorrect output when a tf.distribute.Strategy was used.
  • tf.estimator
    • Added option in tf.estimator.CheckpointSaverHook to not save the GraphDef.
    • Moving the checkpoint reader from swig to pybind11.
  • tf.keras
    • Export depthwise_conv2d in tf.keras.backend.
    • In Keras Layers and Models, Variables in trainable_weights, non_trainable_weights, and weights are explicitly deduplicated.
    • Keras model.load_weights now accepts skip_mismatch as an argument. This was available in external Keras, and has now been copied over to tf.keras.
    • Fix the input shape caching behavior of Keras convolutional layers.
    • Model.fit_generator, Model.evaluate_generator, Model.predict_generator, Model.train_on_batch, Model.test_on_batch, and Model.predict_on_batch methods now respect the run_eagerly property, and will correctly run using tf.function by default. Note that Model.fit_generator, Model.evaluate_generator, and Model.predict_generator are deprecated endpoints. They are subsumed by Model.fit, Model.evaluate, and Model.predict which now support generators and Sequences.
  • tf.lite
    • Legalization for NMS ops in TFLite.
    • add narrow_range and axis to quantize_v2 and dequantize ops.
    • Added support for FusedBatchNormV3 in converter.
    • Add an errno-like field to NNAPI delegate for detecting NNAPI errors for fallback behaviour.
    • Refactors NNAPI Delegate to support detailed reason why an operation is not accelerated.
    • Converts hardswish subgraphs into atomic ops.
  • Other
    • Critical stability updates for TPUs, especially in cases where the XLA compiler produces compilation errors.
    • TPUs can now be re-initialized multiple times, using tf.tpu.experimental.initialize_tpu_system.
    • Add RaggedTensor.merge_dims().
    • Added new uniform_row_length row-partitioning tensor to RaggedTensor.
    • Add shape arg to RaggedTensor.to_tensor; Improve speed of RaggedTensor.to_tensor.
    • tf.io.parse_sequence_example and tf.io.parse_single_sequence_example now support ragged features.
    • Fix while_v2 with variables in custom gradient.
    • Support taking gradients of V2 tf.cond and tf.while_loop using LookupTable.
    • Fix bug where vectorized_map failed on inputs with unknown static shape.
    • Add preliminary support for sparse CSR matrices.
    • Tensor equality with None now behaves as expected.
    • Make calls to tf.function(f)(), tf.function(f).get_concrete_function and tf.function(f).get_initialization_function thread-safe.
    • Extend tf.identity to work with CompositeTensors (such as SparseTensor)
    • Added more dtypes and zero-sized inputs to Einsum Op and improved its performance
    • Enable multi-worker NCCL all-reduce inside functions executing eagerly.
    • Added complex128 support to RFFT, RFFT2D, RFFT3D, IRFFT, IRFFT2D, and IRFFT3D.
    • Add pfor converter for SelfAdjointEigV2.
    • Add tf.math.ndtri and tf.math.erfinv.
    • Add tf.config.experimental.enable_mlir_bridge to allow using MLIR compiler bridge in eager model.
    • Added support for MatrixSolve on Cloud TPU / XLA.
    • Added tf.autodiff.ForwardAccumulator for forward-mode autodiff
    • Add LinearOperatorPermutation.
    • A few performance optimizations on tf.reduce_logsumexp.
    • Added multilabel handling to AUC metric
    • Optimization on zeros_like.
    • Dimension constructor now requires None or types with an __index__ method.
    • Add tf.random.uniform microbenchmark.
    • Use _protogen suffix for proto library targets instead of _cc_protogen suffix.
    • Moving the checkpoint reader from swig to pybind11.
    • tf.device & MirroredStrategy now supports passing in a tf.config.LogicalDevice
    • If you're building Tensorflow from source, consider using bazelisk to automatically download and use the correct Bazel version. Bazelisk reads the .bazelversion file at the root of the project directory.

Thanks to our Contributors

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

8bitmp3, Aaron Ma, AbdüLhamit Yilmaz, Abhai Kollara, aflc, Ag Ramesh, Albert Z. Guo, Alex Torres, amoitra, Andrii Prymostka, angeliand, Anshuman Tripathy, Anthony Barbier, Anton Kachatkou, Anubh-V, Anuja Jakhade, Artem Ryabov, autoih, Bairen Yi, Bas Aarts, Basit Ayantunde, Ben Barsdell, Bhavani Subramanian, Brett Koonce, candy.dc, Captain-Pool, caster, cathy, Chong Yan, Choong Yin Thong, Clayne Robison, Colle, Dan Ganea, David Norman, David Refaeli, dengziming, Diego Caballero, Divyanshu, djshen, Douman, Duncan Riach, EFanZh, Elena Zhelezina, Eric Schweitz, Evgenii Zheltonozhskii, Fei Hu, fo40225, Fred Reiss, Frederic Bastien, Fredrik Knutsson, fsx950223, fwcore, George Grzegorz Pawelczak, George Sterpu, Gian Marco Iodice, Giorgio Arena, giuros01, Gomathi Ramamurthy, Guozhong Zhuang, Haifeng Jin, Haoyu Wu, HarikrishnanBalagopal, HJYOO, Huang Chen-Yi, Ilham Firdausi Putra, Imran Salam, Jared Nielsen, Jason Zaman, Jasper Vicenti, Jeff Daily, Jeff Poznanovic, Jens Elofsson, Jerry Shih, jerryyin, Jesper Dramsch, jim.meyer, Jongwon Lee, Jun Wan, Junyuan Xie, Kaixi Hou, kamalkraj, Kan Chen, Karthik Muthuraman, Keiji Ariyama, Kevin Rose, Kevin Wang, Koan-Sin Tan, kstuedem, Kwabena W. Agyeman, Lakshay Tokas, latyas, Leslie-Fang-Intel, Li, Guizi, Luciano Resende, Lukas Folle, Lukas Geiger, Mahmoud Abuzaina, Manuel Freiberger, Mark Ryan, Martin Mlostek, Masaki Kozuki, Matthew Bentham, Matthew Denton, mbhuiyan, mdfaijul, Muhwan Kim, Nagy Mostafa, nammbash, Nathan Luehr, Nathan Wells, Niranjan Hasabnis, Oleksii Volkovskyi, Olivier Moindrot, olramde, Ouyang Jin, OverLordGoldDragon, Pallavi G, Paul Andrey, Paul Wais, pkanwar23, Pooya Davoodi, Prabindh Sundareson, Rajeshwar Reddy T, Ralovich, Kristof, Refraction-Ray, Richard Barnes, richardbrks, Robert Herbig, Romeo Kienzler, Ryan Mccormick, saishruthi, Saket Khandelwal, Sami Kama, Sana Damani, Satoshi Tanaka, Sergey Mironov, Sergii Khomenko, Shahid, Shawn Presser, ShengYang1, Siddhartha Bagaria, Simon Plovyt, skeydan, srinivasan.narayanamoorthy, Stephen Mugisha, sunway513, Takeshi Watanabe, Taylor Jakobson, TengLu, TheMindVirus, ThisIsIsaac, Tim Gates, Timothy Liu, Tomer Gafner, Trent Lo, Trevor Hickey, Trevor Morris, vcarpani, Wei Wang, Wen-Heng (Jack) Chung, wenshuai, Wenshuai-Xiaomi, wenxizhu, william, William D. Irons, Xinan Jiang, Yannic, Yasir Modak, Yasuhiro Matsumoto, Yong Tang, Yongfeng Gu, Youwei Song, Zaccharie Ramzi, Zhang, Zhenyu Guo, 王振华 (Zhenhua Wang), 韩董, 이중건 Isaac Lee

Release 1.15.0

This is the last 1.x release for TensorFlow. We do not expect to update the 1.x branch with features, although we will issue patch releases to fix vulnerabilities for at least one year.

Major Features and Improvements

  • As announced, tensorflow pip package will by default include GPU support (same as tensorflow-gpu now) for the platforms we currently have GPU support (Linux and Windows). It will work on machines with and without Nvidia GPUs. tensorflow-gpu will still be available, and CPU-only packages can be downloaded at tensorflow-cpu for users who are concerned about package size.
  • TensorFlow 1.15 contains a complete implementation of the 2.0 API in its compat.v2 module. It contains a copy of the 1.15 main module (without contrib) in the compat.v1 module. TensorFlow 1.15 is able to emulate 2.0 behavior using the enable_v2_behavior() function. This enables writing forward compatible code: by explicitly importing either tensorflow.compat.v1 or tensorflow.compat.v2, you can ensure that your code works without modifications against an installation of 1.15 or 2.0.
  • EagerTensor now supports numpy buffer interface for tensors.
  • Add toggles tf.enable_control_flow_v2() and tf.disable_control_flow_v2() for enabling/disabling v2 control flow.
  • Enable v2 control flow as part of tf.enable_v2_behavior() and TF2_BEHAVIOR=1.
  • AutoGraph translates Python control flow into TensorFlow expressions, allowing users to write regular Python inside tf.function-decorated functions. AutoGraph is also applied in functions used with tf.data, tf.distribute and tf.keras APIS.
  • Adds enable_tensor_equality(), which switches the behavior such that:
    • Tensors are no longer hashable.
    • Tensors can be compared with == and !=, yielding a Boolean Tensor with element-wise comparison results. This will be the default behavior in 2.0.

Breaking Changes

  • Tensorflow code now produces 2 different pip packages: tensorflow_core containing all the code (in the future it will contain only the private implementation) and tensorflow which is a virtual pip package doing forwarding to tensorflow_core (and in the future will contain only the public API of tensorflow). We don't expect this to be breaking, unless you were importing directly from the implementation.
  • TensorFlow 1.15 is built using devtoolset7 (GCC7) on Ubuntu 16. This may lead to ABI incompatibilities with extensions built against earlier versions of TensorFlow.
  • Deprecated the use of constraint= and .constraint with ResourceVariable.
  • tf.keras:
    • OMP_NUM_THREADS is no longer used by the default Keras config. To configure the number of threads, use tf.config.threading APIs.
    • tf.keras.model.save_model and model.save now defaults to saving a TensorFlow SavedModel.
    • keras.backend.resize_images (and consequently, keras.layers.Upsampling2D) behavior has changed, a bug in the resizing implementation was fixed.
    • Layers now default to float32, and automatically cast their inputs to the layer's dtype. If you had a model that used float64, it will probably silently use float32 in TensorFlow2, and a warning will be issued that starts with Layer "layer-name" is casting an input tensor from dtype float64 to the layer's dtype of float32. To fix, either set the default dtype to float64 with tf.keras.backend.set_floatx('float64'), or pass dtype='float64' to each of the Layer constructors. See tf.keras.layers.Layer for more information.
    • Some tf.assert_* methods now raise assertions at operation creation time (i.e. when this Python line executes) if the input tensors' values are known at that time, not during the session.run(). When this happens, a noop is returned and the input tensors are marked non-feedable. In other words, if they are used as keys in feed_dict argument to session.run(), an error will be raised. Also, because some assert ops don't make it into the graph, the graph structure changes. A different graph can result in different per-op random seeds when they are not given explicitly (most often).

Bug Fixes and Other Changes

  • tf.estimator:
    • tf.keras.estimator.model_to_estimator now supports exporting to tf.train.Checkpoint format, which allows the saved checkpoints to be compatible with model.load_weights.
    • Fix tests in canned estimators.
    • Expose Head as public API.
    • Fixes critical bugs that help with DenseFeatures usability in TF2
  • tf.data:
    • Promoting unbatch from experimental to core API.
    • Adding support for datasets as inputs to from_tensors and from_tensor_slices and batching and unbatching of nested datasets.
  • tf.keras:
    • tf.keras.estimator.model_to_estimator now supports exporting to tf.train.Checkpoint format, which allows the saved checkpoints to be compatible with model.load_weights.
    • Saving a Keras Model using tf.saved_model.save now saves the list of variables, trainable variables, regularization losses, and the call function.
    • Deprecated tf.keras.experimental.export_saved_model and tf.keras.experimental.function. Please use tf.keras.models.save_model(..., save_format='tf') and tf.keras.models.load_model instead.
    • Add an implementation=3 mode for tf.keras.layers.LocallyConnected2D and tf.keras.layers.LocallyConnected1D layers using tf.SparseTensor to store weights, allowing a dramatic speedup for large sparse models.
    • Enable the Keras compile API experimental_run_tf_function flag by default. This flag enables single training/eval/predict execution path. With this 1. All input types are converted to Dataset. 2. When distribution strategy is not specified this goes through the no-op distribution strategy path. 3. Execution is wrapped in tf.function unless run_eagerly=True is set in compile.
    • Raise error if batch_size argument is used when input is dataset/generator/keras sequence.
  • tf.lite
    • Add GATHER support to NN API delegate.
    • tflite object detection script has a debug mode.
    • Add delegate support for QUANTIZE.
    • Added evaluation script for COCO minival.
    • Add delegate support for QUANTIZED_16BIT_LSTM.
    • Converts hardswish subgraphs into atomic ops.
  • Add support for defaulting the value of cycle_length argument of tf.data.Dataset.interleave to the number of schedulable CPU cores.
  • parallel_for: Add converter for MatrixDiag.
  • Add narrow_range attribute to QuantizeAndDequantizeV2 and V3.
  • Added new op: tf.strings.unsorted_segment_join.
  • Add HW acceleration support for topK_v2.
  • Add new TypeSpec classes.
  • CloudBigtable version updated to v0.10.0.
  • Expose Head as public API.
  • Update docstring for gather to properly describe the non-empty batch_dims case.
  • Added tf.sparse.from_dense utility function.
  • Improved ragged tensor support in TensorFlowTestCase.
  • Makes the a-normal form transformation in Pyct configurable as to which nodes are converted to variables and which are not.
  • ResizeInputTensor now works for all delegates.
  • Add EXPAND_DIMS support to NN API delegate TEST: expand_dims_test
  • tf.cond emits a StatelessIf op if the branch functions are stateless and do not touch any resources.
  • tf.cond, tf.while and if and while in AutoGraph now accept a nonscalar predicate if has a single element. This does not affect non-V2 control flow.
  • tf.while_loop emits a StatelessWhile op if the cond and body functions are stateless and do not touch any resources.
  • Refactors code in Quant8 LSTM support to reduce TFLite binary size.
  • Add support of local soft device placement for eager op.
  • Add HW acceleration support for LogSoftMax.
  • Added a function nested_value_rowids for ragged tensors.
  • Add guard to avoid acceleration of L2 Normalization with input rank != 4
  • Add tf.math.cumulative_logsumexp operation.
  • Add tf.ragged.stack.
  • Fix memory allocation problem when calling AddNewInputConstantTensor.
  • Delegate application failure leaves interpreter in valid state.
  • Add check for correct memory alignment to MemoryAllocation::MemoryAllocation().
  • Extracts NNAPIDelegateKernel from nnapi_delegate.cc
  • Added support for FusedBatchNormV3 in converter.
  • A ragged to dense op for directly calculating tensors.
  • Fix accidental quadratic graph construction cost in graph-mode tf.gradients().

Thanks to our Contributors

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

a6802739, Aaron Ma, Abdullah Selek, Abolfazl Shahbazi, Ag Ramesh, Albert Z. Guo, Albin Joy, Alex Itkes, Alex Sergeev, Alexander Pivovarov, Alexey Romanov, alhkad, Amit Srivastava, amoitra, Andrew Lihonosov, Andrii Prymostka, Anuj Rawat, Astropeak, Ayush Agrawal, Bairen Yi, Bas Aarts, Bastian Eichenberger, Ben Barsdell, Benjamin Peterson, bhack, Bharat Raghunathan, Bhavani Subramanian, Bryan Cutler, candy.dc, Cao Zongyan, Captain-Pool, Casper Da Costa-Luis, Chen Guoyin, Cheng Chang, chengchingwen, Chong Yan, Choong Yin Thong, Christopher Yeh, Clayne Robison, Coady, Patrick, Dan Ganea, David Norman, Denis Khalikov, Deven Desai, Diego Caballero, Duncan Dean, Duncan Riach, Dwight J Lyle, Eamon Ito-Fisher, eashtian3, EFanZh, ejot, Elroy Ashtian Jr, Eric Schweitz, Fangjun Kuang, Fei Hu, fo40225, formath, Fred Reiss, Frederic Bastien, Fredrik Knutsson, G. Hussain Chinoy, Gabriel, gehring, George Grzegorz Pawelczak, Gianluca Varisco, Gleb Popov, Greg Peatfield, Guillaume Klein, Gurpreet Singh, Gustavo Lima Chaves, haison, Haraldur TóMas HallgríMsson, HarikrishnanBalagopal, HåKon Sandsmark, I-Hong, Ilham Firdausi Putra, Imran Salam, Jason Zaman, Jason Zavaglia, jayhpark530, jefby, Jeff Daily, Jeffrey Poznanovic, Jekyll Lai, Jeroen BéDorf, Jerry Shih, jerryyin, jiakai, JiangXIAO, Joe Bowser, Joel Shapiro, Johan Gunnarsson, Jojimon Varghese, Joon, Josh Beal, Julian Niedermeier, Jun Wan, Junqin Zhang, Junyuan Xie, Justin Tunis, Kaixi Hou, Karl Lessard, Karthik Muthuraman, Kbhute-Ibm, khanhlvg, Koock Yoon, kstuedem, Kyuwon Kim, Lakshay Tokas, leike666666, leonard951, Leslie-Fang, Leslie-Fang-Intel, Li, Guizi, Lukas Folle, Lukas Geiger, Mahmoud Abuzaina, Manraj Singh Grover, Margaret Maynard-Reid, Mark Ryan, Matt Conley, Matthew Bentham, Matthew Denton, mbhuiyan, mdfaijul, Mei Jie, merturl, MichaelKonobeev, Michal W. Tarnowski, minds, mpppk, musikisomorphie, Nagy Mostafa, Nayana Thorat, Neil, Niels Ole Salscheider, Niklas SilfverströM, Niranjan Hasabnis, ocjosen, olramde, Pariksheet Pinjari, Patrick J. Lopresti, Patrik Gustavsson, per1234, PeterLee, Phan Van Nguyen Duc, Phillip Kravtsov, Pooya Davoodi, Pranav Marathe, Putra Manggala, Qingqing Cao, Rajeshwar Reddy T, Ramon ViñAs, Rasmus Diederichsen, Reuben Morais, richardbrks, robert, RonLek, Ryan Jiang, saishruthi, Saket Khandelwal, Saleem Abdulrasool, Sami Kama, Sana-Damani, Sergii Khomenko, Severen Redwood, Shubham Goyal, Sigrid Keydana, Siju Samuel, sleighsoft, smilu97, Son Tran, Srini511, srinivasan.narayanamoorthy, Sumesh Udayakumaran, Sungmann Cho, Tae-Hwan Jung, Taehoon Lee, Takeshi Watanabe, TengLu, terryky, TheMindVirus, ThisIsIsaac, Till Hoffmann, Timothy Liu, Tomer Gafner, Tongxuan Liu, Trent Lo, Trevor Morris, Uday Bondhugula, Vasileios Lioutas, vbvg2008, Vishnuvardhan Janapati, Vivek Suryamurthy, Wei Wang, Wen-Heng (Jack) Chung, wenxizhu, William D. Irons, winstonq, wyzhao, Xiaoming (Jason) Cui, Xinan Jiang, Xinping Wang, Yann-Yy, Yasir Modak, Yong Tang, Yongfeng Gu, Yuchen Ying, Yuxin Wu, zyeric, 王振华 (Zhenhua Wang)

Release 2.0.0

Major Features and Improvements

TensorFlow 2.0 focuses on simplicity and ease of use, featuring updates like:

  • Easy model building with Keras and eager execution.
  • Robust model deployment in production on any platform.
  • Powerful experimentation for research.
  • API simplification by reducing duplication and removing deprecated endpoints.

For details on best practices with 2.0, see the Effective 2.0 guide

For information on upgrading your existing TensorFlow 1.x models, please refer to our Upgrade and Migration guides. We have also released a collection of tutorials and getting started guides.

Highlights

  • TF 2.0 delivers Keras as the central high level API used to build and train models. Keras provides several model-building APIs such as Sequential, Functional, and Subclassing along with eager execution, for immediate iteration and intuitive debugging, and tf.data, for building scalable input pipelines. Checkout guide for additional details.
  • Distribution Strategy: TF 2.0 users will be able to use the tf.distribute.Strategy API to distribute training with minimal code changes, yielding great out-of-the-box performance. It supports distributed training with Keras model.fit, as well as with custom training loops. Multi-GPU support is available, along with experimental support for multi worker and Cloud TPUs. Check out the guide for more details.
  • Functions, not Sessions. The traditional declarative programming model of building a graph and executing it via a tf.Session is discouraged, and replaced with by writing regular Python functions. Using the tf.function decorator, such functions can be turned into graphs which can be executed remotely, serialized, and optimized for performance.
  • Unification of tf.train.Optimizers and tf.keras.Optimizers. Use tf.keras.Optimizers for TF2.0. compute_gradients is removed as public API, use GradientTape to compute gradients.
  • AutoGraph translates Python control flow into TensorFlow expressions, allowing users to write regular Python inside tf.function-decorated functions. AutoGraph is also applied in functions used with tf.data, tf.distribute and tf.keras APIs.
  • Unification of exchange formats to SavedModel. All TensorFlow ecosystem projects (TensorFlow Lite, TensorFlow JS, TensorFlow Serving, TensorFlow Hub) accept SavedModels. Model state should be saved to and restored from SavedModels.
  • API Changes: Many API symbols have been renamed or removed, and argument names have changed. Many of these changes are motivated by consistency and clarity. The 1.x API remains available in the compat.v1 module. A list of all symbol changes can be found here.
    • API clean-up, included removing tf.app, tf.flags, and tf.logging in favor of absl-py.
  • No more global variables with helper methods like tf.global_variables_initializer and tf.get_global_step.
  • Add toggles tf.enable_control_flow_v2() and tf.disable_control_flow_v2() for enabling/disabling v2 control flow.
  • Enable v2 control flow as part of tf.enable_v2_behavior() and TF2_BEHAVIOR=1.
  • Fixes autocomplete for most TensorFlow API references by switching to use relative imports in API __init__.py files.
  • Auto Mixed-Precision graph optimizer simplifies converting models to float16 for acceleration on Volta and Turing Tensor Cores. This feature can be enabled by wrapping an optimizer class with tf.train.experimental.enable_mixed_precision_graph_rewrite().
  • Add environment variable TF_CUDNN_DETERMINISTIC. Setting to "true" or "1" forces the selection of deterministic cuDNN convolution and max-pooling algorithms. When this is enabled, the algorithm selection procedure itself is also deterministic.

Breaking Changes

  • Many backwards incompatible API changes have been made to clean up the APIs and make them more consistent.

  • Toolchains:

    • TensorFlow 2.0.0 is built using devtoolset7 (GCC7) on Ubuntu 16. This may lead to ABI incompatibilities with extensions built against earlier versions of TensorFlow.
    • Tensorflow code now produces 2 different pip packages: tensorflow_core containing all the code (in the future it will contain only the private implementation) and tensorflow which is a virtual pip package doing forwarding to tensorflow_core (and in the future will contain only the public API of tensorflow). We don't expect this to be breaking, unless you were importing directly from the implementation. Removed the freeze_graph command line tool; SavedModel should be used in place of frozen graphs.

  • tf.contrib:

    • tf.contrib has been deprecated, and functionality has been either migrated to the core TensorFlow API, to an ecosystem project such as tensorflow/addons or tensorflow/io, or removed entirely.
    • Remove tf.contrib.timeseries dependency on TF distributions.
    • Replace contrib references with tf.estimator.experimental.* for apis in early_stopping.py.

  • tf.estimator:

    • Premade estimators in the tf.estimator.DNN/Linear/DNNLinearCombined family have been updated to use tf.keras.optimizers instead of the tf.compat.v1.train.Optimizers. If you do not pass in an optimizer= arg or if you use a string, the premade estimator will use the Keras optimizer. This is checkpoint breaking, as the optimizers have separate variables. A checkpoint converter tool for converting optimizers is included with the release, but if you want to avoid any change, switch to the v1 version of the estimator: tf.compat.v1.estimator.DNN/Linear/DNNLinearCombined*.
    • Default aggregation for canned Estimators is now SUM_OVER_BATCH_SIZE. To maintain previous default behavior, please pass SUM as the loss aggregation method.
    • Canned Estimators don’t support input_layer_partitioner arg in the API. If you have this arg, you will have to switch to tf.compat.v1 canned Estimators.
    • Estimator.export_savedmodel has been renamed to export_saved_model.
    • When saving to SavedModel, Estimators will strip default op attributes. This is almost always the correct behavior, as it is more forwards compatible, but if you require that default attributes to be saved with the model, please use tf.compat.v1.Estimator.
    • Feature Columns have been upgraded to be more Eager-friendly and to work with Keras. As a result, tf.feature_column.input_layer has been deprecated in favor of tf.keras.layers.DenseFeatures. v1 feature columns have direct analogues in v2 except for shared_embedding_columns, which are not cross-compatible with v1 and v2. Use tf.feature_column.shared_embeddings instead.

  • tf.keras:

    • OMP_NUM_THREADS is no longer used by the default Keras config. To configure the number of threads, use tf.config.threading APIs.
    • tf.keras.model.save_model and model.save now defaults to saving a TensorFlow SavedModel. HDF5 files are still supported.
    • Deprecated tf.keras.experimental.export_saved_model and tf.keras.experimental.function. Please use tf.keras.models.save_model(..., save_format='tf') and tf.keras.models.load_model instead.
    • Layers now default to float32, and automatically cast their inputs to the layer's dtype. If you had a model that used float64, it will probably silently use float32 in TensorFlow 2, and a warning will be issued that starts with Layer <layer-name> is casting an input tensor from dtype float64 to the layer's dtype of float32. To fix, either set the default dtype to float64 with tf.keras.backend.set_floatx('float64'), or pass dtype='float64' to each of the Layer constructors. See tf.keras.layers.Layer for more information.

  • tf.lite:

    • Removed lite.OpHint, lite.experimental, and lite.constant from 2.0 API.

  • Tensors are no longer hashable, but instead compare element-wise with == and !=. Use tf.compat.v1.disable_tensor_equality() to return to the previous behavior.

  • Performing equality operations on Tensors or Variables with incompatible shapes an exception is no longer thrown. Instead __eq__ returns False and __ne__ returns True.

  • Removed tf.string_split from v2 API.

  • Deprecated the use of constraint= and .constraint with ResourceVariable.

