1# Release 2.0.1 2 3## Bug Fixes and Other Changes 4* Fixes a security vulnerability where converting a Python string to a `tf.float16` value produces a segmentation fault ([CVE-2020-5215](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2020-5215)) 5* Updates `curl` to `7.66.0` to handle [CVE-2019-5482](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-5482) and [CVE-2019-5481](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-5481) 6* Updates `sqlite3` to `3.30.01` to handle [CVE-2019-19646](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-19646), [CVE-2019-19645](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-19645) and [CVE-2019-16168](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-16168) 7 8 9# Release 1.15.2 10 11## Bug Fixes and Other Changes 12* Fixes a security vulnerability where converting a Python string to a `tf.float16` value produces a segmentation fault ([CVE-2020-5215](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2020-5215)) 13* Updates `curl` to `7.66.0` to handle [CVE-2019-5482](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-5482) and [CVE-2019-5481](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-5481) 14* Updates `sqlite3` to `3.30.01` to handle [CVE-2019-19646](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-19646), [CVE-2019-19645](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-19645) and [CVE-2019-16168](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2019-16168) 15 16 17# Release 2.1.0 18 19TensorFlow 2.1 will be the last TF release supporting Python 2. Python 2 support [officially ends an January 1, 2020](https://www.python.org/dev/peps/pep-0373/#update). [As announced earlier](https://groups.google.com/a/tensorflow.org/d/msg/announce/gVwS5RC8mds/dCt1ka2XAAAJ), TensorFlow will also stop supporting Python 2 starting January 1, 2020, and no more releases are expected in 2019. 20 21## Major Features and Improvements 22* 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. 23* **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](https://support.microsoft.com/help/2977003/the-latest-supported-visual-c-downloads). 24 * 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`. 25 * 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. 26* The `tensorflow` pip package is built with CUDA 10.1 and cuDNN 7.6. 27* `tf.keras` 28 * Experimental support for mixed precision is available on GPUs and Cloud TPUs. See [usage guide](https://www.tensorflow.org/guide/keras/mixed_precision). 29 * 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](https://colab.research.google.com/drive/1RvCnR7h0_l4Ekn5vINWToI9TNJdpUZB3). 30 * 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. 31 * 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). 32 * Automatic outside compilation is now enabled for Cloud TPUs. This allows `tf.summary` to be used more conveniently with Cloud TPUs. 33 * Dynamic batch sizes with DistributionStrategy and Keras are supported on Cloud TPUs. 34 * Support for `.fit`, `.evaluate`, `.predict` on TPU using numpy data, in addition to `tf.data.Dataset`. 35 * Keras reference implementations for many popular models are available in the TensorFlow [Model Garden](https://github.com/tensorflow/models/tree/master/official). 36* `tf.data` 37 * 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. 38 * `tf.data.Dataset` now supports automatic data distribution and sharding in distributed environments, including on TPU pods. 39 * 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)` 40* `tf.debugging` 41 * Add `tf.debugging.enable_check_numerics()` and `tf.debugging.disable_check_numerics()` to help debugging the root causes of issues involving infinities and `NaN`s. 42* `tf.distribute` 43 * Custom training loop support on TPUs and TPU pods is avaiable through `strategy.experimental_distribute_dataset`, `strategy.experimental_distribute_datasets_from_function`, `strategy.experimental_run_v2`, `strategy.reduce`. 44 * Support for a global distribution strategy through `tf.distribute.experimental_set_strategy(),` in addition to `strategy.scope()`. 45* `TensorRT` 46 * [TensorRT 6.0](https://developer.nvidia.com/tensorrt#tensorrt-whats-new) 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`. 47* 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. 48 49## Breaking Changes 50* Deletes `Operation.traceback_with_start_lines` for which we know of no usages. 51* Removed `id` from `tf.Tensor.__repr__()` as `id` is not useful other than internal debugging. 52* 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). 53* 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`. 54* `tf.config.experimentalVirtualDeviceConfiguration` has been renamed to `tf.config.LogicalDeviceConfiguration`. 55* `tf.config.experimental_list_devices` has been removed, please use 56`tf.config.list_logical_devices`. 57 58## Bug Fixes and Other Changes 59* `tf.data` 60 * Fixes concurrency issue with `tf.data.experimental.parallel_interleave` with `sloppy=True`. 61 * Add `tf.data.experimental.dense_to_ragged_batch()`. 62 * Extend `tf.data` parsing ops to support `RaggedTensors`. 63* `tf.distribute` 64 * Fix issue where GRU would crash or give incorrect output when a `tf.distribute.Strategy` was used. 65* `tf.estimator` 66 * Added option in `tf.estimator.CheckpointSaverHook` to not save the `GraphDef`. 67 * Moving the checkpoint reader from swig to pybind11. 68* `tf.keras` 69 * Export `depthwise_conv2d` in `tf.keras.backend`. 70 * In Keras Layers and Models, Variables in `trainable_weights`, `non_trainable_weights`, and `weights` are explicitly deduplicated. 71 * 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`. 72 * Fix the input shape caching behavior of Keras convolutional layers. 73 * `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. 74* `tf.lite` 75 * Legalization for `NMS` ops in TFLite. 76 * add `narrow_range` and `axis` to `quantize_v2` and `dequantize` ops. 77 * Added support for `FusedBatchNormV3` in converter. 78 * Add an `errno`-like field to `NNAPI` delegate for detecting `NNAPI` errors for fallback behaviour. 79 * Refactors `NNAPI` Delegate to support detailed reason why an operation is not accelerated. 80 * Converts hardswish subgraphs into atomic ops. 81* Other 82 * Critical stability updates for TPUs, especially in cases where the XLA compiler produces compilation errors. 83 * TPUs can now be re-initialized multiple times, using `tf.tpu.experimental.initialize_tpu_system`. 84 * Add `RaggedTensor.merge_dims()`. 85 * Added new `uniform_row_length` row-partitioning tensor to `RaggedTensor`. 86 * Add `shape` arg to `RaggedTensor.to_tensor`; Improve speed of `RaggedTensor.to_tensor`. 87 * `tf.io.parse_sequence_example` and `tf.io.parse_single_sequence_example` now support ragged features. 88 * Fix `while_v2` with variables in custom gradient. 89 * Support taking gradients of V2 `tf.cond` and `tf.while_loop` using `LookupTable`. 90 * Fix bug where `vectorized_map` failed on inputs with unknown static shape. 91 * Add preliminary support for sparse CSR matrices. 92 * Tensor equality with `None` now behaves as expected. 93 * Make calls to `tf.function(f)()`, `tf.function(f).get_concrete_function` and `tf.function(f).get_initialization_function` thread-safe. 94 * Extend `tf.identity` to work with CompositeTensors (such as SparseTensor) 95 * Added more `dtypes` and zero-sized inputs to `Einsum` Op and improved its performance 96 * Enable multi-worker `NCCL` `all-reduce` inside functions executing eagerly. 97 * Added complex128 support to `RFFT`, `RFFT2D`, `RFFT3D`, `IRFFT`, `IRFFT2D`, and `IRFFT3D`. 98 * Add `pfor` converter for `SelfAdjointEigV2`. 99 * Add `tf.math.ndtri` and `tf.math.erfinv`. 100 * Add `tf.config.experimental.enable_mlir_bridge` to allow using MLIR compiler bridge in eager model. 101 * Added support for MatrixSolve on Cloud TPU / XLA. 102 * Added `tf.autodiff.ForwardAccumulator` for forward-mode autodiff 103 * Add `LinearOperatorPermutation`. 104 * A few performance optimizations on `tf.reduce_logsumexp`. 105 * Added multilabel handling to `AUC` metric 106 * Optimization on `zeros_like`. 107 * Dimension constructor now requires `None` or types with an `__index__` method. 108 * Add `tf.random.uniform` microbenchmark. 109 * Use `_protogen` suffix for proto library targets instead of `_cc_protogen` suffix. 110 * Moving the checkpoint reader from `swig` to `pybind11`. 111 * `tf.device` & `MirroredStrategy` now supports passing in a `tf.config.LogicalDevice` 112 * If you're building Tensorflow from source, consider using [bazelisk](https://github.com/bazelbuild/bazelisk) to automatically download and use the correct Bazel version. Bazelisk reads the `.bazelversion` file at the root of the project directory. 113 114## Thanks to our Contributors 115 116This release contains contributions from many people at Google, as well as: 117 1188bitmp3, 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 119 120# Release 1.15.0 121This 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. 122 123## Major Features and Improvements 124* As [announced](https://groups.google.com/a/tensorflow.org/forum/#!topic/developers/iRCt5m4qUz0), `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. 125* 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. 126This 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. 127* EagerTensor now supports numpy buffer interface for tensors. 128* Add toggles `tf.enable_control_flow_v2()` and `tf.disable_control_flow_v2()` for enabling/disabling v2 control flow. 129* Enable v2 control flow as part of `tf.enable_v2_behavior()` and `TF2_BEHAVIOR=1`. 130* 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. 131* Adds `enable_tensor_equality()`, which switches the behavior such that: 132 * Tensors are no longer hashable. 133 * Tensors can be compared with `==` and `!=`, yielding a Boolean Tensor with element-wise comparison results. This will be the default behavior in 2.0. 134 135## Breaking Changes 136* 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. 137* 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. 138* Deprecated the use of `constraint=` and `.constraint` with ResourceVariable. 139* `tf.keras`: 140 * `OMP_NUM_THREADS` is no longer used by the default Keras config. To configure the number of threads, use `tf.config.threading` APIs. 141 * `tf.keras.model.save_model` and `model.save` now defaults to saving a TensorFlow SavedModel. 142 * `keras.backend.resize_images` (and consequently, `keras.layers.Upsampling2D`) behavior has changed, a bug in the resizing implementation was fixed. 143 * 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. 144 * 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). 145 146## Bug Fixes and Other Changes 147* `tf.estimator`: 148 * `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`. 149 * Fix tests in canned estimators. 150 * Expose Head as public API. 151 * Fixes critical bugs that help with `DenseFeatures` usability in TF2 152* `tf.data`: 153 * Promoting `unbatch` from experimental to core API. 154 * Adding support for datasets as inputs to `from_tensors` and `from_tensor_slices` and batching and unbatching of nested datasets. 155* `tf.keras`: 156 * `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`. 157 * Saving a Keras Model using `tf.saved_model.save` now saves the list of variables, trainable variables, regularization losses, and the call function. 158 * 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. 159 * 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. 160 * 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. 161 * Raise error if `batch_size` argument is used when input is dataset/generator/keras sequence. 162* `tf.lite` 163 * Add `GATHER` support to NN API delegate. 164 * tflite object detection script has a debug mode. 165 * Add delegate support for `QUANTIZE`. 166 * Added evaluation script for COCO minival. 167 * Add delegate support for `QUANTIZED_16BIT_LSTM`. 168 * Converts hardswish subgraphs into atomic ops. 169* Add support for defaulting the value of `cycle_length` argument of `tf.data.Dataset.interleave` to the number of schedulable CPU cores. 170* `parallel_for`: Add converter for `MatrixDiag`. 171* Add `narrow_range` attribute to `QuantizeAndDequantizeV2` and V3. 172* Added new op: `tf.strings.unsorted_segment_join`. 173* Add HW acceleration support for `topK_v2`. 174* Add new `TypeSpec` classes. 175* CloudBigtable version updated to v0.10.0. 176* Expose `Head` as public API. 177* Update docstring for gather to properly describe the non-empty `batch_dims` case. 178* Added `tf.sparse.from_dense` utility function. 179* Improved ragged tensor support in `TensorFlowTestCase`. 180* Makes the a-normal form transformation in Pyct configurable as to which nodes are converted to variables and which are not. 181* `ResizeInputTensor` now works for all delegates. 182* Add `EXPAND_DIMS` support to NN API delegate TEST: expand_dims_test 183* `tf.cond` emits a StatelessIf op if the branch functions are stateless and do not touch any resources. 184* `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. 185* `tf.while_loop` emits a StatelessWhile op if the cond and body functions are stateless and do not touch any resources. 186* Refactors code in Quant8 LSTM support to reduce TFLite binary size. 187* Add support of local soft device placement for eager op. 188* Add HW acceleration support for `LogSoftMax`. 189* Added a function `nested_value_rowids` for ragged tensors. 190* Add guard to avoid acceleration of L2 Normalization with input rank != 4 191* Add `tf.math.cumulative_logsumexp operation`. 192* Add `tf.ragged.stack`. 193* Fix memory allocation problem when calling `AddNewInputConstantTensor`. 194* Delegate application failure leaves interpreter in valid state. 195* Add check for correct memory alignment to `MemoryAllocation::MemoryAllocation()`. 196* Extracts `NNAPIDelegateKernel` from nnapi_delegate.cc 197* Added support for `FusedBatchNormV3` in converter. 198* A ragged to dense op for directly calculating tensors. 199* Fix accidental quadratic graph construction cost in graph-mode `tf.gradients()`. 200 201## Thanks to our Contributors 202 203This release contains contributions from many people at Google, as well as: 204 205a6802739, 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) 206 207# Release 2.0.0 208 209## Major Features and Improvements 210 211TensorFlow 2.0 focuses on **simplicity** and **ease of use**, featuring updates like: 212 213* Easy model building with Keras and eager execution. 214* Robust model deployment in production on any platform. 215* Powerful experimentation for research. 216* API simplification by reducing duplication and removing deprecated endpoints. 217 218For details on best practices with 2.0, see [the Effective 2.0 guide](https://www.tensorflow.org/beta/guide/effective_tf2) 219 220 221For information on upgrading your existing TensorFlow 1.x models, please refer to our [Upgrade](https://medium.com/tensorflow/upgrading-your-code-to-tensorflow-2-0-f72c3a4d83b5) and [Migration](https://www.tensorflow.org/beta/guide/migration_guide) guides. We have also released a collection of [tutorials and getting started guides](https://www.tensorflow.org/beta). 222 223## Highlights 224 225* TF 2.0 delivers Keras as the central high level API used to build and train 226 models. Keras provides several model-building APIs such as Sequential, 227 Functional, and Subclassing along with eager execution, for immediate 228 iteration and intuitive debugging, and `tf.data`, for building scalable 229 input pipelines. Checkout 230 [guide](https://www.tensorflow.org/beta/guide/keras/overview) for additional 231 details. 232* Distribution Strategy: TF 2.0 users will be able to use the 233 [`tf.distribute.Strategy`](https://www.tensorflow.org/beta/guide/distribute_strategy) 234 API to distribute training with minimal code changes, yielding great 235 out-of-the-box performance. It supports distributed training with Keras 236 model.fit, as well as with custom training loops. Multi-GPU support is 237 available, along with experimental support for multi worker and Cloud TPUs. 238 Check out the 239 [guide](https://www.tensorflow.org/beta/guide/distribute_strategy) for more 240 details. 241* Functions, not Sessions. The traditional declarative programming model of 242 building a graph and executing it via a `tf.Session` is discouraged, and 243 replaced with by writing regular Python functions. Using the `tf.function` 244 decorator, such functions can be turned into graphs which can be executed 245 remotely, serialized, and optimized for performance. 246* Unification of `tf.train.Optimizers` and `tf.keras.Optimizers`. Use 247 `tf.keras.Optimizers` for TF2.0. `compute_gradients` is removed as public 248 API, use `GradientTape` to compute gradients. 249* AutoGraph translates Python control flow into TensorFlow expressions, 250 allowing users to write regular Python inside `tf.function`-decorated 251 functions. AutoGraph is also applied in functions used with tf.data, 252 tf.distribute and tf.keras APIs. 253* Unification of exchange formats to SavedModel. All TensorFlow ecosystem 254 projects (TensorFlow Lite, TensorFlow JS, TensorFlow Serving, TensorFlow 255 Hub) accept SavedModels. Model state should be saved to and restored from 256 SavedModels. 257* API Changes: Many API symbols have been renamed or removed, and argument 258 names have changed. Many of these changes are motivated by consistency and 259 clarity. The 1.x API remains available in the compat.v1 module. A list of 260 all symbol changes can be found 261 [here](https://docs.google.com/spreadsheets/d/1FLFJLzg7WNP6JHODX5q8BDgptKafq_slHpnHVbJIteQ/edit#gid=0). 262 * API clean-up, included removing `tf.app`, `tf.flags`, and `tf.logging` 263 in favor of [absl-py](https://github.com/abseil/abseil-py). 264* No more global variables with helper methods like 265 `tf.global_variables_initializer` and `tf.get_global_step`. 266* Add toggles `tf.enable_control_flow_v2()` and `tf.disable_control_flow_v2()` 267 for enabling/disabling v2 control flow. 268* Enable v2 control flow as part of `tf.enable_v2_behavior()` and 269 `TF2_BEHAVIOR=1`. 270* Fixes autocomplete for most TensorFlow API references by switching to use 271 relative imports in API `__init__.py` files. 272* Auto Mixed-Precision graph optimizer simplifies converting models to 273 `float16` for acceleration on Volta and Turing Tensor Cores. This feature 274 can be enabled by wrapping an optimizer class with 275 `tf.train.experimental.enable_mixed_precision_graph_rewrite()`. 276* Add environment variable `TF_CUDNN_DETERMINISTIC`. Setting to "true" or "1" 277 forces the selection of deterministic cuDNN convolution and max-pooling 278 algorithms. When this is enabled, the algorithm selection procedure itself 279 is also deterministic. 