  • Add UnifiedGRU as the new GRU implementation for tf2.0. Change the default recurrent activation function for GRU from hard_sigmoid to sigmoid, and reset_after to True in 2.0. Historically recurrent activation is hard_sigmoid since it is fast than 'sigmoid'. With new unified backend between CPU and GPU mode, since the CuDNN kernel is using sigmoid, we change the default for CPU mode to sigmoid as well. With that, the default GRU will be compatible with both CPU and GPU kernel. This will enable user with GPU to use CuDNN kernel by default and get a 10x performance boost in training. Note that this is checkpoint breaking change. If user want to use their 1.x pre-trained checkpoint, please construct the layer with GRU(recurrent_activation='hard_sigmoid', reset_after=False) to fallback to 1.x behavior.

  • CUDNN_INSTALL_PATH, TENSORRT_INSTALL_PATH, NCCL_INSTALL_PATH, NCCL_HDR_PATH are deprecated. Use TF_CUDA_PATHS instead which supports a comma-separated list of base paths that are searched to find CUDA libraries and headers.

Refer to our public project status tracker and issues tagged with 2.0 on GitHub for insight into recent issues and development progress.

If you experience any snags when using TF 2.0, please let us know at the TF 2.0 Testing User Group. We have a support mailing list as well as weekly testing meetings, and would love to hear your migration feedback and questions.

Bug Fixes and Other Changes

  • tf.contrib:

    • Expose tf.contrib.proto.* ops in tf.io (they will exist in TF2)

  • tf.data:

    • Add support for TensorArrays to tf.data Dataset.
    • Integrate Ragged Tensors with tf.data.
    • All core and experimental tf.data transformations that input user-defined functions can span multiple devices now.
    • Extending the TF 2.0 support for shuffle(..., reshuffle_each_iteration=True) and cache() to work across different Python iterators for the same dataset.
    • Removing the experimental_numa_aware option from tf.data.Options.
    • Add num_parallel_reads and passing in a Dataset containing filenames into TextLineDataset and FixedLengthRecordDataset.
    • Add support for defaulting the value of cycle_length argument of tf.data.Dataset.interleave to the number of schedulable CPU cores.
    • Promoting tf.data.experimental.enumerate_dataset to core as tf.data.Dataset.enumerate.
    • Promoting tf.data.experimental.unbatch to core as tf.data.Dataset.unbatch.
    • Adds option for introducing slack in the pipeline to reduce CPU contention, via tf.data.Options().experimental_slack = True
    • Added experimental support for parallel batching to batch() and padded_batch(). This functionality can be enabled through tf.data.Options().
    • Support cancellation of long-running reduce.
    • Now we use dataset node name as prefix instead of the op name, to identify the component correctly in metrics, for pipelines with repeated components.
    • Improve the performance of datasets using from_tensors().
    • Promoting unbatch from experimental to core API.
    • Adding support for datasets as inputs to from_tensors and from_tensor_slices and batching and unbatching of nested datasets.

  • tf.distribute:

    • Enable tf.distribute.experimental.MultiWorkerMirroredStrategy working in eager mode.
    • Callbacks are supported in MultiWorkerMirroredStrategy.
    • Disable run_eagerly and distribution strategy if there are symbolic tensors added to the model using add_metric or add_loss.
    • Loss and gradients should now more reliably be correctly scaled w.r.t. the global batch size when using a tf.distribute.Strategy.
    • Set default loss reduction as AUTO for improving reliability of loss scaling with distribution strategy and custom training loops. AUTO indicates that the reduction option will be determined by the usage context. For almost all cases this defaults to SUM_OVER_BATCH_SIZE. When used in distribution strategy scope, outside of built-in training loops such as tf.keras compile and fit, we expect reduction value to be 'None' or 'SUM'. Using other values will raise an error.
    • Support for multi-host ncclAllReduce in Distribution Strategy.

  • tf.estimator:

    • Replace tf.contrib.estimator.add_metrics with tf.estimator.add_metrics
    • Use tf.compat.v1.estimator.inputs instead of tf.estimator.inputs
    • Replace contrib references with tf.estimator.experimental.* for apis in early_s in Estimator
    • Canned Estimators will now use keras optimizers by default. An error will be raised if tf.train.Optimizers are used, and you will have to switch to tf.keras.optimizers or tf.compat.v1 canned Estimators.
    • A checkpoint converter for canned Estimators has been provided to transition canned Estimators that are warm started from tf.train.Optimizers to tf.keras.optimizers.
    • Losses are scaled in canned estimator v2 and not in the optimizers anymore. If you are using Estimator + distribution strategy + optimikzer v1 then the behavior does not change. This implies that if you are using custom estimator with optimizer v2, you have to scale losses. We have new utilities to help scale losses tf.nn.compute_average_loss, tf.nn.scale_regularization_loss.

  • tf.keras:

    • Premade models (including Linear and WideDeep) have been introduced for the purpose of replacing Premade estimators.
    • Model saving changes
    • model.save and tf.saved_model.save may now save to the TensorFlow SavedModel format. The model can be restored using tf.keras.models.load_model. HDF5 files are still supported, and may be used by specifying save_format="h5" when saving.
    • Raw TensorFlow functions can now be used in conjunction with the Keras Functional API during model creation. This obviates the need for users to create Lambda layers in most cases when using the Functional API. Like Lambda layers, TensorFlow functions that result in Variable creation or assign ops are not supported.
    • Add support for passing list of lists to the metrics argument in Keras compile.
    • Add tf.keras.layers.AbstractRNNCell as the preferred implementation for RNN cells in TF v2. User can use it to implement RNN cells with custom behavior.
    • Keras training and validation curves are shown on the same plot when using the TensorBoard callback.
    • Switched Keras fit/evaluate/predict execution to use only a single unified path by default unless eager execution has been explicitly disabled, regardless of input type. This unified path places an eager-friendly training step inside of a tf.function. With this
    • All input types are converted to Dataset.
    • The path assumes there is always a distribution strategy. when distribution strategy is not specified the path uses a no-op distribution strategy.
    • The training step is wrapped in tf.function unless run_eagerly=True is set in compile. The single path execution code does not yet support all use cases. We fallback to the existing v1 execution paths if your model contains the following:
      1. sample_weight_mode in compile
      2. weighted_metrics in compile
      3. v1 optimizer
      4. target tensors in compile If you are experiencing any issues because of this change, please inform us (file an issue) about your use case and you can unblock yourself by setting experimental_run_tf_function=False in compile meanwhile. We have seen couple of use cases where the model usage pattern is not as expected and would not work with this change.
    • output tensors of one layer is used in the constructor of another.
    • symbolic tensors outside the scope of the model are used in custom loss functions. The flag can be disabled for these cases and ideally the usage pattern will need to be fixed.
    • Mark Keras set_session as compat.v1 only.
    • tf.keras.estimator.model_to_estimator now supports exporting to tf.train.Checkpoint format, which allows the saved checkpoints to be compatible with model.load_weights.
    • keras.backend.resize_images (and consequently, keras.layers.Upsampling2D) behavior has changed, a bug in the resizing implementation was fixed.
    • Add an implementation=3 mode for tf.keras.layers.LocallyConnected2D and tf.keras.layers.LocallyConnected1D layers using tf.SparseTensor to store weights, allowing a dramatic speedup for large sparse models.
    • Raise error if batch_size argument is used when input is dataset/generator/keras sequence.
    • Update TF 2.0 keras.backend.name_scope to use TF 2.0 name_scope.
    • Add v2 module aliases for losses, metrics, initializers and optimizers: tf.losses = tf.keras.losses & tf.metrics = tf.keras.metrics & tf.initializers = tf.keras.initializers & tf.optimizers = tf.keras.optimizers.
    • Updates binary cross entropy logic in Keras when input is probabilities. Instead of converting probabilities to logits, we are using the cross entropy formula for probabilities.
    • Added public APIs for cumsum and cumprod keras backend functions.
    • Add support for temporal sample weight mode in subclassed models.
    • Raise ValueError if an integer is passed to the training APIs.
    • Added fault-tolerance support for training Keras model via model.fit() with MultiWorkerMirroredStrategy, tutorial available.
    • Custom Callback tutorial is now available.
    • To train with tf.distribute, Keras API is recommended over estimator.
    • steps_per_epoch and steps arguments are supported with numpy arrays.
    • New error message when unexpected keys are used in sample_weight/class_weight dictionaries
    • Losses are scaled in Keras compile/fit and not in the optimizers anymore. If you are using custom training loop, we have new utilities to help scale losses tf.nn.compute_average_loss, tf.nn.scale_regularization_loss.
    • Layer apply and add_variable APIs are deprecated.
    • Added support for channels first data format in cross entropy losses with logits and support for tensors with unknown ranks.
    • Error messages will be raised if add_update, add_metric, add_loss, activity regularizers are used inside of a control flow branch.
    • New loss reduction types:
    • AUTO: Indicates that the reduction option will be determined by the usage context. For almost all cases this defaults to SUM_OVER_BATCH_SIZE. When used with tf.distribute.Strategy, outside of built-in training loops such as tf.keras compile and fit, we expect reduction value to be SUM or NONE. Using AUTO in that case will raise an error.
    • NONE: Weighted losses with one dimension reduced (axis=-1, or axis specified by loss function). When this reduction type used with built-in Keras training loops like fit/evaluate, the unreduced vector loss is passed to the optimizer but the reported loss will be a scalar value.
    • SUM: Scalar sum of weighted losses. 4. SUM_OVER_BATCH_SIZE: Scalar SUM divided by number of elements in losses. This reduction type is not supported when used with tf.distribute.Strategy outside of built-in training loops like tf.keras compile/fit.
    • Wraps losses passed to the compile API (strings and v1 losses) which are not instances of v2 Loss class in LossWrapper class. => All losses will now use SUM_OVER_BATCH_SIZE reduction as default.
    • model.add_loss(symbolic_tensor) should work in ambient eager.
    • Update metric name to always reflect what the user has given in compile. Affects following cases
    • When name is given as 'accuracy'/'crossentropy'
    • When an aliased function name is used eg. 'mse'
    • Removing the weighted prefix from weighted metric names.
    • Allow non-Tensors through v2 losses.
    • Add v2 sparse categorical crossentropy metric.
    • Add v2 APIs for AUCCurve and AUCSummationMethod enums.
    • add_update can now be passed a zero-arg callable in order to support turning off the update when setting trainable=False on a Layer of a Model compiled with run_eagerly=True.
    • Standardize the LayerNormalization API by replacing the args norm_axis and params_axis with axis.
    • Fixed critical bugs that help with DenseFeatures usability in TF2

  • tf.lite:

    • Added evaluation script for COCO minival
    • Add delegate support for QUANTIZE.
    • Add GATHER support to NN API delegate.
    • Added support for TFLiteConverter Python API in 2.0. Contains functions from_saved_model, from_keras_file, and from_concrete_functions.
    • Add EXPAND_DIMS support to NN API delegate TEST.
    • Add narrow_range attribute to QuantizeAndDequantizeV2 and V3.
    • Added support for tflite_convert command line tool in 2.0.
    • Post-training quantization tool supports quantizing weights shared by multiple operations. The models made with versions of this tool will use INT8 types for weights and will only be executable interpreters from this version onwards.
    • Post-training quantization tool supports fp16 weights and GPU delegate acceleration for fp16.
    • Add delegate support for QUANTIZED_16BIT_LSTM.
    • Extracts NNAPIDelegateKernel from nnapi_delegate.cc

  • TensorRT

    • Add TensorFlow 2.0-compatible TrtGraphConverterV2 API for TensorRT conversion. TensorRT initialization arguments are now passed wrapped in a named-tuple, TrtConversionParams, rather than as separate arguments as in TrtGraphConverter.
    • Changed API to optimize TensorRT engines during graph optimization. This is now done by calling converter.build() where previously is_dynamic_op=False would be set.
    • converter.convert() no longer returns a tf.function. Now the function must be accessed from the saved model.
    • The converter.calibrate() method has been removed. To trigger calibration, a calibration_input_fn should be provided to converter.convert().

  • Other:

    • Fix accidental quadratic graph construction cost in graph-mode tf.gradients().
    • ResourceVariable's gather op supports batch dimensions.
    • ResourceVariable support for gather_nd.
    • ResourceVariable and Variable no longer accepts constraint in the constructor, nor expose it as a @property.
    • Added gradient for SparseToDense op.
    • Expose a flag that allows the number of threads to vary across Python benchmarks.
    • image.resize in 2.0 now supports gradients for the new resize kernels.
    • image.resize now considers proper pixel centers and has new kernels (incl. anti-aliasing).
    • Renamed tf.image functions to remove duplicate "image" where it is redundant.
    • Variadic reduce is supported on CPU Variadic reduce is supported on CPU
    • Remove unused StringViewVariantWrapper.
    • Delete unused Fingerprint64Map op registration
    • Add broadcasting support to tf.matmul.
    • Add C++ Gradient for BatchMatMulV2.
    • Add tf.math.cumulative_logsumexp operation.
    • Add ellipsis (...) support for tf.einsum().
    • Add expand_composites argument to all nest.* methods.
    • Added strings.byte_split.
    • Add a new "result_type" parameter to tf.strings.split.
    • Add name argument to tf.string_split and tf.strings_split.
    • Extend tf.strings.split to support inputs with any rank.
    • Added tf.random.binomial.
    • Added key and skip methods to random.experimental.Generator.
    • Extend tf.function with basic support for CompositeTensors arguments (such as SparseTensor and RaggedTensor).
    • parallel_for.pfor: add converters for Softmax, LogSoftmax, IsNaN, All, Any, and MatrixSetDiag.
    • parallel_for: add converters for LowerTriangularSolve and Cholesky.
    • parallel_for: add converters for LogMatrixDeterminant and MatrixBandPart.
    • parallel_for: Add converter for MatrixDiag.
    • parallel_for: Add converters for OneHot, LowerBound, UpperBound.
    • parallel_for: add converter for BroadcastTo.
    • Add pfor converter for Squeeze.
    • Add RaggedTensor.placeholder().
    • Add ragged tensor support to tf.squeeze.
    • Update RaggedTensors to support int32 row_splits.
    • Allow LinearOperator.solve to take a LinearOperator.
    • Allow all dtypes for LinearOperatorCirculant.
    • Introduce MaxParallelism method
    • Add LinearOperatorHouseholder.
    • Adds Philox support to new stateful RNG's XLA path.
    • Added TensorSpec support for CompositeTensors.
    • Added tf.linalg.tridiagonal_solve op.
    • Added partial_pivoting input parameter to tf.linalg.tridiagonal_solve.
    • Added gradient to tf.linalg.tridiagonal_solve.
    • Added tf.linalg.tridiagonal_mul op.
    • Added GPU implementation of tf.linalg.tridiagonal_matmul.
    • Added LinearOperatorToeplitz.
    • Upgraded LIBXSMM to version 1.11.
    • Uniform processing of quantized embeddings by Gather and EmbeddingLookup Ops.
    • Correct a misstatement in the documentation of the sparse softmax cross entropy logit parameter.
    • Add tf.ragged.boolean_mask.
    • tf.switch_case added, which selects a branch_fn based on a branch_index.
    • The C++ kernel of gather op supports batch dimensions.
    • Fixed default value and documentation for trainable arg of tf.Variable.
    • EagerTensor now supports numpy buffer interface for tensors.
    • This change bumps the version number of the FullyConnected Op to 5.
    • Added new op: tf.strings.unsorted_segment_join.
    • Added HW acceleration support for topK_v2.
    • CloudBigtable version updated to v0.10.0 BEGIN_PUBLIC CloudBigtable version updated to v0.10.0.
    • Expose Head as public API.
    • Added tf.sparse.from_dense utility function.
    • Improved ragged tensor support in TensorFlowTestCase.
    • Added a function nested_value_rowids for ragged tensors.
    • Added tf.ragged.stack.
    • Makes the a-normal form transformation in Pyct configurable as to which nodes are converted to variables and which are not.
    • ResizeInputTensor now works for all delegates.
    • tf.cond emits a StatelessIf op if the branch functions are stateless and do not touch any resources.
    • Add support of local soft device placement for eager op.
    • Pass partial_pivoting to the _TridiagonalSolveGrad.
    • Add HW acceleration support for LogSoftMax.
    • Add guard to avoid acceleration of L2 Normalization with input rank != 4
    • Fix memory allocation problem when calling AddNewInputConstantTensor.
    • Delegate application failure leaves interpreter in valid state
    • tf.while_loop emits a StatelessWhile op if the cond and body functions are stateless and do not touch any resources.
    • tf.cond, tf.while and if and while in AutoGraph now accept a nonscalar predicate if has a single element. This does not affect non-V2 control flow.
    • Fix potential security vulnerability where decoding variant tensors from proto could result in heap out of bounds memory access.
    • Only create a GCS directory object if the object does not already exist.
    • Introduce dynamic constructor argument in Layer and Model, which should be set to True when using imperative control flow in the call method.
    • Begin adding Go wrapper for C Eager API.
    • XLA HLO graphs can be inspected with interactive_graphviz tool now.
    • Add dataset ops to the graph (or create kernels in Eager execution) during the python Dataset object creation instead doing it during Iterator creation time.
    • Add batch_dims argument to tf.gather.
    • The behavior of tf.gather is now correct when axis=None and batch_dims<0.
    • Update docstring for gather to properly describe the non-empty batch_dims case.
    • Removing of dtype in the constructor of initializers and partition_info in call.
    • Add tf.math.nextafter op.
    • Turn on MKL-DNN contraction kernels by default. MKL-DNN dynamically dispatches the best kernel implementation based on CPU vector architecture. To disable them, build with --define=tensorflow_mkldnn_contraction_kernel=0.
    • tf.linspace(start, stop, num) now always uses "stop" as last value (for num > 1)
    • Added top-k to precision and recall to keras metrics.
    • Add a ragged size op and register it to the op dispatcher
    • Transitive dependencies on :pooling_ops were removed. Some users may need to add explicit dependencies on :pooling_ops if they reference the operators from that library.
    • Add CompositeTensor base class.
    • Malformed gif images could result in an access out of bounds in the color palette of the frame. This has been fixed now
    • Add templates and interfaces for creating lookup tables
    • Tensor::UnsafeCopyFromInternal deprecated in favor Tensor::BitcastFrom.
    • In map_vectorization optimization, reduce the degree of parallelism in the vectorized map node.
    • Add variant wrapper for absl::string_view.
    • Add OpKernels for some stateless maps.
    • DType is no longer convertible to an int. Use dtype.as_datatype_enum instead of int(dtype) to get the same result.
    • Support both binary and -1/1 label input in v2 hinge and squared hinge losses.
    • Added LinearOperator.adjoint and LinearOperator.H (alias).
    • Expose CriticalSection in core as tf.CriticalSection.
    • Enhanced graphviz output.
    • Add opkernel templates for common table operations.
    • Fix callbacks do not log values in eager mode when a deferred build model is used.
    • SignatureDef util functions have been deprecated.
    • Update Fingerprint64Map to use aliases
    • Add legacy string flat hash map op kernels.
    • Add support for add_metric in the graph function mode.
    • Updating cosine similarity loss - removed the negate sign from cosine similarity.
    • Changed default for gradient accumulation for TPU embeddings to true.
    • Adds summary trace API for collecting graph and profile information.
    • The precision_mode argument to TrtGraphConverter is now case insensitive.

Thanks to our Contributors

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

1e100, a6802739, 4d55397500, a6802739, Abdullah Selek, abenmao, Abolfazl Shahbazi, Adam Richter, Adam Weiss, Ag Ramesh, Alan Du, Albin Joy, Alex, Alex Itkes, Alex Sergeev, Alexander Pivovarov, Alexey Romanov, alhkad, Aman Patel, Amit, Amit Kumar Jaiswal, Amit Srivastava, amoitra, Andreas Eberle, Andrew Lihonosov, Andy Craze, Anshuman Tripathy, Anthony Hsu, Anthony Platanios, Anuj Rawat, arp95, Arpit Shah, Armen Poghosov, armenpoghosov, Astropeak, Ashwin Ramaswami, Arpit Shah, Augustina Ragwitz, Aurelien Geron, AuréLien Geron, avasid, aweers, awesomealex1, Ayush Agrawal, Bas Aarts, Bastian Eichenberger, Bairen Yi, Bayberry Z, Ben Barsdell, Benjamin Peterson, bhack, Bharat Raghunathan, Bhavani Subramanian, Bin Fan, blairhan, BléNesi Attila, Bodin-E, Brandon Carter, Bryan Cutler, candy.dc, Cao Zongyan, Casper Da Costa-Luis, Chao Liu, Chen Guoyin, chenchc, chengchingwen, chie8842, Christian Hansen, Christoph Boeddeker, Christopher Yeh, Clayne Robison, Coady, Patrick, crafet, csukuangfj, ctiijima, Dan Jarvis, Dan Lazewatsky, Daniel Ingram, Daniel Rasmussen, Daniel Salvadori, Dave Airlie, David Norman, Dayananda V, delock, Denis Khalikov, Deven Desai, Dheeraj Rajaram Reddy, Diego Caballero, dmitrievanthony, Donovan Ong, Drew Szurko, Duncan Dean, Duncan Riach, Dustin Neighly, Dwight J Lyle, Eamon Ito-Fisher, eashtian3, Edward Forgacs, EFanZh, ejot, Elroy Ashtian Jr, Eric Schweitz, Evgeniy Polyakov, Fangjun Kuang, Federico Martinez, Fei Hu, Felix Lemke, Filip Matzner, FlashTek, fo40225, formath, FrançOis Chollet, frreiss, Fred Reiss, Frederic Bastien, Fredrik Knutsson, G. Hussain Chinoy, Gabriel, Gautam, gehring, Geoffrey Irving, George Grzegorz Pawelczak, Grzegorz Pawelczak, George Sterpu, Gianluca Varisco, Gleb Popov, Greg Peatfield, Guillaume Klein, Gurpreet Singh, Gustavo Lima Chaves, Gyoung-Yoon Ryoo, haison, Hanton Yang, HanGuo97, Haraldur TóMas HallgríMsson, Hari Shankar, hehongliang, Heungsub Lee, Hoeseong Kim, Huan Li (李卓桓), HåKon Sandsmark, I-Hong, I-Hong Jhuo, Ilham Firdausi Putra, Ilango R, Imran Salam, Innovimax, Jacky Ko, Irene Dea, Ivan Habernal, Jakub Lipinski, Jacky, Jason Zaman, Jason Zavaglia, jayhpark530, jcf94, jefby, Jeff Daily, Jeff Poznanovic, Jeffrey Poznanovic, Jekyll Lai, jer, Jeroen BéDorf, jerryyin, jhalakp, jiakai, Jia Qingtong, Jiankang, JiangXIAO, Joe Bowser, Joe Q, Joe Quadrino, Joel Shapiro, Johan Gunnarsson, Jojimon Varghese, Jonas Rauber, Jonathan Kyl, Jonathan, Joon, Joppe Geluykens, Joseph Friedman, Josh Beal, jtressle, Julian Niedermeier, Junqin Zhang, Justin Dujardin, Justin Tunis, jwu, K. Hodges, kaixih, Kaixi Hou, kjopek, Karl Lessard, Karl Weinmeister, Karthik Muthuraman, Kashif Rasul, Kay Zhu, Kbhute-Ibm, KDR, Keno Fischer, Kevin Mader, khanhlvg, Kilaru Yasaswi Sri Chandra Gandhi, Koan-Sin Tan, Koock Yoon, kouml, ktaebum, Kyuwon Kim, Lakshay Tokas, Laurent Le Brun, leike666666, leonard951, Leslie-Fang, Letian Kang, Li, Guizi, Loo Rong Jie, Lucas Hendren, Lukas Folle, Lukas Geiger, Luke Han, luxupu, lvli, Ma, Guokai, Mahmoud Abuzaina, Maksym Kysylov, Mandar Deshpande, manhyuk, Manraj Singh Grover, Marco Gaido, Marek Drozdowski, Margaret Maynard-Reid, Mark Ryan, mars20, Mateusz Chudyk, Matt Conley, mbhuiyan, mdfaijul, Mei Jie, Melissa Grueter, merturl, MichaelKonobeev, Michael KäUfl, Michal W. Tarnowski, MickaëL Schoentgen, Miguel Morin, Mihail Salnikov, Mikalai Drabovich, Mike Arpaia, Mike Holcomb, minds, monklof, Moses Marin, mpppk, Mr. Metal, Mshr-H, musikisomorphie, nammbash, Natalia Gimelshein, Nathan Luehr, Nayana-Ibm, Nayana Thorat, neargye, Neeraj Pradhan, Nehal J Wani, Neil, Nick, Nick Lewycky, Niels Ole Salscheider, Niklas SilfverströM, Niranjan Hasabnis, Nuka-137, Nutti, ocjosen, olicht, omeir1, P Sudeepam, Paige Bailey, Palmer Lao, Pan Daoxin, Pariksheet Pinjari, Pasquale Minervini, Patrick J. Lopresti, Patrik Gustavsson, Pavel Akhtyamov, Pavel Samolysov, PENGWA, per1234, PeterLee, Phan Van Nguyen Duc, Philipp Jund, Phillip Kravtsov, Pooya Davoodi, Pranav Marathe, Putra Manggala, Qingqing Cao, R S Nikhil Krishna, Rajeshwar Reddy T, Ramon ViñAs, Rasmus Diederichsen, Reuben Morais, robert, Rohit Gupta, Roland Zimmermann, Roman Soldatow, RonLek, Ruizhe, Ryan Jiang, saishruthi, Saleem Abdulrasool, Samantha Andow, Sami Kama, Sana-Damani, Saurabh Deoras, sdamani, Sean Morgan, seanshpark, Sebastien Iooss, Serv-Inc, Severen Redwood, Shahzad Lone, Shashank Gupta, shashvat, Shashvat Chand Shahi, Shubham Goyal, Shashi, Sigrid Keydana, Siju, Siju Samuel, sleighsoft, smilu97, Snease-Abq, Son Tran, Spencer Schaber, sremedios, Srini511, srinivasan.narayanamoorthy, Steve Lang, Steve Nesae, Subin, Sumesh Udayakumaran, Sungmann Cho, sunway513, Supriya Rao, sxwang, Tae-Hwan Jung, Taehoon Lee, Takeo Sawada, Taylor Jakobson, Taylor Thornton, Ted Chang, TengLu, terryky, ThisIsIsaac, ThisIsPIRI, Thomas Deegan, Thomas Hagebols, tianyapiaozi, Till Hoffmann, Tim Zaman, tomguluson92, Tongxuan Liu, Trent Lo, Trevor Morris, TungJerry, Tyorden, Uday Bondhugula, v1incent, Vagif, Vasileios Lioutas, vbvg2008, vcarpani, Vijay Ravichandran, Vikram Tiwari,Viktor Gal, Vishwak Srinivasan, Vincent, Vishnuvardhan Janapati, Vitor-Alves, Vivek Suryamurthy, wangsiyu, wateryzephyr, WeberXie, Wei Wang, WeijieSun, Wen-Heng (Jack) Chung, wenxizhu, Will Battel, William D. Irons, winstonq, wyzhao, Xiaoming (Jason) Cui, Xiaoquan Kong, Xin, Xinping Wang, Yan Facai (颜发才), Yann-Yy, Yasir Modak, Yasuhiro Matsumoto, ymodak, Yong Tang, Yongfeng Gu, Younes Khoudli, Yuan Lin, Yuan (Terry) Tang, Yuchen Ying, Yves-Noel Weweler, zhangyujing, zjjott, zyeric, 王振华 (Zhenhua Wang), 黄鑫

Release 1.14.0

Major Features and Improvements

  • This is the first 1.x release containing the compat.v2 module. This module is required to allow libraries to publish code which works in both 1.x and 2.x. After this release, no backwards incompatible changes are allowed in the 2.0 Python API.
  • Turn on MKL-DNN contraction kernels by default. MKL-DNN dynamically dispatches the best kernel implementation based on CPU vector architecture. To disable them, build with --define=tensorflow_mkldnn_contraction_kernel=0.

Behavioral changes

  • Set default loss reduction as AUTO for improving reliability of loss scaling with distribution strategy and custom training loops. AUTO indicates that the reduction option will be determined by the usage context. For almost all cases this defaults to SUM_OVER_BATCH_SIZE. When used in distribution strategy scope, outside of built-in training loops such as tf.keras compile and fit, we expect reduction value to be 'None' or 'SUM'. Using other values will raise an error.
  • Wraps losses passed to the compile API (strings and v1 losses) which are not instances of v2 Loss class in LossWrapper class. => All losses will now use SUM_OVER_BATCH_SIZE reduction as default.
  • Disable run_eagerly and distribution strategy if there are symbolic tensors added to the model using add_metric or add_loss.
  • tf.linspace(start, stop, num) now always uses "stop" as last value (for num > 1)
  • ResourceVariable and Variable no longer accepts constraint in the constructor, nor expose it as a @property.
  • The behavior of tf.gather is now correct when axis=None and batch_dims<0.
  • Only create a GCS directory object if the object does not already exist.
  • In map_vectorization optimization, reduce the degree of parallelism in the vectorized map node.
  • Bug fix: loss and gradients should now more reliably be correctly scaled w.r.t. the global batch size when using a tf.distribute.Strategy.
  • Updating cosine similarity loss - removed the negate sign from cosine similarity.
  • DType is no longer convertible to an int. Use dtype.as_datatype_enum instead of int(dtype) to get the same result.
  • Changed default for gradient accumulation for TPU embeddings to true.
  • Callbacks now log values in eager mode when a deferred build model is used.
  • Transitive dependencies on :pooling_ops were removed. Some users may need to add explicit dependencies on :pooling_ops if they reference the operators from that library.
  • tf.keras.optimizers default learning rate changes:
    • Adadelta: 1.000 to 0.001
    • Adagrad: 0.01 to 0.001
    • Adamax: 0.002 to 0.001
    • NAdam: 0.002 to 0.001

Bug Fixes and Other Changes

  • Documentation
  • Deprecations and Symbol renames.
    • Remove unused StringViewVariantWrapper
    • Delete unused Fingerprint64Map op registration
    • SignatureDef util functions have been deprecated.
    • Renamed tf.image functions to remove duplicate "image" where it is redundant.
    • tf.keras.experimental.export renamed to tf.keras.experimental.export_saved_model
    • Standardize the LayerNormalization API by replacing the args norm_axis and params_axis with axis.
    • Tensor::UnsafeCopyFromInternal deprecated in favor Tensor::BitcastFrom
  • Keras & Python API
    • Add v2 module aliases for:
    • tf.initializers => tf.keras.initializers
    • tf.losses => tf.keras.losses & tf.metrics => tf.keras.metrics
    • tf.optimizers => tf.keras.optimizers
    • Add tf.keras.layers.AbstractRNNCell as the preferred implementation of RNN cell for TF v2. User can use it to implement RNN cell with custom behavior.
    • Adding clear_losses API to be able to clear losses at the end of forward pass in a custom training loop in eager.
    • Add support for passing list of lists to the metrics param in Keras compile.
    • Added top-k to precision and recall to keras metrics.
    • Adding public APIs for cumsum and cumprod keras backend functions.
    • Fix: model.add_loss(symbolic_tensor) should work in ambient eager.
    • Add name argument to tf.string_split and tf.strings_split
    • Minor change to SavedModels exported from Keras using tf.keras.experimental.export. (SignatureDef key for evaluation mode is now "eval" instead of "test"). This will be reverted back to "test" in the near future.
    • Updates binary cross entropy logic in Keras when input is probabilities. Instead of converting probabilities to logits, we are using the cross entropy formula for probabilities.
    • Raw TensorFlow functions can now be used in conjunction with the Keras Functional API during model creation. This obviates the need for users to create Lambda layers in most cases when using the Functional API. Like Lambda layers, TensorFlow functions that result in Variable creation or assign ops are not supported.
    • Keras training and validation curves are shown on the same plot.
    • Introduce dynamic constructor argument in Layer and Model, which should be set to True when using imperative control flow in the call method.
    • Removing of dtype in the constructor of initializers and partition_info in call.
  • New ops and improved op functionality
    • Add OpKernels for some stateless maps
    • Add v2 APIs for AUCCurve and AUCSummationMethod enums. #tf-metrics-convergence
    • Add tf.math.nextafter op.
    • Add CompositeTensor base class.
    • Add tf.linalg.tridiagonal_solve op.
    • Add opkernel templates for common table operations.
    • Added support for TFLite in TensorFlow 2.0.
    • Adds summary trace API for collecting graph and profile information.
    • Add batch_dims argument to tf.gather.
    • Add support for add_metric in the graph function mode.
    • Add C++ Gradient for BatchMatMulV2.
    • Added tf.random.binomial
    • Added gradient for SparseToDense op.
    • Add legacy string flat hash map op kernels
    • Add a ragged size op and register it to the op dispatcher
    • Add broadcasting support to tf.matmul.
    • Add ellipsis (...) support for tf.einsum()
    • Added LinearOperator.adjoint and LinearOperator.H (alias).
    • Added GPU implementation of tf.linalg.tridiagonal_solve.
    • Added strings.byte_split
    • Add RaggedTensor.placeholder()
    • Add a new "result_type" parameter to tf.strings.split
    • add_update can now be passed a zero-arg callable in order to support turning off the update when setting trainable=False on a Layer of a Model compiled with run_eagerly=True.
    • Add variant wrapper for absl::string_view
    • Add expand_composites argument to all nest.* methods.
    • Add pfor converter for Squeeze.
    • Bug fix for tf.tile gradient
    • Expose CriticalSection in core as tf.CriticalSection.
    • Update Fingerprint64Map to use aliases
    • ResourceVariable support for gather_nd.
    • ResourceVariable's gather op supports batch dimensions.
    • Variadic reduce is supported on CPU
    • Extend tf.function with basic support for CompositeTensors arguments (such as SparseTensor and RaggedTensor).
    • Add templates and interfaces for creating lookup tables
    • Post-training quantization tool supports quantizing weights shared by multiple operations. The models made with versions of this tool will use INT8 types for weights and will only be executable interpreters from this version onwards.
    • Malformed gif images could result in an access out of bounds in the color palette of the frame. This has been fixed now
    • image.resize now considers proper pixel centers and has new kernels (incl. anti-aliasing).
    • Added an isotonic regression solver (tf.nn.isotonic_regression).
  • Performance
    • Turn on MKL-DNN contraction kernels by default. MKL-DNN dynamically dispatches the best kernel implementation based on CPU vector architecture. To disable them, build with --define=tensorflow_mkldnn_contraction_kernel=0.
    • Support for multi-host ncclAllReduce in Distribution Strategy.
    • Expose a flag that allows the number of threads to vary across Python benchmarks.
  • TensorFlow 2.0 Development
    • Add v2 sparse categorical crossentropy metric.
    • Allow non-Tensors through v2 losses.
    • Add UnifiedGRU as the new GRU implementation for tf2.0. Change the default recurrent activation function for GRU from 'hard_sigmoid' to 'sigmoid', and 'reset_after' to True in 2.0. Historically recurrent activation is 'hard_sigmoid' since it is fast than 'sigmoid'. With new unified backend between CPU and GPU mode, since the CuDNN kernel is using sigmoid, we change the default for CPU mode to sigmoid as well. With that, the default GRU will be compatible with both CPU and GPU kernel. This will enable user with GPU to use CuDNN kernel by default and get a 10x performance boost in training. Note that this is checkpoint breaking change. If user want to use their 1.x pre-trained checkpoint, please construct the layer with GRU(recurrent_activation='hard_sigmoid', reset_after=False) to fallback to 1.x behavior.
    • TF 2.0 - Update metric name to always reflect what the user has given in compile. Affects following cases 1. When name is given as 'accuracy'/'crossentropy' 2. When an aliased function name is used eg. 'mse' 3. Removing the weighted prefix from weighted metric names.
    • Begin adding Go wrapper for C Eager API
    • image.resize in 2.0 now supports gradients for the new resize kernels.
    • removed tf.string_split from v2 API
    • Expose tf.contrib.proto.* ops in tf.io (they will exist in TF2)
    • "Updates the TFLiteConverter API in 2.0. Changes from_concrete_function to from_concrete_functions."
    • Enable tf.distribute.experimental.MultiWorkerMirroredStrategy working in eager mode.
    • Support both binary and -1/1 label input in v2 hinge and squared hinge losses.
  • TensorFlow Lite
    • "Adds support for tflite_convert in 2.0."
    • "Remove lite.OpHint, lite.experimental, and lite.constant from 2.0 API."
  • tf.contrib
  • tf.data
    • Add num_parallel_reads and passing in a Dataset containing filenames into TextLineDataset and FixedLengthRecordDataset
    • Going forward we operate in TF 2.0, this change is part of the effort to slowly converting XYZDataset to DatasetV2 type which is the official version going to be used in TF 2.0 and motivated by some compatibility issue found, _BigtableXYZDataset (of type DatasetV2) does not implement the _as_variant_tensor() of DatasetV1, when moving contrib.bigtable to tensorflow_io. Converting into DatasetV2 removes the overheads to maintain V1 while we are moving into TF 2.0.
    • Add dataset ops to the graph (or create kernels in Eager execution) during the python Dataset object creation instead doing it during Iterator creation time.
    • Add support for TensorArrays to tf.data Dataset.
    • Switching tf.data functions to use defun, providing an escape hatch to continue using the legacy Defun.
  • Toolchains
    • CUDNN_INSTALL_PATH, TENSORRT_INSTALL_PATH, NCCL_INSTALL_PATH, NCCL_HDR_PATH are deprecated. Use TF_CUDA_PATHS instead which supports a comma-separated list of base paths that are searched to find CUDA libraries and headers.
    • TF code now resides in tensorflow_core and tensorflow is just a virtual pip package. No code changes are needed for projects using TensorFlow, the change is transparent
  • XLA
    • XLA HLO graphs can be inspected with interactive_graphviz tool now.
  • Estimator
    • Use tf.compat.v1.estimator.inputs instead of tf.estimator.inputs
    • Replace contrib references with tf.estimator.experimental.* for apis in early_stopping.py

Thanks to our Contributors

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

1e100, 4d55397500, a6802739, abenmao, Adam Weiss, Ag Ramesh, Alan Du, Albin Joy, Alex, Aman Patel, Amit, Amit Kumar Jaiswal, Amit Srivastava, Andreas Eberle, Andy Craze, Anthony Platanios, Armen Poghosov, armenpoghosov, arp95, Arpit Shah, Ashwin Ramaswami, Aurelien Geron, AuréLien Geron, aweers, awesomealex1, Ayush Agrawal, Ben Barsdell, Bharat Raghunathan, Bhavani Subramanian, blairhan, BléNesi Attila, Brandon Carter, candy.dc, Chao Liu, chenchc, chie8842, Christian Hansen, Christian Sigg, Clayne Robison, crafet, csukuangfj, ctiijima, Dan Jarvis, Dan Lazewatsky, Daniel Ingram, Daniel Salvadori, Dave Airlie, David Norman, Dayananda V, Dayananda-V, delock, Denis Khalikov, Deven Desai, Dheeraj Rajaram Reddy, dmitrievanthony, Donovan Ong, Drew Szurko, Duncan Riach, Dustin Neighly, Edward Forgacs, EFanZh, Fei Hu, Felix Lemke, Filip Matzner, fo40225, frreiss, Gautam, gehring, Geoffrey Irving, Grzegorz George Pawelczak, Grzegorz Pawelczak, Gyoung-Yoon Ryoo, HanGuo97, Hanton Yang, Hari Shankar, hehongliang, Heungsub Lee, Hoeseong Kim, I-Hong Jhuo, Ilango R, Innovimax, Irene Dea, Jacky Ko, Jakub Lipinski, Jason Zaman, jcf94, Jeffrey Poznanovic, Jens Elofsson, Jeroen BéDorf, Jia Qingtong, Jiankang, Joe Q, Joe Quadrino, Joeran Beel, Jonas Rauber, Jonathan, Jonathan Kyl, Joppe Geluykens, Joseph Friedman, jtressle, jwu, K Yasaswi Sri Chandra Gandhi, K. Hodges, Kaixi Hou, Karl Lessard, Karl Weinmeister, Karthik Muthuraman, Kashif Rasul, KDR, Keno Fischer, Kevin Mader, kjopek, Koan-Sin Tan, kouml, ktaebum, Lakshay Tokas, Laurent Le Brun, Letian Kang, Li, Guizi, Loo Rong Jie, Lucas Hendren, Lukas Geiger, Luke Han, luxupu, Ma, Guokai, Mahmoud Abuzaina, Mandar Deshpande, manhyuk, Marco Gaido, Marek Drozdowski, Mark Collier, Mark Ryan, mars20, Mateusz Chudyk, Matt Conley, MattConley, mbhuiyan, mdfaijul, Melissa Grueter, Michael KäUfl, MickaëL Schoentgen, Miguel Morin, Mihail Salnikov, Mike Arpaia, Mike Holcomb, monklof, Moses Marin, Mshr-H, nammbash, Natalia Gimelshein, Nayana-Ibm, neargye, Neeraj Pradhan, Nehal J Wani, Nick, Niels Ole Salscheider, Niranjan Hasabnis, nlewycky, Nuka-137, Nutti, olicht, P Sudeepam, Palmer Lao, Pan Daoxin, Pariksheet Pinjari, Pavel Samolysov, PENGWA, Pooya Davoodi, R S Nikhil Krishna, Rohit Gupta, Roman Soldatow, rthadur, Ruizhe, Ryan Jiang, Samantha Andow, Sami Kama, Sana-Damani, Saurabh Deoras, sdamani, seanshpark, Sebastien Iooss, Serv-Inc, Shahzad Lone, Shashank Gupta, Shashi, shashvat, shashvatshahi1998, Siju, Siju Samuel, Snease-Abq, Spencer Schaber, sremedios, srinivasan.narayanamoorthy, Steve Lang, Steve Nesae, Sumesh Udayakumaran, Supriya Rao, Taylor Jakobson, Taylor Thornton, Ted Chang, ThisIsPIRI, Thomas Deegan, Thomas Hagebols, tianyapiaozi, Tim Zaman, tomguluson92, Tongxuan Liu, TungJerry, v1incent, Vagif, vcarpani, Vikram Tiwari, Vishwak Srinivasan, Vitor-Alves, wangsiyu, wateryzephyr, WeberXie, WeijieSun, Wen-Heng (Jack) Chung, wenxizhu, Will Battel, William D. Irons, wyzhao, Xin, Yasuhiro Matsumoto, ymodak, Yong Tang, Younes Khoudli, Yuan Lin, Yves-Noel Weweler, Zantares, zjjott, 卜居, 王振华 (Wang Zhenhua), 黄鑫

Release 1.12.3

Bug Fixes and Other Changes

  • Updates png_archive dependency to 1.6.37 to not be affected by CVE-2019-7317, CVE-2018-13785, and CVE-2018-14048.
  • Updates sqlite dependency to 3.28.0 to not be affected by CVE-2018-20506, CVE-2018-20346, and CVE-2018-20505.

Release 1.12.2

Bug Fixes and Other Changes

  • Fixes a potential security vulnerability where carefully crafted GIF images can produce a null pointer dereference during decoding.

Release 1.13.0

Major Features and Improvements

  • TensorFlow Lite has moved from contrib to core. This means that Python modules are under tf.lite and source code is now under tensorflow/lite rather than tensorflow/contrib/lite.
  • TensorFlow GPU binaries are now built against CUDA 10 and TensorRT 5.0.
  • Support for Python3.7 on all operating systems.
  • Moved NCCL to core.

Behavioral changes

  • Disallow conversion of python floating types to uint32/64 (matching behavior of other integer types) in tf.constant.
  • Make the gain argument of convolutional orthogonal initializers (convolutional_delta_orthogonal, convolutional_orthogonal_1D, convolutional_orthogonal_2D, convolutional_orthogonal_3D) have consistent behavior with the tf.initializers.orthogonal initializer, i.e. scale the output l2-norm by gain and NOT by sqrt(gain). (Note that these functions are currently in tf.contrib which is not guaranteed backward compatible).

Bug Fixes and Other Changes

  • Documentation
    • Update the doc with the details about the rounding mode used in quantize_and_dequantize_v2.
    • Clarify that tensorflow::port::InitMain() should be called before using the TensorFlow library. Programs failing to do this are not portable to all platforms.
  • Deprecations and Symbol renames.
    • Removing deprecations for the following endpoints: tf.acos, tf.acosh, tf.add, tf.as_string, tf.asin, tf.asinh, tf.atan, tf.atan2, tf.atanh, tf.cos, tf.cosh, tf.equal, tf.exp, tf.floor, tf.greater, tf.greater_equal, tf.less, tf.less_equal, tf.log, tf.logp1, tf.logical_and, tf.logical_not, tf.logical_or, tf.maximum, tf.minimum, tf.not_equal, tf.sin, tf.sinh, tf.tan
    • Deprecate tf.data.Dataset.shard.
    • Deprecate saved_model.loader.load which is replaced by saved_model.load and saved_model.main_op, which will be replaced by saved_model.main_op in V2.
    • Deprecate tf.QUANTIZED_DTYPES. The official new symbol is tf.dtypes.QUANTIZED_DTYPES.
    • Update sklearn imports for deprecated packages.
    • Deprecate Variable.count_up_to and tf.count_up_to in favor of Dataset.range.
    • Export confusion_matrix op as tf.math.confusion_matrix instead of tf.train.confusion_matrix.
    • Add tf.dtypes. endpoint for every constant in dtypes.py. Moving endpoints in versions.py to corresponding endpoints in tf.sysconfig. and tf.version.. Moving all constants under tf.saved_model submodules to tf.saved_model module. New endpoints are added in V1 and V2 but existing endpoint removals are only applied in V2.
    • Deprecates behavior where device assignment overrides collocation constraints inside a collocation context manager.
  • Keras & Python API
    • Add to Keras functionality analogous to tf.register_tensor_conversion_function.
    • Subclassed Keras models can now be saved through tf.contrib.saved_model.save_keras_model.
    • LinearOperator.matmul now returns a new LinearOperator.
  • New ops and improved op functionality
    • Add a Nearest Neighbor Resize op.
    • Add an ignore_unknown argument to parse_values which suppresses ValueError for unknown hyperparameter types. Such * Add tf.linalg.matvec convenience function.
    • tf.einsum()raises ValueError for unsupported equations like "ii->".
    • Add DCT-I and IDCT-I in tf.signal.dct and tf.signal.idct.
    • Add LU decomposition op.
    • Add quantile loss to gradient boosted trees in estimator.
    • Add round_mode to QuantizeAndDequantizeV2 op to select rounding algorithm.
    • Add unicode_encode, unicode_decode, unicode_decode_with_offsets, unicode_split, unicode_split_with_offset, and unicode_transcode ops. Amongst other things, this Op adds the ability to encode, decode, and transcode a variety of input text encoding formats into the main Unicode encodings (UTF-8, UTF-16-BE, UTF-32-BE)
    • Add "unit" attribute to the substr op, which allows obtaining the substring of a string containing unicode characters.
    • Broadcasting support for Ragged Tensors.
    • SpaceToDepth supports uint8 data type.
    • Support multi-label quantile regression in estimator.
    • We now use "div" as the default partition_strategy in tf.nn.safe_embedding_lookup_sparse, tf.nn.sampled_softmax and tf.nn.nce_loss. hyperparameter are ignored.
  • Performance
    • Improve performance of GPU cumsum/cumprod by up to 300x.
    • Added support for weight decay in most TPU embedding optimizers, including AdamW and MomentumW.
  • TensorFlow 2.0 Development
    • Add a command line tool to convert to TF2.0, tf_upgrade_v2
    • Merge tf.spectral into tf.signal for TensorFlow 2.0.
    • Change the default recurrent activation function for LSTM from 'hard_sigmoid' to 'sigmoid' in 2.0. Historically recurrent activation is 'hard_sigmoid' since it is fast than 'sigmoid'. With new unified backend between CPU and GPU mode, since the CuDNN kernel is using sigmoid, we change the default for CPU mode to sigmoid as well. With that, the default LSTM will be compatible with both CPU and GPU kernel. This will enable user with GPU to use CuDNN kernel by default and get a 10x performance boost in training. Note that this is checkpoint breaking change. If user want to use their 1.x pre-trained checkpoint, please construct the layer with LSTM(recurrent_activation='hard_sigmoid') to fallback to 1.x behavior.
  • TensorFlow Lite
    • Move from tensorflow/contrib/lite to tensorflow/lite.
    • Add experimental Java API for injecting TensorFlow Lite delegates
    • Add support for strings in TensorFlow Lite Java API.
  • tf.contrib:
    • Add Apache Ignite Filesystem plugin to support accessing Apache IGFS.
    • Dropout now takes rate argument, keep_prob is deprecated.
    • Estimator occurrences references tf.contrib.estimator were changed to tf.estimator:
    • tf.contrib.estimator.BaselineEstimator with tf.estimator.BaselineEstimator
    • tf.contrib.estimator.DNNLinearCombinedEstimator with tf.estimator.DNNLinearCombinedEstimator
    • tf.contrib.estimator.DNNEstimator with tf.estimator.DNNEstimator
    • tf.contrib.estimator.LinearEstimator with tf.estimator.LinearEstimator
    • tf.contrib.estimator.InMemoryEvaluatorHook and tf.estimator.experimental.InMemoryEvaluatorHook`.
    • tf.contrib.estimator.make_stop_at_checkpoint_step_hook with tf.estimator.experimental.make_stop_at_checkpoint_step_hook.
    • Expose `tf.distribute.Strategy as the new name for tf.contrib.distribute.DistributionStrategy.
    • Migrate linear optimizer from contrib to core.
    • Move tf.contrib.signal to tf.signal (preserving aliases in tf.contrib.signal).
    • Users of tf.contrib.estimator.export_all_saved_models and related should switch to tf.estimator.Estimator.experimental_export_all_saved_models.
  • tf.data:
    • Add tf.data.experimental.StatsOptions(), to configure options to collect statistics from tf.data.Dataset pipeline using StatsAggregator. Add nested option, experimental_stats (which takes a tf.data.experimen tal.StatsOptions object), to tf.data.Options. Deprecates tf.data.experimental.set_stats_agregator.
    • Performance optimizations:
    • Add tf.data.experimental.OptimizationOptions(), to configure options to enable tf.data performance optimizations. Add nested option, experimental_optimization (which takes a tf.data.experimental.OptimizationOptions object), to tf.data.Options. Remove performance optimization options from tf.data.Options, and add them under tf.data.experimental.OptimizationOptions instead.
    • Enable map_and_batch_fusion and noop_elimination optimizations by default. They can be disabled by configuring tf.data.experimental.OptimizationOptions to set map_and_batch = False or noop_elimination = False respectively. To disable all default optimizations, set apply_default_optimizations = False.
    • Support parallel map in map_and_filter_fusion.
    • Disable static optimizations for input pipelines that use non-resource tf.Variables.
    • Add NUMA-aware MapAndBatch dataset.
    • Deprecate tf.data.Dataset.make_one_shot_iterator() in V1, removed it from V2, and added tf.compat.v1.data.make_one_shot_iterator()`.
    • Deprecate tf.data.Dataset.make_initializable_iterator() in V1, removed it from V2, and added tf.compat.v1.data.make_initializable_iterator().
    • Enable nested dataset support in core tf.data transformations.
    • For tf.data.Dataset implementers: Added tf.data.Dataset._element_structured property to replace Dataset.output_{types,shapes,classes}.
    • Make num_parallel_calls of tf.data.Dataset.interleave and tf.data.Dataset.map work in Eager mode.
  • Toolchains
    • Fixed OpenSSL compatibility by avoiding EVP_MD_CTX_destroy.
    • Added bounds checking to printing deprecation warnings.
    • Upgraded CUDA dependency to 10.0
    • To build with Android NDK r14b, add "#include <linux/compiler.h>" to android-ndk-r14b/platforms/android-14/arch-*/usr/include/linux/futex.h
    • Removed :android_tensorflow_lib_selective_registration* targets, use :android_tensorflow_lib_lite* targets instead.
  • XLA
    • Move RoundToEven function to xla/client/lib/math.h.
    • A new environment variable TF_XLA_DEBUG_OPTIONS_PASSTHROUGH set to "1" or "true" allows the debug options passed within an XRTCompile op to be passed directly to the XLA compilation backend. If such variable is not set (service side), only a restricted set will be passed through.
    • Allow the XRTCompile op to return the ProgramShape resulted form the XLA compilation as a second return argument.
    • XLA HLO graphs can now be rendered as SVG/HTML.
  • Estimator
    • Replace all occurrences of tf.contrib.estimator.BaselineEstimator with tf.estimator.BaselineEstimator
    • Replace all occurrences of tf.contrib.estimator.DNNLinearCombinedEstimator with tf.estimator.DNNLinearCombinedEstimator
    • Replace all occurrences of tf.contrib.estimator.DNNEstimator with tf.estimator.DNNEstimator
    • Replace all occurrences of tf.contrib.estimator.LinearEstimator with tf.estimator.LinearEstimator
    • Users of tf.contrib.estimator.export_all_saved_models and related should switch to tf.estimator.Estimator.experimental_export_all_saved_models.
    • Update regression_head to the new Head API for Canned Estimator V2.
    • Switch multi_class_head to Head API for Canned Estimator V2.
    • Replace all occurrences of tf.contrib.estimator.InMemoryEvaluatorHook and tf.contrib.estimator.make_stop_at_checkpoint_step_hook with tf.estimator.experimental.InMemoryEvaluatorHook and tf.estimator.experimental.make_stop_at_checkpoint_step_hook
    • Migrate linear optimizer from contrib to core.