280 281## Breaking Changes 282* Many backwards incompatible API changes have been made to clean up the APIs and make them more consistent. 283* Toolchains: 284 * 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. 285 * 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. 286 Removed the `freeze_graph` command line tool; `SavedModel` should be used in place of frozen graphs. 287 288* `tf.contrib`: 289 * `tf.contrib` has been deprecated, and functionality has been either migrated to the core TensorFlow API, to an ecosystem project such as [tensorflow/addons](https://www.github.com/tensorflow/addons) or [tensorflow/io](https://www.github.com/tensorflow/io), or removed entirely. 290 * Remove `tf.contrib.timeseries` dependency on TF distributions. 291 * Replace contrib references with `tf.estimator.experimental.*` for apis in `early_stopping.py`. 292 293* `tf.estimator`: 294 * 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.Optimizer`s. 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*`. 295 * Default aggregation for canned Estimators is now `SUM_OVER_BATCH_SIZE`. To maintain previous default behavior, please pass `SUM` as the loss aggregation method. 296 * 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`. 297 * `Estimator.export_savedmodel` has been renamed to `export_saved_model`. 298 * 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`. 299 * 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. 300 301* `tf.keras`: 302 * `OMP_NUM_THREADS` is no longer used by the default Keras config. To configure the number of threads, use `tf.config.threading` APIs. 303 * `tf.keras.model.save_model` and `model.save` now defaults to saving a TensorFlow SavedModel. HDF5 files are still supported. 304 * 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. 305 * 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. 306 307* `tf.lite`: 308 * Removed `lite.OpHint`, `lite.experimental`, and `lite.constant` from 2.0 API. 309* 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. 310* Performing equality operations on Tensors or Variables with incompatible shapes an exception is no longer thrown. Instead `__eq__` returns False and `__ne__` returns True. 311* Removed `tf.string_split` from v2 API. 312* Deprecated the use of `constraint=` and `.constraint` with ResourceVariable. 313* 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. 314* `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. 315 316Refer to our [public project status tracker](https://github.com/orgs/tensorflow/projects/4) and [issues tagged with `2.0`](https://github.com/tensorflow/tensorflow/issues?q=is%3Aopen+is%3Aissue+label%3A2.0) on GitHub for insight into recent issues and development progress. 317 318If you experience any snags when using TF 2.0, please let us know at the [TF 2.0 Testing User Group](https://groups.google.com/a/tensorflow.org/forum/?utm_medium=email&utm_source=footer#!forum/testing). We have a support mailing list as well as weekly testing meetings, and would love to hear your migration feedback and questions. 319 320 321## Bug Fixes and Other Changes 322 323* `tf.contrib`: 324 325 * Expose `tf.contrib.proto.*` ops in `tf.io` (they will exist in TF2) 326 327* `tf.data`: 328 329 * Add support for TensorArrays to `tf.data Dataset`. 330 * Integrate Ragged Tensors with `tf.data`. 331 * All core and experimental tf.data transformations that input 332 user-defined functions can span multiple devices now. 333 * Extending the TF 2.0 support for `shuffle(..., 334 reshuffle_each_iteration=True)` and `cache()` to work across different 335 Python iterators for the same dataset. 336 * Removing the `experimental_numa_aware` option from `tf.data.Options`. 337 * Add `num_parallel_reads` and passing in a Dataset containing filenames 338 into `TextLineDataset` and `FixedLengthRecordDataset`. 339 * Add support for defaulting the value of `cycle_length` argument of 340 `tf.data.Dataset.interleave` to the number of schedulable CPU cores. 341 * Promoting `tf.data.experimental.enumerate_dataset` to core as 342 `tf.data.Dataset.enumerate`. 343 * Promoting `tf.data.experimental.unbatch` to core as 344 `tf.data.Dataset.unbatch`. 345 * Adds option for introducing slack in the pipeline to reduce CPU 346 contention, via `tf.data.Options().experimental_slack = True` 347 * Added experimental support for parallel batching to `batch()` and 348 `padded_batch()`. This functionality can be enabled through 349 `tf.data.Options()`. 350 * Support cancellation of long-running `reduce`. 351 * Now we use `dataset` node name as prefix instead of the op name, to 352 identify the component correctly in metrics, for pipelines with repeated 353 components. 354 * Improve the performance of datasets using `from_tensors()`. 355 * Promoting `unbatch` from experimental to core API. 356 * Adding support for datasets as inputs to `from_tensors` and 357 `from_tensor_slices` and batching and unbatching of nested datasets. 358 359* `tf.distribute`: 360 361 * Enable `tf.distribute.experimental.MultiWorkerMirroredStrategy` working 362 in eager mode. 363 * Callbacks are supported in `MultiWorkerMirroredStrategy`. 364 * Disable `run_eagerly` and distribution strategy if there are symbolic 365 tensors added to the model using `add_metric` or `add_loss`. 366 * Loss and gradients should now more reliably be correctly scaled w.r.t. 367 the global batch size when using a `tf.distribute.Strategy`. 368 * Set default loss reduction as `AUTO` for improving reliability of loss 369 scaling with distribution strategy and custom training loops. `AUTO` 370 indicates that the reduction option will be determined by the usage 371 context. For almost all cases this defaults to `SUM_OVER_BATCH_SIZE`. 372 When used in distribution strategy scope, outside of built-in training 373 loops such as `tf.keras` `compile` and `fit`, we expect reduction value 374 to be 'None' or 'SUM'. Using other values will raise an error. 375 * Support for multi-host `ncclAllReduce` in Distribution Strategy. 376 377* `tf.estimator`: 378 379 * Replace `tf.contrib.estimator.add_metrics` with 380 `tf.estimator.add_metrics` 381 * Use `tf.compat.v1.estimator.inputs` instead of `tf.estimator.inputs` 382 * Replace contrib references with `tf.estimator.experimental.*` for apis 383 in early_s in Estimator 384 * Canned Estimators will now use keras optimizers by default. An error 385 will be raised if tf.train.Optimizers are used, and you will have to 386 switch to tf.keras.optimizers or tf.compat.v1 canned Estimators. 387 * A checkpoint converter for canned Estimators has been provided to 388 transition canned Estimators that are warm started from 389 `tf.train.Optimizers` to `tf.keras.optimizers`. 390 * Losses are scaled in canned estimator v2 and not in the optimizers 391 anymore. If you are using Estimator + distribution strategy + optimikzer 392 v1 then the behavior does not change. This implies that if you are using 393 custom estimator with optimizer v2, you have to scale losses. We have 394 new utilities to help scale losses `tf.nn.compute_average_loss`, 395 `tf.nn.scale_regularization_loss`. 396 397* `tf.keras`: 398 399 * Premade models (including Linear and WideDeep) have been introduced for 400 the purpose of replacing Premade estimators. 401 * Model saving changes 402 * `model.save` and `tf.saved_model.save` may now save to the TensorFlow 403 SavedModel format. The model can be restored using 404 `tf.keras.models.load_model`. HDF5 files are still supported, and may be 405 used by specifying `save_format="h5"` when saving. 406 * Raw TensorFlow functions can now be used in conjunction with the Keras 407 Functional API during model creation. This obviates the need for users 408 to create Lambda layers in most cases when using the Functional API. 409 Like Lambda layers, TensorFlow functions that result in Variable 410 creation or assign ops are not supported. 411 * Add support for passing list of lists to the `metrics` argument in Keras 412 `compile`. 413 * Add `tf.keras.layers.AbstractRNNCell` as the preferred implementation 414 for RNN cells in TF v2. User can use it to implement RNN cells with 415 custom behavior. 416 * Keras training and validation curves are shown on the same plot when 417 using the TensorBoard callback. 418 * Switched Keras `fit/evaluate/predict` execution to use only a single 419 unified path by default unless eager execution has been explicitly 420 disabled, regardless of input type. This unified path places an 421 eager-friendly training step inside of a `tf.function`. With this 422 * All input types are converted to `Dataset`. 423 * The path assumes there is always a distribution strategy. when 424 distribution strategy is not specified the path uses a no-op 425 distribution strategy. 426 * The training step is wrapped in `tf.function` unless `run_eagerly=True` 427 is set in compile. The single path execution code does not yet support 428 all use cases. We fallback to the existing v1 execution paths if your 429 model contains the following: 430 1. `sample_weight_mode` in compile 431 2. `weighted_metrics` in compile 432 3. v1 optimizer 433 4. target tensors in compile If you are experiencing any issues because 434 of this change, please inform us (file an issue) about your use case 435 and you can unblock yourself by setting 436 `experimental_run_tf_function=False` in compile meanwhile. We have 437 seen couple of use cases where the model usage pattern is not as 438 expected and would not work with this change. 439 * output tensors of one layer is used in the constructor of another. 440 * symbolic tensors outside the scope of the model are used in custom loss 441 functions. The flag can be disabled for these cases and ideally the 442 usage pattern will need to be fixed. 443 * Mark Keras `set_session` as `compat.v1` only. 444 * `tf.keras.estimator.model_to_estimator` now supports exporting to 445 `tf.train.Checkpoint format`, which allows the saved checkpoints to be 446 compatible with `model.load_weights`. 447 * `keras.backend.resize_images` (and consequently, 448 `keras.layers.Upsampling2D`) behavior has changed, a bug in the resizing 449 implementation was fixed. 450 * Add an `implementation=3` mode for `tf.keras.layers.LocallyConnected2D` 451 and `tf.keras.layers.LocallyConnected1D` layers using `tf.SparseTensor` 452 to store weights, allowing a dramatic speedup for large sparse models. 453 * Raise error if `batch_size` argument is used when input is 454 dataset/generator/keras sequence. 455 * Update TF 2.0 `keras.backend.name_scope` to use TF 2.0 `name_scope`. 456 * Add v2 module aliases for losses, metrics, initializers and optimizers: 457 `tf.losses = tf.keras.losses` & `tf.metrics = tf.keras.metrics` & 458 `tf.initializers = tf.keras.initializers` & `tf.optimizers = 459 tf.keras.optimizers`. 460 * Updates binary cross entropy logic in Keras when input is probabilities. 461 Instead of converting probabilities to logits, we are using the cross 462 entropy formula for probabilities. 463 * Added public APIs for `cumsum` and `cumprod` keras backend functions. 464 * Add support for temporal sample weight mode in subclassed models. 465 * Raise `ValueError` if an integer is passed to the training APIs. 466 * Added fault-tolerance support for training Keras model via `model.fit()` 467 with `MultiWorkerMirroredStrategy`, tutorial available. 468 * Custom Callback tutorial is now available. 469 * To train with `tf.distribute`, Keras API is recommended over estimator. 470 * `steps_per_epoch` and `steps` arguments are supported with numpy arrays. 471 * New error message when unexpected keys are used in 472 sample_weight/class_weight dictionaries 473 * Losses are scaled in Keras compile/fit and not in the optimizers 474 anymore. If you are using custom training loop, we have new utilities to 475 help scale losses `tf.nn.compute_average_loss`, 476 `tf.nn.scale_regularization_loss`. 477 * `Layer` apply and add_variable APIs are deprecated. 478 * Added support for channels first data format in cross entropy losses 479 with logits and support for tensors with unknown ranks. 480 * Error messages will be raised if `add_update`, `add_metric`, `add_loss`, 481 activity regularizers are used inside of a control flow branch. 482 * New loss reduction types: 483 * `AUTO`: Indicates that the reduction option will be determined by the 484 usage context. For almost all cases this defaults to 485 `SUM_OVER_BATCH_SIZE`. When used with `tf.distribute.Strategy`, outside 486 of built-in training loops such as `tf.keras` `compile` and `fit`, we 487 expect reduction value to be `SUM` or `NONE`. Using `AUTO` in that case 488 will raise an error. 489 * `NONE`: Weighted losses with one dimension reduced (axis=-1, or axis 490 specified by loss function). When this reduction type used with built-in 491 Keras training loops like `fit`/`evaluate`, the unreduced vector loss is 492 passed to the optimizer but the reported loss will be a scalar value. 493 * `SUM`: Scalar sum of weighted losses. 4. `SUM_OVER_BATCH_SIZE`: Scalar 494 `SUM` divided by number of elements in losses. This reduction type is 495 not supported when used with `tf.distribute.Strategy` outside of 496 built-in training loops like `tf.keras` `compile`/`fit`. 497 * Wraps losses passed to the `compile` API (strings and v1 losses) which 498 are not instances of v2 `Loss` class in `LossWrapper` class. => All 499 losses will now use `SUM_OVER_BATCH_SIZE` reduction as default. 500 * `model.add_loss(symbolic_tensor)` should work in ambient eager. 501 * Update metric name to always reflect what the user has given in compile. 502 Affects following cases 503 * When name is given as 'accuracy'/'crossentropy' 504 * When an aliased function name is used eg. 'mse' 505 * Removing the `weighted` prefix from weighted metric names. 506 * Allow non-Tensors through v2 losses. 507 * Add v2 sparse categorical crossentropy metric. 508 * Add v2 APIs for `AUCCurve` and `AUCSummationMethod` enums. 509 * `add_update` can now be passed a zero-arg callable in order to support 510 turning off the update when setting `trainable=False` on a Layer of a 511 Model compiled with `run_eagerly=True`. 512 * Standardize the LayerNormalization API by replacing the args `norm_axis` 513 and `params_axis` with `axis`. 514 * Fixed critical bugs that help with DenseFeatures usability in TF2 515 516* `tf.lite`: 517 518 * Added evaluation script for `COCO` minival 519 * Add delegate support for `QUANTIZE`. 520 * Add `GATHER` support to NN API delegate. 521 * Added support for TFLiteConverter Python API in 2.0. Contains functions 522 from_saved_model, from_keras_file, and from_concrete_functions. 523 * Add `EXPAND_DIMS` support to NN API delegate TEST. 524 * Add `narrow_range` attribute to QuantizeAndDequantizeV2 and V3. 525 * Added support for `tflite_convert` command line tool in 2.0. 526 * Post-training quantization tool supports quantizing weights shared by 527 multiple operations. The models made with versions of this tool will use 528 INT8 types for weights and will only be executable interpreters from 529 this version onwards. 530 * Post-training quantization tool supports fp16 weights and GPU delegate 531 acceleration for fp16. 532 * Add delegate support for `QUANTIZED_16BIT_LSTM`. 533 * Extracts `NNAPIDelegateKernel` from nnapi_delegate.cc 534 535* TensorRT 536 537 * Add TensorFlow 2.0-compatible `TrtGraphConverterV2` API for TensorRT 538 conversion. TensorRT initialization arguments are now passed wrapped in 539 a named-tuple, `TrtConversionParams`, rather than as separate arguments 540 as in `TrtGraphConverter`. 541 * Changed API to optimize TensorRT enginges during graph optimization. 542 This is now done by calling `converter.build()` where previously 543 `is_dynamic_op=False` would be set. 544 * `converter.convert()` no longer returns a `tf.function`. Now the 545 function must be accessed from the saved model. 546 * The `converter.calibrate()` method has been removed. To trigger 547 calibration, a `calibration_input_fn` should be provided to 548 `converter.convert()`. 549 550* Other: 551 552 * Fix accidental quadratic graph construction cost in graph-mode 553 `tf.gradients()`. 554 * ResourceVariable's gather op supports batch dimensions. 555 * ResourceVariable support for `gather_nd`. 556 * `ResourceVariable` and `Variable` no longer accepts `constraint` in the 557 constructor, nor expose it as a @property. 558 * Added gradient for `SparseToDense` op. 559 * Expose a flag that allows the number of threads to vary across Python 560 benchmarks. 561 * `image.resize` in 2.0 now supports gradients for the new resize kernels. 562 * `image.resize` now considers proper pixel centers and has new kernels 563 (incl. anti-aliasing). 564 * Renamed `tf.image` functions to remove duplicate "image" where it is 565 redundant. 566 * Variadic reduce is supported on CPU Variadic reduce is supported on CPU 567 * Remove unused `StringViewVariantWrapper`. 568 * Delete unused `Fingerprint64Map` op registration 569 * Add broadcasting support to `tf.matmul`. 570 * Add C++ Gradient for `BatchMatMulV2`. 571 * Add `tf.math.cumulative_logsumexp` operation. 572 * Add ellipsis (...) support for `tf.einsum()`. 573 * Add expand_composites argument to all `nest.*` methods. 574 * Added `strings.byte_split`. 575 * Add a new "result_type" parameter to `tf.strings.split`. 576 * Add name argument to `tf.string_split` and `tf.strings_split`. 577 * Extend `tf.strings.split` to support inputs with any rank. 578 * Added `tf.random.binomial`. 579 * Added `key` and `skip` methods to `random.experimental.Generator`. 580 * Extend `tf.function` with basic support for CompositeTensors arguments 581 (such as `SparseTensor` and `RaggedTensor`). 582 * `parallel_for.pfor`: add converters for Softmax, LogSoftmax, IsNaN, All, 583 Any, and MatrixSetDiag. 584 * `parallel_for`: add converters for LowerTriangularSolve and Cholesky. 585 * `parallel_for`: add converters for `LogMatrixDeterminant` and 586 `MatrixBandPart`. 587 * `parallel_for`: Add converter for `MatrixDiag`. 588 * `parallel_for`: Add converters for `OneHot`, `LowerBound`, `UpperBound`. 589 * `parallel_for`: add converter for `BroadcastTo`. 590 * Add `pfor` converter for `Squeeze`. 591 * Add `RaggedTensor.placeholder()`. 592 * Add ragged tensor support to `tf.squeeze`. 593 * Update RaggedTensors to support int32 row_splits. 594 * Allow `LinearOperator.solve` to take a `LinearOperator`. 595 * Allow all dtypes for `LinearOperatorCirculant`. 596 * Introduce MaxParallelism method 597 * Add `LinearOperatorHouseholder`. 598 * Adds Philox support to new stateful RNG's XLA path. 599 * Added `TensorSpec` support for CompositeTensors. 600 * Added `tf.linalg.tridiagonal_solve` op. 601 * Added partial_pivoting input parameter to `tf.linalg.tridiagonal_solve`. 602 * Added gradient to `tf.linalg.tridiagonal_solve`. 603 * Added `tf.linalg.tridiagonal_mul op`. 604 * Added GPU implementation of `tf.linalg.tridiagonal_matmul`. 