Thanks to our Contributors

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

Abhinav Upadhyay, Ag Ramesh, akikaaa, Alexis Louis, Anders Huss, Andreas Madsen, Andrew Banchich, Andy Craze, Anton Dmitriev, Artem Malykh, Avijit-Nervana, Balint Cristian, Benjamin Tan Wei Hao, Bhavani Subramanian, Brendan Finan, Brian Nemsick, Bryan Cutler, By Shen, Cao Zongyan, Castiel, Chris Antaki, Christian Goll, Cibifang, Clayne Robison, Codrut Grosu, Cong Xu, Dalmo Cirne, Daniel Hunter, Dougal J. Sutherland, Edvard Fagerholm, EFanZh, Erik Smistad, Evgeniy Polyakov, Feiyang Chen, franklin5, Fred Reiss, Gautam, gehring, Geoffrey Irving, George Sterpu, Gitea, Grzegorz George Pawelczak, Guozhong Zhuang, himkt, Hoeseong Kim, Huan Li (李卓桓), HuiyangFei, hyunyoung, Isaac Burbank, jackonan, Jacky Ko, Jason Furmanek, Jason Zaman, Javier Luraschi, Jiang,Zhoulong, joaak, John Lin, Jonathan Wyatt Hoech, josephyearsley, Josh Gordon, Julian Niedermeier, Karl Lessard, Keno Fischer, lanhin, Leon Graser, leondgarse, Li, Guizi, Li, Yiqiang, lxl910915, Mahmoud Abuzaina, manhyuk, Marcela Morales Quispe, margaretmz, Matt Conley, Max Pumperla, mbhuiyan, mdfaijul, Meng, Peng, Michael, Michael Gielda, mrTsjolder, Muhammad Wildan, neargye, Nehal J Wani, NEWPLAN, Niranjan Hasabnis, Nutti, olicht, Pan Daoxin, Pedro Monreal, Peng Yu, pillarpond, Pooya Davoodi, qiezi, Rholais Lii, Richard Yu, Rin Arakaki, Roger Iyengar, sahilbadyal, Sami Kama, Sandip Giri, Scott Leishman, Serge Panev, Seunghoon Park, Shafi Dayatar, shengfuintel, Shimin Guo, Siju, silent567, Stefan Dyulgerov, steven, Tao Wei, Thor Johnsen, Tingbo Lu, tomguluson92, Tongxuan Liu, Trevor Morris, Ubuntu, Vadim Borisov, vanderliang, wangsiyu, Wen Yun, Wen-Heng (Jack) Chung, wenxizhu, William D. Irons, Xiaoming (Jason) Cui, Yan Facai (颜发才), Yanbo Liang, Yaniv Blumenfeld, Yash Gaurkar, Yicheng Fan, Yong Tang, Yongjoon Lee, Yuan (Terry) Tang, Yuxin Wu, zldrobit

Release 1.12.0

Major Features and Improvements

  • Keras models can now be directly exported to the SavedModel format(tf.contrib.saved_model.save_keras_model()) and used with Tensorflow Serving.
  • Keras models now support evaluating with a tf.data.Dataset.
  • TensorFlow binaries are built with XLA support linked in by default.
  • Ignite Dataset added to contrib/ignite that allows to work with Apache Ignite.

Bug Fixes and Other Changes

  • tf.data:
    • tf.data users can now represent, get, and set options of TensorFlow input pipelines using tf.data.Options(), tf.data.Dataset.options(), and tf.data.Dataset.with_options() respectively.
    • New tf.data.Dataset.reduce() API allows users to reduce a finite dataset to a single element using a user-provided reduce function.
    • New tf.data.Dataset.window() API allows users to create finite windows of input dataset; when combined with the tf.data.Dataset.reduce() API, this allows users to implement customized batching.
    • All C++ code moves to the tensorflow::data namespace.
    • Add support for num_parallel_calls to tf.data.Dataset.interleave.
  • tf.contrib:
    • Remove tf.contrib.linalg. tf.linalg should be used instead.
    • Replace any calls to tf.contrib.get_signature_def_by_key(metagraph_def, signature_def_key) with meta_graph_def.signature_def[signature_def_key]. Catching a ValueError exception thrown by tf.contrib.get_signature_def_by_key should be replaced by catching a KeyError exception.
  • tf.contrib.data
    • Deprecate, and replace by tf.data.experimental.
  • Other:
    • Instead of jemalloc, revert back to using system malloc since it simplifies build and has comparable performance.
    • Remove integer types from tf.nn.softplus and tf.nn.softsign OpDefs. This is a bugfix; these ops were never meant to support integers.
    • Allow subslicing Tensors with a single dimension.
    • Add option to calculate string length in Unicode characters.
    • Add functionality to SubSlice a tensor.
    • Add searchsorted (ie lower/upper_bound) op.
    • Add model explainability to Boosted Trees.
    • Support negative positions for tf.substr.
    • There was previously a bug in the bijector_impl where the _reduce_jacobian_det_over_event does not handle scalar ILDJ implementations properly.
    • In tf eager execution, allow re-entering a GradientTape context.
    • Add tf_api_version flag. If --define=tf_api_version=2 flag is passed in, then bazel will build TensorFlow API version 2.0. Note that TensorFlow 2.0 is under active development and has no guarantees at this point.
    • Add additional compression options to TfRecordWriter.
    • Performance improvements for regex full match operations.
    • Replace tf.GraphKeys.VARIABLES with tf.GraphKeys.GLOBAL_VARIABLES.
    • Remove unused dynamic learning rate support.

Thanks to our Contributors

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

(David) Siu-Kei Muk, Ag Ramesh, Anton Dmitriev, Artem Sobolev, Avijit-Nervana, Bairen Yi, Bruno Goncalves, By Shen, candy.dc, Cheng Chen, Clayne Robison, coder3101, Dao Zhang, Elms, Fei Hu, feiquan, Geoffrey Irving, Guozhong Zhuang, hellcom, Hoeseong Kim, imsheridan, Jason Furmanek, Jason Zaman, Jenny Sahng, jiefangxuanyan, Johannes Bannhofer, Jonathan Homer, Koan-Sin Tan, kouml, Loo Rong Jie, Lukas Geiger, manipopopo, Ming Li, Moritz KröGer, Naurril, Niranjan Hasabnis, Pan Daoxin, Peng Yu, pengwa, rasmi, Roger Xin, Roland Fernandez, Sami Kama, Samuel Matzek, Sangjung Woo, Sergei Lebedev, Sergii Khomenko, shaohua, Shaohua Zhang, Shujian2015, Sunitha Kambhampati, tomguluson92, ViníCius Camargo, wangsiyu, weidankong, Wen-Heng (Jack) Chung, William D. Irons, Xin Jin, Yan Facai (颜发才), Yanbo Liang, Yash Katariya, Yong Tang, 在原佐为

Release 1.11.0

Major Features and Improvements

  • Nvidia GPU:
  • Google Cloud TPU:
    • Experimental tf.data integration for Keras on Google Cloud TPUs.
    • Experimental / preview support for eager execution on Google Cloud TPUs.
  • DistributionStrategy:
    • Add multi-GPU DistributionStrategy support in tf.keras. Users can now use fit, evaluate and predict to distribute their model on multiple GPUs.
    • Add multi-worker DistributionStrategy and standalone client support in Estimator. See README for more details.
  • Add C, C++, and Python functions for querying kernels.

Breaking Changes

  • Keras:
    • The default values for tf.keras RandomUniform, RandomNormal, and TruncatedNormal initializers have been changed to match those in external Keras.
    • Breaking change: model.get_config() on a Sequential model now returns a config dictionary (consistent with other Model instances) instead of a list of configs for the underlying layers.

Bug Fixes and Other Changes

  • C++:
    • Changed the signature of SessionFactory::NewSession so that it can return a meaningful error message on failure.
  • tf.data:
    • Remove num_parallel_parser_calls argument from tf.contrib.data.make_csv_dataset(). [tf.data] Remove num_parallel_parser_calls argument from tf.contrib.data.make_csv_dataset().
    • tf.data.Dataset.list_files() raises an exception at initialization time if the argument matches no files.
    • Renamed BigTable class to BigtableTable for clarity
    • Document use of the Cloud Bigtable API
    • Add tf.contrib.data.reduce_dataset which can be used to reduce a dataset to a single element.
    • Generalization of tf.contrib.data.sliding_window_batch.
  • INC:
    • Runtime improvements to triangular solve.
  • tf.contrib:
    • Add an implementation argument to tf.keras.layers.LocallyConnected2D and tf.keras.layers.LocallyConnected1D. The new mode (implementation=2) performs forward pass as a single dense matrix multiplication, allowing dramatic speedups in certain scenarios (but worse performance in others - see docstring). The option also allows to use padding=same.
    • Add documentation clarifying the differences between tf.fill and tf.constant.
    • Add experimental IndexedDatasets.
    • Add selective registration target using the lite proto runtime.
    • Add simple Tensor and DataType classes to TensorFlow Lite Java
    • Add support for bitcasting to/from uint32 and uint64.
    • Added a subclass of Estimator that can be created from a SavedModel (SavedModelEstimator).
    • Adds leaf index modes as an argument.
    • Allow a different output shape from the input in tf.contrib.image.transform.
    • Change the state_size order of the StackedRNNCell to be natural order. To keep the existing behavior, user can add reverse_state_order=True when constructing the StackedRNNCells.
    • Deprecate self.test_session() in favor of self.session() or self.cached_session().
    • Directly import tensor.proto.h (the transitive import will be removed from tensor.h soon).
    • Estimator.train() now supports tf.contrib.summary.* summaries out of the box; each call to .train() will now create a separate tfevents file rather than re-using a shared one.
    • Fix FTRL L2-shrinkage behavior: the gradient from the L2 shrinkage term should not end up in the accumulator.
    • Fix toco compilation/execution on Windows.
    • GoogleZoneProvider class added to detect which Google Cloud Engine zone tensorflow is running in.
    • It is now safe to call any of the C API's TF_Delete* functions on nullptr.
    • Log some errors on Android to logcat.
    • Match FakeQuant numerics in TFLite to improve accuracy of TFLite quantized inference models.
    • Optional bucket location check for the GCS Filesystem.
    • Performance enhancements for StringSplitOp & StringSplitV2Op.
    • Performance improvements for regex replace operations.
    • TFRecordWriter now raises an error if .write() fails.
    • TPU: More helpful error messages in TPUClusterResolvers.
    • The legacy_init_op argument to SavedModelBuilder methods for adding MetaGraphs has been deprecated. Please use the equivalent main_op argument instead. As part of this, we now explicitly check for a single main_op or legacy_init_op at the time of SavedModel building, whereas the check on main_op was previously only done at load time.
    • The protocol used for Estimator training is now configurable in RunConfig.
    • Triangular solve performance improvements.
    • Unify RNN cell interface between TF and Keras. Add new get_initial_state() to Keras and TF RNN cell, which will use to replace the existing zero_state() method.
    • Update initialization of variables in Keras.
    • Updates to "constrained_optimization" in tensorflow/contrib.
    • boosted trees: adding pruning mode.
    • tf.train.Checkpoint does not delete old checkpoints by default.
    • tfdbg: Limit the total disk space occupied by dumped tensor data to 100 GBytes. Add environment variable TFDBG_DISK_BYTES_LIMIT to allow adjustment of this upper limit.

Thanks to our Contributors

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

Aapeli, adoda, Ag Ramesh, Amogh Mannekote, Andrew Gibiansky, Andy Craze, Anirudh Koul, Aurelien Geron, Avijit, Avijit-Nervana, Ben, Benjamin H. Myara, bhack, Brett Koonce, Cao Zongyan, cbockman, cheerss, Chikanaga Tomoyuki, Clayne Robison, cosine0, Cui Wei, Dan J, David, David Norman, Dmitry Klimenkov, Eliel Hojman, Florian Courtial, fo40225, formath, Geoffrey Irving, gracehoney, Grzegorz Pawelczak, Guoliang Hua, Guozhong Zhuang, Herman Zvonimir DošIlović, HuiyangFei, Jacker, Jan HüNnemeyer, Jason Taylor, Jason Zaman, Jesse, Jiang,Zhoulong, Jiawei Zhang, Jie, Joe Yearsley, Johannes Schmitz, Jon Perl, Jon Triebenbach, Jonathan, Jonathan Hseu, Jongmin Park, Justin Shenk, karl@kubx.ca, Kate Hodesdon, Kb Sriram, Keishi Hattori, Kenneth Blomqvist, Koan-Sin Tan, Li Liangbin, Li, Yiqiang, Loo Rong Jie, Madiyar, Mahmoud Abuzaina, Mark Ryan, Matt Dodge, mbhuiyan, melvinljy96, Miguel Mota, Nafis Sadat, Nathan Luehr, naurril, Nehal J Wani, Niall Moran, Niranjan Hasabnis, Nishidha Panpaliya, npow, olicht, Pei Zhang, Peng Wang (Simpeng), Peng Yu, Philipp Jund, Pradeep Banavara, Pratik Kalshetti, qwertWZ, Rakesh Chada, Randy West, Ray Kim, Rholais Lii, Robin Richtsfeld, Rodrigo Silveira, Ruizhi, Santosh Kumar, Seb Bro, Sergei Lebedev, sfujiwara, Shaba Abhiram, Shashi, SneakyFish5, Soila Kavulya, Stefan Dyulgerov, Steven Winston, Sunitha Kambhampati, Surry Shome, Taehoon Lee, Thor Johnsen, Tristan Rice, TShapinsky, tucan, tucan9389, Vicente Reyes, Vilmar-Hillow, Vitaly Lavrukhin, wangershi, weidan.kong, weidankong, Wen-Heng (Jack) Chung, William D. Irons, Wim Glenn, XFeiF, Yan Facai (颜发才), Yanbo Liang, Yong Tang, Yoshihiro Yamazaki, Yuan (Terry) Tang, Yuan, Man, zhaoyongke, ÁRon Ricardo Perez-Lopez, 张天启, 张晓飞

Release 1.10.1

Bug Fixes and Other Changes

  • tf.keras:
    • Fixing keras on Cloud TPUs. No new binaries will be built for Windows.

Release 1.10.0

Major Features And Improvements

  • The tf.lite runtime now supports complex64.
  • Initial Google Cloud Bigtable integration for tf.data.
  • Improved local run behavior in tf.estimator.train_and_evaluate which does not reload checkpoints for evaluation.
  • RunConfig now sets device_filters to restrict how workers and PS can communicate. This can speed up training and ensure clean shutdowns in some situations. But if you have jobs that require communication between workers, you will have to set custom session_options in your RunConfig.
  • Moved Distributions and Bijectors from tf.contrib.distributions to Tensorflow Probability (TFP). tf.contrib.distributions is now deprecated and will be removed by the end of 2018.
  • Adding new endpoints for existing tensorflow symbols. These endpoints are going to be the preferred endpoints going forward and may replace some of the existing endpoints in the future. See below for the complete list. New symbols have been added to the following modules: tf.debugging, tf.dtypes, tf.image, tf.io, tf.linalg, tf.manip, tf.math, tf.quantization, tf.strings

Breaking Changes

  • Prebuilt binaries are now (as of TensorFlow 1.10) built against NCCL 2.2 and no longer include NCCL in the binary install. TensorFlow usage with multiple GPUs and NCCL requires upgrade to NCCL 2.2. See updated install guides: TensorFlow GPU support and Build TensorFlow from source.
  • Starting from TensorFlow 1.11, Windows builds will use Bazel. Therefore, we will drop official support for cmake.

Bug Fixes and Other Changes

  • tf.data:
    • tf.contrib.data.group_by_reducer() is now available via the public API.
    • tf.contrib.data.choose_from_datasets() is now available via the public API.
    • Adding drop_remainder argument to tf.data.Dataset.batch() and tf.data.Dataset.padded_batch(), deprecating tf.contrib.data.batch_and_drop_remainder() and tf.contrib.data.padded_batch_and_drop_remainder().
  • tf.estimator:
    • Estimators now use custom savers included in EstimatorSpec scaffolds for saving SavedModels during export.
    • EstimatorSpec will now add a default prediction output for export if no export_output is provided, eliminating the need to explicitly include a PredictOutput object in the model_fn for simple use-cases.
    • Support sparse_combiner in canned Linear Estimators.
    • Added batch normalization to DNNClassifier, DNNRegressor, and DNNEstimator.
    • Adding ranking support for boosted trees.
    • Adding center bias option for boosted trees.
  • Add synchronization and aggregation args to get_variable(). These args will be used for distributed variables.
  • Add synchronization and aggregation args to the layer add_weight() API. These args will be used for distributed variables.
  • tf.losses.* do not add to the global collection when executing eagerly (to avoid leaking memory).
  • Support different summary and checkpoint directories in tf.train.MonitoredTrainingSession().
  • Added IndRNN, IndyGRU, and IndyLSTM cells to tf.contrib.rnn.
  • Add safe static factory functions for SparseTensor and convert all CHECKs to DCHECKs. Using the constructor directly is unsafe and deprecated.
  • Make the Bigtable client connection pool configurable & increase the default # of connections for performance.
  • Added derivative of tf.random_gamma with respect to the alpha parameter.
  • Added derivative of tf.igamma(a, x) and tf.igammac(a, x) with respect to a.
  • Modified Bessel functions of order zero and one.
  • Add FillTriangular Bijector to create triangular matrices.
  • Added support for Type III DCT, and tf.spectral.idct(type=2|3).
  • Correctly handle CuDNN RNN weight loaded when nest in TimeDistributed.
  • Adding per-element weight support for WALSComputePartialLhsAndRhsOp.
  • ZerosLike and OnesLike ops treated as constants by Graph Transform Tool.
  • Gamma distribution and the derived distributions (Beta, Dirichlet, Student's t, inverse Gamma) now fully reparameterized.
  • Java: Experimental wrapper classes to make graph generation easier. Thanks @karllessard and @kbsriram
  • Build & link in secure gRPC components (switch from the insecure grpc dependency to secure grpc dependency).
  • Adding new endpoints for existing tensorflow symbols. These endpoints are going to be the preferred endpoints going forward and may replace some of the existing endpoints in the future. List of new endpoints:
    • New endpoints in tf.image namespace: tf.image.extract_image_patches
    • New endpoints in tf.debugging namespace: tf.debugging.check_numerics, tf.debugging.is_finite, tf.debugging.is_inf, tf.debugging.is_nan.
    • New endpoints in tf.dtypes namespace: tf.dtypes.as_string.
    • New endpoints in tf.io namespace: tf.io.decode_base64, tf.io.decode_compressed, tf.io.decode_json_example, tf.io.decode_raw, tf.io.encode_base64, tf.io.matching_files, tf.io.parse_tensor, tf.io.read_file, tf.io.write_file`.
    • New endpoints in tf.linalg namespace: tf.linalg.cross, tf.linalg.tensor_diag (corresponds to tf.diag), tf.linalg.tensor_diag_part (corresponds to tf.diag_part).
    • New endpoints in tf.manip namespace: tf.manip.batch_to_space_nd, tf.manip.gather_nd, tf.manip.reshape, tf.manip.reverse, tf.manip.scatter_nd, tf.manip.space_to_batch_nd, tf.manip.tile
    • New endpoints in tf.math namespace: tf.math.acos, tf.math.acosh, tf.math.add, tf.math.asin, tf.math.asinh, tf.math.atan, tf.math.atan2, tf.math.atanh, tf.math.betainc, tf.math.ceil, tf.math.cos, tf.math.cosh, tf.math.digamma, tf.math.equal, tf.math.erfc, tf.math.exp, tf.math.expm1, tf.math.floor, tf.math.greater, tf.math.greater_equal, tf.math.igamma, tf.math.igammac, tf.math.invert_permutation, tf.math.less, tf.math.less_equal, tf.math.lgamma, tf.math.log, tf.math.log1p, tf.math.logical_and, tf.math.logical_not, tf.math.logical_or, tf.math.maximum, tf.math.minimum, tf.math.not_equal, tf.math.polygamma, tf.math.reciprocal, tf.math.rint, tf.math.rsqrt, tf.math.segment_max, tf.math.segment_mean, tf.math.segment_min, tf.math.segment_prod, tf.math.segment_sum, tf.math.sin, tf.math.sinh, tf.math.softplus, tf.math.softsign, tf.math.squared_difference, tf.math.tan, tf.math.unsorted_segment_max, tf.math.unsorted_segment_min, tf.math.unsorted_segment_prod, tf.math.unsorted_segment_sum, tf.math.zeta.
    • New endpoints in tf.quantization namespace: tf.quantization.dequantize, tf.quantization.fake_quant_with_min_max_args, tf.quantization.fake_quant_with_min_max_args_gradient, tf.quantization.fake_quant_with_min_max_vars, tf.quantization.fake_quant_with_min_max_vars_gradient, tf.quantization.fake_quant_with_min_max_vars_per_channel, tf.quantization.fake_quant_with_min_max_vars_per_channel_gradient.
    • New endpoints in tf.strings namespace: tf.strings.join (corresponds to tf.string_join), tf.strings.regex_replace, tf.strings.to_number (corresponds to tf.string_to_number), tf.strings.strip (corresponds to tf.string_strip), tf.strings.substr, tf.strings.to_hash_bucket (corresponds to tf.string_to_hash_bucket), tf.strings.to_hash_bucket_fast (corresponds to tf.string_to_hash_bucket_fast), tf.strings.to_hash_bucket_strong (corresponds to tf.string_to_hash_bucket_strong).

Thanks to our Contributors

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

Ag Ramesh, Alex Wiltschko, Alexander Pantyukhin, Amogh Mannekote, An Jiaoyang, Andrei Nigmatulin, Andrew Ginns, BjøRn Moholt, Brett Koonce, Chengzhi Chen, Chinmay Das, Christian Ertler, Christoph Boeddeker, Clayne Robison, Courtial Florian, ctiijima, Dan Douthit, Dan J, Dan Ringwalt, EFanZh, Emanuele Ballarin, eqy, Evgeniy Zheltonozhskiy, Freedom" Koan-Sin Tan, FréDéRic Branchaud-Charron, G K, gracehoney, Guillaume Klein, Guozhong Zhuang, Hsien-Yang Li, hsm207, ImSheridan, Jayaram Bobba, Jiandong Ruan, Jie, Joel Shor, Jonas Rauber, Jongmin Baek, jsawruk, Karan Kaw, Karl Lessard, karl@kubx.ca, Kb Sriram, KinmanLam, leiiwang, Li, Yiqiang, Loo Rong Jie, Mahmoud Abuzaina, Mahmoud Aslan, ManHyuk, Martin Patz, Martin Zeitler, mktozk, Mohammad Ashraf Bhuiyan, mrTsjolder, Naman Bhalla, Nick Felt, Nicolas Lopez, Niranjan Hasabnis, Nishidha Panpaliya, Nitish, nrstott, Nutti, Parag Jain, PeterLee, Philipp Jund, Rach L, Rafal Wojdyla, Roland Zimmermann, Sergei Lebedev, SneakyFish5, Soila Kavulya, Sriram Veturi, Steven Schmatz, Taehoon Lee, Tang, Wenyi, Taras Sereda, Ted Chang, Tim Zaman, Tristan Rice, tucan, vchigrin, Vikram Tiwari, Vincent, WeberXie, William D. Irons, Yan Facai (颜发才), Yong Tang, Yu Yi, Yuxin Wu, Zé ViníCius

Release 1.9.0

Major Features And Improvements

Breaking Changes

  • If you're opening empty variable scopes; replace variable_scope('', ...) by variable_scope(tf.get_variable_scope(), ...).
  • Headers used for building custom ops have been moved from site-packages/external into site-packages/tensorflow/include/external.

Bug Fixes and Other Changes

  • tfe.Network is deprecated. Please inherit from tf.keras.Model.
  • Layered variable names have changed in the following conditions:
    • Using tf.keras.layers with custom variable scopes.
    • Using tf.layers in a subclassed tf.keras.Model class. See here for more details
  • tf.data:
    • Dataset.from_generator() now accepts an args list, in order to create nested generators.
    • Dataset.list_files() now produces deterministic results when shuffle=False or a seed is passed.
    • tf.contrib.data.sample_from_datasets() and tf.contrib.data.choose_from_datasets() make it easier to sample or deterministically choose elements from multiple datasets.
    • tf.contrib.data.make_csv_dataset() now supports line breaks in quoted strings, and two infrequently used arguments removed.
    • (C++) DatasetBase::DebugString() is now const.
    • (C++) DatasetBase::MakeIterator() has been renamed to DatasetBase::MakeIteratorInternal().
    • (C++) IteratorBase::Initialize() method was added to support raising errors during iterator construction.
  • Eager Execution:
    • Added the ability to pause recording operations for gradient computation via tf.GradientTape.stop_recording.
    • Updated documentation, introductory notebooks.
  • tf.keras:
    • Move Keras code out of _impl folder and remove API files.
    • tf.keras.Model.save_weights now saves in TensorFlow format by default.
    • Enable dataset iterators to be passed to tf.keras.Model training/eval methods.
  • TensorFlow Debugger (tfdbg) CLI: fix an issue in which the TensorBoard Debugger Plugin could not handle total source file size exceeding gRPC message size limit (4 MB).
  • tf.contrib:
    • tf.contrib.framework.zero_initializer supports ResourceVariable.
    • Adding "constrained_optimization" to tensorflow/contrib.
  • Other:
    • Add GCS Configuration Ops.
    • Changing signature of MakeIterator to enable propagating error status.
    • KL divergence for two Dirichlet distributions.
    • More consistent GcsFileSystem behavior for certain reads past EOF.
    • Update benchmark for tf.scan to match ranges across eager and graph modes.
    • Fixed bug in tf.reduce_prod gradient for complex dtypes.
    • Allow the use of '.' in variables (e.g. "hparams.parse('a.b=1.0')"), which would previously raise an error. This will correspond to an attribute name with an embedded '.' symbol (e.g. 'a.b'), which can only be accessed indirectly (e.g. through getattr and setattr). To set this up the user will first need to explicitly add the variable to the hparam object (e.g. "hparams.add_hparam(name='a.b', value=0.0)").
    • Benchmark for tf.scan in graph and eager modes.
    • Added complex128 support to FFT, FFT2D, FFT3D, IFFT, IFFT2D, and IFFT3D.
    • Making ids unique in nn.embedding_lookup_sparse. This helps to reduce RPC calls for looking up the embeddings when there are repeated ids in the batch.
    • Support indicator column in boosted trees.
    • Prevent tf.gradients() from backpropagating through integer tensors.
    • LinearOperator[1D,2D,3D]Circulant added to tensorflow.linalg.
    • Conv3D, Conv3DBackpropInput, Conv3DBackpropFilter now supports arbitrary.
    • Added tf.train.Checkpoint for reading/writing object-based checkpoints.
    • Added LinearOperatorKronecker, a dense-free implementation of the Kronecker Product.
    • Allow LinearOperator to broadcast.
    • SavedModelBuilder will now deduplicate asset names that point to files with the same basename and the same contents. Note that this may result in new asset files included in SavedModels in cases where assets with the same name but different contents were previously overwriting each other.

Thanks to our Contributors

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

Abdullah Alrasheed, Achal Shah, Ad-530, ADiegoCAlonso, Aditya Yogi, Ag Ramesh, akindyakov, Andy Kernahan, Anya Petrova, Aurelien Geron, Ben, Ben Barsdell, Bhavani-Subramanian, braincodercn, Brett Koonce, Brian Nemsick, Brian Zier, Bryan Heden, candy.dc, cclauss, Clayne Robison, ctiijima, Dalmo Cirne, David Norman, David T.H. Kao, DosLin, ekelsen, Elson Rodriguez, Erik Smistad, Felix Abecassis, Fergal Cotter, fo40225, foo0x29a, Freedom" Koan-Sin Tan, FréDéRic Branchaud-Charron, gdh1995, Geoffrey Irving, Giuseppe, gracehoney, Guido Zuidhof, Guillaume Klein, Guozhong Zhuang, Haggai, Harald Husum, imsheridan, Ivan Zhang, Jan Zikes, Jayaram Bobba, Jesse Benson, Jesse Gumz, Jiajia Li, Jie, jinghuangintel, Jingwen, jjsjann123, Joe Yearsley, Joel Hestness, Joel Shor, josephyearsley, Junpeng Lao, Karol M. Langner, Kb Sriram, krantideep95, Krish Ravindranath, Letian Feng, Loo Rong Jie, Lukas Geiger, Maciej, Mahmoud Abuzaina, ManHyuk, Mark Ryan, mbhuiyan, Michal Turek, Mostafa Alaa, Myungsung Kwak, Nand Dalal, Nehal J Wani, Neil Tenenholtz, ngc92, Nicholas Nadeau, P.Eng., Avs, Niranjan Hasabnis, P-Hidringer, Paul Van Eck, Peng Yu, Qing Zhao, Qingying Chen, Quanlong, Rajendra Arora, Rholais Lii, rmanyari, Robin Richtsfeld, Russell Klopfer, Sagi, Sam Sendelbach, Sandeep N Gupta, Sandip Giri, Sarah Edkins, Scott Tseng, Sdalbsoo, Sergii Khomenko, Seungwoo Choi (Biggie), Seyed Majid Azimi, Shaoning Zeng, shengfuintel, Siu Kei, Muk, Smit Shilu, soonson, Stefan Schweter, Sukhwan Kim, Sunitha Kambhampati, Taehoon Lee, tamimaddari82, Tang, Wenyi, Ted Chang, u2takey, Utkarsh Upadhyay, Vadim Markovtsev, voegtlel, Wai Hon Law, wangsiyu, Wenhao Hu, wenhao.hu, William D. Irons, Yan Facai (颜发才), Yanbo Liang, Yihong Wang, Yilei (Dolee) Yang, Yong Tang, Yuan (Terry) Tang

Release 1.8.0

Major Features And Improvements

  • Can now pass tf.contrib.distribute.MirroredStrategy() to tf.estimator.RunConfig() to run an Estimator model on multiple GPUs on one machine.
  • Add tf.contrib.data.prefetch_to_device(), which supports prefetching to GPU memory.
  • Added Gradient Boosted Trees as pre-made Estimators: BoostedTreesClassifier, BoostedTreesRegressor.
  • Add 3rd generation pipeline config for Cloud TPUs which improves performance and usability.
  • tf.contrib.bayesflow is moving out to it's own repo.
  • Added tf.contrib.{proto,rpc} to allow generic proto parsing and RPC communication1.

Bug Fixes and Other Changes

  • tf.data:
    • Add tf.contrib.data.prefetch_to_device, which enables prefetching dataset elements to GPU memory.
    • Add tf.contrib.data.AUTOTUNE, which allows the tf.data runtime to automatically tune the prefetch buffer sizes based on your system and environment.
    • Add tf.contrib.data.make_csv_dataset for building datasets of CSV files.
  • Eager Execution:
    • With eager execution Datasets can now be used as standard python iterators (for batch in dataset:). Both Dataset.__iter__() and Dataset.make_one_shot_iterator() can now be used to create iterators when eager execution is enabled.
    • Automatic device placement has been enabled (i.e., use a GPU if available automatically, without requiring an explicit with tf.device(“/gpu:0”)) (Fixes #14133)
    • tf.GradientTape has moved out of contrib.
  • tf.keras:
    • Added the fashion mnist dataset.
    • New data preprocessing functions: image/random_brightness, sequence/TimeseriesGenerator, and text/hashing_trick.
  • Accelerated Linear Algebra (XLA):
    • Select and scatter in reference util and evaluator now use lexicographical order to break ties.
  • TensorFlow Debugger (tfdbg) CLI:
    • During tensor-filter operations, allow exclusion of nodes by regular expressions.
    • Fix spurious background colors in some text terminals.
  • tf.contrib:
    • Add meta-distribution BatchReshape which reshapes batch dimensions.
    • tf.contrib.layers.recompute_grad works for explicit gradient checkpointing on TPU.
    • Add tf.contrib.framework.argsort.
    • Allow DNNBoostedTreeCombinedEstimator to work with core versions of feature columns and losses.
    • Add non-linear image warping ops: tf.contrib.image.sparse_image_warp, tf.contrib.image.dense_image_warp, and tf.contrib.image.interpolate_spline.
    • Fix bug in tf.contrib.opt.MultitaskOptimizerWrapper where types of tensors were mismatched.
  • Other:
    • Low-level graph construction now calls the TensorFlow C API. This change should be invisible to most users, but can be disabled by setting the environment variable TF_C_API_GRAPH_CONSTRUCTION=0 in this release. Future releases will remove the ability to disable this change. Please file a bug if you find yourself using this escape hatch.
    • Add description of shapes and a pointer to tutorial notebook in tf.distributions.Distribution.
    • Update scatter operations:
      • Add tf.scatter_min and tf.scatter_max
      • Extend scatter operations to work with a scalar update parameter.
    • Move cuDNN RNN ops to core for use in TensorFlow codebase only.
    • Add float64 support for Conv2d, Conv2dBackpropInput, and Conv2dBackpropFilter.
    • Add float64 support for AvgPool/AvgPoolGrad.
    • Make graph name scope thread local so that they work correctly in multi-threaded environments.
    • Update nsync synchronization library to avoid slow primitives on Linux.
    • Removed need to put nsync/public on C include path when building custom ops.
    • Add tf.image.psnr, tf.image.ssim, tf.image.ssim_multiscale, tf.image.image_gradients, tf.image.sobel_edges.
    • Add links to https://js.tensorflow.org.
    • Fix non-uniformity of orthogonal matrices.
    • Fix bug where multi-image Estimator eval summaries were not displayed correctly.

1 The cancellation logic of the RPC op contains a concurrency error. A fix has been submitted to master and will be part of the next release.

Thanks to our Contributors

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

4d55397500, Aghasy, Alan Du, Alan Lee, Alan Yee, Alex Wiltschko, Animesh Karnewar, Ankit Gupta, Anton Matosov, Aris L, Ben Barsdell, Brent Yi, Brett Koonce, Carl Thomé, cbockman, Chikanaga Tomoyuki, Chris Tava, CéDric Deltheil, Dahan Gong, Dalmo Cirne, Daniel Erenrich, David Norman, DavidNorman, Edd Wilder-James, Fanjin Zeng, Felix Abecassis, fo40225, George Sterpu, Giovanni Terlingen, Gor Baghdasaryan, Guillaume Klein, Hanchen Li, Ilya Polenov, Jakub Kolodziejczyk, Jason Sadler, Jayaram Bobba, Jerry Liu, jinghuangintel, Jiongyan Zhang (张炯衍), Joel Shor, Jong Wook Kim, Julian Eisenschlos, Karl Lessard, Krish Ravindranath, Loo Rong Jie, Lukas Geiger, Luke Iwanski, Mahmoud Abuzaina, ManHyuk, Marvin Richter, Maximilian Mitchell, Mohammad Ashraf Bhuiyan, msofka, Mustafa Kasap, Nathan Burnham, Nathan Luehr, Naveen Marri, ngc92, nio1814, Oleg Zabluda, Ou Changkun, Panos Ipeirotis, Paul Van Eck, Peter Lee, Piotr Czapla, qjivy, Rholais Lii, Rodrigo Formigone, Russell Klopfer, ryantimjohn, Sang Han, SebastiáN RamíRez, shengfuintel, Siby Jose Plathottam, Silver Chan, Stanislaw Antol, Taehoon Lee, Tarang Chugh, Ted Chang, Thomas Bastiani, Xian Xu, Xiaoming (Jason) Cui, Yan Facai (颜发才), yaox12, Yashal Shakti Kanungo, Yong Tang, Yuan (Terry) Tang, Yuxin Wu, Ziyue(Louis) Lu

Release 1.7.0

Major Features And Improvements

  • Eager mode is moving out of contrib, try tf.enable_eager_execution().
  • Graph rewrites emulating fixed-point quantization compatible with TensorFlow Lite, supported by new tf.contrib.quantize package.
  • Easily customize gradient computation with tf.custom_gradient.
  • TensorBoard Debugger Plugin, the graphical user interface (GUI) of TensorFlow Debugger (tfdbg), is now in alpha.
  • Experimental support for reading a sqlite database as a Dataset with new tf.contrib.data.SqlDataset.
  • Distributed Mutex / CriticalSection added to tf.contrib.framework.CriticalSection.
  • Better text processing with tf.regex_replace.
  • Easy, efficient sequence input with tf.contrib.data.bucket_by_sequence_length
  • Initial support for tf.contrib.tensorrt that enables native TensorRT in TensorFlow.

Bug Fixes and Other Changes

  • Accelerated Linear Algebra (XLA):
    • Add MaxPoolGradGrad support for XLA
    • CSE pass from Tensorflow is now disabled in XLA.
  • tf.data:
    • tf.data.Dataset
      • Add support for building C++ Dataset op kernels as external libraries, using the tf.load_op_library() mechanism.
      • Dataset.list_files() now shuffles its output by default.
      • Dataset.shuffle(..., seed=tf.constant(0, dtype=tf.int64)) now yields the same sequence of elements as Dataset.shuffle(..., seed=0).
    • Add num_parallel_reads argument to tf.data.TFRecordDataset.
  • tf.contrib:
    • tf.contrib.bayesflow.halton_sequence now supports randomization.
    • Add support for scalars in tf.contrib.all_reduce.
    • Add effective_sample_size to tf.contrib.bayesflow.mcmc_diagnostics.
    • Add potential_scale_reduction to tf.contrib.bayesflow.mcmc_diagnostics.
    • Add BatchNormalization, Kumaraswamy bijectors.
    • Deprecate tf.contrib.learn. Please check contrib/learn/README.md for instructions on how to convert existing code.
    • tf.contrib.data
      • Remove deprecated tf.contrib.data.Dataset, tf.contrib.data.Iterator, tf.contrib.data.FixedLengthRecordDataset, tf.contrib.data.TextLineDataset, and tf.contrib.data.TFRecordDataset classes.
      • Added bucket_by_sequence_length, sliding_window_batch, and make_batched_features_dataset
    • Remove unmaintained tf.contrib.ndlstm. You can find it externally at https://github.com/tmbarchive/tfndlstm.
    • Moved most of tf.contrib.bayesflow to its own repo: tfp
  • Other:
    • tf.py_func now reports the full stack trace if an exception occurs.
    • Integrate TPUClusterResolver with GKE's integration for Cloud TPUs.
    • Add a library for statistical testing of samplers.
    • Add Helpers to stream data from the GCE VM to a Cloud TPU.
    • Integrate ClusterResolvers with TPUEstimator.
    • Unify metropolis_hastings interface with HMC kernel.
    • Move LIBXSMM convolutions to a separate --define flag so that they are disabled by default.
    • Fix MomentumOptimizer lambda.
    • Reduce tfp.layers boilerplate via programmable docstrings.
    • Add auc_with_confidence_intervals, a method for computing the AUC and confidence interval with linearithmic time complexity.
    • regression_head now accepts customized link function, to satisfy the usage that user can define their own link function if the array_ops.identity does not meet the requirement.
    • Fix initialized_value and initial_value behaviors for ResourceVariables created from VariableDef protos.
    • Add TensorSpec to represent the specification of Tensors.
    • Constant folding pass is now deterministic.
    • Support float16 dtype in tf.linalg.*.
    • Add tf.estimator.export.TensorServingInputReceiver that allows tf.estimator.Estimator.export_savedmodel to pass raw tensors to model functions.

Deprecations

  • TensorFlow 1.7 may be the last time we support Cuda versions below 8.0. Starting with TensorFlow 1.8 release, 8.0 will be the minimum supported version.
  • TensorFlow 1.7 may be the last time we support cuDNN versions below 6.0. Starting with TensorFlow 1.8 release, 6.0 will be the minimum supported version.

Thanks to our Contributors

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

4d55397500, Abe, Alistair Low, Andy Kernahan, Appledore, Ben, Ben Barsdell, Boris Pfahringer, Brad Wannow, Brett Koonce, Carl Thomé, cclauss, Chengzhi Chen, Chris Drake, Christopher Yeh, Clayne Robison, Codrut Grosu, Daniel Trebbien, Danny Goodman, David Goodwin, David Norman, Deron Eriksson, Donggeon Lim, Donny Viszneki, DosLin, DylanDmitri, Francisco Guerrero, Fred Reiss, gdh1995, Giuseppe, Glenn Weidner, gracehoney, Guozhong Zhuang, Haichen "Hc" Li, Harald Husum, harumitsu.nobuta, Henry Spivey, hsm207, Jekyll Song, Jerome, Jiongyan Zhang, jjsjann123, John Sungjin Park, Johnson145, JoshVarty, Julian Wolff, Jun Wang, June-One, Kamil Sindi, Kb Sriram, Kdavis-Mozilla, Kenji, lazypanda1, Liang-Chi Hsieh, Loo Rong Jie, Mahesh Bhosale, MandarJKulkarni, ManHyuk, Marcus Ong, Marshal Hayes, Martin Pool, matthieudelaro, mdfaijul, mholzel, Michael Zhou, Ming Li, Minmin Sun, Myungjoo Ham, MyungsungKwak, Naman Kamra, Peng Yu, Penghao Cen, Phil, Raghuraman-K, resec, Rohin Mohanadas, Sandeep N Gupta, Scott Tseng, seaotterman, Seo Sanghyeon, Sergei Lebedev, Ted Chang, terrytangyuan, Tim H, tkunic, Tod, vihanjain, Yan Facai (颜发才), Yin Li, Yong Tang, Yukun Chen, Yusuke Yamada

Release 1.6.0

Breaking Changes

  • Prebuilt binaries are now built against CUDA 9.0 and cuDNN 7.
  • Prebuilt binaries will use AVX instructions. This may break TF on older CPUs.

Major Features And Improvements

  • New Optimizer internal API for non-slot variables. Descendants of AdamOptimizer that access _beta[12]_power will need to be updated.
  • tf.estimator.{FinalExporter,LatestExporter} now export stripped SavedModels. This improves forward compatibility of the SavedModel.
  • FFT support added to XLA CPU/GPU.

Bug Fixes and Other Changes

  • Documentation updates:
    • Added a second version of Getting Started, which is aimed at ML newcomers.
    • Clarified documentation on resize_images.align_corners parameter.
    • Additional documentation for TPUs.
  • Google Cloud Storage (GCS):
    • Add client-side throttle.
    • Add a FlushCaches() method to the FileSystem interface, with an implementation for GcsFileSystem.
  • Other:
    • Add tf.contrib.distributions.Kumaraswamy.
    • RetryingFileSystem::FlushCaches() calls the base FileSystem's FlushCaches().
    • Add auto_correlation to distributions.
    • Add tf.contrib.distributions.Autoregressive.
    • Add SeparableConv1D layer.
    • Add convolutional Flipout layers.
    • When both inputs of tf.matmul are bfloat16, it returns bfloat16, instead of float32.
    • Added tf.contrib.image.connected_components.
    • Add tf.contrib.framework.CriticalSection that allows atomic variable access.
    • Output variance over trees predictions for classifications tasks.
    • For pt and eval commands, allow writing tensor values to filesystem as numpy files.
    • gRPC: Propagate truncated errors (instead of returning gRPC internal error).
    • Augment parallel_interleave to support 2 kinds of prefetching.
    • Improved XLA support for C64-related ops log, pow, atan2, tanh.
    • Add probabilistic convolutional layers.

API Changes

  • Introducing prepare_variance boolean with default setting to False for backward compatibility.
  • Move layers_dense_variational_impl.py to layers_dense_variational.py.

Known Bugs

  • Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or CUDA_ILLEGAL_ADDRESS failures.

    Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA 9 and CUDA 9.1 sometimes does not properly compute the carry bit when decomposing 64-bit address calculations with large offsets (e.g. load [x + large_constant]) into 32-bit arithmetic in SASS.

    As a result, these versions of ptxas miscompile most XLA programs which use more than 4GB of temp memory. This results in garbage results and/or CUDA_ERROR_ILLEGAL_ADDRESS failures.

    A fix in CUDA 9.1.121 is expected in late February 2018. We do not expect a fix for CUDA 9.0.x. Until the fix is available, the only workaround is to downgrade to CUDA 8.0.x or disable XLA:GPU.

    TensorFlow will print a warning if you use XLA:GPU with a known-bad version of CUDA; see e00ba24c4038e7644da417ddc639169b6ea59122.

Thanks to our Contributors

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

4d55397500, Ag Ramesh, Aiden Scandella, Akimasa Kimura, Alex Rothberg, Allen Goodman, amilioto, Andrei Costinescu, Andrei Nigmatulin, Anjum Sayed, Anthony Platanios, Anush Elangovan, Armando Fandango, Ashish Kumar Ram, Ashwini Shukla, Ben, Bhavani Subramanian, Brett Koonce, Carl Thomé, cclauss, Cesc, Changming Sun, Christoph Boeddeker, Clayne Robison, Clemens Schulz, Clint (Woonhyuk Baek), codrut3, Cole Gerdemann, Colin Raffel, Daniel Trebbien, Daniel Ylitalo, Daniel Zhang, Daniyar, Darjan Salaj, Dave Maclachlan, David Norman, Dong--Jian, dongsamb, dssgsra, Edward H, eladweiss, elilienstein, Eric Lilienstein, error.d, Eunji Jeong, fanlu, Florian Courtial, fo40225, Fred, Gregg Helt, Guozhong Zhuang, Hanchen Li, hsm207, hyunyoung2, ImSheridan, Ishant Mrinal Haloi, Jacky Ko, Jay Young, Jean Flaherty, Jerome, JerrikEph, Jesse Kinkead, jfaath, Jian Lin, jinghuangintel, Jiongyan Zhang, Joel Hestness, Joel Shor, Johnny Chan, Julian Niedermeier, Julian Wolff, JxKing, K-W-W, Karl Lessard, Kasper Marstal, Keiji Ariyama, Koan-Sin Tan, Loki Der Quaeler, Loo Rong Jie, Luke Schaefer, Lynn Jackson, ManHyuk, Matt Basta, Matt Smith, Matthew Schulkind, Michael, michaelkhan3, Miguel Piedrafita, Mikalai Drabovich, Mike Knapp, mjwen, mktozk, Mohamed Aly, Mohammad Ashraf Bhuiyan, Myungjoo Ham, Naman Bhalla, Namrata-Ibm, Nathan Luehr, nathansilberman, Netzeband, Niranjan Hasabnis, Omar Aflak, Ozge Yalcinkaya, Parth P Panchal, patrickzzy, Patryk Chrabaszcz, Paul Van Eck, Paweł Kapica, Peng Yu, Philip Yang, Pierre Blondeau, Po-Hsien Chu, powderluv, Puyu Wang, Rajendra Arora, Rasmus, Renat Idrisov, resec, Robin Richtsfeld, Ronald Eddy Jr, Sahil Singh, Sam Matzek, Sami Kama, sandipmgiri, Santiago Castro, Sayed Hadi Hashemi, Scott Tseng, Sergii Khomenko, Shahid, Shengpeng Liu, Shreyash Sharma, Shrinidhi Kl, Simone Cirillo, simsicon, Stanislav Levental, starsblinking, Stephen Lumenta, Steven Hickson, Su Tang, Taehoon Lee, Takuya Wakisaka, Ted Chang, Ted Ying, Tijmen Verhulsdonck, Timofey Kondrashov, vade, vaibhav, Valentin Khrulkov, vchigrin, Victor Costan, Viraj Navkal, Vivek Rane, wagonhelm, Yan Facai (颜发才), Yanbo Liang, Yaroslav Bulatov, yegord, Yong Tang, Yoni Tsafir, yordun, Yuan (Terry) Tang, Yuxin Wu, zhengdi, Zhengsheng Wei, 田传武

Release 1.5.0

Breaking Changes

  • Prebuilt binaries are now built against CUDA 9.0 and cuDNN 7.
  • Starting from 1.6 release, our prebuilt binaries will use AVX instructions. This may break TF on older CPUs.

Major Features And Improvements

  • Eager execution preview version is now available.
  • TensorFlow Lite dev preview is now available.
  • CUDA 9.0 and cuDNN 7 support.
  • Accelerated Linear Algebra (XLA):
    • Add complex64 support to XLA compiler.
    • bfloat support is now added to XLA infrastructure.
    • Make ClusterSpec propagation work with XLA devices.
    • Use a deterministic executor to generate XLA graph.
  • tf.contrib:
    • tf.contrib.distributions:
      • Add tf.contrib.distributions.Autoregressive.
      • Make tf.contrib.distributions QuadratureCompound classes support batch
      • Infer tf.contrib.distributions.RelaxedOneHotCategorical dtype from arguments.
      • Make tf.contrib.distributions quadrature family parameterized by quadrature_grid_and_prob vs quadrature_degree.
      • auto_correlation added to tf.contrib.distributions
    • Add tf.contrib.bayesflow.layers, a collection of probabilistic (neural) layers.
    • Add tf.contrib.bayesflow.halton_sequence.
    • Add tf.contrib.data.make_saveable_from_iterator.
    • Add tf.contrib.data.shuffle_and_repeat.
    • Add new custom transformation: tf.contrib.data.scan().
    • tf.contrib.distributions.bijectors:
      • Add tf.contrib.distributions.bijectors.MaskedAutoregressiveFlow.
      • Add tf.contrib.distributions.bijectors.Permute.
      • Add tf.contrib.distributions.bijectors.Gumbel.
      • Add tf.contrib.distributions.bijectors.Reshape.
      • Support shape inference (i.e., shapes containing -1) in the Reshape bijector.
  • Add streaming_precision_recall_at_equal_thresholds, a method for computing streaming precision and recall with O(num_thresholds + size of predictions) time and space complexity.
  • Change RunConfig default behavior to not set a random seed, making random behavior independently random on distributed workers. We expect this to generally improve training performance. Models that do rely on determinism should set a random seed explicitly.
  • Replaced the implementation of tf.flags with absl.flags.
  • Add support for CUBLAS_TENSOR_OP_MATH in fp16 GEMM
  • Add support for CUDA on NVIDIA Tegra devices

Bug Fixes and Other Changes

  • Documentation updates:
    • Clarified that you can only install TensorFlow on 64-bit machines.
    • Added a short doc explaining how Estimators save checkpoints.
    • Add documentation for ops supported by the tf2xla bridge.
    • Fix minor typos in the doc of SpaceToDepth and DepthToSpace.
    • Updated documentation comments in mfcc_mel_filterbank.h and mfcc.h to clarify that the input domain is squared magnitude spectra and the weighting is done on linear magnitude spectra (sqrt of inputs).
    • Change tf.contrib.distributions docstring examples to use tfd alias rather than ds, bs.
    • Fix docstring typos in tf.distributions.bijectors.Bijector.
    • tf.assert_equal no longer raises ValueError. It now raises InvalidArgumentError, as documented.
    • Update Getting Started docs and API intro.
  • Google Cloud Storage (GCS):
    • Add userspace DNS caching for the GCS client.
    • Customize request timeouts for the GCS filesystem.
    • Improve GCS filesystem caching.
  • Bug Fixes:
    • Fix bug where partitioned integer variables got their wrong shapes. Before
    • Fix correctness bug in CPU and GPU implementations of Adadelta.
    • Fix a bug in import_meta_graph's handling of partitioned variables when importing into a scope. WARNING: This may break loading checkpoints of graphs with partitioned variables saved after using import_meta_graph with a non-empty import_scope argument.
    • Fix bug in offline debugger which prevented viewing events.
    • Added the WorkerService.DeleteWorkerSession method to the gRPC interface, to fix a memory leak. Ensure that your master and worker servers are running the same version of TensorFlow to avoid compatibility issues.
    • Fix bug in peephole implementation of BlockLSTM cell.
    • Fix bug by casting dtype of log_det_jacobian to match log_prob in TransformedDistribution.
    • Fix a bug in import_meta_graph's handling of partitioned variables when
    • Ensure tf.distributions.Multinomial doesn't underflow in log_prob. Before this change, all partitions of an integer variable were initialized with the shape of the unpartitioned variable; after this change they are initialized correctly.
  • Other:
    • Add necessary shape util support for bfloat16.
    • Add a way to run ops using a step function to MonitoredSession.
    • Add DenseFlipout probabilistic layer.
    • A new flag ignore_live_threads is available on train. If set to True, it will ignore threads that remain running when tearing down infrastructure after successfully completing training, instead of throwing a RuntimeError.
    • Restandardize DenseVariational as simpler template for other probabilistic layers.
    • tf.data now supports tf.SparseTensor components in dataset elements.
    • It is now possible to iterate over Tensors.
    • Allow SparseSegmentReduction ops to have missing segment IDs.
    • Modify custom export strategy to account for multidimensional sparse float splits.
    • Conv2D, Conv2DBackpropInput, Conv2DBackpropFilter now supports arbitrary dilations with GPU and cuDNNv6 support.
    • Estimator now supports Dataset: input_fn can return a Dataset instead of Tensors.
    • Add RevBlock, a memory-efficient implementation of reversible residual layers.
    • Reduce BFCAllocator internal fragmentation.
    • Add cross_entropy and kl_divergence to tf.distributions.Distribution.
    • Add tf.nn.softmax_cross_entropy_with_logits_v2 which enables backprop w.r.t. the labels.
    • GPU back-end now uses ptxas to compile generated PTX.
    • BufferAssignment's protocol buffer dump is now deterministic.
    • Change embedding op to use parallel version of DynamicStitch.
    • Add support for sparse multidimensional feature columns.
    • Speed up the case for sparse float columns that have only 1 value.
    • Allow sparse float splits to support multivalent feature columns.
    • Add quantile to tf.distributions.TransformedDistribution.
    • Add NCHW_VECT_C support for tf.depth_to_space on GPU.
    • Add NCHW_VECT_C support for tf.space_to_depth on GPU.