605 * Added `LinearOperatorToeplitz`. 606 * Upgraded LIBXSMM to version 1.11. 607 * Uniform processing of quantized embeddings by Gather and EmbeddingLookup 608 Ops. 609 * Correct a misstatement in the documentation of the sparse softmax cross 610 entropy logit parameter. 611 * Add `tf.ragged.boolean_mask`. 612 * `tf.switch_case` added, which selects a branch_fn based on a 613 branch_index. 614 * The C++ kernel of gather op supports batch dimensions. 615 * Fixed default value and documentation for `trainable` arg of 616 tf.Variable. 617 * `EagerTensor` now supports numpy buffer interface for tensors. 618 * This change bumps the version number of the `FullyConnected` Op to 5. 619 * Added new op: `tf.strings.unsorted_segment_join`. 620 * Added HW acceleration support for `topK_v2`. 621 * CloudBigtable version updated to v0.10.0 BEGIN_PUBLIC CloudBigtable 622 version updated to v0.10.0. 623 * Expose `Head` as public API. 624 * Added `tf.sparse.from_dense` utility function. 625 * Improved ragged tensor support in `TensorFlowTestCase`. 626 * Added a function `nested_value_rowids` for ragged tensors. 627 * Added `tf.ragged.stack`. 628 * Makes the a-normal form transformation in Pyct configurable as to which 629 nodes are converted to variables and which are not. 630 * `ResizeInputTensor` now works for all delegates. 631 * `tf.cond` emits a StatelessIf op if the branch functions are stateless 632 and do not touch any resources. 633 * Add support of local soft device placement for eager op. 634 * Pass partial_pivoting to the `_TridiagonalSolveGrad`. 635 * Add HW acceleration support for `LogSoftMax`. 636 * Add guard to avoid acceleration of L2 Normalization with input rank != 4 637 * Fix memory allocation problem when calling `AddNewInputConstantTensor`. 638 * Delegate application failure leaves interpreter in valid state 639 * `tf.while_loop` emits a StatelessWhile op if the cond and body functions 640 are stateless and do not touch any resources. 641 * `tf.cond`, `tf.while` and if and while in AutoGraph now accept a 642 nonscalar predicate if has a single element. This does not affect non-V2 643 control flow. 644 * Fix potential security vulnerability where decoding variant tensors from 645 proto could result in heap out of bounds memory access. 646 * Only create a GCS directory object if the object does not already exist. 647 * Introduce `dynamic` constructor argument in Layer and Model, which 648 should be set to `True` when using imperative control flow in the `call` 649 method. 650 * Begin adding Go wrapper for C Eager API. 651 * XLA HLO graphs can be inspected with interactive_graphviz tool now. 652 * Add dataset ops to the graph (or create kernels in Eager execution) 653 during the python Dataset object creation instead doing it during 654 Iterator creation time. 655 * Add `batch_dims` argument to `tf.gather`. 656 * The behavior of `tf.gather` is now correct when `axis=None` and 657 `batch_dims<0`. 658 * Update docstring for gather to properly describe the non-empty 659 `batch_dims` case. 660 * Removing of dtype in the constructor of initializers and partition_info 661 in call. 662 * Add `tf.math.nextafter` op. 663 * Turn on MKL-DNN contraction kernels by default. MKL-DNN dynamically 664 dispatches the best kernel implementation based on CPU vector 665 architecture. To disable them, build with 666 `--define=tensorflow_mkldnn_contraction_kernel=0`. 667 * `tf.linspace(start, stop, num)` now always uses "stop" as last value 668 (for num > 1) 669 * Added top-k to precision and recall to keras metrics. 670 * Add a ragged size op and register it to the op dispatcher 671 * Transitive dependencies on :`pooling_ops` were removed. Some users may 672 need to add explicit dependencies on :`pooling_ops` if they reference 673 the operators from that library. 674 * Add `CompositeTensor` base class. 675 * Malformed gif images could result in an access out of bounds in the 676 color palette of the frame. This has been fixed now 677 * Add templates and interfaces for creating lookup tables 678 * `Tensor::UnsafeCopyFromInternal` deprecated in favor 679 `Tensor::BitcastFrom`. 680 * In `map_vectorization` optimization, reduce the degree of parallelism in 681 the vectorized map node. 682 * Add variant wrapper for `absl::string_view`. 683 * Add OpKernels for some stateless maps. 684 * DType is no longer convertible to an int. Use `dtype.as_datatype_enum` 685 instead of `int(dtype)` to get the same result. 686 * Support both binary and -1/1 label input in v2 hinge and squared hinge 687 losses. 688 * Added `LinearOperator.adjoint` and `LinearOperator.H` (alias). 689 * Expose CriticalSection in core as `tf.CriticalSection`. 690 * Enhanced graphviz output. 691 * Add opkernel templates for common table operations. 692 * Fix callbacks do not log values in eager mode when a deferred build 693 model is used. 694 * `SignatureDef` util functions have been deprecated. 695 * Update `Fingerprint64Map` to use aliases 696 * Add legacy string flat hash map op kernels. 697 * Add support for `add_metric` in the graph function mode. 698 * Updating cosine similarity loss - removed the negate sign from cosine 699 similarity. 700 * Changed default for gradient accumulation for TPU embeddings to true. 701 * Adds summary trace API for collecting graph and profile information. 702 * The `precision_mode` argument to `TrtGraphConverter` is now case 703 insensitive. 704 705## Thanks to our Contributors 706 707This release contains contributions from many people at Google, as well as: 708 7091e100, a6802739, 4d55397500, a6802739, Abdullah Selek, abenmao, Abolfazl 710Shahbazi, Adam Richter, Adam Weiss, Ag Ramesh, Alan Du, Albin Joy, Alex, Alex 711Itkes, Alex Sergeev, Alexander Pivovarov, Alexey Romanov, alhkad, Aman Patel, 712Amit, Amit Kumar Jaiswal, Amit Srivastava, amoitra, Andreas Eberle, Andrew 713Lihonosov, Andy Craze, Anshuman Tripathy, Anthony Hsu, Anthony Platanios, Anuj 714Rawat, arp95, Arpit Shah, Armen Poghosov, armenpoghosov, Astropeak, Ashwin 715Ramaswami, Arpit Shah, Augustina Ragwitz, Aurelien Geron, AuréLien Geron, 716avasid, aweers, awesomealex1, Ayush Agrawal, Bas Aarts, Bastian Eichenberger, 717Bairen Yi, Bayberry Z, Ben Barsdell, Benjamin Peterson, bhack, Bharat 718Raghunathan, Bhavani Subramanian, Bin Fan, blairhan, BléNesi Attila, Bodin-E, 719Brandon Carter, Bryan Cutler, candy.dc, Cao Zongyan, Casper Da Costa-Luis, Chao 720Liu, Chen Guoyin, chenchc, chengchingwen, chie8842, Christian Hansen, Christoph 721Boeddeker, Christopher Yeh, Clayne Robison, Coady, Patrick, crafet, csukuangfj, 722ctiijima, Dan Jarvis, Dan Lazewatsky, Daniel Ingram, Daniel Rasmussen, Daniel 723Salvadori, Dave Airlie, David Norman, Dayananda V, delock, Denis Khalikov, Deven 724Desai, Dheeraj Rajaram Reddy, Diego Caballero, dmitrievanthony, Donovan Ong, 725Drew Szurko, Duncan Dean, Duncan Riach, Dustin Neighly, Dwight J Lyle, Eamon 726Ito-Fisher, eashtian3, Edward Forgacs, EFanZh, ejot, Elroy Ashtian Jr, Eric 727Schweitz, Evgeniy Polyakov, Fangjun Kuang, Federico Martinez, Fei Hu, Felix 728Lemke, Filip Matzner, FlashTek, fo40225, formath, FrançOis Chollet, frreiss, 729Fred Reiss, Frederic Bastien, Fredrik Knutsson, G. Hussain Chinoy, Gabriel, 730Gautam, gehring, Geoffrey Irving, George Grzegorz Pawelczak, Grzegorz Pawelczak, 731George Sterpu, Gianluca Varisco, Gleb Popov, Greg Peatfield, Guillaume Klein, 732Gurpreet Singh, Gustavo Lima Chaves, Gyoung-Yoon Ryoo, haison, Hanton Yang, 733HanGuo97, Haraldur TóMas HallgríMsson, Hari Shankar, hehongliang, Heungsub Lee, 734Hoeseong Kim, Huan Li (李卓桓), HåKon Sandsmark, I-Hong, I-Hong Jhuo, Ilham 735Firdausi Putra, Ilango R, Imran Salam, Innovimax, Jacky Ko, Irene Dea, Ivan 736Habernal, Jakub Lipinski, Jacky, Jason Zaman, Jason Zavaglia, jayhpark530, 737jcf94, jefby, Jeff Daily, Jeff Poznanovic, Jeffrey Poznanovic, Jekyll Lai, jer, 738Jeroen BéDorf, jerryyin, jhalakp, jiakai, Jia Qingtong, Jiankang, JiangXIAO, Joe 739Bowser, Joe Q, Joe Quadrino, Joel Shapiro, Johan Gunnarsson, Jojimon Varghese, 740Jonas Rauber, Jonathan Kyl, Jonathan, Joon, Joppe Geluykens, Joseph Friedman, 741Josh Beal, jtressle, Julian Niedermeier, Junqin Zhang, Justin Dujardin, Justin 742Tunis, jwu, K. Hodges, kaixih, Kaixi Hou, kjopek, Karl Lessard, Karl 743Weinmeister, Karthik Muthuraman, Kashif Rasul, Kay Zhu, Kbhute-Ibm, KDR, Keno 744Fischer, Kevin Mader, khanhlvg, Kilaru Yasaswi Sri Chandra Gandhi, Koan-Sin Tan, 745Koock Yoon, kouml, ktaebum, Kyuwon Kim, Lakshay Tokas, Laurent Le Brun, 746leike666666, leonard951, Leslie-Fang, Letian Kang, Li, Guizi, Loo Rong Jie, 747Lucas Hendren, Lukas Folle, Lukas Geiger, Luke Han, luxupu, lvli, Ma, Guokai, 748Mahmoud Abuzaina, Maksym Kysylov, Mandar Deshpande, manhyuk, Manraj Singh 749Grover, Marco Gaido, Marek Drozdowski, Margaret Maynard-Reid, Mark Ryan, mars20, 750Mateusz Chudyk, Matt Conley, mbhuiyan, mdfaijul, Mei Jie, Melissa Grueter, 751merturl, MichaelKonobeev, Michael KäUfl, Michal W. Tarnowski, MickaëL 752Schoentgen, Miguel Morin, Mihail Salnikov, Mikalai Drabovich, Mike Arpaia, Mike 753Holcomb, minds, monklof, Moses Marin, mpppk, Mr. Metal, Mshr-H, musikisomorphie, 754nammbash, Natalia Gimelshein, Nathan Luehr, Nayana-Ibm, Nayana Thorat, neargye, 755Neeraj Pradhan, Nehal J Wani, Neil, Nick, Nick Lewycky, Niels Ole Salscheider, 756Niklas SilfverströM, Niranjan Hasabnis, Nuka-137, Nutti, ocjosen, olicht, 757omeir1, P Sudeepam, Paige Bailey, Palmer Lao, Pan Daoxin, Pariksheet Pinjari, 758Pasquale Minervini, Patrick J. Lopresti, Patrik Gustavsson, Pavel Akhtyamov, 759Pavel Samolysov, PENGWA, per1234, PeterLee, Phan Van Nguyen Duc, Philipp Jund, 760Phillip Kravtsov, Pooya Davoodi, Pranav Marathe, Putra Manggala, Qingqing Cao, R 761S Nikhil Krishna, Rajeshwar Reddy T, Ramon ViñAs, Rasmus Diederichsen, Reuben 762Morais, robert, Rohit Gupta, Roland Zimmermann, Roman Soldatow, RonLek, Ruizhe, 763Ryan Jiang, saishruthi, Saleem Abdulrasool, Samantha Andow, Sami Kama, 764Sana-Damani, Saurabh Deoras, sdamani, Sean Morgan, seanshpark, Sebastien Iooss, 765Serv-Inc, Severen Redwood, Shahzad Lone, Shashank Gupta, shashvat, Shashvat 766Chand Shahi, Shubham Goyal, Shashi, Sigrid Keydana, Siju, Siju Samuel, 767sleighsoft, smilu97, Snease-Abq, Son Tran, Spencer Schaber, sremedios, Srini511, 768srinivasan.narayanamoorthy, Steve Lang, Steve Nesae, Subin, Sumesh Udayakumaran, 769Sungmann Cho, sunway513, Supriya Rao, sxwang, Tae-Hwan Jung, Taehoon Lee, Takeo 770Sawada, Taylor Jakobson, Taylor Thornton, Ted Chang, TengLu, terryky, 771ThisIsIsaac, ThisIsPIRI, Thomas Deegan, Thomas Hagebols, tianyapiaozi, Till 772Hoffmann, Tim Zaman, tomguluson92, Tongxuan Liu, Trent Lo, Trevor Morris, 773TungJerry, Tyorden, Uday Bondhugula, v1incent, Vagif, Vasileios Lioutas, 774vbvg2008, vcarpani, Vijay Ravichandran, Vikram Tiwari,Viktor Gal, Vishwak 775Srinivasan, Vincent, Vishnuvardhan Janapati, Vitor-Alves, Vivek Suryamurthy, 776wangsiyu, wateryzephyr, WeberXie, Wei Wang, WeijieSun, Wen-Heng (Jack) Chung, 777wenxizhu, Will Battel, William D. Irons, winstonq, wyzhao, Xiaoming (Jason) Cui, 778Xiaoquan Kong, Xin, Xinping Wang, Yan Facai (颜发才), Yann-Yy, Yasir Modak, 779Yasuhiro Matsumoto, ymodak, Yong Tang, Yongfeng Gu, Younes Khoudli, Yuan Lin, 780Yuan (Terry) Tang, Yuchen Ying, Yves-Noel Weweler, zhangyujing, zjjott, zyeric, 781王振华 (Zhenhua Wang), 黄鑫 782 783# Release 1.14.0 784 785## Major Features and Improvements 786 787* This is the first 1.x release containing the compat.v2 module. This module 788 is required to allow libraries to publish code which works in both 1.x and 789 2.x. After this release, no backwards incompatible changes are allowed in 790 the 2.0 Python API. 791* Turn on MKL-DNN contraction kernels by default. MKL-DNN dynamically 792 dispatches the best kernel implementation based on CPU vector architecture. 793 To disable them, build with --define=tensorflow_mkldnn_contraction_kernel=0. 794 795## Behavioral changes 796 797* Set default loss reduction as `AUTO` for improving reliability of loss 798 scaling with distribution strategy and custom training loops. `AUTO` 799 indicates that the reduction option will be determined by the usage context. 800 For almost all cases this defaults to `SUM_OVER_BATCH_SIZE`. When used in 801 distribution strategy scope, outside of built-in training loops such as 802 `tf.keras` `compile` and `fit`, we expect reduction value to be 'None' or 803 'SUM'. Using other values will raise an error. 804* Wraps losses passed to the `compile` API (strings and v1 losses) which are 805 not instances of v2 `Loss` class in `LossWrapper` class. => All losses will 806 now use `SUM_OVER_BATCH_SIZE` reduction as default. 807* Disable `run_eagerly` and distribution strategy if there are symbolic 808 tensors added to the model using `add_metric` or `add_loss`. 809* tf.linspace(start, stop, num) now always uses "stop" as last value (for 810 num > 1) 811* `ResourceVariable` and `Variable` no longer accepts `constraint` in the 812 constructor, nor expose it as a @property. 813* The behavior of tf.gather is now correct when axis=None and batch_dims<0. 814* Only create a GCS directory object if the object does not already exist. 815* In `map_vectorization` optimization, reduce the degree of parallelism in the 816 vectorized map node. 817* Bug fix: loss and gradients should now more reliably be correctly scaled 818 w.r.t. the global batch size when using a tf.distribute.Strategy. 819* Updating cosine similarity loss - removed the negate sign from cosine 820 similarity. 821* DType is no longer convertible to an int. Use dtype.as_datatype_enum instead 822 of int(dtype) to get the same result. 823* Changed default for gradient accumulation for TPU embeddings to true. 824* Callbacks now log values in eager mode when a deferred build model is used. 825* Transitive dependencies on :pooling_ops were removed. Some users may need to 826 add explicit dependencies on :pooling_ops if they reference the operators 827 from that library. 828* tf.keras.optimizers default learning rate changes: 829 * Adadelta: 1.000 to 0.001 830 * Adagrad: 0.01 to 0.001 831 * Adamax: 0.002 to 0.001 832 * NAdam: 0.002 to 0.001 833 834## Bug Fixes and Other Changes 835 836* Documentation 837* Deprecations and Symbol renames. 838 * Remove unused StringViewVariantWrapper 839 * Delete unused Fingerprint64Map op registration 840 * SignatureDef util functions have been deprecated. 841 * Renamed tf.image functions to remove duplicate "image" where it is 842 redundant. 843 * tf.keras.experimental.export renamed to 844 tf.keras.experimental.export_saved_model 845 * Standardize the LayerNormalization API by replacing the args `norm_axis` 846 and `params_axis` with `axis`. 847 * Tensor::UnsafeCopyFromInternal deprecated in favor Tensor::BitcastFrom 848* Keras & Python API 849 * Add v2 module aliases for: 850 * tf.initializers => tf.keras.initializers 851 * tf.losses => tf.keras.losses & tf.metrics => tf.keras.metrics 852 * tf.optimizers => tf.keras.optimizers 853 * Add tf.keras.layers.AbstractRNNCell as the preferred implementation of 854 RNN cell for TF v2. User can use it to implement RNN cell with custom 855 behavior. 856 * Adding `clear_losses` API to be able to clear losses at the end of 857 forward pass in a custom training loop in eager. 858 * Add support for passing list of lists to the `metrics` param in Keras 859 `compile`. 860 * Added top-k to precision and recall to keras metrics. 861 * Adding public APIs for `cumsum` and `cumprod` keras backend functions. 862 * Fix: model.add_loss(symbolic_tensor) should work in ambient eager. 863 * Add name argument to tf.string_split and tf.strings_split 864 * Minor change to SavedModels exported from Keras using 865 tf.keras.experimental.export. (SignatureDef key for evaluation mode is 866 now "eval" instead of "test"). This will be reverted back to "test" in 867 the near future. 868 * Updates binary cross entropy logic in Keras when input is probabilities. 869 Instead of converting probabilities to logits, we are using the cross 870 entropy formula for probabilities. 871 * Raw TensorFlow functions can now be used in conjunction with the Keras 872 Functional API during model creation. This obviates the need for users 873 to create Lambda layers in most cases when using the Functional API. 874 Like Lambda layers, TensorFlow functions that result in Variable 875 creation or assign ops are not supported. 876 * Keras training and validation curves are shown on the same plot. 877 * Introduce `dynamic` constructor argument in Layer and Model, which 878 should be set to True when using imperative control flow in the `call` 879 method. 880 * Removing of dtype in the constructor of initializers and partition_info 881 in call. 882* New ops and improved op functionality 883 * Add OpKernels for some stateless maps 884 * Add v2 APIs for AUCCurve and AUCSummationMethod 885 enums. #tf-metrics-convergence 886 * Add tf.math.nextafter op. 887 * Add CompositeTensor base class. 888 * Add tf.linalg.tridiagonal_solve op. 889 * Add opkernel templates for common table operations. 890 * Added support for TFLite in TensorFlow 2.0. 891 * Adds summary trace API for collecting graph and profile information. 892 * Add batch_dims argument to tf.gather. 893 * Add support for `add_metric` in the graph function mode. 894 * Add C++ Gradient for BatchMatMulV2. 895 * Added tf.random.binomial 896 * Added gradient for SparseToDense op. 897 * Add legacy string flat hash map op kernels 898 * Add a ragged size op and register it to the op dispatcher 899 * Add broadcasting support to tf.matmul. 900 * Add ellipsis (...) support for tf.einsum() 901 * Added LinearOperator.adjoint and LinearOperator.H (alias). 902 * Added GPU implementation of tf.linalg.tridiagonal_solve. 903 * Added strings.byte_split 904 * Add RaggedTensor.placeholder() 905 * Add a new "result_type" parameter to tf.strings.split 906 * `add_update` can now be passed a zero-arg callable in order to support 907 turning off the update when setting `trainable=False` on a Layer of a 908 Model compiled with `run_eagerly=True`. 909 * Add variant wrapper for absl::string_view 910 * Add expand_composites argument to all nest.* methods. 911 * Add pfor converter for Squeeze. 912 * Bug fix for tf.tile gradient 913 * Expose CriticalSection in core as tf.CriticalSection. 914 * Update Fingerprint64Map to use aliases 915 * ResourceVariable support for gather_nd. 916 * ResourceVariable's gather op supports batch dimensions. 917 * Variadic reduce is supported on CPU 918 * Extend tf.function with basic support for CompositeTensors arguments 919 (such as SparseTensor and RaggedTensor). 920 * Add templates and interfaces for creating lookup tables 921 * Post-training quantization tool supports quantizing weights shared by 922 multiple operations. The models made with versions of this tool will use 923 INT8 types for weights and will only be executable interpreters from 924 this version onwards. 925 * Malformed gif images could result in an access out of bounds in the 926 color palette of the frame. This has been fixed now 927 * image.resize now considers proper pixel centers and has new kernels 928 (incl. anti-aliasing). 929* Performance 930 * Turn on MKL-DNN contraction kernels by default. MKL-DNN dynamically 931 dispatches the best kernel implementation based on CPU vector 932 architecture. To disable them, build with 933 --define=tensorflow_mkldnn_contraction_kernel=0. 934 * Support for multi-host ncclAllReduce in Distribution Strategy. 935 * Expose a flag that allows the number of threads to vary across Python 936 benchmarks. 937* TensorFlow 2.0 Development 938 * Add v2 sparse categorical crossentropy metric. 939 * Allow non-Tensors through v2 losses. 940 * Add UnifiedGRU as the new GRU implementation for tf2.0. Change the 941 default recurrent activation function for GRU from 'hard_sigmoid' to 942 'sigmoid', and 'reset_after' to True in 2.0. Historically recurrent 943 activation is 'hard_sigmoid' since it is fast than 'sigmoid'. With new 944 unified backend between CPU and GPU mode, since the CuDNN kernel is 945 using sigmoid, we change the default for CPU mode to sigmoid as well. 946 With that, the default GRU will be compatible with both CPU and GPU 947 kernel. This will enable user with GPU to use CuDNN kernel by default 948 and get a 10x performance boost in training. Note that this is 949 checkpoint breaking change. If user want to use their 1.x pre-trained 950 checkpoint, please construct the layer with 951 GRU(recurrent_activation='hard_sigmoid', reset_after=False) to fallback 952 to 1.x behavior. 953 * TF 2.0 - Update metric name to always reflect what the user has given in 954 compile. Affects following cases 1. When name is given as 955 'accuracy'/'crossentropy' 2. When an aliased function name is used eg. 956 'mse' 3. Removing the `weighted` prefix from weighted metric names. 957 * Begin adding Go wrapper for C Eager API 958 * image.resize in 2.0 now supports gradients for the new resize kernels. 959 * removed tf.string_split from v2 API 960 * Expose tf.contrib.proto.* ops in tf.io (they will exist in TF2) 961 * "Updates the TFLiteConverter API in 2.0. Changes from_concrete_function 962 to from_concrete_functions." 963 * Enable tf.distribute.experimental.MultiWorkerMirroredStrategy working in 964 eager mode. 965 * Support both binary and -1/1 label input in v2 hinge and squared hinge 966 losses. 967* TensorFlow Lite 968 * "Adds support for tflite_convert in 2.0." 969 * "Remove lite.OpHint, lite.experimental, and lite.constant from 2.0 API." 970* tf.contrib 971 * Added Neural Turing Implementation as described in 972 https://arxiv.org/abs/1807.08518. 973 * Remove tf.contrib.timeseries dependency on TF distributions. 974* tf.data 975 * Add num_parallel_reads and passing in a Dataset containing filenames 976 into TextLineDataset and FixedLengthRecordDataset 977 * Going forward we operate in TF 2.0, this change is part of the effort to 978 slowly converting XYZDataset to DatasetV2 type which is the official 979 version going to be used in TF 2.0 and motivated by some compatibility 980 issue found, _BigtableXYZDataset (of type DatasetV2) does not implement 981 the _as_variant_tensor() of DatasetV1, when moving contrib.bigtable to 982 tensorflow_io. Converting into DatasetV2 removes the overheads to 983 maintain V1 while we are moving into TF 2.0. 984 * Add dataset ops to the graph (or create kernels in Eager execution) 985 during the python Dataset object creation instead doing it during 986 Iterator creation time. 987 * Add support for TensorArrays to tf.data Dataset. 988 * Switching tf.data functions to use `defun`, providing an escape hatch to 989 continue using the legacy `Defun`. 990* Toolchains 991 * CUDNN_INSTALL_PATH, TENSORRT_INSTALL_PATH, NCCL_INSTALL_PATH, 992 NCCL_HDR_PATH are deprecated. Use TF_CUDA_PATHS instead which supports a 993 comma-separated list of base paths that are searched to find CUDA 994 libraries and headers. 995 * TF code now resides in `tensorflow_core` and `tensorflow` is just a 996 virtual pip package. No code changes are needed for projects using 997 TensorFlow, the change is transparent 998* XLA 999 * XLA HLO graphs can be inspected with interactive_graphviz tool now. 1000* Estimator 1001 * Use tf.compat.v1.estimator.inputs instead of tf.estimator.inputs 1002 * Replace contrib references with tf.estimator.experimental.* for apis in 1003 early_stopping.py 1004 1005## Thanks to our Contributors 1006 1007This release contains contributions from many people at Google, as well as: 1008 10091e100, 4d55397500, a6802739, abenmao, Adam Weiss, Ag Ramesh, Alan Du, Albin Joy, 1010Alex, Aman Patel, Amit, Amit Kumar Jaiswal, Amit Srivastava, Andreas Eberle, 1011Andy Craze, Anthony Platanios, Armen Poghosov, armenpoghosov, arp95, Arpit Shah, 1012Ashwin Ramaswami, Aurelien Geron, AuréLien Geron, aweers, awesomealex1, Ayush 1013Agrawal, Ben Barsdell, Bharat Raghunathan, Bhavani Subramanian, blairhan, 1014BléNesi Attila, Brandon Carter, candy.dc, Chao Liu, chenchc, chie8842, Christian 1015Hansen, Christian Sigg, Clayne Robison, crafet, csukuangfj, ctiijima, Dan 1016Jarvis, Dan Lazewatsky, Daniel Ingram, Daniel Salvadori, Dave Airlie, David 1017Norman, Dayananda V, Dayananda-V, delock, Denis Khalikov, Deven Desai, Dheeraj 1018Rajaram Reddy, dmitrievanthony, Donovan Ong, Drew Szurko, Duncan Riach, Dustin 1019Neighly, Edward Forgacs, EFanZh, Fei Hu, Felix Lemke, Filip Matzner, fo40225, 1020frreiss, Gautam, gehring, Geoffrey Irving, Grzegorz George Pawelczak, Grzegorz 1021Pawelczak, Gyoung-Yoon Ryoo, HanGuo97, Hanton Yang, Hari Shankar, hehongliang, 1022Heungsub Lee, Hoeseong Kim, I-Hong Jhuo, Ilango R, Innovimax, Irene Dea, Jacky 1023Ko, Jakub Lipinski, Jason Zaman, jcf94, Jeffrey Poznanovic, Jens Elofsson, 1024Jeroen BéDorf, Jia Qingtong, Jiankang, Joe Q, Joe Quadrino, Joeran Beel, Jonas 1025Rauber, Jonathan, Jonathan Kyl, Joppe Geluykens, Joseph Friedman, jtressle, jwu, 1026K Yasaswi Sri Chandra Gandhi, K. Hodges, Kaixi Hou, Karl Lessard, Karl 1027Weinmeister, Karthik Muthuraman, Kashif Rasul, KDR, Keno Fischer, Kevin Mader, 1028kjopek, Koan-Sin Tan, kouml, ktaebum, Lakshay Tokas, Laurent Le Brun, Letian 1029Kang, Li, Guizi, Loo Rong Jie, Lucas Hendren, Lukas Geiger, Luke Han, luxupu, 1030Ma, Guokai, Mahmoud Abuzaina, Mandar Deshpande, manhyuk, Marco Gaido, Marek 1031Drozdowski, Mark Collier, Mark Ryan, mars20, Mateusz Chudyk, Matt Conley, 1032MattConley, mbhuiyan, mdfaijul, Melissa Grueter, Michael KäUfl, MickaëL 1033Schoentgen, Miguel Morin, Mihail Salnikov, Mike Arpaia, Mike Holcomb, monklof, 1034Moses Marin, Mshr-H, nammbash, Natalia Gimelshein, Nayana-Ibm, neargye, Neeraj 1035Pradhan, Nehal J Wani, Nick, Niels Ole Salscheider, Niranjan Hasabnis, nlewycky, 1036Nuka-137, Nutti, olicht, P Sudeepam, Palmer Lao, Pan Daoxin, Pariksheet Pinjari, 1037Pavel Samolysov, PENGWA, Pooya Davoodi, R S Nikhil Krishna, Rohit Gupta, Roman 1038Soldatow, rthadur, Ruizhe, Ryan Jiang, Samantha Andow, Sami Kama, Sana-Damani, 1039Saurabh Deoras, sdamani, seanshpark, Sebastien Iooss, Serv-Inc, Shahzad Lone, 1040Shashank Gupta, Shashi, shashvat, shashvatshahi1998, Siju, Siju Samuel, 1041Snease-Abq, Spencer Schaber, sremedios, srinivasan.narayanamoorthy, Steve Lang, 1042Steve Nesae, Sumesh Udayakumaran, Supriya Rao, Taylor Jakobson, Taylor Thornton, 1043Ted Chang, ThisIsPIRI, Thomas Deegan, Thomas Hagebols, tianyapiaozi, Tim Zaman, 1044tomguluson92, Tongxuan Liu, TungJerry, v1incent, Vagif, vcarpani, Vikram Tiwari, 1045Vishwak Srinivasan, Vitor-Alves, wangsiyu, wateryzephyr, WeberXie, WeijieSun, 1046Wen-Heng (Jack) Chung, wenxizhu, Will Battel, William D. Irons, wyzhao, Xin, 1047Yasuhiro Matsumoto, ymodak, Yong Tang, Younes Khoudli, Yuan Lin, Yves-Noel 1048Weweler, Zantares, zjjott, 卜居, 王振华 (Wang Zhenhua), 黄鑫 1049 1050# Release 1.12.3 1051 1052## Bug Fixes and Other Changes 1053 1054* Updates `png_archive` dependency to 1.6.37 to not be affected by 1055 CVE-2019-7317, CVE-2018-13785, and CVE-2018-14048. 1056* Updates `sqlite` dependency to 3.28.0 to not be affected by CVE-2018-20506, 1057 CVE-2018-20346, and CVE-2018-20505. 1058 1059# Release 1.12.2 1060 1061## Bug Fixes and Other Changes 1062 1063* Fixes a potential security vulnerability where carefully crafted GIF images 1064 can produce a null pointer dereference during decoding. 1065 1066# Release 1.13.0 1067 1068## Major Features and Improvements 1069 1070* 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`. 1071* TensorFlow GPU binaries are now built against CUDA 10 and TensorRT 5.0. 1072* Support for Python3.7 on all operating systems. 1073* Moved NCCL to core. 1074 1075## Behavioral changes 1076 1077* Disallow conversion of python floating types to uint32/64 (matching behavior of other integer types) in `tf.constant`. 1078* 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). 1079 1080## Bug Fixes and Other Changes 1081 1082* Documentation 1083 * Update the doc with the details about the rounding mode used in 1084 quantize_and_dequantize_v2. 1085 * Clarify that tensorflow::port::InitMain() _should_ be called before 1086 using the TensorFlow library. Programs failing to do this are not 1087 portable to all platforms. 1088* Deprecations and Symbol renames. 1089 * Removing deprecations for the following endpoints: `tf.acos`, 1090 `tf.acosh`, `tf.add`, `tf.as_string`, `tf.asin`, `tf.asinh`, `tf.atan`, 1091 `tf.atan2`, `tf.atanh`, `tf.cos`, `tf.cosh`, `tf.equal`, `tf.exp`, 1092 `tf.floor`, `tf.greater`, `tf.greater_equal`, `tf.less`, 1093 `tf.less_equal`, `tf.log`, `tf.logp1`, `tf.logical_and`, 1094 `tf.logical_not`, `tf.logical_or`, `tf.maximum`, `tf.minimum`, 1095 `tf.not_equal`, `tf.sin`, `tf.sinh`, `tf.tan` 1096 * Deprecate `tf.data.Dataset.shard`. 1097 * Deprecate `saved_model.loader.load` which is replaced by 1098 `saved_model.load` and `saved_model.main_op`, which will be replaced by 1099 `saved_model.main_op` in V2. 1100 * Deprecate tf.QUANTIZED_DTYPES. The official new symbol is 1101 tf.dtypes.QUANTIZED_DTYPES. 1102 * Update sklearn imports for deprecated packages. 1103 * Deprecate `Variable.count_up_to` and `tf.count_up_to` in favor of 1104 `Dataset.range`. 1105 * Export `confusion_matrix` op as `tf.math.confusion_matrix` instead of 1106 `tf.train.confusion_matrix`. 1107 * Add `tf.dtypes.` endpoint for every constant in dtypes.py. Moving 1108 endpoints in versions.py to corresponding endpoints in `tf.sysconfig.` 1109 and `tf.version.`. Moving all constants under `tf.saved_model` 1110 submodules to `tf.saved_model` module. New endpoints are added in V1 and 1111 V2 but existing endpoint removals are only applied in V2. 1112 * Deprecates behavior where device assignment overrides collocation 1113 constraints inside a collocation context manager. 1114* Keras & Python API 1115 * Add to Keras functionality analogous to 1116 `tf.register_tensor_conversion_function`. 1117 * Subclassed Keras models can now be saved through 1118 `tf.contrib.saved_model.save_keras_model`. 1119 * `LinearOperator.matmul` now returns a new `LinearOperator`. 1120* New ops and improved op functionality 1121 * Add a Nearest Neighbor Resize op. 1122 * Add an `ignore_unknown` argument to `parse_values` which suppresses 1123 ValueError for unknown hyperparameter types. Such * Add 1124 `tf.linalg.matvec` convenience function. 1125 * `tf.einsum()`raises `ValueError` for unsupported equations like 1126 `"ii->"`. 1127 * Add DCT-I and IDCT-I in `tf.signal.dct` and `tf.signal.idct`. 1128 * Add LU decomposition op. 1129 * Add quantile loss to gradient boosted trees in estimator. 1130 * Add `round_mode` to `QuantizeAndDequantizeV2` op to select rounding 1131 algorithm. 1132 * Add `unicode_encode`, `unicode_decode`, `unicode_decode_with_offsets`, 1133 `unicode_split`, `unicode_split_with_offset`, and `unicode_transcode` 1134 ops. Amongst other things, this Op adds the ability to encode, decode, 1135 and transcode a variety of input text encoding formats into the main 1136 Unicode encodings (UTF-8, UTF-16-BE, UTF-32-BE) 1137 * Add "unit" attribute to the substr op, which allows obtaining the 1138 substring of a string containing unicode characters. 1139 * Broadcasting support for Ragged Tensors. 1140 * `SpaceToDepth` supports uint8 data type. 1141 * Support multi-label quantile regression in estimator. 1142 * We now use "div" as the default partition_strategy in 1143 `tf.nn.safe_embedding_lookup_sparse`, `tf.nn.sampled_softmax` and 1144 `tf.nn.nce_loss`. hyperparameter are ignored. 1145* Performance 1146 * Improve performance of GPU cumsum/cumprod by up to 300x. 1147 * Added support for weight decay in most TPU embedding optimizers, 1148 including AdamW and MomentumW. 1149* TensorFlow 2.0 Development 1150 * Add a command line tool to convert to TF2.0, tf_upgrade_v2 1151 * Merge `tf.spectral` into `tf.signal` for TensorFlow 2.0. 1152 * Change the default recurrent activation function for LSTM from 1153 'hard_sigmoid' to 'sigmoid' in 2.0. Historically recurrent activation is 1154 'hard_sigmoid' since it is fast than 'sigmoid'. With new unified backend 1155 between CPU and GPU mode, since the CuDNN kernel is using sigmoid, we 1156 change the default for CPU mode to sigmoid as well. With that, the 1157 default LSTM will be compatible with both CPU and GPU kernel. This will 1158 enable user with GPU to use CuDNN kernel by default and get a 10x 1159 performance boost in training. Note that this is checkpoint breaking 1160 change. If user want to use their 1.x pre-trained checkpoint, please 1161 construct the layer with LSTM(recurrent_activation='hard_sigmoid') to 1162 fallback to 1.x behavior. 1163* TensorFlow Lite 1164 * Move from `tensorflow/contrib/lite` to `tensorflow/lite`. 1165 * Add experimental Java API for injecting TensorFlow Lite delegates 1166 * Add support for strings in TensorFlow Lite Java API. 1167* `tf.contrib`: 1168 * Add Apache Ignite Filesystem plugin to support accessing Apache IGFS. 1169 * Dropout now takes `rate` argument, `keep_prob` is deprecated. 1170 * Estimator occurrences references `tf.contrib.estimator` were changed to 1171 `tf.estimator`: 1172 * `tf.contrib.estimator.BaselineEstimator` with 1173 `tf.estimator.BaselineEstimator` 1174 * `tf.contrib.estimator.DNNLinearCombinedEstimator` with 1175 `tf.estimator.DNNLinearCombinedEstimator` 1176 * `tf.contrib.estimator.DNNEstimator` with `tf.estimator.DNNEstimator` 1177 * `tf.contrib.estimator.LinearEstimator` with 1178 `tf.estimator.LinearEstimator` 1179 * `tf.contrib.estimator.InMemoryEvaluatorHook` and 1180 tf.estimator.experimental.InMemoryEvaluatorHook`. 1181 * `tf.contrib.estimator.make_stop_at_checkpoint_step_hook` with 1182 `tf.estimator.experimental.make_stop_at_checkpoint_step_hook`. 1183 * Expose `tf.distribute.Strategy as the new name for 1184 tf.contrib.distribute.DistributionStrategy. 1185 * Migrate linear optimizer from contrib to core. 1186 * Move `tf.contrib.signal` to `tf.signal` (preserving aliases in 1187 tf.contrib.signal). 1188 * Users of `tf.contrib.estimator.export_all_saved_models` and related 1189 should switch to 1190 `tf.estimator.Estimator.experimental_export_all_saved_models`. 1191* tf.data: 1192 * Add `tf.data.experimental.StatsOptions()`, to configure options to 1193 collect statistics from `tf.data.Dataset` pipeline using 1194 `StatsAggregator`. Add nested option, `experimental_stats` (which takes 1195 a `tf.data.experimen tal.StatsOptions` object), to `tf.data.Options`. 1196 Deprecates `tf.data.experimental.set_stats_agregator`. 1197 * Performance optimizations: 1198 * Add `tf.data.experimental.OptimizationOptions()`, to configure options 1199 to enable `tf.data` performance optimizations. Add nested option, 1200 `experimental_optimization` (which takes a 1201 `tf.data.experimental.OptimizationOptions` object), to 1202 `tf.data.Options`. Remove performance optimization options from 1203 `tf.data.Options`, and add them under 1204 `tf.data.experimental.OptimizationOptions` instead. 1205 * Enable `map_and_batch_fusion` and `noop_elimination` optimizations by 1206 default. They can be disabled by configuring 1207 `tf.data.experimental.OptimizationOptions` to set `map_and_batch = 1208 False` or `noop_elimination = False` respectively. To disable all 1209 default optimizations, set `apply_default_optimizations = False`. 1210 * Support parallel map in `map_and_filter_fusion`. 1211 * Disable static optimizations for input pipelines that use non-resource 1212 `tf.Variable`s. 1213 * Add NUMA-aware MapAndBatch dataset. 1214 * Deprecate `tf.data.Dataset.make_one_shot_iterator()` in V1, removed it 1215 from V2, and added tf.compat.v1.data.make_one_shot_iterator()`. 1216 * Deprecate `tf.data.Dataset.make_initializable_iterator()` in V1, removed 1217 it from V2, and added `tf.compat.v1.data.make_initializable_iterator()`. 1218 * Enable nested dataset support in core `tf.data` transformations. 1219 * For `tf.data.Dataset` implementers: Added 1220 `tf.data.Dataset._element_structured property` to replace 1221 `Dataset.output_{types,shapes,classes}`. 1222 * Make `num_parallel_calls` of `tf.data.Dataset.interleave` and 1223 `tf.data.Dataset.map` work in Eager mode. 1224* Toolchains 1225 * Fixed OpenSSL compatibility by avoiding `EVP_MD_CTX_destroy`. 1226 * Added bounds checking to printing deprecation warnings. 1227 * Upgraded CUDA dependency to 10.0 1228 * To build with Android NDK r14b, add "#include <linux/compiler.h>" to 1229 android-ndk-r14b/platforms/android-14/arch-*/usr/include/linux/futex.h 1230 * Removed `:android_tensorflow_lib_selective_registration*` targets, use 1231 `:android_tensorflow_lib_lite*` targets instead. 1232* XLA 1233 * Move `RoundToEven` function to xla/client/lib/math.h. 1234 * A new environment variable `TF_XLA_DEBUG_OPTIONS_PASSTHROUGH` set to "1" 1235 or "true" allows the debug options passed within an XRTCompile op to be 1236 passed directly to the XLA compilation backend. If such variable is not 1237 set (service side), only a restricted set will be passed through. 1238 * Allow the XRTCompile op to return the ProgramShape resulted form the XLA 1239 compilation as a second return argument. 1240 * XLA HLO graphs can now be rendered as SVG/HTML. 1241* Estimator 1242 * Replace all occurrences of `tf.contrib.estimator.BaselineEstimator` with 1243 `tf.estimator.BaselineEstimator` 1244 * Replace all occurrences of 1245 `tf.contrib.estimator.DNNLinearCombinedEstimator` with 1246 `tf.estimator.DNNLinearCombinedEstimator` 1247 * Replace all occurrences of `tf.contrib.estimator.DNNEstimator` with 1248 `tf.estimator.DNNEstimator` 1249 * Replace all occurrences of `tf.contrib.estimator.LinearEstimator` with 1250 `tf.estimator.LinearEstimator` 1251 * Users of `tf.contrib.estimator.export_all_saved_models` and related 1252 should switch to 1253 `tf.estimator.Estimator.experimental_export_all_saved_models`. 