API Changes

  • Rename SqueezeDims attribute to Axis in C++ API for Squeeze op.
  • Stream::BlockHostUntilDone now returns Status rather than bool.
  • Minor refactor: move stats files from stochastic to common and remove stochastic.

Known Bugs

  • Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or CUDA_ILLEGAL_ADDRESS failures.

    Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA 9 and CUDA 9.1 sometimes does not properly compute the carry bit when decomposing 64-bit address calculations with large offsets (e.g. load [x + large_constant]) into 32-bit arithmetic in SASS.

    As a result, these versions of ptxas miscompile most XLA programs which use more than 4GB of temp memory. This results in garbage results and/or CUDA_ERROR_ILLEGAL_ADDRESS failures.

    A fix in CUDA 9.1.121 is expected in late February 2018. We do not expect a fix for CUDA 9.0.x. Until the fix is available, the only workaround is to downgrade to CUDA 8.0.x or disable XLA:GPU.

    TensorFlow will print a warning if you use XLA:GPU with a known-bad version of CUDA; see e00ba24c4038e7644da417ddc639169b6ea59122.

Thanks to our Contributors

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

Adam Zahran, Ag Ramesh, Alan Lee, Alan Yee, Alex Sergeev, Alexander, Amir H. Jadidinejad, Amy, Anastasios Doumoulakis, Andrei Costinescu, Andrei Nigmatulin, Anthony Platanios, Anush Elangovan, arixlin, Armen Donigian, ArtëM Sobolev, Atlas7, Ben Barsdell, Bill Prin, Bo Wang, Brett Koonce, Cameron Thomas, Carl Thomé, Cem Eteke, cglewis, Changming Sun, Charles Shenton, Chi-Hung, Chris Donahue, Chris Filo Gorgolewski, Chris Hoyean Song, Chris Tava, Christian Grail, Christoph Boeddeker, cinqS, Clayne Robison, codrut3, concerttttt, CQY, Dan Becker, Dan Jarvis, Daniel Zhang, David Norman, dmaclach, Dmitry Trifonov, Donggeon Lim, dongpilYu, Dr. Kashif Rasul, Edd Wilder-James, Eric Lv, fcharras, Felix Abecassis, FirefoxMetzger, formath, FredZhang, Gaojin Cao, Gary Deer, Guenther Schmuelling, Hanchen Li, Hanmin Qin, hannesa2, hyunyoung2, Ilya Edrenkin, Jackson Kontny, Jan, Javier Luraschi, Jay Young, Jayaram Bobba, Jeff, Jeff Carpenter, Jeremy Sharpe, Jeroen BéDorf, Jimmy Jia, Jinze Bai, Jiongyan Zhang, Joe Castagneri, Johan Ju, Josh Varty, Julian Niedermeier, JxKing, Karl Lessard, Kb Sriram, Keven Wang, Koan-Sin Tan, Kyle Mills, lanhin, LevineHuang, Loki Der Quaeler, Loo Rong Jie, Luke Iwanski, LáSzló Csomor, Mahdi Abavisani, Mahmoud Abuzaina, ManHyuk, Marek ŠUppa, MathSquared, Mats Linander, Matt Wytock, Matthew Daley, Maximilian Bachl, mdymczyk, melvyniandrag, Michael Case, Mike Traynor, miqlas, Namrata-Ibm, Nathan Luehr, Nathan Van Doorn, Noa Ezra, Nolan Liu, Oleg Zabluda, opensourcemattress, Ouwen Huang, Paul Van Eck, peisong, Peng Yu, PinkySan, pks, powderluv, Qiao Hai-Jun, Qiao Longfei, Rajendra Arora, Ralph Tang, resec, Robin Richtsfeld, Rohan Varma, Ryohei Kuroki, SaintNazaire, Samuel He, Sandeep Dcunha, sandipmgiri, Sang Han, scott, Scott Mudge, Se-Won Kim, Simon Perkins, Simone Cirillo, Steffen Schmitz, Suvojit Manna, Sylvus, Taehoon Lee, Ted Chang, Thomas Deegan, Till Hoffmann, Tim, Toni Kunic, Toon Verstraelen, Tristan Rice, Urs KöSter, Utkarsh Upadhyay, Vish (Ishaya) Abrams, Winnie Tsang, Yan Chen, Yan Facai (颜发才), Yi Yang, Yong Tang, Youssef Hesham, Yuan (Terry) Tang, Zhengsheng Wei, zxcqwe4906, 张志豪, 田传武

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 1.4.1

Bug Fixes and Other Changes

  • LinearClassifier fix.

Release 1.4.0

Major Features And Improvements

  • tf.keras is now part of the core TensorFlow API.
  • tf.data is now part of the core TensorFlow API.
    • The API is now subject to backwards compatibility guarantees.
    • For a guide to migrating from the tf.contrib.data API, see the README.
    • Major new features include Dataset.from_generator() (for building an input pipeline from a Python generator), and the Dataset.apply() method for applying custom transformation functions.
    • Several custom transformation functions have been added, including tf.contrib.data.batch_and_drop_remainder() and tf.contrib.data.sloppy_interleave().
  • Add train_and_evaluate for simple distributed Estimator training.
  • Add tf.spectral.dct for computing the DCT-II.
  • Add Mel-Frequency Cepstral Coefficient support to tf.contrib.signal (with GPU and gradient support).
  • Add a self-check on import tensorflow for Windows DLL issues.
  • Add NCHW support to tf.depth_to_space on GPU.
  • TensorFlow Debugger (tfdbg):
    • Add eval command to allow evaluation of arbitrary Python/numpy expressions in tfdbg command-line interface. See Debugging TensorFlow Programs for more details.
    • Usability improvement: The frequently used tensor filter has_inf_or_nan is now added to Session wrappers and hooks by default. So there is no need for clients to call .add_tensor_filter(tf_debug.has_inf_or_nan) anymore.
  • SinhArcsinh (scalar) distribution added to contrib.distributions.
  • Make GANEstimator opensource.
  • Estimator.export_savedmodel() now includes all valid serving signatures that can be constructed from the Serving Input Receiver and all available ExportOutputs. For instance, a classifier may provide regression- and prediction-flavored outputs, in addition to the classification-flavored one. Building signatures from these allows TF Serving to honor requests using the different APIs (Classify, Regress, and Predict). Furthermore, serving_input_receiver_fn() may now specify alternative subsets of nodes that may act as inputs. This allows, for instance, producing a prediction signature for a classifier that accepts raw Tensors instead of a serialized tf.Example.
  • Add tf.contrib.bayesflow.hmc.
  • Add tf.contrib.distributions.MixtureSameFamily.
  • Make Dataset.shuffle() always reshuffles after each iteration by default.
  • Add tf.contrib.bayesflow.metropolis_hastings.
  • Add log_rate parameter to tf.contrib.distributions.Poisson.
  • Extend tf.contrib.distributions.bijector API to handle some non-injective transforms.
  • Java:
    • Generics (e.g., Tensor<Integer>) for improved type-safety (courtesy @andrewcmyers).
    • Support for multi-dimensional string tensors.
    • Support loading of custom operations (e.g. many in tf.contrib) on Linux and OS X
  • All our prebuilt binaries have been built with CUDA 8 and cuDNN 6. We anticipate releasing TensorFlow 1.5 with CUDA 9 and cuDNN 7.

Bug Fixes and Other Changes

  • tf.nn.rnn_cell.DropoutWrapper is now more careful about dropping out LSTM states. Specifically, it no longer ever drops the c (memory) state of an LSTMStateTuple. The new behavior leads to proper dropout behavior for LSTMs and stacked LSTMs. This bug fix follows recommendations from published literature, but is a behavioral change. State dropout behavior may be customized via the new dropout_state_filter_visitor argument.
  • Removed tf.contrib.training.python_input. The same behavior, in a more flexible and reproducible package, is available via the new tf.contrib.data.Dataset.from_generator method!
  • Fix tf.contrib.distributions.Affine incorrectly computing log-det-jacobian.
  • Fix tf.random_gamma incorrectly handling non-batch, scalar draws.
  • Resolved a race condition in TensorForest TreePredictionsV4Op.
  • Google Cloud Storage file system, Amazon S3 file system, and Hadoop file system support are now default build options.
  • Custom op libraries must link against libtensorflow_framework.so (installed at tf.sysconfig.get_lib()).
  • Change RunConfig default behavior to not set a random seed, making random behavior independently random on distributed workers. We expect this to generally improve training performance. Models that do rely on determinism should set a random seed explicitly.

Breaking Changes to the API

  • The signature of the tf.contrib.data.rejection_resample() function has been changed. It now returns a function that can be used as an argument to Dataset.apply().
  • Remove tf.contrib.data.Iterator.from_dataset() method. Use Dataset.make_initializable_iterator() instead.
  • Remove seldom used and unnecessary tf.contrib.data.Iterator.dispose_op().
  • Reorder some TF-GAN loss functions in a non-backwards compatible way.

Known Issues

  • In Python 3, Dataset.from_generator() does not support Unicode strings. You must convert any strings to bytes objects before yielding them from the generator.

Thanks to our Contributors

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

4d55397500, Abdullah Alrasheed, abenmao, Adam Salvail, Aditya Dhulipala, Ag Ramesh, Akimasa Kimura, Alan Du, Alan Yee, Alexander, Amit Kushwaha, Amy, Andrei Costinescu, Andrei Nigmatulin, Andrew Erlichson, Andrew Myers, Andrew Stepanov, Androbin, AngryPowman, Anish Shah, Anton Daitche, Artsiom Chapialiou, asdf2014, Aseem Raj Baranwal, Ash Hall, Bart Kiers, Batchu Venkat Vishal, ben, Ben Barsdell, Bill Piel, Carl Thomé, Catalin Voss, Changming Sun, Chengzhi Chen, Chi Zeng, Chris Antaki, Chris Donahue, Chris Oelmueller, Chris Tava, Clayne Robison, Codrut, Courtial Florian, Dalmo Cirne, Dan J, Darren Garvey, David Kristoffersson, David Norman, David RöThlisberger, DavidNorman, Dhruv, DimanNe, Dorokhov, Duncan Mac-Vicar P, EdwardDixon, EMCP, error.d, FAIJUL, Fan Xia, Francois Xavier, Fred Reiss, Freedom" Koan-Sin Tan, Fritz Obermeyer, Gao, Xiang, Guenther Schmuelling, Guo Yejun (郭叶军), Hans Gaiser, HectorSVC, Hyungsuk Yoon, James Pruegsanusak, Jay Young, Jean Wanka, Jeff Carpenter, Jeremy Rutman, Jeroen BéDorf, Jett Jones, Jimmy Jia, jinghuangintel, jinze1994, JKurland, Joel Hestness, joetoth, John B Nelson, John Impallomeni, John Lawson, Jonas, Jonathan Dekhtiar, joshkyh, Jun Luan, Jun Mei, Kai Sasaki, Karl Lessard, karl@kubx.ca, Kb Sriram, Kenichi Ueno, Kevin Slagle, Kongsea, Lakshay Garg, lhlmgr, Lin Min, liu.guangcong, Loki Der Quaeler, Louie Helm, lucasmoura, Luke Iwanski, Lyndon White, Mahmoud Abuzaina, Marcel Puyat, Mark Aaron Shirley, Michele Colombo, MtDersvan, Namrata-Ibm, Nathan Luehr, Naurril, Nayana Thorat, Nicolas Lopez, Niranjan Hasabnis, Nolan Liu, Nouce, Oliver Hennigh, osdamv, Patrik Erdes, Patryk Chrabaszcz, Pavel Christof, Penghao Cen, postBG, Qingqing Cao, Qingying Chen, qjivy, Raphael, Rasmi, raymondxyang, Renze Yu, resec, Roffel, Ruben Vereecken, Ryohei Kuroki, sandipmgiri, Santiago Castro, Scott Kirkland, Sean Vig, Sebastian Raschka, Sebastian Weiss, Sergey Kolesnikov, Sergii Khomenko, Shahid, Shivam Kotwalia, Stuart Berg, Sumit Gouthaman, superzerg, Sven Mayer, tetris, Ti Zhou, Tiago Freitas Pereira, Tian Jin, Tomoaki Oiki, Vaibhav Sood, vfdev, Vivek Rane, Vladimir Moskva, wangqr, Weber Xie, Will Frey, Yan Facai (颜发才), yanivbl6, Yaroslav Bulatov, Yixing Lao, Yong Tang, youkaichao, Yuan (Terry) Tang, Yue Zhang, Yuxin Wu, Ziming Dong, ZxYuan, 黄璞

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 1.3.0

See also TensorBoard 0.1.4 release notes.

Major Features and Improvements

  • Added canned estimators to Tensorflow library. List of added estimators:
    • DNNClassifier
    • DNNRegressor
    • LinearClassifier
    • LinearRegressor
    • DNNLinearCombinedClassifier
    • DNNLinearCombinedRegressor.
  • All our prebuilt binaries have been built with cuDNN 6. We anticipate releasing TensorFlow 1.4 with cuDNN 7.
  • import tensorflow now goes much faster.
  • Adds a file cache to the GCS filesystem with configurable max staleness for file contents. This permits caching of file contents across close/open boundaries.
  • Added an axis parameter to tf.gather.
  • Added a constant_values keyword argument to tf.pad.
  • Adds Dataset.interleave transformation.
  • Add ConcatenateDataset to concatenate two datasets.
  • Added Mobilenet support to TensorFlow for Poets training script.
  • Adds a block cache to the GCS filesystem with configurable block size and count.
  • SinhArcSinh bijector added.
  • Added Dataset.list_files API.
  • Introduces new operations and Python bindings for the Cloud TPU.
  • Adding TensorFlow-iOS CocoaPod for symmetry with tensorflow-android.
  • Introduces base implementations of ClusterResolvers.
  • Unify memory representations of TensorShape and PartialTensorShape. As a consequence, tensors now have a maximum of 254 dimensions, not 255.
  • Changed references to LIBXSMM to use version 1.8.1.
  • TensorFlow Debugger (tfdbg):
    • Display summaries of numeric tensor values with the -s flag to command print_tensor or pt.
    • Display feed values with the print_feed or pf command and clickable links in the curses UI.
    • Runtime profiler at the op level and the Python source line level with the run -p command.
  • Initial release of the statistical distribution library tf.distributions.
  • GPU kernels and speed improvements for unary tf.where and tf.nn.top_k.
  • Monotonic Attention wrappers added to tf.contrib.seq2seq.
  • Added tf.contrib.signal, a library for signal processing primitives.
  • Added tf.contrib.resampler, containing CPU and GPU ops for differentiable resampling of images.

Breaking Changes to the API

  • tf.RewriterConfig was removed from the Python API after being available in 1.2 release candidates (it was never in an actual release). Graph rewriting is still available, just not as tf.RewriterConfig. Instead add an explicit import.
  • Breaking change to tf.contrib.data.Dataset APIs that expect a nested structure. Lists are now converted to tf.Tensor implicitly. You may need to change uses of lists to tuples in existing code. In addition, dicts are now supported as a nested structure.

Changes to contrib APIs

  • Adds tf.contrib.nn.rank_sampled_softmax_loss, a sampled-softmax variant that can improve rank loss.
  • tf.contrib.metrics.{streaming_covariance,streaming_pearson_correlation} modified to return nan when they have seen less or equal to 1 unit of weight.
  • Adds time series models to contrib. See contrib/timeseries/README.md for details.
  • Adds FULLY_CONNECTED Op to tensorflow/lite/schema.fbs

Known Issues

  • Tensorflow_gpu compilation fails with Bazel 0.5.3.

Bug Fixes and Other Changes

  • Fixes strides and begin dtype mismatch when slicing using int64 Tensor index in python.
  • Improved convolution padding documentation.
  • Add a tag constant, gpu, to present graph with GPU support.
  • saved_model.utils now support SparseTensors transparently.
  • A more efficient implementation of non-max suppression.
  • Add support for the shrinkage-type L2 to FtrlOptimizer in addition to the online L2 it already supports.
  • Fix negative variance in moments calculation.
  • Expand UniqueOp Benchmark Tests to cover more collision cases.
  • Improves stability of GCS filesystem on Mac.
  • Add time estimation to HloCostAnalysis.
  • Fixed the bug in Estimator that params in constructor was not a deepcopy of the user provided one. This bugs inadvertently enabled user to mutate the params after the creation of Estimator, leading to potentially undefined behavior.
  • Added None check for save_path in saver.restore.
  • Register devices under their legacy names in device_mgr to ease the transition to clusterspec-propagated configurations.
  • VectorExponential added to distributions.
  • Add a bitwise module with bitwise_and, bitwise_or, bitwise_xor, and invert functions.
  • Add fixed-grid ODE integration routines.
  • Allow passing bounds to ScipyOptimizerInterface.
  • Correctness fixes for fft_length parameter to tf.spectral.rfft & tf.spectral.irfft.
  • Exported model signatures using the 'predict' method will no longer have their input and output keys silently ignored and rewritten to 'inputs' and 'outputs'. If a model was exported with different names before 1.2, and is now served with tensorflow/serving, it will accept requests using 'inputs' and 'outputs'. Starting at 1.2, such a model will accept the keys specified during export. Therefore, inference requests using 'inputs' and 'outputs' may start to fail. To fix this, either update any inference clients to send requests with the actual input and output keys used by the trainer code, or conversely, update the trainer code to name the input and output Tensors 'inputs' and 'outputs', respectively. Signatures using the 'classify' and 'regress' methods are not affected by this change; they will continue to standardize their input and output keys as before.
  • Add in-memory caching to the Dataset API.
  • Set default end_of_sequence variable in datasets iterators to false.
  • [Performance] Increase performance of tf.layers.conv2d when setting use_bias=True by 2x by using nn.bias_add.
  • Update iOS examples to use CocoaPods, and moved to tensorflow/examples/ios.
  • Adds a family= attribute in tf.summary ops to allow controlling the tab name used in Tensorboard for organizing summaries.
  • When GPU is configured, do not require --config=cuda, instead, automatically build for GPU if this is requested in the configure script.
  • Fix incorrect sampling of small probabilities in CPU/GPU multinomial.
  • Add a list_devices() API on sessions to list devices within a cluster. Additionally, this change augment the ListDevices master API to support specifying a session.
  • Allow uses of over-parameterized separable convolution.
  • TensorForest multi-regression bug fix.
  • Framework now supports armv7, cocoapods.org now displays correct page.
  • Script to create iOS framework for CocoaPods.
  • Android releases of TensorFlow are now pushed to jcenter for easier integration into apps. See https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/android/inference_interface/README.md for more details.
  • TensorFlow Debugger (tfdbg):
    • Fixed a bug that prevented tfdbg from functioning with multi-GPU setups.
    • Fixed a bug that prevented tfdbg from working with tf.Session.make_callable.

Thanks to our Contributors

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

4F2E4A2E, Adriano Carmezim, Adrià Arrufat, Alan Yee, Alex Lattas, Alex Rothberg, Alexandr Baranezky, Ali Siddiqui, Andreas Solleder, Andrei Costinescu, Andrew Hundt, Androbin, Andy Kernahan, Anish Shah, Anthony Platanios, Arvinds-Ds, b1rd, Baptiste Arnaud, Ben Mabey, Benedikt Linse, Beomsu Kim, Bo Wang, Boyuan Deng, Brett Koonce, Bruno Rosa, Carl Thomé, Changming Sun, Chase Roberts, Chirag Bhatia, Chris Antaki, Chris Hoyean Song, Chris Tava, Christos Nikolaou, Croath Liu, cxx, Czxck001, Daniel Ylitalo, Danny Goodman, Darren Garvey, David Brailovsky, David Norman, DavidNorman, davidpham87, ddurham2, Dhruv, DimanNe, Drew Hintz, Dustin Tran, Earthson Lu, ethiraj, Fabian Winnen, Fei Sun, Freedom" Koan-Sin Tan, Fritz Obermeyer, Gao, Xiang, Gautam, Guenther Schmuelling, Gyu-Ho Lee, Hauke Brammer, horance, Humanity123, J Alammar, Jayeol Chun, Jeroen BéDorf, Jianfei Wang, jiefangxuanyan, Jing Jun Yin, Joan Puigcerver, Joel Hestness, Johannes Mayer, John Lawson, Johnson145, Jon Malmaud, Jonathan Alvarez-Gutierrez, Juang, Yi-Lin, Julian Viereck, Kaarthik Sivashanmugam, Karl Lessard, karl@kubx.ca, Kevin Carbone, Kevin Van Der Burgt, Kongsea, ksellesk, lanhin, Lef Ioannidis, Liangliang He, Louis Tiao, Luke Iwanski, LáSzló Csomor, magixsno, Mahmoud Abuzaina, Marcel Hlopko, Mark Neumann, Maxwell Paul Brickner, mdfaijul, MichaëL Defferrard, Michał JastrzęBski, Michele Colombo, Mike Brodie, Mosnoi Ion, mouradmourafiq, myPrecious, Nayana Thorat, Neeraj Kashyap, Nelson Liu, Niranjan Hasabnis, Olivier Moindrot, orome, Pankaj Gupta, Paul Van Eck, peeyush18, Peng Yu, Pierre, preciousdp11, qjivy, Raingo, raoqiyu, ribx, Richard S. Imaoka, Rishabh Patel, Robert Walecki, Rockford Wei, Ryan Kung, Sahil Dua, Sandip Giri, Sayed Hadi Hashemi, sgt101, Shitian Ni, Shuolongbj, Siim PõDer, Simon Perkins, sj6077, SOLARIS, Spotlight0xff, Steffen Eberbach, Stephen Fox, superryanguo, Sven Mayer, Tapan Prakash, Tiago Morais Morgado, Till Hoffmann, Tj Rana, Vadim Markovtsev, vhasanov, Wei Wu, windead, Yan (Asta) Li, Yan Chen, Yann Henon, Yi Wang, Yong Tang, yorkie, Yuan (Terry) Tang, Yuxin Wu, zhengjiajin, zhongzyd, 黄璞

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 1.2.1

Bug Fixes and Other Changes

  • Updating markdown version required to >= 2.6.8.
  • Support tensors as dropout rates again, by removing the min(max(..))

Release 1.2.0

Major Features and Improvements

  • Python 3.6 support on Windows.

  • Added tf.layers.conv3d_transpose layer for spatio temporal deconvolution.

  • Added tf.Session.make_callable(), which provides a lower overhead means of running a similar step multiple times.

  • Added libverbs-based RDMA support to contrib (courtesy @junshi15 from Yahoo).

  • Bring tf.feature_column.* into the API. Non-deprecated functionality from tf.contrib.layers.* is moved to tf.feature_column.* with cosmetic changes.

  • RNNCell objects now subclass tf.layers.Layer. The strictness described in the TensorFlow 1.1 release is gone: The first time an RNNCell is used, it caches its scope. All future uses of the RNNCell will reuse variables from that same scope. This is a breaking change from the behavior of RNNCells in TensorFlow versions <= 1.0.1. TensorFlow 1.1 had checks in place to ensure old code works correctly with the new semantics; this version allows more flexible uses of RNNCell but can lead to subtle errors if using code meant for TensorFlow <= 1.0.1. For example, writing: MultiRNNCell([lstm] * 5) will now build a 5-layer LSTM stack where each layer shares the same parameters. To get 5 layers each with their own parameters, write: MultiRNNCell([LSTMCell(...) for _ in range(5)]). If at all unsure, first test your code with TF 1.1; ensure it raises no errors, and then upgrade to TF 1.2.

  • RNNCells' variable names have been renamed for consistency with Keras layers. Specifically, the previous variable names "weights" and "biases" have been changed to "kernel" and "bias", respectively. This may cause backward incompatibility with regard to your old checkpoints containing such RNN cells, in which case you can use the tool checkpoint_convert script to convert the variable names in your old checkpoints.

  • Many of the RNN functions and classes that were in the tf.nn namespace before the 1.0 release and which were moved to tf.contrib.rnn have now been moved back to the core namespace. This includes RNNCell, LSTMCell, GRUCell, and a number of other cells. These now reside in tf.nn.rnn_cell (with aliases in tf.contrib.rnn for backwards compatibility). The original tf.nn.rnn function is now tf.nn.static_rnn, and the bidirectional static and state saving static rnn functions are also now back in the tf.nn namespace.

    Notable exceptions are the EmbeddingWrapper, InputProjectionWrapper and OutputProjectionWrapper, which will slowly be moved to deprecation in tf.contrib.rnn. These are inefficient wrappers that should often be replaced by calling embedding_lookup or layers.dense as pre- or post- processing of the rnn. For RNN decoding, this functionality has been replaced with an alternative API in tf.contrib.seq2seq.

  • Intel MKL Integration (https://software.intel.com/en-us/articles/tensorflow-optimizations-on-modern-intel-architecture). Intel developed a number of optimized deep learning primitives: In addition to matrix multiplication and convolution, these building blocks include: Direct batched convolution Pooling: maximum, minimum, average Normalization: LRN, batch normalization Activation: rectified linear unit (ReLU) Data manipulation: multi-dimensional transposition (conversion), split, concat, sum and scale.