1254 * Update `regression_head` to the new Head API for Canned Estimator V2. 1255 * Switch `multi_class_head` to Head API for Canned Estimator V2. 1256 * Replace all occurrences of `tf.contrib.estimator.InMemoryEvaluatorHook` 1257 and `tf.contrib.estimator.make_stop_at_checkpoint_step_hook` with 1258 `tf.estimator.experimental.InMemoryEvaluatorHook` and 1259 `tf.estimator.experimental.make_stop_at_checkpoint_step_hook` 1260 * Migrate linear optimizer from contrib to core. 1261 1262## Thanks to our Contributors 1263 1264This release contains contributions from many people at Google, as well as: 1265 1266Abhinav 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 1267 1268# Release 1.12.0 1269 1270## Major Features and Improvements 1271 1272* Keras models can now be directly exported to the SavedModel 1273 format(`tf.contrib.saved_model.save_keras_model()`) and used with Tensorflow 1274 Serving. 1275* Keras models now support evaluating with a `tf.data.Dataset`. 1276* TensorFlow binaries are built with XLA support linked in by default. 1277* Ignite Dataset added to contrib/ignite that allows to work with Apache 1278 Ignite. 1279 1280## Bug Fixes and Other Changes 1281 1282* tf.data: 1283 * tf.data users can now represent, get, and set options of TensorFlow 1284 input pipelines using `tf.data.Options()`, `tf.data.Dataset.options()`, 1285 and `tf.data.Dataset.with_options()` respectively. 1286 * New `tf.data.Dataset.reduce()` API allows users to reduce a finite 1287 dataset to a single element using a user-provided reduce function. 1288 * New `tf.data.Dataset.window()` API allows users to create finite windows 1289 of input dataset; when combined with the `tf.data.Dataset.reduce()` API, 1290 this allows users to implement customized batching. 1291 * All C++ code moves to the `tensorflow::data` namespace. 1292 * Add support for `num_parallel_calls` to `tf.data.Dataset.interleave`. 1293* `tf.contrib`: 1294 * Remove `tf.contrib.linalg`. `tf.linalg` should be used instead. 1295 * Replace any calls to `tf.contrib.get_signature_def_by_key(metagraph_def, 1296 signature_def_key)` with 1297 `meta_graph_def.signature_def[signature_def_key]`. Catching a ValueError 1298 exception thrown by `tf.contrib.get_signature_def_by_key` should be 1299 replaced by catching a KeyError exception. 1300* `tf.contrib.data` 1301 * Deprecate, and replace by tf.data.experimental. 1302* Other: 1303 * Instead of jemalloc, revert back to using system malloc since it 1304 simplifies build and has comparable performance. 1305 * Remove integer types from `tf.nn.softplus` and `tf.nn.softsign` OpDefs. 1306 This is a bugfix; these ops were never meant to support integers. 1307 * Allow subslicing Tensors with a single dimension. 1308 * Add option to calculate string length in Unicode characters. 1309 * Add functionality to SubSlice a tensor. 1310 * Add searchsorted (ie lower/upper_bound) op. 1311 * Add model explainability to Boosted Trees. 1312 * Support negative positions for tf.substr. 1313 * There was previously a bug in the bijector_impl where the 1314 _reduce_jacobian_det_over_event does not handle scalar ILDJ 1315 implementations properly. 1316 * In tf eager execution, allow re-entering a GradientTape context. 1317 * Add tf_api_version flag. If --define=tf_api_version=2 flag is passed in, 1318 then bazel will build TensorFlow API version 2.0. Note that TensorFlow 1319 2.0 is under active development and has no guarantees at this point. 1320 * Add additional compression options to TfRecordWriter. 1321 * Performance improvements for regex full match operations. 1322 * Replace tf.GraphKeys.VARIABLES with `tf.GraphKeys.GLOBAL_VARIABLES`. 1323 * Remove unused dynamic learning rate support. 1324 1325## Thanks to our Contributors 1326 1327This release contains contributions from many people at Google, as well as: 1328 1329(David) Siu-Kei Muk, Ag Ramesh, Anton Dmitriev, Artem Sobolev, Avijit-Nervana, 1330Bairen Yi, Bruno Goncalves, By Shen, candy.dc, Cheng Chen, Clayne Robison, 1331coder3101, Dao Zhang, Elms, Fei Hu, feiquan, Geoffrey Irving, Guozhong Zhuang, 1332hellcom, Hoeseong Kim, imsheridan, Jason Furmanek, Jason Zaman, Jenny Sahng, 1333jiefangxuanyan, Johannes Bannhofer, Jonathan Homer, Koan-Sin Tan, kouml, Loo 1334Rong Jie, Lukas Geiger, manipopopo, Ming Li, Moritz KröGer, Naurril, Niranjan 1335Hasabnis, Pan Daoxin, Peng Yu, pengwa, rasmi, Roger Xin, Roland Fernandez, Sami 1336Kama, Samuel Matzek, Sangjung Woo, Sergei Lebedev, Sergii Khomenko, shaohua, 1337Shaohua Zhang, Shujian2015, Sunitha Kambhampati, tomguluson92, ViníCius Camargo, 1338wangsiyu, weidankong, Wen-Heng (Jack) Chung, William D. Irons, Xin Jin, Yan 1339Facai (颜发才), Yanbo Liang, Yash Katariya, Yong Tang, 在原佐为 1340 1341# Release 1.11.0 1342 1343## Major Features and Improvements 1344 1345* Nvidia GPU: 1346 * Prebuilt binaries are now (as of TensorFlow 1.11) built against cuDNN 1347 7.2 and TensorRT 4. See updated install guides: 1348 [Installing TensorFlow on Ubuntu](https://www.tensorflow.org/install/install_linux#tensorflow_gpu_support) 1349* Google Cloud TPU: 1350 * Experimental tf.data integration for Keras on Google Cloud TPUs. 1351 * Experimental / preview support for eager execution on Google Cloud TPUs. 1352* DistributionStrategy: 1353 * Add multi-GPU DistributionStrategy support in tf.keras. Users can now 1354 use `fit`, `evaluate` and `predict` to distribute their model on 1355 multiple GPUs. 1356 * Add multi-worker DistributionStrategy and standalone client support in 1357 Estimator. See 1358 [README](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/distribute) 1359 for more details. 1360* Add C, C++, and Python functions for querying kernels. 1361 1362## Breaking Changes 1363 1364* Keras: 1365 * The default values for tf.keras `RandomUniform`, `RandomNormal`, and `TruncatedNormal` initializers have been changed to match those in external Keras. 1366 * 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. 1367 1368## Bug Fixes and Other Changes 1369 1370* C++: 1371 * Changed the signature of SessionFactory::NewSession so that it can 1372 return a meaningful error message on failure. 1373* tf.data: 1374 * Remove `num_parallel_parser_calls` argument from 1375 `tf.contrib.data.make_csv_dataset()`. [tf.data] Remove 1376 `num_parallel_parser_calls` argument from 1377 `tf.contrib.data.make_csv_dataset()`. 1378 * `tf.data.Dataset.list_files()` raises an exception at initialization 1379 time if the argument matches no files. 1380 * Renamed BigTable class to BigtableTable for clarity 1381 * Document use of the Cloud Bigtable API 1382 * Add `tf.contrib.data.reduce_dataset` which can be used to reduce a 1383 dataset to a single element. 1384 * Generalization of `tf.contrib.data.sliding_window_batch`. 1385* INC: 1386 * Runtime improvements to triangular solve. 1387* `tf.contrib`: 1388 * Add an `implementation` argument to `tf.keras.layers.LocallyConnected2D` 1389 and `tf.keras.layers.LocallyConnected1D`. The new mode 1390 (`implementation=2`) performs forward pass as a single dense matrix 1391 multiplication, allowing dramatic speedups in certain scenarios (but 1392 worse performance in others - see docstring). The option also allows to 1393 use `padding=same`. 1394 * Add documentation clarifying the differences between tf.fill and 1395 tf.constant. 1396 * Add experimental IndexedDatasets. 1397 * Add selective registration target using the lite proto runtime. 1398 * Add simple Tensor and DataType classes to TensorFlow Lite Java 1399 * Add support for bitcasting to/from uint32 and uint64. 1400 * Added a subclass of Estimator that can be created from a SavedModel 1401 (SavedModelEstimator). 1402 * Adds leaf index modes as an argument. 1403 * Allow a different output shape from the input in 1404 tf.contrib.image.transform. 1405 * Change the state_size order of the StackedRNNCell to be natural order. 1406 To keep the existing behavior, user can add reverse_state_order=True 1407 when constructing the StackedRNNCells. 1408 * Deprecate self.test_session() in favor of self.session() or 1409 self.cached_session(). 1410 * Directly import tensor.proto.h (the transitive import will be removed 1411 from tensor.h soon). 1412 * Estimator.train() now supports tf.contrib.summary.\* summaries out of 1413 the box; each call to .train() will now create a separate tfevents file 1414 rather than re-using a shared one. 1415 * Fix FTRL L2-shrinkage behavior: the gradient from the L2 shrinkage term 1416 should not end up in the accumulator. 1417 * Fix toco compilation/execution on Windows. 1418 * GoogleZoneProvider class added to detect which Google Cloud Engine zone 1419 tensorflow is running in. 1420 * It is now safe to call any of the C API's TF_Delete\* functions on 1421 nullptr. 1422 * Log some errors on Android to logcat. 1423 * Match FakeQuant numerics in TFLite to improve accuracy of TFLite 1424 quantized inference models. 1425 * Optional bucket location check for the GCS Filesystem. 1426 * Performance enhancements for StringSplitOp & StringSplitV2Op. 1427 * Performance improvements for regex replace operations. 1428 * TFRecordWriter now raises an error if .write() fails. 1429 * TPU: More helpful error messages in TPUClusterResolvers. 1430 * The legacy_init_op argument to SavedModelBuilder methods for adding 1431 MetaGraphs has been deprecated. Please use the equivalent main_op 1432 argument instead. As part of this, we now explicitly check for a single 1433 main_op or legacy_init_op at the time of SavedModel building, whereas 1434 the check on main_op was previously only done at load time. 1435 * The protocol used for Estimator training is now configurable in 1436 RunConfig. 1437 * Triangular solve performance improvements. 1438 * Unify RNN cell interface between TF and Keras. Add new 1439 get_initial_state() to Keras and TF RNN cell, which will use to replace 1440 the existing zero_state() method. 1441 * Update initialization of variables in Keras. 1442 * Updates to "constrained_optimization" in tensorflow/contrib. 1443 * boosted trees: adding pruning mode. 1444 * tf.train.Checkpoint does not delete old checkpoints by default. 1445 * tfdbg: Limit the total disk space occupied by dumped tensor data to 100 1446 GBytes. Add environment variable `TFDBG_DISK_BYTES_LIMIT` to allow 1447 adjustment of this upper limit. 1448 1449## Thanks to our Contributors 1450 1451This release contains contributions from many people at Google, as well as: 1452 1453Aapeli, 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 1454Ricardo Perez-Lopez, 张天启, 张晓飞 1455 1456 1457# Release 1.10.1 1458## Bug Fixes and Other Changes 1459 1460* `tf.keras`: 1461 * Fixing keras on Cloud TPUs. No new binaries will be built for Windows. 1462 1463 1464# Release 1.10.0 1465 1466## Major Features And Improvements 1467 1468* The `tf.lite` runtime now supports `complex64`. 1469* Initial [Google Cloud Bigtable integration](https://github.com/tensorflow/tensorflow/tree/r1.10/tensorflow/contrib/bigtable) for `tf.data`. 1470* Improved local run behavior in `tf.estimator.train_and_evaluate` which does not reload checkpoints for evaluation. 1471* `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`. 1472* Moved Distributions and Bijectors from `tf.contrib.distributions` to [Tensorflow Probability (TFP)](https://github.com/tensorflow/probability). `tf.contrib.distributions` is now deprecated and will be removed by the end of 2018. 1473* 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`](https://www.tensorflow.org/versions/master/api_docs/python/tf/debugging), [`tf.dtypes`](https://www.tensorflow.org/versions/master/api_docs/python/tf/dtypes), [`tf.image`](https://www.tensorflow.org/versions/master/api_docs/python/tf/image), [`tf.io`](https://www.tensorflow.org/versions/master/api_docs/python/tf/io), [`tf.linalg`](https://www.tensorflow.org/versions/master/api_docs/python/tf/linalg), [`tf.manip`](https://www.tensorflow.org/versions/master/api_docs/python/tf/manip), [`tf.math`](https://www.tensorflow.org/versions/master/api_docs/python/tf/math), [`tf.quantization`](https://www.tensorflow.org/versions/master/api_docs/python/tf/quantization), [`tf.strings`](https://www.tensorflow.org/versions/master/api_docs/python/tf/strings) 1474 1475## Breaking Changes 1476 1477* 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](https://developer.nvidia.com/nccl). See updated install guides: [TensorFlow GPU support](https://www.tensorflow.org/install/gpu) and [Build TensorFlow from source](https://www.tensorflow.org/install/source). 1478* Starting from TensorFlow 1.11, Windows builds will use Bazel. Therefore, we will drop official support for cmake. 1479 1480## Bug Fixes and Other Changes 1481 1482* `tf.data`: 1483 * `tf.contrib.data.group_by_reducer()` is now available via the public API. 1484 * `tf.contrib.data.choose_from_datasets()` is now available via the public API. 1485 * 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()`. 1486* `tf.estimator`: 1487 * `Estimator`s now use custom savers included in `EstimatorSpec` scaffolds for saving SavedModels during export. 1488 * `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. 1489 * Support sparse_combiner in canned Linear Estimators. 1490 * Added batch normalization to `DNNClassifier`, `DNNRegressor`, and `DNNEstimator`. 1491 * Adding ranking support for boosted trees. 1492 * Adding center bias option for boosted trees. 1493* Add `synchronization` and `aggregation` args to get_variable(). These args will be used for distributed variables. 1494* Add `synchronization` and `aggregation` args to the layer `add_weight()` API. These args will be used for distributed variables. 1495* `tf.losses.*` do not add to the global collection when executing eagerly (to avoid leaking memory). 1496* Support different summary and checkpoint directories in `tf.train.MonitoredTrainingSession()`. 1497* Added IndRNN, IndyGRU, and IndyLSTM cells to `tf.contrib.rnn`. 1498* Add safe static factory functions for SparseTensor and convert all CHECKs to DCHECKs. Using the constructor directly is unsafe and deprecated. 1499* Make the Bigtable client connection pool configurable & increase the default # of connections for performance. 1500* Added derivative of `tf.random_gamma` with respect to the alpha parameter. 1501* Added derivative of `tf.igamma(a, x)` and `tf.igammac(a, x)` with respect to a. 1502* Modified Bessel functions of order zero and one. 1503* Add FillTriangular Bijector to create triangular matrices. 1504* Added support for Type III DCT, and `tf.spectral.idct(type=2|3)`. 1505* Correctly handle CuDNN RNN weight loaded when nest in `TimeDistributed`. 1506* Adding per-element weight support for `WALSComputePartialLhsAndRhsOp`. 1507* ZerosLike and OnesLike ops treated as constants by Graph Transform Tool. 1508* Gamma distribution and the derived distributions (Beta, Dirichlet, Student's t, inverse Gamma) now fully reparameterized. 1509* Java: Experimental wrapper classes to make graph generation easier. Thanks @karllessard and @kbsriram 1510* Build & link in secure gRPC components (switch from the insecure grpc dependency to secure grpc dependency). 1511* 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: 1512 * New endpoints in `tf.image` namespace: `tf.image.extract_image_patches` 1513 * New endpoints in `tf.debugging` namespace: `tf.debugging.check_numerics`, `tf.debugging.is_finite`, `tf.debugging.is_inf`, `tf.debugging.is_nan`. 1514 * New endpoints in `tf.dtypes` namespace: `tf.dtypes.as_string`. 1515 * 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`. 1516 * 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`). 1517 * 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` 1518 * 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`. 1519 * 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`. 1520 * 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`). 1521 1522 1523## Thanks to our Contributors 1524 1525This release contains contributions from many people at Google, as well as: 1526 1527Ag 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 1528 1529# Release 1.9.0 1530 1531## Major Features And Improvements 1532* Updated docs for `tf.keras`: New Keras-based [get started](http://tensorflow.org/versions/r1.9/get_started), 1533 and [programmers guide page](http://tensorflow.org/versions/r1.9/programmers_guide/keras). 1534* Update `tf.keras` to the Keras 2.1.6 API. 1535* Added [`tf.keras.layers.CuDNNGRU`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/keras/layers/CuDNNGRU) and [`tf.keras.layers.CuDNNLSTM`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/keras/layers/CuDNNLSTM) layers. [Try it](https://colab.sandbox.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb?linkId=53292082). 1536* Adding support of core [feature columns](https://www.tensorflow.org/get_started/feature_columns) and [losses](https://www.tensorflow.org/api_docs/python/tf/losses) to [gradient boosted trees estimators](https://github.com/tensorflow/models/tree/master/official/r1/boosted_trees). 1537* The [python interface](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/lite) 1538 for the [TFLite Optimizing Converter](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/toco/README.md) 1539 has been expanded, and the command line interface (AKA: `toco`, `tflite_convert`) is once again 1540 included in the standard `pip` installation. 1541* Improved data-loading and text processing with: 1542 * [`tf.decode_compressed`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/decode_compressed) 1543 * [`tf.string_strip`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/string_strip) 1544 * [`tf.strings.regex_full_match`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/strings/regex_full_match) 1545* Added experimental support for new pre-made Estimators: 1546 * [`tf.contrib.estimator.BaselineEstimator`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/estimator/BaselineEstimator) 1547 * [`tf.contrib.estimator.RNNClassifier`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/estimator/RNNEstimator) 1548 * [`tf.contrib.estimator.RNNEstimator`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/estimator/RNNClassifier) 1549* The [distributions.Bijector](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/distributions/bijectors/Bijector) 1550 API supports broadcasting for Bijectors with new API changes. 1551 1552## Breaking Changes 1553 * If you're opening empty variable scopes; replace `variable_scope('', ...)` by 1554 `variable_scope(tf.get_variable_scope(), ...)`. 1555 * Headers used for building custom ops have been moved from site-packages/external into site-packages/tensorflow/include/external. 1556 1557## Bug Fixes and Other Changes 1558 1559* `tfe.Network` is deprecated. Please inherit from `tf.keras.Model`. 1560* Layered variable names have changed in the following conditions: 1561 * Using `tf.keras.layers` with custom variable scopes. 