  • TensorForest Estimator now supports SavedModel export for serving.

  • Support client-provided ClusterSpec's and propagate them to all workers to enable the creation of dynamic TensorFlow clusters.

  • TensorFlow C library now available for Windows.

  • We released a new open-source version of TensorBoard.

  • SavedModel CLI tool available to inspect and execute MetaGraph in SavedModel

  • Android releases of TensorFlow are now pushed to jcenter for easier integration into apps. See https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/android/inference_interface/README.md for more details.

Deprecations

  • TensorFlow 1.2 may be the last time we build with cuDNN 5.1. Starting with TensorFlow 1.3, we will try to build all our prebuilt binaries with cuDNN 6.0. While we will try to keep our source code compatible with cuDNN 5.1, it will be best effort.

Breaking Changes to the API

  • org.tensorflow.contrib.android.TensorFlowInferenceInterface now throws exceptions where possible and has simplified method signatures.

Changes to contrib APIs

  • Added tf.contrib.util.create_example.
  • Added bilinear interpolation to tf.contrib.image.
  • Add tf.contrib.stateless for random ops with custom seed control.
  • MultivariateNormalFullCovariance added to contrib/distributions/
  • tensorflow/contrib/rnn undergoes RNN cell variable renaming for consistency with Keras layers. Specifically, the previous variable names "weights" and "biases" are changed to "kernel" and "bias", respectively. This may cause backward incompatibility with regard to your old checkpoints containing such RNN cells, in which case you can use the checkpoint_convert script to convert the variable names in your old checkpoints.
  • Added tf.contrib.kernel_methods module with Ops and estimators for primal (explicit) kernel methods in TensorFlow.

Bug Fixes and Other Changes

  • In python, Operation.get_attr on type attributes returns the Python DType version of the type to match expected get_attr documentation rather than the protobuf enum.
  • tensorflow/contrib/rnn undergoes RNN cell variable renaming for consistency with Keras layers. Specifically, the previous variable names "weights" and "biases" are changed to "kernel" and "bias", respectively.
  • Changed MIN_SDK version to 8.0 when building iOS libraries.
  • Fixed LIBXSMM integration.
  • Make decode_jpeg/decode_png/decode_gif handle all formats, since users frequently try to decode an image as the wrong type.
  • Improve implicit broadcasting lowering.
  • Improving stability of GCS/BigQuery clients by a faster retrying of stale transmissions.
  • Remove OpKernelConstruction::op_def() as part of minimizing proto dependencies.
  • VectorLaplaceDiag distribution added.
  • Android demo no longer requires libtensorflow_demo.so to run (libtensorflow_inference.so still required)
  • Added categorical_column_with_vocabulary_file.
  • Introduce ops for batching/unbatching tensors across Session::Run() calls.
  • Add tf.log_sigmoid(x) = tf.log(tf.sigmoid(x)) = -tf.nn.softplus(-x).
  • Changed hooks lists to immutable tuples, and now allow any iterable for the associated arguments.
  • Introduce TFDecorator.
  • Added an Mfcc op for speech feature generation.
  • Improved DirectSession::Run() overhead and error checking. Feeding a value of the wrong type will now synchronously raise an INVALID_ARGUMENT error instead of asynchronously raising an INTERNAL error. Code that depends on the (undefined) behavior when feeding a tensor of the wrong type may need to be updated.
  • Added unreduced NONE, and reduced MEAN options for losses. Removed "WEIGHTED_" prefix from other Reduction constants.
  • assertAllClose now handles dicts.
  • Added Gmock matcher for HloInstructions.
  • Add var name to errors on variable restore.
  • Added an AudioSpectrogram op for audio feature generation.
  • Added reduction arg to losses.
  • tf.placeholder can represent scalar shapes and partially known.
  • Remove estimator_spec(mode) argument.
  • Added an AudioSpectrogram op for audio feature generation.
  • TensorBoard disables all runs by default if there are more than 40 runs.
  • Removed old doc generator code.
  • GCS file system integration now supports domain buckets, e.g gs://bucket.domain.com/path.
  • Add tf.summary.text for outputting text to TensorBoard.
  • The "run" command of tfdbg's command-line interface now supports filtering of tensors by node name, op type and tensor dtype.
  • tf.string_to_number now supports int64 and float64 outputs.

Thanks to our Contributors

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

4F2E4A2E, Aaron Schumacher, Abhi Agg, admcrae, Adriano Carmezim, Adrià Arrufat, agramesh1, Akimitsu Seo, Alan Mosca, Alex Egg, Alex Rothberg, Alexander Heinecke, Alexander Matyasko, Alexandr Baranezky, Alexandre Caulier, Ali Siddiqui, Anand Venkat, Andrew Hundt, Androbin, Anmol Sharma, Arie, Arno Leist, Arron Cao, AuréLien Geron, Bairen Yi, Beomsu Kim, Carl Thomé, cfperez, Changming Sun, Corey Wharton, critiqjo, Dalei Li, Daniel Rasmussen, Daniel Trebbien, DaríO Hereñú, David Eng, David Norman, David Y. Zhang, Davy Song, ddurham2, Deepak Subburam, Dmytro Kyrychuk, Dominic Rossi, Dominik SchlöSser, Dustin Tran, Eduardo Pinho, Egil Martinsson, Elliot Saba, Eric Bigelow, Erik Smistad, Evan Klitzke, Fabrizio Milo, Falcon Dai, Fei Gao, FloopCZ, Fung Lam, Gautam, GBLin5566, Greg Peatfield, Gu Wang, Guenther Schmuelling, Hans Pabst, Harun Gunaydin, Huaizheng, Ido Shamay, Ikaro Silva, Ilya Edrenkin, Immexxx, James Mishra, Jamie Cooke, Jay Young, Jayaram Bobba, Jianfei Wang, jinghua2, Joey Meyer, John Maidens, Jonghoon Jin, Julian Villella, Jun Kim, Jun Shi, Junwei Pan, jyegerlehner, Karan Desai, Karel Van De Plassche, Kb Sriram, KhabarlakKonstantin, Koan-Sin Tan, krivard, Kwotsin, Leandro Gracia Gil, Li Chen, Liangliang He, Louie Helm, lspvic, Luiz Henrique Soares, LáSzló Csomor, Mark Wong, Mathew Wicks, Matthew Rahtz, Maxwell Paul Brickner, Michael Hofmann, Miguel Flores Ruiz De Eguino, MikeTam1021, Mortada Mehyar, Mycosynth, Namnamseo, Nate Harada, Neven Miculinic, Nghia Tran, Nick Lyu, Niranjan Hasabnis, Nishidha, Oleksii Kuchaiev, Oyesh Mann Singh, Panmari, Patrick, Paul Van Eck, Piyush Chaudhary, Quim Llimona, Raingo, Richard Davies, Ruben Vereecken, Sahit Chintalapudi, Sam Abrahams, Santiago Castro, Scott Sievert, Sean O'Keefe, Sebastian Schlecht, Shane, Shubhankar Deshpande, Spencer Schaber, Sunyeop Lee, t13m, td2014, Thomas H. P. Andersen, Toby Petty, Umang Mehta, Vadim Markovtsev, Valentin Iovene, Vincent Zhao, Vit Stepanovs, Vivek Rane, Vu Pham, wannabesrevenge, weipingpku, wuhaixutab, wydwww, Xiang Gao, Xiaolin Lin, xiaoyaozhuzi, Yaroslav Bulatov, Yi Liu, Yoshihiro Sugi, Yuan (Terry) Tang, Yuming Wang, Yuxin Wu, Zader Zheng, Zhaojun Zhang, zhengjiajin, ZhipengShen, Ziming Dong, zjj2wry

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 1.1.0

Major Features and Improvements

  • Added Java API support for Windows.
  • Added tf.spectral module. Moved existing FFT ops to tf.spectral while keeping an alias in the old location (tf.*).
  • Added 1D, 2D and 3D Fourier transform ops for real signals to tf.spectral.
  • Added a tf.bincount function.
  • Added Keras 2 API to contrib.
  • Added a new lightweight queue-like object - RecordInput.
  • Added tf.contrib.image.compose_transforms function.
  • Bring tf.estimator.* into the API. Non-deprecated functionality from tf.contrib.learn.Estimator is moved to tf.estimator.Estimator with cosmetic changes.
  • Docker images: TF images on gcr.io and Docker Hub are upgraded to ubuntu:16.04.
  • Added the following features to TensorFlow Debugger (tfdbg):
    • Ability to inspect Python source file against TF ops and tensors (command print_source / ps)
    • New navigation bar in Curses-based UI
    • NodeStepper (command invoke_stepper) now uses intermediate tensor dumps. It also uses TensorHandles as direct feeds during successive cont calls for improved performance and reduced memory consumption.
  • Initial release of installation guides for Java, C, and Go.
  • Added Text Dashboard to TensorBoard.

Deprecations

  • TensorFlow 1.1.0 will be the last time we release a binary with Mac GPU support. Going forward, we will stop testing on Mac GPU systems. We continue to welcome patches that maintain Mac GPU support, and we will try to keep the Mac GPU build working.

Changes to contrib APIs

  • The behavior of RNNCells is now stricter due to the transition towards making RNNCells act more like Keras layers.
    • If an RNNCell is used twice in two different variable scopes, an error is raised describing how to avoid this behavior.
    • If an RNNCell is used in a variable scope with existing conflicting variables, an error is raised showing that the RNNCell must be constructed with argument reuse=True.
  • Deprecated contrib/distributions pmf, pdf, log_pmf, log_pdf.
  • Moved bayesflow.special_math to distributions.
  • tf.contrib.tensor_forest.python.tensor_forest.RandomForestDeviceAssigner removed.
  • Changed some MVN classes and parameters:
    • tf.contrib.distributions.MultivariateNormalFull replaced by tf.contrib.distributions.MultivariateNormalTriL.
    • tf.contrib.distributions.MultivariateNormalCholesky replaced by tf.contrib.distributions.MultivariateNormalTriL
    • tf.contrib.distributions.MultivariateNormalDiagWithSoftplusStDev replaced by tf.contrib.distributions.MultivariateNormalDiagWithSoftplusScale
    • tf.contrib.distributions.MultivariateNormalDiag arguments changed from mu, diag_stddev to log, scale_diag.
    • tf.contrib.distributions.MultivariateNormalDiagPlusVDVT removed.
    • tf.contrib.distributions.MultivariateNormalDiagPlusLowRank added.

Bug Fixes and Other Changes

  • Java: Support for loading models exported using the SavedModel API (courtesy @EronWright).
  • Go: Added support for incremental graph execution.
  • Fix a bug in the WALS solver when single-threaded.
  • Added support for integer sparse feature values in tf.contrib.layers.sparse_column_with_keys.
  • Fixed tf.set_random_seed(0) to be deterministic for all ops.
  • Stability improvements for the GCS file system support.
  • Improved TensorForest performance.
  • Added support for multiple filename globs in tf.matching_files.
  • LogMessage now includes a timestamp as beginning of a message.
  • Added MultiBox person detector example standalone binary.
  • Android demo: Makefile build functionality added to build.gradle to fully support building TensorFlow demo in Android on Windows.
  • Android demo: read MultiBox priors from txt file rather than protobuf.
  • Added colocation constraints to StagingArea.
  • sparse_matmul_op reenabled for Android builds.
  • Restrict weights rank to be the same as the broadcast target, to avoid ambiguity on broadcast rules.
  • Upgraded libxsmm to 1.7.1 and applied other changes for performance and memory usage.
  • Fixed bfloat16 integration of LIBXSMM sparse mat-mul.
  • Improved performance and reduce memory usage by allowing ops to forward input buffers to output buffers and perform computations in-place.
  • Improved the performance of CPU assignment for strings.
  • Speed up matrix * vector multiplication and matrix * matrix with unknown shapes.
  • C API: Graph imports now support input remapping, control dependencies, and returning imported nodes (see TF_GraphImportGraphDefWithReturnOutputs())
  • Multiple C++ API updates.
  • Multiple TensorBoard updates including:
    • Users can now view image summaries at various sampled steps (instead of just the last step).
    • Bugs involving switching runs as well as the image dashboard are fixed.
    • Removed data download links from TensorBoard.
    • TensorBoard uses a relative data directory, for easier embedding.
    • TensorBoard automatically ignores outliers for domain calculation, and formats proportional values consistently.
  • Multiple tfdbg bug fixes:
    • Fixed Windows compatibility issues.
    • Command history now persists across runs.
    • Bug fix in graph validation related to tf.while_loops.
  • Java Maven fixes for bugs with Windows installation.
  • Backport fixes and improvements from external keras.
  • Keras config file handling fix.

Thanks to our Contributors

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

A. Besir Kurtulmus, Adal Chiriliuc, @akash, Alec-Desouza, Alex Rothberg, Alex Sergeev, Alexander Heinecke, Allen Guo, Andreas Madsen, Ankesh Anand, Anton Loss, @Aravind, @Arie, Ashutosh Das, AuréLien Geron, Bairen Yi, @bakunyo, Ben Visser, Brady Zhou, Calpa Liu, Changming Sun, Chih Cheng Liang, Christopher Berner, Clark Zinzow, @Conchylicultor, Dan Ellis, Dan J, Dan Jarvis, Daniel Ylitalo, Darren Garvey, David Norman, David Truong, @DavidNorman, Dimitar Pavlov, Dmitry Persiyanov, @Eddie, @elirex, Erfan Noury, Eron Wright, Evgeny Mazovetskiy, Fabrizio (Misto) Milo, @fanlu, Fisher Coder, Florian Courtial, Franck Dernoncourt, Gagan Goel, Gao, Xiang, @Gautam, Gefu Tang, @guilherme, @guschmue, Hannah Provenza, Hans Pabst, @hartb, Hsiao Yi, Huazuo Gao, Igor ChorążEwicz, Ivan Smirnov, Jakub Kolodziejczyk, Jason Gavris, Jason Morton, Jay Young, Jayaram Bobba, Jeremy Sawruk, Jiaming Liu, Jihun Choi, @jiqiu, Joan Thibault, John C F, Jojy George Varghese, Jon Malmaud, Julian Berman, Julian Niedermeier, Junpeng Lao, Kai Sasaki, @Kankroc, Karl Lessard, Kyle Bostelmann, @Lezcano, Li Yi, Luo Yun, @lurker, Mahmoud-Abuzaina, Mandeep Singh, Marek Kolodziej, Mark Szepieniec, Martial Hue, Medhat Omr, Memo Akten, Michael Gharbi, MichaëL Defferrard, Milan Straka, @MircoT, @mlucool, Muammar Ibn Faisal, Nayana Thorat, @nghiattran, Nicholas Connor, Nikolaas Steenbergen, Niraj Patel, Niranjan Hasabnis, @Panmari, Pavel Bulanov, Philip Pries Henningsen, Philipp Jund, @polonez, Prayag Verma, Rahul Kavi, Raphael Gontijo Lopes, @rasbt, Raven Iqqe, Reid Pryzant, Richard Shin, Rizwan Asif, Russell Kaplan, Ryo Asakura, RüDiger Busche, Saisai Shao, Sam Abrahams, @sanosay, Sean Papay, @seaotterman, @selay01, Shaurya Sharma, Sriram Narayanamoorthy, Stefano Probst, @taknevski, @tbonza, @teldridge11, Tim Anglade, Tomas Reimers, Tomer Gafner, Valentin Iovene, Vamsi Sripathi, Viktor Malyi, Vit Stepanovs, Vivek Rane, Vlad Firoiu, @wangg12, @will, Xiaoyu Tao, Yaroslav Bulatov, Yi Liu, Yuan (Terry) Tang, @Yufeng, Yuming Wang, Yuxin Wu, Zafar Takhirov, Ziming Dong

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 1.0.1

Bug Fixes and Other Changes

  • Change GraphConstructor to not increase the version when importing, but instead take the min of all versions.
  • Google Cloud Storage fixes.
  • Removed tf.core and tf.python modules from the API. These were never intended to be exposed. Please use the same objects through top-level tf module instead.

Release 1.0.0

Major Features and Improvements

  • XLA (experimental): initial release of XLA, a domain-specific compiler for TensorFlow graphs, that targets CPUs and GPUs.
  • TensorFlow Debugger (tfdbg): command-line interface and API.
  • New python 3 docker images added.
  • Made pip packages pypi compliant. TensorFlow can now be installed by pip install tensorflow command.
  • Several python API calls have been changed to resemble NumPy more closely.
  • Android: person detection + tracking demo implementing Scalable Object Detection using Deep Neural Networks.
  • New (experimental) Java API.
  • Add new Android image stylization demo based on "A Learned Representation For Artistic Style", and add YOLO object detector support.

Breaking Changes to the API

To help you upgrade your existing TensorFlow Python code to match the API changes below, we have prepared a conversion script.

  • TensorFlow/models have been moved to a separate github repository.
  • Division and modulus operators (/, //, %) now match Python (flooring) semantics. This applies to tf.div and tf.mod as well. To obtain forced integer truncation based behaviors you can use tf.truncatediv and tf.truncatemod.
  • tf.divide() is now the recommended division function. tf.div() will remain, but its semantics do not respond to Python 3 or from future mechanisms.
  • tf.reverse() now takes indices of axes to be reversed. E.g. tf.reverse(a, [True, False, True]) must now be written as tf.reverse(a, [0, 2]). tf.reverse_v2() will remain until 1.0 final.
  • tf.mul, tf.sub and tf.neg are deprecated in favor of tf.multiply, tf.subtract and tf.negative.
  • tf.pack and tf.unpack are deprecated in favor of tf.stack and tf.unstack.
  • TensorArray.pack and TensorArray.unpack are getting deprecated in favor of TensorArray.stack and TensorArray.unstack.
  • The following Python functions have had their arguments changed to use axis when referring to specific dimensions. We have kept the old keyword arguments for compatibility currently, but we will be removing them well before the final 1.0.
    • tf.argmax: dimension becomes axis
    • tf.argmin: dimension becomes axis
    • tf.count_nonzero: reduction_indices becomes axis
    • tf.expand_dims: dim becomes axis
    • tf.reduce_all: reduction_indices becomes axis
    • tf.reduce_any: reduction_indices becomes axis
    • tf.reduce_join: reduction_indices becomes axis
    • tf.reduce_logsumexp: reduction_indices becomes axis
    • tf.reduce_max: reduction_indices becomes axis
    • tf.reduce_mean: reduction_indices becomes axis
    • tf.reduce_min: reduction_indices becomes axis
    • tf.reduce_prod: reduction_indices becomes axis
    • tf.reduce_sum: reduction_indices becomes axis
    • tf.reverse_sequence: batch_dim becomes batch_axis, seq_dim becomes seq_axis
    • tf.sparse_concat: concat_dim becomes axis
    • tf.sparse_reduce_sum: reduction_axes becomes axis
    • tf.sparse_reduce_sum_sparse: reduction_axes becomes axis
    • tf.sparse_split: split_dim becomes axis
  • tf.listdiff has been renamed to tf.setdiff1d to match NumPy naming.
  • tf.inv has been renamed to be tf.reciprocal (component-wise reciprocal) to avoid confusion with np.inv which is matrix inversion
  • tf.round now uses banker's rounding (round to even) semantics to match NumPy.
  • tf.split now takes arguments in a reversed order and with different keywords. In particular, we now match NumPy order as tf.split(value, num_or_size_splits, axis).
  • tf.sparse_split now takes arguments in reversed order and with different keywords. In particular we now match NumPy order as tf.sparse_split(sp_input, num_split, axis). NOTE: we have temporarily made tf.sparse_split require keyword arguments.
  • tf.concat now takes arguments in reversed order and with different keywords. In particular we now match NumPy order as tf.concat(values, axis, name).
  • tf.image.decode_jpeg by default uses the faster DCT method, sacrificing a little fidelity for improved speed. One can revert to the old behavior by specifying the attribute dct_method='INTEGER_ACCURATE'.
  • tf.complex_abs has been removed from the Python interface. tf.abs supports complex tensors and should be used instead.
  • In the C++ API (in tensorflow/cc), Input, Output, etc. have moved from the tensorflow::ops namespace to tensorflow.
  • Template.var_scope property renamed to .variable_scope
  • SyncReplicasOptimizer is removed and SyncReplicasOptimizerV2 renamed to SyncReplicasOptimizer.
  • tf.zeros_initializer() and tf.ones_initializer() now return a callable that must be called with initializer arguments, in your code replace tf.zeros_initializer with tf.zeros_initializer().
  • SparseTensor.shape has been renamed to SparseTensor.dense_shape. Same for SparseTensorValue.shape.
  • Replace tf.scalar_summary, tf.histogram_summary, tf.audio_summary, tf.image_summary with tf.summary.scalar, tf.summary.histogram, tf.summary.audio, tf.summary.image, respectively. The new summary ops take name rather than tag as their first argument, meaning summary ops now respect TensorFlow name scopes.
  • Replace tf.train.SummaryWriter and tf.train.SummaryWriterCache with tf.summary.FileWriter and tf.summary.FileWriterCache.
  • Removes RegisterShape from public API. Use C++ shape function registration instead.
  • Deprecated _ref dtypes from the python API.
  • In the C++ API (in tensorflow/cc), Input, Output, etc. have moved from the tensorflow::ops namespace to tensorflow.
  • Change arg order for {softmax,sparse_softmax,sigmoid}_cross_entropy_with_logits to be (labels, predictions), and force use of named args.
  • tf.nn.rnn_cell.* and most functions in tf.nn.rnn.* (with the exception of dynamic_rnn and raw_rnn) are temporarily in tf.contrib.rnn. They will be moved back into core for TF 1.2.
  • tf.nn.sampled_softmax_loss and tf.nn.nce_loss have both changed their API such that you need to switch the inputs, labels to labels, inputs parameters.
  • The shape keyword argument of the SparseTensor constructor changes its name to dense_shape between Tensorflow 0.12 and Tensorflow 1.0.

Bug Fixes and Other Changes

  • Numerous C++ API updates.
  • New op: parallel_stack.
  • Introducing common tf io compression options constants for RecordReader/RecordWriter.
  • Add sparse_column_with_vocabulary_file, to specify a feature column that transform string features to IDs, where the mapping is defined by a vocabulary file.
  • Added index_to_string_table which returns a lookup table that maps indices to strings.
  • Add string_to_index_table, which returns a lookup table that matches strings to indices.
  • Add a ParallelForWithWorkerId function.
  • Add string_to_index_table, which returns a lookup table that matches strings to indices.
  • Support restore session from checkpoint files in v2 in contrib/session_bundle.
  • Added a tf.contrib.image.rotate function for arbitrary angles.
  • Added tf.contrib.framework.filter_variables as a convenience function to filter lists of variables based on regular expressions.
  • make_template() takes an optional custom_getter_ param.
  • Added comment about how existing directories are handled by recursive_create_dir.
  • Added an op for QR factorizations.
  • Divides and mods in Python API now use flooring (Python) semantics.
  • Android: pre-built libs are now built nightly.
  • Android: cmake/gradle build for TensorFlow Inference library under contrib/android/cmake
  • Android: Much more robust Session initialization code.
  • Android: TF stats now exposed directly in demo and log when debug mode is active
  • Android: new/better README.md documentation
  • saved_model is available as tf.saved_model.
  • Empty op is now stateful.
  • Improve speed of scatter_update on the cpu for ASSIGN operations.
  • Change reduce_join to treat reduction_indices in the same way as other reduce_ ops.
  • Move TensorForestEstimator to contrib/tensor_forest.
  • Enable compiler optimizations by default and allow configuration in configure.
  • tf.divide now honors the name field.
  • Make metrics weight broadcasting more strict.
  • Add new queue-like StagingArea and new ops: stage and unstage.
  • Enable inplace update ops for strings on CPU. Speed up string concat.

Thanks to our Contributors

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

Aaron Hu, Abhishek Aggarwal, Adam Michael, Adriano Carmezim, @AfirSraftGarrier, Alexander Novikov, Alexander Rosenberg Johansen, Andrew Gibiansky, Andrew Hundt, Anish Shah, Anton Loss, @b0noI, @BoyuanJiang, Carl Thomé, Chad Kennedy, Comic Chang, Connor Braa, Daniel N. Lang, Daniel Trebbien, @danielgordon10, Darcy Liu, Darren Garvey, Dmitri Lapin, Eron Wright, Evan Cofer, Fabrizio Milo, Finbarr Timbers, Franck Dernoncourt, Garrett Smith, @guschmue, Hao Wei, Henrik Holst, Huazuo Gao, @Ian, @Issac, Jacob Israel, Jangsoo Park, Jin Kim, Jingtian Peng, John Pope, Kye Bostelmann, Liangliang He, Ling Zhang, Luheng He, Luke Iwanski, @lvli, Michael Basilyan, Mihir Patel, Mikalai Drabovich, Morten Just, @newge, Nick Butlin, Nishant Shukla, Pengfei Ni, Przemyslaw Tredak, @rasbt, @Ronny, Rudolf Rosa, @RustingSword, Sam Abrahams, Sam Putnam, @SeongAhJo, Shi Jiaxin, @skavulya, Steffen MüLler, @TheUSER123, @tiriplicamihai, @vhasanov, Victor Costan, Vit Stepanovs, Wangda Tan, Wenjian Huang, Xingdong Zuo, Yaroslav Bulatov, Yota Toyama, Yuan (Terry) Tang, Yuxin Wu

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 0.12.0

Major Features and Improvements

  • TensorFlow now builds and runs on Microsoft Windows (tested on Windows 10, Windows 7, and Windows Server 2016). Supported languages include Python (via a pip package) and C++. CUDA 8.0 and cuDNN 5.1 are supported for GPU acceleration. Known limitations include: It is not currently possible to load a custom op library. The GCS and HDFS file systems are not currently supported. The following ops are not currently implemented: Dequantize, QuantizeAndDequantize, QuantizedAvgPool, QuantizedBatchNomWithGlobalNormalization, QuantizedBiasAdd, QuantizedConcat, QuantizedConv2D, QuantizedMatmul, QuantizedMaxPool, QuantizeDownAndShrinkRange, QuantizedRelu, QuantizedRelu6, QuantizedReshape, QuantizeV2, RequantizationRange, and Requantize.
  • Go: Experimental API in Go to create and execute graphs (https://godoc.org/github.com/tensorflow/tensorflow/tensorflow/go)
  • New checkpoint format becomes the default in tf.train.Saver. Old V1 checkpoints continue to be readable; controlled by the write_version argument, tf.train.Saver now by default writes out in the new V2 format. It significantly reduces the peak memory required and latency incurred during restore.
  • Added a new library for library of matrix-free (iterative) solvers for linear equations, linear least-squares, eigenvalues and singular values in tensorflow/contrib/solvers. Initial version has lanczos bidiagonalization, conjugate gradients and CGLS.
  • Added gradients for matrix_solve_ls and self_adjoint_eig.
  • Large cleanup to add second order gradient for ops with C++ gradients and improve existing gradients such that most ops can now be differentiated multiple times.
  • Added a solver for ordinary differential equations, tf.contrib.integrate.odeint.
  • New contrib module for tensors with named axes, tf.contrib.labeled_tensor.
  • Visualization of embeddings in TensorBoard.