1562 * Using `tf.layers` in a subclassed `tf.keras.Model` class. See 1563 [here](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/layers) 1564 for more details 1565* `tf.data`: 1566 * `Dataset.from_generator()` now accepts an `args` list, in order to 1567 create nested generators. 1568 * `Dataset.list_files()` now produces deterministic results when 1569 `shuffle=False` or a `seed` is passed. 1570 * `tf.contrib.data.sample_from_datasets()` and 1571 `tf.contrib.data.choose_from_datasets()` make it easier to sample or 1572 deterministically choose elements from multiple datasets. 1573 * `tf.contrib.data.make_csv_dataset()` now supports line breaks in quoted 1574 strings, and two infrequently used arguments removed. 1575 * (C++) `DatasetBase::DebugString()` is now `const`. 1576 * (C++) `DatasetBase::MakeIterator()` has been renamed to 1577 `DatasetBase::MakeIteratorInternal()`. 1578 * (C++) `IteratorBase::Initialize()` method was added to support raising 1579 errors during iterator construction. 1580* Eager Execution: 1581 * Added the ability to pause recording operations for gradient computation 1582 via `tf.GradientTape.stop_recording`. 1583 * Updated documentation, introductory notebooks. 1584* `tf.keras`: 1585 * Move Keras code out of _impl folder and remove API files. 1586 * `tf.keras.Model.save_weights` now saves in TensorFlow format by default. 1587 * Enable dataset iterators to be passed to `tf.keras.Model` training/eval 1588 methods. 1589* TensorFlow Debugger (tfdbg) CLI: fix an issue in which the TensorBoard 1590 Debugger Plugin could not handle total source file size exceeding gRPC 1591 message size limit (4 MB). 1592* `tf.contrib`: 1593 * `tf.contrib.framework.zero_initializer` supports ResourceVariable. 1594 * Adding "constrained_optimization" to tensorflow/contrib. 1595* Other: 1596 * Add GCS Configuration Ops. 1597 * Changing signature of `MakeIterator` to enable propagating error status. 1598 * KL divergence for two Dirichlet distributions. 1599 * More consistent GcsFileSystem behavior for certain reads past EOF. 1600 * Update benchmark for tf.scan to match ranges across eager and graph 1601 modes. 1602 * Fixed bug in `tf.reduce_prod gradient` for complex dtypes. 1603 * Allow the use of '.' in variables (e.g. "hparams.parse('a.b=1.0')"), 1604 which would previously raise an error. This will correspond to an 1605 attribute name with an embedded '.' symbol (e.g. 'a.b'), which can only 1606 be accessed indirectly (e.g. through getattr and setattr). To set this 1607 up the user will first need to explicitly add the variable to the hparam 1608 object (e.g. "hparams.add_hparam(name='a.b', value=0.0)"). 1609 * Benchmark for tf.scan in graph and eager modes. 1610 * Added complex128 support to FFT, FFT2D, FFT3D, IFFT, IFFT2D, and IFFT3D. 1611 * Making ids unique in `nn.embedding_lookup_sparse`. This helps to reduce 1612 RPC calls for looking up the embeddings when there are repeated ids in 1613 the batch. 1614 * Support indicator column in boosted trees. 1615 * Prevent `tf.gradients()` from backpropagating through integer tensors. 1616 * LinearOperator[1D,2D,3D]Circulant added to `tensorflow.linalg`. 1617 * Conv3D, Conv3DBackpropInput, Conv3DBackpropFilter now supports 1618 arbitrary. 1619 * Added `tf.train.Checkpoint` for reading/writing object-based 1620 checkpoints. 1621 * Added LinearOperatorKronecker, a dense-free implementation of the 1622 Kronecker Product. 1623 * Allow LinearOperator to broadcast. 1624 * SavedModelBuilder will now deduplicate asset names that point to files 1625 with the same basename and the same contents. Note that this may result 1626 in new asset files included in SavedModels in cases where assets with 1627 the same name but different contents were previously overwriting each 1628 other. 1629 1630## Thanks to our Contributors 1631 1632This release contains contributions from many people at Google, as well as: 1633 1634Abdullah 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 1635 1636# Release 1.8.0 1637 1638## Major Features And Improvements 1639* Can now pass `tf.contrib.distribute.MirroredStrategy()` to `tf.estimator.RunConfig()` to run an Estimator model on multiple GPUs on one machine. 1640* Add `tf.contrib.data.prefetch_to_device()`, which supports prefetching to GPU memory. 1641* Added Gradient Boosted Trees as pre-made Estimators: BoostedTreesClassifier, BoostedTreesRegressor. 1642* Add 3rd generation pipeline config for Cloud TPUs which improves performance and usability. 1643* `tf.contrib.bayesflow` is moving out to it's own repo. 1644* Added `tf.contrib.{proto,rpc}` to allow generic proto parsing and RPC communication<sup>[1](#rpc-issue)</sup>. 1645 1646## Bug Fixes and Other Changes 1647* `tf.data`: 1648 * Add `tf.contrib.data.prefetch_to_device`, which enables prefetching dataset elements to GPU memory. 1649 * Add `tf.contrib.data.AUTOTUNE`, which allows the tf.data runtime to automatically tune the prefetch buffer sizes based on your system and environment. 1650 * Add `tf.contrib.data.make_csv_dataset` for building datasets of CSV files. 1651* Eager Execution: 1652 * 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. 1653 * 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) 1654 * `tf.GradientTape` has moved out of contrib. 1655* `tf.keras`: 1656 * Added the fashion mnist dataset. 1657 * New data preprocessing functions: `image/random_brightness`, `sequence/TimeseriesGenerator`, and `text/hashing_trick`. 1658* Accelerated Linear Algebra (XLA): 1659 * Select and scatter in reference util and evaluator now use lexicographical order to break ties. 1660* TensorFlow Debugger (tfdbg) CLI: 1661 * During tensor-filter operations, allow exclusion of nodes by regular expressions. 1662 * Fix spurious background colors in some text terminals. 1663* `tf.contrib`: 1664 * Add meta-distribution BatchReshape which reshapes batch dimensions. 1665 * `tf.contrib.layers.recompute_grad` works for explicit gradient checkpointing on TPU. 1666 * Add `tf.contrib.framework.argsort`. 1667 * Allow `DNNBoostedTreeCombinedEstimator` to work with core versions of feature columns and losses. 1668 * Add non-linear image warping ops: `tf.contrib.image.sparse_image_warp`, `tf.contrib.image.dense_image_warp`, and `tf.contrib.image.interpolate_spline`. 1669 * Fix bug in `tf.contrib.opt.MultitaskOptimizerWrapper` where types of tensors were mismatched. 1670* Other: 1671 * 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](https://github.com/tensorflow/tensorflow/issues/new) if you find yourself using this escape hatch. 1672 * Add description of shapes and a pointer to tutorial notebook in `tf.distributions.Distribution`. 1673 * Update scatter operations: 1674 * Add `tf.scatter_min` and `tf.scatter_max` 1675 * Extend scatter operations to work with a scalar update parameter. 1676 * Move cuDNN RNN ops to core for use in TensorFlow codebase only. 1677 * Add `float64` support for `Conv2d`, `Conv2dBackpropInput`, and `Conv2dBackpropFilter`. 1678 * Add `float64` support for `AvgPool`/`AvgPoolGrad`. 1679 * Make graph name scope thread local so that they work correctly in multi-threaded environments. 1680 * Update nsync synchronization library to avoid slow primitives on Linux. 1681 * Removed need to put nsync/public on C include path when building custom ops. 1682 * Add `tf.image.psnr`, `tf.image.ssim`, `tf.image.ssim_multiscale`, `tf.image.image_gradients`, `tf.image.sobel_edges`. 1683 * Add links to https://js.tensorflow.org. 1684 * Fix non-uniformity of orthogonal matrices. 1685 * Fix bug where multi-image Estimator eval summaries were not displayed correctly. 1686 1687<a name="rpc-issue"><sup>1</sup></a> 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. 1688 1689## Thanks to our Contributors 1690 1691This release contains contributions from many people at Google, as well as: 1692 16934d55397500, 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 1694 1695# Release 1.7.0 1696 1697## Major Features And Improvements 1698* Eager mode is moving out of contrib, try `tf.enable_eager_execution()`. 1699* Graph rewrites emulating fixed-point quantization compatible with TensorFlow Lite, supported by new `tf.contrib.quantize` package. 1700* Easily customize gradient computation with `tf.custom_gradient`. 1701* [TensorBoard Debugger Plugin](https://github.com/tensorflow/tensorboard/blob/master/tensorboard/plugins/debugger/README.md), the graphical user interface (GUI) of TensorFlow Debugger (tfdbg), is now in alpha. 1702* Experimental support for reading a sqlite database as a `Dataset` with new `tf.contrib.data.SqlDataset`. 1703* Distributed Mutex / CriticalSection added to `tf.contrib.framework.CriticalSection`. 1704* Better text processing with `tf.regex_replace`. 1705* Easy, efficient sequence input with `tf.contrib.data.bucket_by_sequence_length` 1706* Initial support for `tf.contrib.tensorrt` that enables native TensorRT in 1707 TensorFlow. 1708 1709## Bug Fixes and Other Changes 1710* Accelerated Linear Algebra (XLA): 1711 * Add `MaxPoolGradGrad` support for XLA 1712 * CSE pass from Tensorflow is now disabled in XLA. 1713* `tf.data`: 1714 * `tf.data.Dataset` 1715 * Add support for building C++ Dataset op kernels as external libraries, using the `tf.load_op_library()` mechanism. 1716 * `Dataset.list_files()` now shuffles its output by default. 1717 * `Dataset.shuffle(..., seed=tf.constant(0, dtype=tf.int64))` now yields the same sequence of elements as `Dataset.shuffle(..., seed=0)`. 1718 * Add `num_parallel_reads` argument to `tf.data.TFRecordDataset`. 1719* `tf.contrib`: 1720 * `tf.contrib.bayesflow.halton_sequence` now supports randomization. 1721 * Add support for scalars in `tf.contrib.all_reduce`. 1722 * Add `effective_sample_size` to `tf.contrib.bayesflow.mcmc_diagnostics`. 1723 * Add `potential_scale_reduction` to `tf.contrib.bayesflow.mcmc_diagnostics`. 1724 * Add `BatchNormalization`, `Kumaraswamy` bijectors. 1725 * Deprecate `tf.contrib.learn`. Please check contrib/learn/README.md for instructions on how to convert existing code. 1726 * `tf.contrib.data` 1727 * Remove deprecated `tf.contrib.data.Dataset`, `tf.contrib.data.Iterator`, `tf.contrib.data.FixedLengthRecordDataset`, `tf.contrib.data.TextLineDataset`, and `tf.contrib.data.TFRecordDataset` classes. 1728 * Added `bucket_by_sequence_length`, `sliding_window_batch`, and `make_batched_features_dataset` 1729 * Remove unmaintained `tf.contrib.ndlstm`. You can find it externally at https://github.com/tmbarchive/tfndlstm. 1730 * Moved most of `tf.contrib.bayesflow` to its own repo: `tfp` 1731* Other: 1732 * tf.py_func now reports the full stack trace if an exception occurs. 1733 * Integrate `TPUClusterResolver` with GKE's integration for Cloud TPUs. 1734 * Add a library for statistical testing of samplers. 1735 * Add Helpers to stream data from the GCE VM to a Cloud TPU. 1736 * Integrate ClusterResolvers with TPUEstimator. 1737 * Unify metropolis_hastings interface with HMC kernel. 1738 * Move LIBXSMM convolutions to a separate --define flag so that they are disabled by default. 1739 * Fix `MomentumOptimizer` lambda. 1740 * Reduce `tfp.layers` boilerplate via programmable docstrings. 1741 * Add `auc_with_confidence_intervals`, a method for computing the AUC and confidence interval with linearithmic time complexity. 1742 * `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. 1743 * Fix `initialized_value` and `initial_value` behaviors for `ResourceVariables` created from `VariableDef` protos. 1744 * Add TensorSpec to represent the specification of Tensors. 1745 * Constant folding pass is now deterministic. 1746 * Support `float16` `dtype` in `tf.linalg.*`. 1747 * Add `tf.estimator.export.TensorServingInputReceiver` that allows `tf.estimator.Estimator.export_savedmodel` to pass raw tensors to model functions. 1748 1749## Deprecations 1750 1751* TensorFlow 1.7 may be the last time we support Cuda versions below 8.0. 1752 Starting with TensorFlow 1.8 release, 8.0 will be the minimum supported 1753 version. 1754* TensorFlow 1.7 may be the last time we support cuDNN versions below 6.0. 1755 Starting with TensorFlow 1.8 release, 6.0 will be the minimum supported 1756 version. 1757 1758## Thanks to our Contributors 1759 1760This release contains contributions from many people at Google, as well as: 1761 17624d55397500, 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 1763 1764 1765 1766# Release 1.6.0 1767 1768## Breaking Changes 1769* Prebuilt binaries are now built against CUDA 9.0 and cuDNN 7. 1770* Prebuilt binaries will use AVX instructions. This may break TF on older CPUs. 1771 1772## Major Features And Improvements 1773* New Optimizer internal API for non-slot variables. Descendants of AdamOptimizer that access _beta[12]_power will need to be updated. 1774* `tf.estimator.{FinalExporter,LatestExporter}` now export stripped SavedModels. This improves forward compatibility of the SavedModel. 1775* FFT support added to XLA CPU/GPU. 1776 1777## Bug Fixes and Other Changes 1778* Documentation updates: 1779 * Added a second version of Getting Started, which is aimed at ML 1780newcomers. 1781 * Clarified documentation on `resize_images.align_corners` parameter. 1782 * Additional documentation for TPUs. 1783* Google Cloud Storage (GCS): 1784 * Add client-side throttle. 1785 * Add a `FlushCaches()` method to the FileSystem interface, with an implementation for GcsFileSystem. 1786* Other: 1787 * Add `tf.contrib.distributions.Kumaraswamy`. 1788 * `RetryingFileSystem::FlushCaches()` calls the base FileSystem's `FlushCaches()`. 1789 * Add `auto_correlation` to distributions. 1790 * Add `tf.contrib.distributions.Autoregressive`. 1791 * Add SeparableConv1D layer. 1792 * Add convolutional Flipout layers. 1793 * When both inputs of `tf.matmul` are bfloat16, it returns bfloat16, instead of float32. 1794 * Added `tf.contrib.image.connected_components`. 1795 * Add `tf.contrib.framework.CriticalSection` that allows atomic variable access. 1796 * Output variance over trees predictions for classifications tasks. 1797 * For `pt` and `eval` commands, allow writing tensor values to filesystem as numpy files. 1798 * gRPC: Propagate truncated errors (instead of returning gRPC internal error). 1799 * Augment `parallel_interleave` to support 2 kinds of prefetching. 1800 * Improved XLA support for C64-related ops log, pow, atan2, tanh. 1801 * Add probabilistic convolutional layers. 1802 1803## API Changes 1804* Introducing `prepare_variance` boolean with default setting to False for backward compatibility. 1805* Move `layers_dense_variational_impl.py` to `layers_dense_variational.py`. 1806 1807## Known Bugs 1808* Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or 1809 `CUDA_ILLEGAL_ADDRESS` failures. 1810 1811 Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA 9 1812 and CUDA 9.1 sometimes does not properly compute the carry bit when 1813 decomposing 64-bit address calculations with large offsets (e.g. `load [x + 1814 large_constant]`) into 32-bit arithmetic in SASS. 1815 1816 As a result, these versions of `ptxas` miscompile most XLA programs which use 1817 more than 4GB of temp memory. This results in garbage results and/or 1818 `CUDA_ERROR_ILLEGAL_ADDRESS` failures. 1819 1820 A fix in CUDA 9.1.121 is expected in late February 2018. We do not expect a 1821 fix for CUDA 9.0.x. Until the fix is available, the only workaround is to 1822 [downgrade](https://developer.nvidia.com/cuda-toolkit-archive) to CUDA 8.0.x 1823 or disable XLA:GPU. 1824 1825 TensorFlow will print a warning if you use XLA:GPU with a known-bad version of 1826 CUDA; see e00ba24c4038e7644da417ddc639169b6ea59122. 1827 1828## Thanks to our Contributors 1829 1830This release contains contributions from many people at Google, as well as: 1831 18324d55397500, Ag Ramesh, Aiden Scandella, Akimasa Kimura, Alex Rothberg, Allen Goodman, 1833amilioto, Andrei Costinescu, Andrei Nigmatulin, Anjum Sayed, Anthony Platanios, 1834Anush Elangovan, Armando Fandango, Ashish Kumar Ram, Ashwini Shukla, Ben, Bhavani Subramanian, 1835Brett Koonce, Carl Thomé, cclauss, Cesc, Changming Sun, Christoph Boeddeker, Clayne Robison, 1836Clemens Schulz, Clint (Woonhyuk Baek), codrut3, Cole Gerdemann, Colin Raffel, Daniel Trebbien, 1837Daniel Ylitalo, Daniel Zhang, Daniyar, Darjan Salaj, Dave Maclachlan, David Norman, Dong--Jian, 1838dongsamb, dssgsra, Edward H, eladweiss, elilienstein, Eric Lilienstein, error.d, Eunji Jeong, fanlu, 1839Florian Courtial, fo40225, Fred, Gregg Helt, Guozhong Zhuang, Hanchen Li, hsm207, hyunyoung2, 1840ImSheridan, Ishant Mrinal Haloi, Jacky Ko, Jay Young, Jean Flaherty, Jerome, JerrikEph, Jesse 1841Kinkead, jfaath, Jian Lin, jinghuangintel, Jiongyan Zhang, Joel Hestness, Joel Shor, Johnny Chan, 1842Julian Niedermeier, Julian Wolff, JxKing, K-W-W, Karl Lessard, Kasper Marstal, Keiji Ariyama, 1843Koan-Sin Tan, Loki Der Quaeler, Loo Rong Jie, Luke Schaefer, Lynn Jackson, ManHyuk, Matt Basta, 1844Matt Smith, Matthew Schulkind, Michael, michaelkhan3, Miguel Piedrafita, Mikalai Drabovich, 1845Mike Knapp, mjwen, mktozk, Mohamed Aly, Mohammad Ashraf Bhuiyan, Myungjoo Ham, Naman Bhalla, 1846Namrata-Ibm, Nathan Luehr, nathansilberman, Netzeband, Niranjan Hasabnis, Omar Aflak, Ozge 1847Yalcinkaya, Parth P Panchal, patrickzzy, Patryk Chrabaszcz, Paul Van Eck, Paweł Kapica, Peng Yu, 1848Philip Yang, Pierre Blondeau, Po-Hsien Chu, powderluv, Puyu Wang, Rajendra Arora, Rasmus, Renat 1849Idrisov, resec, Robin Richtsfeld, Ronald Eddy Jr, Sahil Singh, Sam Matzek, Sami Kama, sandipmgiri, 1850Santiago Castro, Sayed Hadi Hashemi, Scott Tseng, Sergii Khomenko, Shahid, Shengpeng Liu, Shreyash 1851Sharma, Shrinidhi Kl, Simone Cirillo, simsicon, Stanislav Levental, starsblinking, Stephen Lumenta, 1852Steven Hickson, Su Tang, Taehoon Lee, Takuya Wakisaka, Ted Chang, Ted Ying, Tijmen Verhulsdonck, 1853Timofey Kondrashov, vade, vaibhav, Valentin Khrulkov, vchigrin, Victor Costan, Viraj Navkal, 1854Vivek Rane, wagonhelm, Yan Facai (颜发才), Yanbo Liang, Yaroslav Bulatov, yegord, Yong Tang, 1855Yoni Tsafir, yordun, Yuan (Terry) Tang, Yuxin Wu, zhengdi, Zhengsheng Wei, 田传武 1856 1857# Release 1.5.0 1858 1859## Breaking Changes 1860* Prebuilt binaries are now built against CUDA 9.0 and cuDNN 7. 1861* Starting from 1.6 release, our prebuilt binaries will use AVX instructions. 1862 This may break TF on older CPUs. 1863 1864## Major Features And Improvements 1865* [Eager execution](https://github.com/tensorflow/tensorflow/tree/r1.5/tensorflow/contrib/eager) 1866 preview version is now available. 1867* [TensorFlow Lite](https://github.com/tensorflow/tensorflow/tree/r1.5/tensorflow/lite) 1868 dev preview is now available. 1869* CUDA 9.0 and cuDNN 7 support. 1870* Accelerated Linear Algebra (XLA): 1871 * Add `complex64` support to XLA compiler. 1872 * `bfloat` support is now added to XLA infrastructure. 1873 * Make `ClusterSpec` propagation work with XLA devices. 1874 * Use a deterministic executor to generate XLA graph. 1875* `tf.contrib`: 1876 * `tf.contrib.distributions`: 1877 * Add `tf.contrib.distributions.Autoregressive`. 1878 * Make `tf.contrib.distributions` QuadratureCompound classes support batch 1879 * Infer `tf.contrib.distributions.RelaxedOneHotCategorical` `dtype` from arguments. 1880 * Make `tf.contrib.distributions` quadrature family parameterized by 1881 `quadrature_grid_and_prob` vs `quadrature_degree`. 1882 * `auto_correlation` added to `tf.contrib.distributions` 1883 * Add `tf.contrib.bayesflow.layers`, a collection of probabilistic (neural) layers. 1884 * Add `tf.contrib.bayesflow.halton_sequence`. 1885 * Add `tf.contrib.data.make_saveable_from_iterator.` 1886 * Add `tf.contrib.data.shuffle_and_repeat`. 1887 * Add new custom transformation: `tf.contrib.data.scan()`. 1888 * `tf.contrib.distributions.bijectors`: 1889 * Add `tf.contrib.distributions.bijectors.MaskedAutoregressiveFlow`. 1890 * Add `tf.contrib.distributions.bijectors.Permute`. 1891 * Add `tf.contrib.distributions.bijectors.Gumbel`. 1892 * Add `tf.contrib.distributions.bijectors.Reshape`. 1893 * Support shape inference (i.e., shapes containing -1) in the Reshape bijector. 1894* Add `streaming_precision_recall_at_equal_thresholds,` a method for computing 1895 streaming precision and recall with `O(num_thresholds + size of predictions)` 1896 time and space complexity. 1897* Change `RunConfig` default behavior to not set a random seed, making random 1898 behavior independently random on distributed workers. We expect this to 1899 generally improve training performance. Models that do rely on determinism 1900 should set a random seed explicitly. 1901* Replaced the implementation of `tf.flags` with `absl.flags`. 1902* Add support for `CUBLAS_TENSOR_OP_MATH` in fp16 GEMM 1903* Add support for CUDA on NVIDIA Tegra devices 1904 1905## Bug Fixes and Other Changes 1906* Documentation updates: 1907 * Clarified that you can only install TensorFlow on 64-bit machines. 1908 * Added a short doc explaining how `Estimator`s save checkpoints. 1909 * Add documentation for ops supported by the `tf2xla` bridge. 1910 * Fix minor typos in the doc of `SpaceToDepth` and `DepthToSpace`. 1911 * Updated documentation comments in `mfcc_mel_filterbank.h` and `mfcc.h` to 1912 clarify that the input domain is squared magnitude spectra and the weighting 1913 is done on linear magnitude spectra (sqrt of inputs). 1914 * Change `tf.contrib.distributions` docstring examples to use `tfd` alias 1915 rather than `ds`, `bs`. 1916 * Fix docstring typos in `tf.distributions.bijectors.Bijector`. 1917 * `tf.assert_equal` no longer raises `ValueError.` It now raises 1918 `InvalidArgumentError,` as documented. 1919 * Update Getting Started docs and API intro. 1920* Google Cloud Storage (GCS): 1921 * Add userspace DNS caching for the GCS client. 1922 * Customize request timeouts for the GCS filesystem. 1923 * Improve GCS filesystem caching. 1924* Bug Fixes: 1925 * Fix bug where partitioned integer variables got their wrong shapes. Before 1926 * Fix correctness bug in CPU and GPU implementations of Adadelta. 1927 * Fix a bug in `import_meta_graph`'s handling of partitioned variables when 1928 importing into a scope. WARNING: This may break loading checkpoints of 1929 graphs with partitioned variables saved after using `import_meta_graph` with 1930 a non-empty `import_scope` argument. 1931 * Fix bug in offline debugger which prevented viewing events. 1932 * Added the `WorkerService.DeleteWorkerSession` method to the gRPC interface, 1933 to fix a memory leak. Ensure that your master and worker servers are running 1934 the same version of TensorFlow to avoid compatibility issues. 1935 * Fix bug in peephole implementation of BlockLSTM cell. 1936 * Fix bug by casting dtype of `log_det_jacobian` to match `log_prob` in 1937 `TransformedDistribution`. 1938 * Fix a bug in `import_meta_graph`'s handling of partitioned variables when 1939 * Ensure `tf.distributions.Multinomial` doesn't underflow in `log_prob`. 1940 Before this change, all partitions of an integer variable were initialized 1941 with the shape of the unpartitioned variable; after this change they are 1942 initialized correctly. 1943* Other: 1944 * Add necessary shape util support for bfloat16. 1945 * Add a way to run ops using a step function to MonitoredSession. 1946 * Add `DenseFlipout` probabilistic layer. 1947 * A new flag `ignore_live_threads` is available on train. If set to `True`, it 1948 will ignore threads that remain running when tearing down infrastructure 1949 after successfully completing training, instead of throwing a RuntimeError. 1950 * Restandardize `DenseVariational` as simpler template for other probabilistic 1951 layers. 1952 * `tf.data` now supports `tf.SparseTensor` components in dataset elements. 1953 * It is now possible to iterate over `Tensor`s. 1954 * Allow `SparseSegmentReduction` ops to have missing segment IDs. 1955 * Modify custom export strategy to account for multidimensional sparse float 1956 splits. 1957 * `Conv2D`, `Conv2DBackpropInput`, `Conv2DBackpropFilter` now supports arbitrary 1958 dilations with GPU and cuDNNv6 support. 1959 * `Estimator` now supports `Dataset`: `input_fn` can return a `Dataset` 1960 instead of `Tensor`s. 1961 * Add `RevBlock`, a memory-efficient implementation of reversible residual layers. 1962 * Reduce BFCAllocator internal fragmentation. 1963 * Add `cross_entropy` and `kl_divergence` to `tf.distributions.Distribution`. 1964 * Add `tf.nn.softmax_cross_entropy_with_logits_v2` which enables backprop 1965 w.r.t. the labels. 1966 * GPU back-end now uses `ptxas` to compile generated PTX. 1967 * `BufferAssignment`'s protocol buffer dump is now deterministic. 1968 * Change embedding op to use parallel version of `DynamicStitch`. 1969 * Add support for sparse multidimensional feature columns. 1970 * Speed up the case for sparse float columns that have only 1 value. 1971 * Allow sparse float splits to support multivalent feature columns. 1972 * Add `quantile` to `tf.distributions.TransformedDistribution`. 1973 * Add `NCHW_VECT_C` support for `tf.depth_to_space` on GPU. 1974 * Add `NCHW_VECT_C` support for `tf.space_to_depth` on GPU. 1975 1976## API Changes 1977* Rename `SqueezeDims` attribute to `Axis` in C++ API for Squeeze op. 1978* `Stream::BlockHostUntilDone` now returns Status rather than bool. 1979* Minor refactor: move stats files from `stochastic` to `common` and remove 1980 `stochastic`. 1981 1982## Known Bugs 1983* Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or 1984 `CUDA_ILLEGAL_ADDRESS` failures. 1985 1986 Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA 9 1987 and CUDA 9.1 sometimes does not properly compute the carry bit when 1988 decomposing 64-bit address calculations with large offsets (e.g. `load [x + 1989 large_constant]`) into 32-bit arithmetic in SASS. 1990 1991 As a result, these versions of `ptxas` miscompile most XLA programs which use 1992 more than 4GB of temp memory. This results in garbage results and/or 1993 `CUDA_ERROR_ILLEGAL_ADDRESS` failures. 1994 1995 A fix in CUDA 9.1.121 is expected in late February 2018. We do not expect a 1996 fix for CUDA 9.0.x. Until the fix is available, the only workaround is to 1997 [downgrade](https://developer.nvidia.com/cuda-toolkit-archive) to CUDA 8.0.x 1998 or disable XLA:GPU. 1999 2000 TensorFlow will print a warning if you use XLA:GPU with a known-bad version of 2001 CUDA; see e00ba24c4038e7644da417ddc639169b6ea59122. 2002 2003## Thanks to our Contributors 2004 2005This release contains contributions from many people at Google, as well as: 2006 2007Adam Zahran, Ag Ramesh, Alan Lee, Alan Yee, Alex Sergeev, Alexander, Amir H. Jadidinejad, 2008Amy, Anastasios Doumoulakis, Andrei Costinescu, Andrei Nigmatulin, Anthony Platanios, 2009Anush Elangovan, arixlin, Armen Donigian, ArtëM Sobolev, Atlas7, Ben Barsdell, Bill Prin, 2010Bo Wang, Brett Koonce, Cameron Thomas, Carl Thomé, Cem Eteke, cglewis, Changming Sun, 2011Charles Shenton, Chi-Hung, Chris Donahue, Chris Filo Gorgolewski, Chris Hoyean Song, 2012Chris Tava, Christian Grail, Christoph Boeddeker, cinqS, Clayne Robison, codrut3, concerttttt, 2013CQY, Dan Becker, Dan Jarvis, Daniel Zhang, David Norman, dmaclach, Dmitry Trifonov, 2014Donggeon Lim, dongpilYu, Dr. Kashif Rasul, Edd Wilder-James, Eric Lv, fcharras, Felix Abecassis, 2015FirefoxMetzger, formath, FredZhang, Gaojin Cao, Gary Deer, Guenther Schmuelling, Hanchen Li, 2016Hanmin Qin, hannesa2, hyunyoung2, Ilya Edrenkin, Jackson Kontny, Jan, Javier Luraschi, 2017Jay Young, Jayaram Bobba, Jeff, Jeff Carpenter, Jeremy Sharpe, Jeroen BéDorf, Jimmy Jia, 2018Jinze Bai, Jiongyan Zhang, Joe Castagneri, Johan Ju, Josh Varty, Julian Niedermeier, 2019JxKing, Karl Lessard, Kb Sriram, Keven Wang, Koan-Sin Tan, Kyle Mills, lanhin, LevineHuang, 2020Loki Der Quaeler, Loo Rong Jie, Luke Iwanski, LáSzló Csomor, Mahdi Abavisani, Mahmoud Abuzaina, 2021ManHyuk, Marek ŠUppa, MathSquared, Mats Linander, Matt Wytock, Matthew Daley, Maximilian Bachl, 2022mdymczyk, melvyniandrag, Michael Case, Mike Traynor, miqlas, Namrata-Ibm, Nathan Luehr, 2023Nathan Van Doorn, Noa Ezra, Nolan Liu, Oleg Zabluda, opensourcemattress, Ouwen Huang, 2024Paul Van Eck, peisong, Peng Yu, PinkySan, pks, powderluv, Qiao Hai-Jun, Qiao Longfei, 2025Rajendra Arora, Ralph Tang, resec, Robin Richtsfeld, Rohan Varma, Ryohei Kuroki, SaintNazaire, 2026Samuel He, Sandeep Dcunha, sandipmgiri, Sang Han, scott, Scott Mudge, Se-Won Kim, Simon Perkins, 2027Simone Cirillo, Steffen Schmitz, Suvojit Manna, Sylvus, Taehoon Lee, Ted Chang, Thomas Deegan, 2028Till Hoffmann, Tim, Toni Kunic, Toon Verstraelen, Tristan Rice, Urs KöSter, Utkarsh Upadhyay, 2029Vish (Ishaya) Abrams, Winnie Tsang, Yan Chen, Yan Facai (颜发才), Yi Yang, Yong Tang, 2030Youssef Hesham, Yuan (Terry) Tang, Zhengsheng Wei, zxcqwe4906, 张志豪, 田传武 2031 2032We are also grateful to all who filed issues or helped resolve them, asked and 2033answered questions, and were part of inspiring discussions. 2034 2035# Release 1.4.1 2036 2037## Bug Fixes and Other Changes 2038* `LinearClassifier` fix. 2039 2040# Release 1.4.0 2041 2042## Major Features And Improvements 2043* `tf.keras` is now part of the core TensorFlow API. 2044* [`tf.data`](http://tensorflow.org/guide/data) is now part of 2045 the core TensorFlow API. 2046 * The API is now subject to backwards compatibility guarantees. 2047 * For a guide to migrating from the `tf.contrib.data` API, see the 2048 [README](https://github.com/tensorflow/tensorflow/blob/r1.4/tensorflow/contrib/data/README.md). 2049 * Major new features include `Dataset.from_generator()` (for building an input 2050 pipeline from a Python generator), and the `Dataset.apply()` method for 2051 applying custom transformation functions. 2052 * Several custom transformation functions have been added, including 2053 `tf.contrib.data.batch_and_drop_remainder()` and 2054 `tf.contrib.data.sloppy_interleave()`. 2055* Add `train_and_evaluate` for simple distributed `Estimator` training. 2056* Add `tf.spectral.dct` for computing the DCT-II. 2057* Add Mel-Frequency Cepstral Coefficient support to `tf.contrib.signal` 2058 (with GPU and gradient support). 2059* Add a self-check on `import tensorflow` for Windows DLL issues. 2060* Add NCHW support to `tf.depth_to_space` on GPU. 2061* TensorFlow Debugger (tfdbg): 2062 * Add `eval` command to allow evaluation of arbitrary Python/numpy expressions 2063 in tfdbg command-line interface. See 2064 [Debugging TensorFlow Programs](https://www.tensorflow.org/guide/debugger) 2065 for more details. 2066 * Usability improvement: The frequently used tensor filter `has_inf_or_nan` is 2067 now added to `Session` wrappers and hooks by default. So there is no need 2068 for clients to call `.add_tensor_filter(tf_debug.has_inf_or_nan)` anymore. 2069* SinhArcsinh (scalar) distribution added to `contrib.distributions`. 2070* Make `GANEstimator` opensource. 2071* `Estimator.export_savedmodel()` now includes all valid serving signatures 2072 that can be constructed from the Serving Input Receiver and all available 2073 ExportOutputs. For instance, a classifier may provide regression- and 2074 prediction-flavored outputs, in addition to the classification-flavored one. 2075 Building signatures from these allows TF Serving to honor requests using the 2076 different APIs (Classify, Regress, and Predict). Furthermore, 2077 `serving_input_receiver_fn()` may now specify alternative subsets of nodes 2078 that may act as inputs. This allows, for instance, producing a prediction 2079 signature for a classifier that accepts raw `Tensors` instead of a serialized 2080 `tf.Example`. 2081* Add `tf.contrib.bayesflow.hmc`. 2082* Add `tf.contrib.distributions.MixtureSameFamily`. 2083* Make `Dataset.shuffle()` always reshuffles after each iteration by default. 2084* Add `tf.contrib.bayesflow.metropolis_hastings`. 2085* Add `log_rate` parameter to `tf.contrib.distributions.Poisson`. 2086* Extend `tf.contrib.distributions.bijector` API to handle some non-injective 2087 transforms. 2088* Java: 2089 * Generics (e.g., `Tensor<Integer>`) for improved type-safety 2090 (courtesy @andrewcmyers). 2091 * Support for multi-dimensional string tensors. 2092 * Support loading of custom operations (e.g. many in `tf.contrib`) on Linux 2093 and OS X 2094* All our prebuilt binaries have been built with CUDA 8 and cuDNN 6. 2095 We anticipate releasing TensorFlow 1.5 with CUDA 9 and cuDNN 7. 2096 2097## Bug Fixes and Other Changes 2098* `tf.nn.rnn_cell.DropoutWrapper` is now more careful about dropping out LSTM 2099 states. Specifically, it no longer ever drops the `c` (memory) state of an 2100 `LSTMStateTuple`. The new behavior leads to proper dropout behavior 2101 for LSTMs and stacked LSTMs. This bug fix follows recommendations from 2102 published literature, but is a behavioral change. State dropout behavior 2103 may be customized via the new `dropout_state_filter_visitor` argument. 2104* Removed `tf.contrib.training.python_input`. The same behavior, in a more 2105 flexible and reproducible package, is available via the new 2106 `tf.contrib.data.Dataset.from_generator` method! 2107* Fix `tf.contrib.distributions.Affine` incorrectly computing log-det-jacobian. 2108* Fix `tf.random_gamma` incorrectly handling non-batch, scalar draws. 2109* Resolved a race condition in TensorForest TreePredictionsV4Op. 2110* Google Cloud Storage file system, Amazon S3 file system, and Hadoop file 2111 system support are now default build options. 2112* Custom op libraries must link against libtensorflow_framework.so 2113 (installed at `tf.sysconfig.get_lib()`). 2114* Change `RunConfig` default behavior to not set a random seed, making random 2115 behavior independently random on distributed workers. We expect this to 2116 generally improve training performance. Models that do rely on determinism 2117 should set a random seed explicitly. 2118 2119## Breaking Changes to the API 2120* The signature of the `tf.contrib.data.rejection_resample()` function has been 2121 changed. It now returns a function that can be used as an argument to 2122 `Dataset.apply()`. 2123* Remove `tf.contrib.data.Iterator.from_dataset()` method. Use 2124 `Dataset.make_initializable_iterator()` instead. 2125* Remove seldom used and unnecessary `tf.contrib.data.Iterator.dispose_op()`. 2126* Reorder some TF-GAN loss functions in a non-backwards compatible way. 2127 2128## Known Issues 2129* In Python 3, `Dataset.from_generator()` does not support Unicode strings. 2130 You must convert any strings to bytes objects before yielding them from 2131 the generator. 2132 2133## Thanks to our Contributors 2134 2135This release contains contributions from many people at Google, as well as: 2136 21374d55397500, Abdullah Alrasheed, abenmao, Adam Salvail, Aditya Dhulipala, Ag Ramesh, 2138Akimasa Kimura, Alan Du, Alan Yee, Alexander, Amit Kushwaha, Amy, Andrei Costinescu, 2139Andrei Nigmatulin, Andrew Erlichson, Andrew Myers, Andrew Stepanov, Androbin, AngryPowman, 2140Anish Shah, Anton Daitche, Artsiom Chapialiou, asdf2014, Aseem Raj Baranwal, Ash Hall, 2141Bart Kiers, Batchu Venkat Vishal, ben, Ben Barsdell, Bill Piel, Carl Thomé, Catalin Voss, 2142Changming Sun, Chengzhi Chen, Chi Zeng, Chris Antaki, Chris Donahue, Chris Oelmueller, 2143Chris Tava, Clayne Robison, Codrut, Courtial Florian, Dalmo Cirne, Dan J, Darren Garvey, 2144David Kristoffersson, David Norman, David RöThlisberger, DavidNorman, Dhruv, DimanNe, 2145Dorokhov, Duncan Mac-Vicar P, EdwardDixon, EMCP, error.d, FAIJUL, Fan Xia, 2146Francois Xavier, Fred Reiss, Freedom" Koan-Sin Tan, Fritz Obermeyer, Gao, Xiang, 2147Guenther Schmuelling, Guo Yejun (郭叶军), Hans Gaiser, HectorSVC, Hyungsuk Yoon, 2148James Pruegsanusak, Jay Young, Jean Wanka, Jeff Carpenter, Jeremy Rutman, Jeroen BéDorf, 2149Jett Jones, Jimmy Jia, jinghuangintel, jinze1994, JKurland, Joel Hestness, joetoth, 2150John B Nelson, John Impallomeni, John Lawson, Jonas, Jonathan Dekhtiar, joshkyh, Jun Luan, 2151Jun Mei, Kai Sasaki, Karl Lessard, karl@kubx.ca, Kb Sriram, Kenichi Ueno, Kevin Slagle, 2152Kongsea, Lakshay Garg, lhlmgr, Lin Min, liu.guangcong, Loki Der Quaeler, Louie Helm, 2153lucasmoura, Luke Iwanski, Lyndon White, Mahmoud Abuzaina, Marcel Puyat, Mark Aaron Shirley, 2154Michele Colombo, MtDersvan, Namrata-Ibm, Nathan Luehr, Naurril, Nayana Thorat, Nicolas Lopez, 2155Niranjan Hasabnis, Nolan Liu, Nouce, Oliver Hennigh, osdamv, Patrik Erdes, 2156Patryk Chrabaszcz, Pavel Christof, Penghao Cen, postBG, Qingqing Cao, Qingying Chen, qjivy, 2157Raphael, Rasmi, raymondxyang, Renze Yu, resec, Roffel, Ruben Vereecken, Ryohei Kuroki, 2158sandipmgiri, Santiago Castro, Scott Kirkland, Sean Vig, Sebastian Raschka, Sebastian Weiss, 2159Sergey Kolesnikov, Sergii Khomenko, Shahid, Shivam Kotwalia, Stuart Berg, Sumit Gouthaman, 2160superzerg, Sven Mayer, tetris, Ti Zhou, Tiago Freitas Pereira, Tian Jin, Tomoaki Oiki, 2161Vaibhav Sood, vfdev, Vivek Rane, Vladimir Moskva, wangqr, Weber Xie, Will Frey, 2162Yan Facai (颜发才), yanivbl6, Yaroslav Bulatov, Yixing Lao, Yong Tang, youkaichao, 2163Yuan (Terry) Tang, Yue Zhang, Yuxin Wu, Ziming Dong, ZxYuan, 黄璞 2164 2165We are also grateful to all who filed issues or helped resolve them, asked and 2166answered questions, and were part of inspiring discussions. 2167 2168# Release 1.3.0 2169 2170See also [TensorBoard 0.1.4](https://github.com/tensorflow/tensorboard/releases/tag/0.1.4) release notes. 2171 2172## Major Features and Improvements 2173* Added canned estimators to Tensorflow library. List of added estimators: 2174 * `DNNClassifier` 2175 * `DNNRegressor` 2176 * `LinearClassifier` 2177 * `LinearRegressor` 2178 * `DNNLinearCombinedClassifier` 2179 * `DNNLinearCombinedRegressor`. 2180* All our prebuilt binaries have been built with cuDNN 6. We anticipate releasing TensorFlow 1.4 with cuDNN 7. 