Breaking Changes to the API

  • BusAdjacency enum replaced with a protocol buffer DeviceLocality. PCI bus indexing now starts from 1 instead of 0, and bus_id==0 is used where previously BUS_ANY was used.
  • Env::FileExists and FileSystem::FileExists now return a tensorflow::Status instead of a bool. Any callers to this function can be converted to a bool by adding .ok() to the call.
  • The C API type TF_SessionWithGraph has been renamed to TF_Session, indicating its preferred use in language bindings for TensorFlow. What was previously TF_Session has been renamed to TF_DeprecatedSession.
  • Renamed TF_Port to TF_Output in the C API.
  • Removes RegisterShape from public API. Use C++ shape function registration instead. indexing now starts from 1 instead of 0, and bus_id==0 is used where previously BUS_ANY was used.
  • Most RNN cells and RNN functions now use different variable scopes to be consistent with layers (tf.contrib.layers). This means old checkpoints written using this code will not load after this change without providing Saver a list of variable renames. Examples of variable scope changes include RNN -> rnn in tf.nn.rnn, tf.nn.dynamic_rnn and moving from Linear/Matrix -> weights and Linear/Bias -> biases in most RNN cells.
  • Deprecated tf.select op. tf.where should be used instead.
  • SparseTensor.shape has been renamed to SparseTensor.dense_shape. Same for SparseTensorValue.shape.
  • Env::FileExists and FileSystem::FileExists now return a tensorflow::Status instead of a bool. Any callers to this function can be converted to a bool by adding .ok() to the call.
  • C API: Type TF_SessionWithGraph has been renamed to TF_Session, indicating its preferred use in language bindings for TensorFlow. What was previously TF_Session has been renamed to TF_DeprecatedSession.
  • C API: Renamed TF_Port to TF_Output.
  • C API: The caller retains ownership of TF_Tensor objects provided to TF_Run, TF_SessionRun, TF_SetAttrTensor etc.
  • Renamed tf.image.per_image_whitening() to tf.image.per_image_standardization()
  • Move Summary protobuf constructors to tf.summary submodule.
  • Deprecate histogram_summary, audio_summary, scalar_summary, image_summary, merge_summary, and merge_all_summaries.
  • Combined batch_* and regular version of linear algebra and FFT ops. The regular op now handles batches as well. All batch_* Python interfaces were removed.
  • tf.all_variables, tf.VARIABLES and tf.initialize_all_variables renamed to tf.global_variables, tf.GLOBAL_VARIABLES and tf.global_variables_initializer respectively.
  • tf.zeros_initializer() and tf.ones_initializer() now return a callable that must be called with initializer arguments, in your code replace tf.zeros_initializer with tf.zeros_initializer()

Bug Fixes and Other Changes

  • Use threadsafe version of lgamma function.
  • Fix tf.sqrt handling of negative arguments.
  • Fixed bug causing incorrect number of threads to be used for multi-threaded benchmarks.
  • Performance optimizations for batch_matmul on multi-core CPUs.
  • Improve trace, matrix_set_diag, matrix_diag_part and their gradients to work for rectangular matrices.
  • Support for SVD of complex valued matrices.

Thanks to our Contributors

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

@a7744hsc, Abhi Agg, @admcrae, Adriano Carmezim, Aki Sukegawa, Alex Kendall, Alexander Rosenberg Johansen, @amcrae, Amlan Kar, Andre Simpelo, Andreas Eberle, Andrew Hundt, Arnaud Lenglet, @b0noI, Balachander Ramachandran, Ben Barsdell, Ben Guidarelli, Benjamin Mularczyk, Burness Duan, @c0g, Changming Sun, @chanis, Corey Wharton, Dan J, Daniel Trebbien, Darren Garvey, David Brailovsky, David Jones, Di Zeng, @DjangoPeng, Dr. Kashif Rasul, @drag0, Fabrizio (Misto) Milo, FabríCio Ceschin, @fp, @Ghedeon, @guschmue, Gökçen Eraslan, Haosdent Huang, Haroen Viaene, Harold Cooper, Henrik Holst, @hoangmit, Ivan Ukhov, Javier Dehesa, Jingtian Peng, Jithin Odattu, Joan Pastor, Johan Mathe, Johannes Mayer, Jongwook Choi, Justus Schwabedal, Kai Wolf, Kamil Hryniewicz, Kamran Amini, Karen Brems, Karl Lattimer, @kborer, Ken Shirriff, Kevin Rose, Larissa Laich, Laurent Mazare, Leonard Lee, Liang-Chi Hsieh, Liangliang He, Luke Iwanski, Marek Kolodziej, Moustafa Alzantot, @MrQianjinsi, @nagachika, Neil Han, Nick Meehan, Niels Ole Salscheider, Nikhil Mishra, @nschuc, Ondrej Skopek, OndřEj Filip, @OscarDPan, Pablo Moyano, Przemyslaw Tredak, @qitaishui, @Quarazy, @raix852, Philipp Helo, Sam Abrahams, @SriramRamesh, Till Hoffmann, Tushar Soni, @tvn, @tyfkda, Uwe Schmidt, Victor Villas, Vit Stepanovs, Vladislav Gubarev, @wujingyue, Xuesong Yang, Yi Liu, Yilei Yang, @youyou3, Yuan (Terry) Tang, Yuming Wang, Zafar Takhirov, @zhongyuk, Ziming Dong, @guotong1988

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 0.11.0

Major Features and Improvements

  • CUDA 8 support.
  • cuDNN 5 support.
  • HDFS Support.
  • Adds Fused LSTM support via cuDNN 5 in tensorflow/contrib/cudnn_rnn.
  • Improved support for NumPy style basic slicing including non-1 strides, ellipses, newaxis, and negative indices. For example complicated expressions like foo[1, 2:4, tf.newaxis, ..., :-3:-1, :] are now supported. In addition we have preliminary (non-broadcasting) support for sliced assignment to variables. In particular one can write var[1:3].assign([1,11,111]).
  • Deprecated tf.op_scope and tf.variable_op_scope in favor of a unified tf.name_scope and tf.variable_scope. The new argument order of tf.variable_scope is incompatible with previous versions.
  • Introducing core/util/tensor_bundle module: a module to efficiently serialize/deserialize tensors to disk. Will be used in TF's new checkpoint format.
  • Added tf.svd for computing the singular value decomposition (SVD) of dense matrices or batches of matrices (CPU only).
  • Added gradients for eigenvalues and eigenvectors computed using self_adjoint_eig or self_adjoint_eigvals.
  • Eliminated batch_* methods for most linear algebra and FFT ops and promoted the non-batch version of the ops to handle batches of matrices.
  • Tracing/timeline support for distributed runtime (no GPU profiler yet).
  • C API gives access to inferred shapes with TF_GraphGetTensorNumDims and TF_GraphGetTensorShape.
  • Shape functions for core ops have moved to C++ via REGISTER_OP(...).SetShapeFn(...). Python shape inference RegisterShape calls use the C++ shape functions with common_shapes.call_cpp_shape_fn. A future release will remove RegisterShape from python.

Bug Fixes and Other Changes

  • Documentation now includes operator overloads on Tensor and Variable.
  • tensorflow.__git_version__ now allows users to identify the version of the code that TensorFlow was compiled with. We also have tensorflow.__git_compiler__ which identifies the compiler used to compile TensorFlow's core.
  • Improved multi-threaded performance of batch_matmul.
  • LSTMCell, BasicLSTMCell, and MultiRNNCell constructors now default to state_is_tuple=True. For a quick fix while transitioning to the new default, simply pass the argument state_is_tuple=False.
  • DeviceFactory's AddDevices and CreateDevices functions now return a Status instead of void.
  • Int32 elements of list(type) arguments are no longer placed in host memory by default. If necessary, a list(type) argument to a kernel can be placed in host memory using a HostMemory annotation.
  • uniform_unit_scaling_initializer() no longer takes a full_shape arg, instead relying on the partition info passed to the initializer function when it's called.
  • The NodeDef protocol message is now defined in its own file node_def.proto instead of graph.proto.
  • ops.NoGradient was renamed ops.NotDifferentiable. ops.NoGradient will be removed soon.
  • dot.h / DotGraph was removed (it was an early analysis tool prior to TensorBoard, no longer that useful). It remains in history should someone find the code useful.
  • re2 / regexp.h was removed from being a public interface of TF. Should users need regular expressions, they should depend on the RE2 library directly rather than via TensorFlow.

Thanks to our Contributors

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

Abid K, @afshinrahimi, @AidanGG, Ajay Rao, Aki Sukegawa, Alex Rothberg, Alexander Rosenberg Johansen, Andrew Gibiansky, Andrew Thomas, @Appleholic, Bastiaan Quast, Ben Dilday, Bofu Chen, Brandon Amos, Bryon Gloden, Cissp®, @chanis, Chenyang Liu, Corey Wharton, Daeyun Shin, Daniel Julius Lasiman, Daniel Waterworth, Danijar Hafner, Darren Garvey, Denis Gorbachev, @DjangoPeng, Egor-Krivov, Elia Palme, Eric Platon, Fabrizio Milo, Gaetan Semet, Georg Nebehay, Gu Wang, Gustav Larsson, @haosdent, Harold Cooper, Hw-Zz, @ichuang, Igor Babuschkin, Igor Macedo Quintanilha, Ilya Edrenkin, @ironhead, Jakub Kolodziejczyk, Jennifer Guo, Jihun Choi, Jonas Rauber, Josh Bleecher Snyder, @jpangburn, Jules Gagnon-Marchand, Karen Brems, @kborer, Kirill Bobyrev, Laurent Mazare, Longqi Yang, Malith Yapa, Maniteja Nandana, Martin Englund, Matthias Winkelmann, @mecab, Mu-Ik Jeon, Nand Dalal, Niels Ole Salscheider, Nikhil Mishra, Park Jiin, Pieter De Rijk, @raix852, Ritwik Gupta, Sahil Sharma, Sangheum Hwang, @SergejsRk, Shinichiro Hamaji, Simon Denel, @Steve, @suiyuan2009, Tiago Jorge, Tijmen Tieleman, @tvn, @tyfkda, Wang Yang, Wei-Ting Kuo, Wenjian Huang, Yan Chen, @YenChenLin, Yuan (Terry) Tang, Yuncheng Li, Yunfeng Wang, Zack Polizzi, @zhongzyd, Ziming Dong, @perhapszzy

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 0.10.0

Major Features and Improvements

  • Added support for C++ shape inference
  • Added graph-construction C API
  • Major revision to the graph-construction C++ API
  • Support makefile build for iOS
  • Added Mac GPU support
  • Full version of TF-Slim available as tf.contrib.slim
  • Added k-Means clustering and WALS matrix factorization

Bug Fixes and Other Changes

  • Allow gradient computation for scalar values.
  • Performance improvements for gRPC
  • Improved support for fp16
  • New high-level ops in tf.contrib.{layers,metrics}
  • New features for TensorBoard, such as shape display, exponential smoothing
  • Faster and more stable Google Cloud Storage (GCS) filesystem support
  • Support for zlib compression and decompression for TFRecordReader and TFRecordWriter
  • Support for reading (animated) GIFs
  • Improved support for SparseTensor
  • Added support for more probability distributions (Dirichlet, Beta, Bernoulli, etc.)
  • Added Python interfaces to reset resource containers.
  • Many bugfixes and performance improvements
  • Many documentation fixes

Thanks to our Contributors

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

Alex Rothberg, Andrew Royer, Austin Marshall, @BlackCoal, Bob Adolf, Brian Diesel, Charles-Emmanuel Dias, @chemelnucfin, Chris Lesniewski, Daeyun Shin, Daniel Rodriguez, Danijar Hafner, Darcy Liu, Kristinn R. Thórisson, Daniel Castro, Dmitry Savintsev, Kashif Rasul, Dylan Paiton, Emmanuel T. Odeke, Ernest Grzybowski, Gavin Sherry, Gideon Dresdner, Gregory King, Harold Cooper, @heinzbeinz, Henry Saputra, Huarong Huo, Huazuo Gao, Igor Babuschkin, Igor Macedo Quintanilha, Ivan Ukhov, James Fysh, Jan Wilken Dörrie, Jihun Choi, Johnny Lim, Jonathan Raiman, Justin Francis, @lilac, Li Yi, Marc Khoury, Marco Marchesi, Max Melnick, Micael Carvalho, @mikowals, Mostafa Gazar, Nico Galoppo, Nishant Agrawal, Petr Janda, Yuncheng Li, @raix852, Robert Rose, @Robin-des-Bois, Rohit Girdhar, Sam Abrahams, satok16, Sergey Kishchenko, Sharkd Tu, @shotat, Siddharth Agrawal, Simon Denel, @sono-bfio, SunYeop Lee, Thijs Vogels, @tobegit3hub, @Undo1, Wang Yang, Wenjian Huang, Yaroslav Bulatov, Yuan Tang, Yunfeng Wang, Ziming Dong

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 0.9.0

Major Features and Improvements

  • Python 3.5 support and binaries
  • Added iOS support
  • Added support for processing on GPUs on MacOS
  • Added makefile for better cross-platform build support (C API only)
  • fp16 support and improved complex128 support for many ops
  • Higher level functionality in contrib.{layers,losses,metrics,learn}
  • More features to Tensorboard
  • Improved support for string embedding and sparse features
  • The RNN api is finally "official" (see, e.g., tf.nn.dynamic_rnn, tf.nn.rnn, and the classes in tf.nn.rnn_cell).
  • TensorBoard now has an Audio Dashboard, with associated audio summaries.

Bug Fixes and Other Changes

  • Turned on CuDNN Autotune.
  • Added support for using third-party Python optimization algorithms (contrib.opt).
  • Google Cloud Storage filesystem support.
  • HDF5 support
  • Add support for 3d convolutions and pooling.
  • Update gRPC release to 0.14.
  • Eigen version upgrade.
  • Switch to eigen thread pool
  • tf.nn.moments() now accepts a shift argument. Shifting by a good estimate of the mean improves numerical stability. Also changes the behavior of the shift argument to tf.nn.sufficient_statistics().
  • Performance improvements
  • Many bugfixes
  • Many documentation fixes
  • TensorBoard fixes: graphs with only one data point, Nan values, reload button and auto-reload, tooltips in scalar charts, run filtering, stable colors
  • Tensorboard graph visualizer now supports run metadata. Clicking on nodes while viewing a stats for a particular run will show runtime statistics, such as memory or compute usage. Unused nodes will be faded out.

Thanks to our Contributors

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

Aaron Schumacher, Aidan Dang, Akihiko ITOH, Aki Sukegawa, Arbit Chen, Aziz Alto, Danijar Hafner, Erik Erwitt, Fabrizio Milo, Felix Maximilian Möller, Henry Saputra, Sung Kim, Igor Babuschkin, Jan Zikes, Jeremy Barnes, Jesper Steen Møller, Johannes Mayer, Justin Harris, Kashif Rasul, Kevin Robinson, Loo Rong Jie, Lucas Moura, Łukasz Bieniasz-Krzywiec, Mario Cho, Maxim Grechkin, Michael Heilman, Mostafa Rahmani, Mourad Mourafiq, @ninotoshi, Orion Reblitz-Richardson, Yuncheng Li, @raoqiyu, Robert DiPietro, Sam Abrahams, Sebastian Raschka, Siddharth Agrawal, @snakecharmer1024, Stephen Roller, Sung Kim, SunYeop Lee, Thijs Vogels, Till Hoffmann, Victor Melo, Ville Kallioniemi, Waleed Abdulla, Wenjian Huang, Yaroslav Bulatov, Yeison Rodriguez, Yuan Tang, Yuxin Wu, @zhongzyd, Ziming Dong, Zohar Jackson

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 0.8.0

Major Features and Improvements

  • Added a distributed runtime using GRPC
  • Move skflow to contrib/learn
  • Better linear optimizer in contrib/linear_optimizer
  • Random forest implementation in contrib/tensor_forest
  • CTC loss and decoders in contrib/ctc
  • Basic support for half data type
  • Better support for loading user ops (see examples in contrib/)
  • Allow use of (non-blocking) Eigen threadpool with TENSORFLOW_USE_EIGEN_THREADPOOL define
  • Add an extension mechanism for adding network file system support
  • TensorBoard displays metadata stats (running time, memory usage and device used) and tensor shapes

Bug Fixes and Other Changes

  • Utility for inspecting checkpoints
  • Basic tracing and timeline support
  • Allow building against cuDNN 5 (not incl. RNN/LSTM support)
  • Added instructions and binaries for ProtoBuf library with fast serialization and without 64MB limit
  • Added special functions
  • bool-strictness: Tensors have to be explicitly compared to None
  • Shape strictness: all fed values must have a shape that is compatible with the tensor they are replacing
  • Exposed tf.while_loop (deprecated control_flow_ops.While)
  • run() now takes RunOptions and RunMetadata, which enable timing stats
  • Fixed lots of potential overflow problems in op kernels
  • Various performance improvements, especially for RNNs and convolutions
  • Many bugfixes
  • Nightly builds, tutorial tests, many test improvements
  • New examples: transfer learning and deepdream ipython notebook
  • Added tutorials, many documentation fixes.

Thanks to our Contributors

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

Abhinav Upadhyay, Aggelos Avgerinos, Alan Wu, Alexander G. de G. Matthews, Aleksandr Yahnev, @amchercashin, Andy Kitchen, Aurelien Geron, Awni Hannun, @BanditCat, Bas Veeling, Cameron Chen, @cg31, Cheng-Lung Sung, Christopher Bonnett, Dan Becker, Dan Van Boxel, Daniel Golden, Danijar Hafner, Danny Goodman, Dave Decker, David Dao, David Kretch, Dongjoon Hyun, Dustin Dorroh, @e-lin, Eurico Doirado, Erik Erwitt, Fabrizio Milo, @gaohuazuo, Iblis Lin, Igor Babuschkin, Isaac Hodes, Isaac Turner, Iván Vallés, J Yegerlehner, Jack Zhang, James Wexler, Jan Zikes, Jay Young, Jeff Hodges, @jmtatsch, Johnny Lim, Jonas Meinertz Hansen, Kanit Wongsuphasawat, Kashif Rasul, Ken Shirriff, Kenneth Mitchner, Kenta Yonekura, Konrad Magnusson, Konstantin Lopuhin, @lahwran, @lekaha, @liyongsea, Lucas Adams, @makseq, Mandeep Singh, @manipopopo, Mark Amery, Memo Akten, Michael Heilman, Michael Peteuil, Nathan Daly, Nicolas Fauchereau, @ninotoshi, Olav Nymoen, @panmari, @papelita1234, Pedro Lopes, Pranav Sailesh Mani, RJ Ryan, Rob Culliton, Robert DiPietro, @ronrest, Sam Abrahams, Sarath Shekkizhar, Scott Graham, Sebastian Raschka, Sung Kim, Surya Bhupatiraju, Syed Ahmed, Till Hoffmann, @timsl, @urimend, @vesnica, Vlad Frolov, Vlad Zagorodniy, Wei-Ting Kuo, Wenjian Huang, William Dmitri Breaden Madden, Wladimir Schmidt, Yuan Tang, Yuwen Yan, Yuxin Wu, Yuya Kusakabe, @zhongzyd, @znah.

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 0.7.1

Bug Fixes and Other Changes

  • Added gfile.Open and gfile.Copy, used by input_data.py.
  • Fixed Saver bug when MakeDirs tried to create empty directory.
  • GPU Pip wheels are built with cuda 7.5 and cudnn-v4, making them required for the binary releases. Lower versions of cuda/cudnn can be supported by installing from sources and setting the options during ./configure
  • Fix dataset encoding example for Python3 (@danijar)
  • Fix PIP installation by not packaging protobuf as part of wheel, require protobuf 3.0.0b2.
  • Fix Mac pip installation of numpy by requiring pip >= 1.10.1.
  • Improvements and fixes to Docker image.

Release 0.7.0

Major Features and Improvements

  • Allow using any installed Cuda >= 7.0 and cuDNN >= R2, and add support for cuDNN R4
  • Added a contrib/ directory for unsupported or experimental features, including higher level layers module
  • Added an easy way to add and dynamically load user-defined ops
  • Built out a good suite of tests, things should break less!
  • Added MetaGraphDef which makes it easier to save graphs with metadata
  • Added assignments for "Deep Learning with TensorFlow" udacity course

Bug Fixes and Other Changes

  • Added a versioning framework for GraphDefs to ensure compatibility
  • Enforced Python 3 compatibility
  • Internal changes now show up as sensibly separated commits
  • Open-sourced the doc generator
  • Un-fork Eigen
  • Simplified the BUILD files and cleaned up C++ headers
  • TensorFlow can now be used as a submodule in another bazel build
  • New ops (e.g., *fft, *_matrix_solve)
  • Support for more data types in many ops
  • Performance improvements
  • Various bugfixes
  • Documentation fixes and improvements

Breaking Changes to the API

  • AdjustContrast kernel deprecated, new kernel AdjustContrastv2 takes and outputs float only. adjust_contrast now takes all data types.
  • adjust_brightness's delta argument is now always assumed to be in [0,1] (as is the norm for images in floating point formats), independent of the data type of the input image.
  • The image processing ops do not take min and max inputs any more, casting safety is handled by saturate_cast, which makes sure over- and underflows are handled before casting to data types with smaller ranges.
  • For C++ API users: IsLegacyScalar and IsLegacyVector are now gone from TensorShapeUtils since TensorFlow is scalar strict within Google (for example, the shape argument to tf.reshape can't be a scalar anymore). The open source release was already scalar strict, so outside Google IsScalar and IsVector are exact replacements.
  • The following files are being removed from tensorflow/core/public/:
    • env.h -> ../platform/env.h
    • status.h -> ../lib/core/status.h
    • tensor.h -> ../framework/tensor.h
    • tensor_shape.h -> ../framework/tensor_shape.h
    • partial_tensor_shape.h -> ../framework/partial_tensor_shape.h
    • tensorflow_server.h deleted
  • For C++ API users: TensorShape::ShortDebugString has been renamed to DebugString, and the previous DebugString behavior is gone (it was needlessly verbose and produced a confusing empty string for scalars).
  • GraphOptions.skip_common_subexpression_elimination has been removed. All graph optimizer options are now specified via GraphOptions.OptimizerOptions.
  • ASSERT_OK / EXPECT_OK macros conflicted with external projects, so they were renamed TF_ASSERT_OK, TF_EXPECT_OK. The existing macros are currently maintained for short-term compatibility but will be removed.
  • The non-public nn.rnn and the various nn.seq2seq methods now return just the final state instead of the list of all states.
  • tf.scatter_update now no longer guarantees that lexicographically largest index be used for update when duplicate entries exist.
  • tf.image.random_crop(image, [height, width]) is now tf.random_crop(image, [height, width, depth]), and tf.random_crop works for any rank (not just 3-D images). The C++ RandomCrop op has been replaced with pure Python.
  • Renamed tf.test.GetTempDir and tf.test.IsBuiltWithCuda to tf.test.get_temp_dir and tf.test.is_built_with_cuda for PEP-8 compatibility.
  • parse_example's interface has changed, the old interface is accessible in legacy_parse_example (same for related functions).
  • New Variables are not added to the same collection several times even if a list with duplicates is passed to the constructor.
  • The Python API will now properly set the list member of AttrValue in constructed GraphDef messages for empty lists. The serialization of some graphs will change, but the change is both forwards and backwards compatible. It will break tests that compare a generated GraphDef to a golden serialized GraphDef (which is discouraged).

Thanks to our Contributors

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

Akiomi Kamakura, Alex Vig, Alexander Rosenberg Johansen, Andre Cruz, Arun Ahuja, Bart Coppens, Bernardo Pires, Carl Vondrick, Cesar Salgado, Chen Yu, Christian Jauvin, Damien Aymeric, Dan Vanderkam, Denny Britz, Dongjoon Hyun, Eren Güven, Erik Erwitt, Fabrizio Milo, G. Hussain Chinoy, Jim Fleming, Joao Felipe Santos, Jonas Meinertz Hansen, Joshi Rekha, Julian Viereck, Keiji Ariyama, Kenton Lee, Krishna Sankar, Kristina Chodorow, Linchao Zhu, Lukas Krecan, Mark Borgerding, Mark Daoust, Moussa Taifi, Nathan Howell, Naveen Sundar Govindarajulu, Nick Sweeting, Niklas Riekenbrauck, Olivier Grisel, Patrick Christ, Povilas Liubauskas, Rainer Wasserfuhr, Romain Thouvenin, Sagan Bolliger, Sam Abrahams, Taehoon Kim, Timothy J Laurent, Vlad Zavidovych, Yangqing Jia, Yi-Lin Juang, Yuxin Wu, Zachary Lipton, Zero Chen, Alan Wu, @brchiu, @emmjaykay, @jalammar, @Mandar-Shinde, @nsipplswezey, @ninotoshi, @panmari, @prolearner and @rizzomichaelg.

We are also grateful to all who filed issues or helped resolve them, asked and answered questions, and were part of inspiring discussions.

Release 0.6.0

Major Features and Improvements

  • Python 3.3+ support via changes to python codebase and ability to specify python version via ./configure.

  • Some improvements to GPU performance and memory usage: convnet benchmarks roughly equivalent with native cudnn v2 performance. Improvements mostly due to moving to 32-bit indices, faster shuffling kernels. More improvements to come in later releases.

Bug Fixes

  • Lots of fixes to documentation and tutorials, many contributed by the public.

  • 271 closed issues on github issues.

Backwards-Incompatible Changes

  • tf.nn.fixed_unigram_candidate_sampler changed its default 'distortion' attribute from 0.0 to 1.0. This was a bug in the original release that is now fixed.

Release 0.5.0

Initial release of TensorFlow.