2181* `import tensorflow` now goes much faster. 2182* 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. 2183* Added an axis parameter to `tf.gather`. 2184* Added a `constant_values` keyword argument to `tf.pad`. 2185* Adds `Dataset.interleave` transformation. 2186* Add `ConcatenateDataset` to concatenate two datasets. 2187* Added Mobilenet support to TensorFlow for Poets training script. 2188* Adds a block cache to the GCS filesystem with configurable block size and count. 2189* SinhArcSinh bijector added. 2190* Added `Dataset.list_files` API. 2191* Introduces new operations and Python bindings for the Cloud TPU. 2192* Adding TensorFlow-iOS CocoaPod for symmetry with tensorflow-android. 2193* Introduces base implementations of ClusterResolvers. 2194* Unify memory representations of TensorShape and PartialTensorShape. As a consequence, tensors now have a maximum of 254 dimensions, not 255. 2195* Changed references to LIBXSMM to use version 1.8.1. 2196* TensorFlow Debugger (tfdbg): 2197 * Display summaries of numeric tensor values with the `-s` flag to command `print_tensor` or `pt`. 2198 * Display feed values with the `print_feed` or `pf` command and clickable links in the curses UI. 2199 * Runtime profiler at the op level and the Python source line level with the `run -p` command. 2200* Initial release of the statistical distribution library `tf.distributions`. 2201* GPU kernels and speed improvements for unary `tf.where` and `tf.nn.top_k`. 2202* Monotonic Attention wrappers added to `tf.contrib.seq2seq`. 2203* Added `tf.contrib.signal`, a library for signal processing primitives. 2204* Added `tf.contrib.resampler`, containing CPU and GPU ops for differentiable resampling of images. 2205 2206## Breaking Changes to the API 2207* `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. 2208* 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. 2209 2210## Changes to contrib APIs 2211* Adds tf.contrib.nn.rank_sampled_softmax_loss, a sampled-softmax variant that can improve rank loss. 2212* `tf.contrib.metrics`.{streaming_covariance,streaming_pearson_correlation} modified to return nan when they have seen less or equal to 1 unit of weight. 2213* Adds time series models to contrib. See contrib/timeseries/README.md for details. 2214* Adds FULLY_CONNECTED Op to tensorflow/lite/schema.fbs 2215 2216## Known Issues 2217* Tensorflow_gpu compilation fails with Bazel 0.5.3. 2218 2219## Bug Fixes and Other Changes 2220* Fixes `strides` and `begin` dtype mismatch when slicing using int64 Tensor index in python. 2221* Improved convolution padding documentation. 2222* Add a tag constant, gpu, to present graph with GPU support. 2223* `saved_model.utils` now support SparseTensors transparently. 2224* A more efficient implementation of non-max suppression. 2225* Add support for the shrinkage-type L2 to FtrlOptimizer in addition to the online L2 it already supports. 2226* Fix negative variance in moments calculation. 2227* Expand UniqueOp Benchmark Tests to cover more collision cases. 2228* Improves stability of GCS filesystem on Mac. 2229* Add time estimation to HloCostAnalysis. 2230* 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. 2231* Added None check for save_path in `saver.restore`. 2232* Register devices under their legacy names in device_mgr to ease the transition to clusterspec-propagated configurations. 2233* VectorExponential added to distributions. 2234* Add a bitwise module with bitwise_and, bitwise_or, bitwise_xor, and invert functions. 2235* Add fixed-grid ODE integration routines. 2236* Allow passing bounds to ScipyOptimizerInterface. 2237* Correctness fixes for fft_length parameter to `tf.spectral.rfft` & `tf.spectral.irfft`. 2238* 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. 2239* Add in-memory caching to the Dataset API. 2240* Set default end_of_sequence variable in datasets iterators to false. 2241* [Performance] Increase performance of `tf.layers.conv2d` when setting use_bias=True by 2x by using nn.bias_add. 2242* Update iOS examples to use CocoaPods, and moved to tensorflow/examples/ios. 2243* Adds a family= attribute in `tf.summary` ops to allow controlling the tab name used in Tensorboard for organizing summaries. 2244* When GPU is configured, do not require --config=cuda, instead, automatically build for GPU if this is requested in the configure script. 2245* Fix incorrect sampling of small probabilities in CPU/GPU multinomial. 2246* 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. 2247* Allow uses of over-parameterized separable convolution. 2248* TensorForest multi-regression bug fix. 2249* Framework now supports armv7, cocoapods.org now displays correct page. 2250* Script to create iOS framework for CocoaPods. 2251* 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. 2252* TensorFlow Debugger (tfdbg): 2253 * Fixed a bug that prevented tfdbg from functioning with multi-GPU setups. 2254 * Fixed a bug that prevented tfdbg from working with `tf.Session.make_callable`. 2255 2256## Thanks to our Contributors 2257 2258This release contains contributions from many people at Google, as well as: 2259 22604F2E4A2E, Adriano Carmezim, Adrià Arrufat, Alan Yee, Alex Lattas, Alex Rothberg, 2261Alexandr Baranezky, Ali Siddiqui, Andreas Solleder, Andrei Costinescu, Andrew Hundt, 2262Androbin, Andy Kernahan, Anish Shah, Anthony Platanios, Arvinds-Ds, b1rd, Baptiste 2263Arnaud, Ben Mabey, Benedikt Linse, Beomsu Kim, Bo Wang, Boyuan Deng, Brett Koonce, 2264Bruno Rosa, Carl Thomé, Changming Sun, Chase Roberts, Chirag Bhatia, Chris Antaki, 2265Chris Hoyean Song, Chris Tava, Christos Nikolaou, Croath Liu, cxx, Czxck001, Daniel 2266Ylitalo, Danny Goodman, Darren Garvey, David Brailovsky, David Norman, DavidNorman, 2267davidpham87, ddurham2, Dhruv, DimanNe, Drew Hintz, Dustin Tran, Earthson Lu, ethiraj, 2268Fabian Winnen, Fei Sun, Freedom" Koan-Sin Tan, Fritz Obermeyer, Gao, Xiang, Gautam, 2269Guenther Schmuelling, Gyu-Ho Lee, Hauke Brammer, horance, Humanity123, J Alammar, 2270Jayeol Chun, Jeroen BéDorf, Jianfei Wang, jiefangxuanyan, Jing Jun Yin, Joan Puigcerver, 2271Joel Hestness, Johannes Mayer, John Lawson, Johnson145, Jon Malmaud, Jonathan Alvarez-Gutierrez, 2272Juang, Yi-Lin, Julian Viereck, Kaarthik Sivashanmugam, Karl Lessard, karl@kubx.ca, Kevin 2273Carbone, Kevin Van Der Burgt, Kongsea, ksellesk, lanhin, Lef Ioannidis, Liangliang He, 2274Louis Tiao, Luke Iwanski, LáSzló Csomor, magixsno, Mahmoud Abuzaina, Marcel Hlopko, Mark 2275Neumann, Maxwell Paul Brickner, mdfaijul, MichaëL Defferrard, Michał JastrzęBski, Michele 2276Colombo, Mike Brodie, Mosnoi Ion, mouradmourafiq, myPrecious, Nayana Thorat, 2277Neeraj Kashyap, Nelson Liu, Niranjan Hasabnis, Olivier Moindrot, orome, Pankaj Gupta, Paul 2278Van Eck, peeyush18, Peng Yu, Pierre, preciousdp11, qjivy, Raingo, raoqiyu, ribx, Richard S. 2279Imaoka, Rishabh Patel, Robert Walecki, Rockford Wei, Ryan Kung, Sahil Dua, Sandip Giri, Sayed 2280Hadi Hashemi, sgt101, Shitian Ni, Shuolongbj, Siim PõDer, Simon Perkins, sj6077, SOLARIS, 2281Spotlight0xff, Steffen Eberbach, Stephen Fox, superryanguo, Sven Mayer, Tapan Prakash, 2282Tiago Morais Morgado, Till Hoffmann, Tj Rana, Vadim Markovtsev, vhasanov, Wei Wu, 2283windead, Yan (Asta) Li, Yan Chen, Yann Henon, Yi Wang, Yong Tang, yorkie, Yuan (Terry) 2284Tang, Yuxin Wu, zhengjiajin, zhongzyd, 黄璞 2285 2286We are also grateful to all who filed issues or helped resolve them, asked and 2287answered questions, and were part of inspiring discussions. 2288 2289# Release 1.2.1 2290 2291## Bug Fixes and Other Changes 2292* Updating markdown version required to >= 2.6.8. 2293* Support tensors as dropout rates again, by removing the min(max(..)) 2294 2295# Release 1.2.0 2296 2297## Major Features and Improvements 2298* Python 3.6 support on Windows. 2299* Added `tf.layers.conv3d_transpose` layer for spatio temporal deconvolution. 2300* Added `tf.Session.make_callable()`, which provides a lower overhead means of running a similar step multiple times. 2301* Added libverbs-based RDMA support to contrib (courtesy @junshi15 from Yahoo). 2302* Bring `tf.feature_column.*` into the API. Non-deprecated functionality from `tf.contrib.layers.*` is moved to `tf.feature_column.*` with cosmetic changes. 2303* `RNNCell` objects now subclass `tf.layers.Layer`. The strictness described 2304 in the TensorFlow 1.1 release is gone: The first time an RNNCell is used, 2305 it caches its scope. All future uses of the RNNCell will reuse variables from 2306 that same scope. This is a breaking change from the behavior of RNNCells 2307 in TensorFlow versions <= 1.0.1. TensorFlow 1.1 had checks in place to 2308 ensure old code works correctly with the new semantics; this version 2309 allows more flexible uses of RNNCell but can lead to subtle errors if 2310 using code meant for TensorFlow <= 1.0.1. For example, writing: 2311 `MultiRNNCell([lstm] * 5)` will now build a 5-layer LSTM stack where each 2312 layer shares the **same** parameters. To get 5 layers each with their own 2313 parameters, write: `MultiRNNCell([LSTMCell(...) for _ in range(5)])`. 2314 If at all unsure, first test your code with TF 1.1; ensure it raises no 2315 errors, and then upgrade to TF 1.2. 2316* RNNCells' variable names have been renamed for consistency with Keras layers. 2317 Specifically, the previous variable names "weights" and "biases" have 2318 been changed to "kernel" and "bias", respectively. 2319 This may cause backward incompatibility with regard to your old 2320 checkpoints containing such RNN cells, in which case you can use the tool 2321 [checkpoint_convert script](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/rnn/python/tools/checkpoint_convert.py) 2322 to convert the variable names in your old checkpoints. 2323* Many of the RNN functions and classes that were in the `tf.nn` namespace 2324 before the 1.0 release and which were moved to `tf.contrib.rnn` have now 2325 been moved back to the core namespace. This includes 2326 `RNNCell`, `LSTMCell`, `GRUCell`, and a number of other cells. These 2327 now reside in `tf.nn.rnn_cell` (with aliases in `tf.contrib.rnn` for backwards 2328 compatibility). The original `tf.nn.rnn` function is now `tf.nn.static_rnn`, 2329 and the bidirectional static and state saving static rnn functions are also 2330 now back in the `tf.nn` namespace. 2331 2332 Notable exceptions are the `EmbeddingWrapper`, `InputProjectionWrapper` and 2333 `OutputProjectionWrapper`, which will slowly be moved to deprecation 2334 in `tf.contrib.rnn`. These are inefficient wrappers that should often 2335 be replaced by calling `embedding_lookup` or `layers.dense` as pre- or post- 2336 processing of the rnn. For RNN decoding, this functionality has been replaced 2337 with an alternative API in `tf.contrib.seq2seq`. 2338* Intel MKL Integration (https://software.intel.com/en-us/articles/tensorflow-optimizations-on-modern-intel-architecture). Intel developed a number of 2339 optimized deep learning primitives: In addition to matrix multiplication and 2340 convolution, these building blocks include: 2341 Direct batched convolution 2342 Pooling: maximum, minimum, average 2343 Normalization: LRN, batch normalization 2344 Activation: rectified linear unit (ReLU) 2345 Data manipulation: multi-dimensional transposition (conversion), split, 2346 concat, sum and scale. 2347* TensorForest Estimator now supports SavedModel export for serving. 2348* Support client-provided ClusterSpec's and propagate them to all workers to enable the creation of dynamic TensorFlow clusters. 2349* TensorFlow C library now available for Windows. 2350* We released a new open-source version of TensorBoard. 2351* [`SavedModel CLI`](https://www.tensorflow.org/versions/master/guide/saved_model_cli) tool available to inspect and execute MetaGraph in SavedModel 2352* Android releases of TensorFlow are now pushed to jcenter for easier 2353 integration into apps. See 2354 https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/android/inference_interface/README.md 2355 for more details. 2356 2357## Deprecations 2358 2359* TensorFlow 1.2 may be the last time we build with cuDNN 5.1. Starting with 2360 TensorFlow 1.3, we will try to build all our prebuilt binaries with cuDNN 6.0. 2361 While we will try to keep our source code compatible with cuDNN 5.1, it will 2362 be best effort. 2363 2364## Breaking Changes to the API 2365* `org.tensorflow.contrib.android.TensorFlowInferenceInterface` now throws exceptions where possible and has simplified method signatures. 2366 2367## Changes to contrib APIs 2368* Added `tf.contrib.util.create_example`. 2369* Added bilinear interpolation to `tf.contrib.image`. 2370* Add `tf.contrib.stateless` for random ops with custom seed control. 2371* MultivariateNormalFullCovariance added to contrib/distributions/ 2372* tensorflow/contrib/rnn undergoes RNN cell variable renaming for 2373 consistency with Keras layers. Specifically, the previous variable names 2374 "weights" and "biases" are changed to "kernel" and "bias", respectively. 2375 This may cause backward incompatibility with regard to your old 2376 checkpoints containing such RNN cells, in which case you can use the 2377 [checkpoint_convert script](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/rnn/python/tools/checkpoint_convert.py) 2378 to convert the variable names in your old checkpoints. 2379* Added `tf.contrib.kernel_methods` module with Ops and estimators for primal 2380 (explicit) kernel methods in TensorFlow. 2381 2382## Bug Fixes and Other Changes 2383* In python, `Operation.get_attr` on type attributes returns the Python DType 2384 version of the type to match expected get_attr documentation rather than the 2385 protobuf enum. 2386* tensorflow/contrib/rnn undergoes RNN cell variable renaming for 2387 consistency with Keras layers. Specifically, the previous variable names 2388 "weights" and "biases" are changed to "kernel" and "bias", respectively. 2389* Changed MIN_SDK version to 8.0 when building iOS libraries. 2390* Fixed LIBXSMM integration. 2391* Make decode_jpeg/decode_png/decode_gif handle all formats, since users frequently try to decode an image as the wrong type. 2392* Improve implicit broadcasting lowering. 2393* Improving stability of GCS/BigQuery clients by a faster retrying of stale transmissions. 2394* Remove OpKernelConstruction::op_def() as part of minimizing proto dependencies. 2395* VectorLaplaceDiag distribution added. 2396* Android demo no longer requires libtensorflow_demo.so to run (libtensorflow_inference.so still required) 2397* Added `categorical_column_with_vocabulary_file`. 2398* Introduce ops for batching/unbatching tensors across Session::Run() calls. 2399* Add tf.log_sigmoid(x) = tf.log(tf.sigmoid(x)) = -tf.nn.softplus(-x). 2400* Changed hooks lists to immutable tuples, and now allow any iterable for the associated arguments. 2401* Introduce TFDecorator. 2402* Added an Mfcc op for speech feature generation. 2403* 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. 2404* Added unreduced NONE, and reduced MEAN options for losses. Removed "WEIGHTED_" prefix from other Reduction constants. 2405* assertAllClose now handles dicts. 2406* Added Gmock matcher for HloInstructions. 2407* Add var name to errors on variable restore. 2408* Added an AudioSpectrogram op for audio feature generation. 2409* Added `reduction` arg to losses. 2410* `tf.placeholder` can represent scalar shapes and partially known. 2411* Remove estimator_spec(mode) argument. 2412* Added an AudioSpectrogram op for audio feature generation. 2413* TensorBoard disables all runs by default if there are more than 40 runs. 2414* Removed old doc generator code. 2415* GCS file system integration now supports domain buckets, e.g gs://bucket.domain.com/path. 2416* Add `tf.summary.text` for outputting text to TensorBoard. 2417* The "run" command of tfdbg's command-line interface now supports filtering of tensors by node name, op type and tensor dtype. 2418* `tf.string_to_number` now supports int64 and float64 outputs. 2419 2420## Thanks to our Contributors 2421 2422This release contains contributions from many people at Google, as well as: 2423 24244F2E4A2E, Aaron Schumacher, Abhi Agg, admcrae, Adriano Carmezim, Adrià Arrufat, 2425agramesh1, Akimitsu Seo, Alan Mosca, Alex Egg, Alex Rothberg, Alexander Heinecke, 2426Alexander Matyasko, Alexandr Baranezky, Alexandre Caulier, Ali Siddiqui, Anand Venkat, 2427Andrew Hundt, Androbin, Anmol Sharma, Arie, Arno Leist, Arron Cao, AuréLien Geron, Bairen Yi, 2428Beomsu Kim, Carl Thomé, cfperez, Changming Sun, Corey Wharton, critiqjo, Dalei Li, Daniel 2429Rasmussen, Daniel Trebbien, DaríO Hereñú, David Eng, David Norman, David Y. Zhang, Davy Song, ddurham2, 2430Deepak Subburam, Dmytro Kyrychuk, Dominic Rossi, Dominik SchlöSser, Dustin Tran, 2431Eduardo Pinho, Egil Martinsson, Elliot Saba, Eric Bigelow, Erik Smistad, Evan Klitzke, 2432Fabrizio Milo, Falcon Dai, Fei Gao, FloopCZ, Fung Lam, Gautam, GBLin5566, Greg Peatfield, 2433Gu Wang, Guenther Schmuelling, Hans Pabst, Harun Gunaydin, Huaizheng, Ido Shamay, Ikaro 2434Silva, Ilya Edrenkin, Immexxx, James Mishra, Jamie Cooke, Jay Young, Jayaram Bobba, 2435Jianfei Wang, jinghua2, Joey Meyer, John Maidens, Jonghoon Jin, Julian Villella, 2436Jun Kim, Jun Shi, Junwei Pan, jyegerlehner, Karan Desai, Karel Van De Plassche, 2437Kb Sriram, KhabarlakKonstantin, Koan-Sin Tan, krivard, Kwotsin, Leandro Gracia Gil, 2438Li Chen, Liangliang He, Louie Helm, lspvic, Luiz Henrique Soares, LáSzló Csomor, 2439Mark Wong, Mathew Wicks, Matthew Rahtz, Maxwell Paul Brickner, Michael Hofmann, Miguel 2440Flores Ruiz De Eguino, MikeTam1021, Mortada Mehyar, Mycosynth, Namnamseo, 2441Nate Harada, Neven Miculinic, Nghia Tran, Nick Lyu, Niranjan Hasabnis, Nishidha, Oleksii 2442Kuchaiev, Oyesh Mann Singh, Panmari, Patrick, Paul Van Eck, Piyush Chaudhary, Quim Llimona, 2443Raingo, Richard Davies, Ruben Vereecken, Sahit Chintalapudi, Sam Abrahams, Santiago Castro, 2444Scott Sievert, Sean O'Keefe, Sebastian Schlecht, Shane, Shubhankar Deshpande, Spencer Schaber, 2445Sunyeop Lee, t13m, td2014, Thomas H. P. Andersen, Toby Petty, Umang Mehta, 2446Vadim Markovtsev, Valentin Iovene, Vincent Zhao, Vit Stepanovs, Vivek Rane, Vu Pham, wannabesrevenge, 2447weipingpku, wuhaixutab, wydwww, Xiang Gao, Xiaolin Lin, xiaoyaozhuzi, Yaroslav Bulatov, Yi Liu, 2448Yoshihiro Sugi, Yuan (Terry) Tang, Yuming Wang, Yuxin Wu, Zader Zheng, Zhaojun Zhang, zhengjiajin, 2449ZhipengShen, Ziming Dong, zjj2wry 2450 2451We are also grateful to all who filed issues or helped resolve them, asked and 2452answered questions, and were part of inspiring discussions. 2453 2454# Release 1.1.0 2455 2456## Major Features and Improvements 2457* Added Java API support for Windows. 2458* Added `tf.spectral` module. Moved existing FFT ops to `tf.spectral` while 2459 keeping an alias in the old location (`tf.*`). 2460* Added 1D, 2D and 3D Fourier transform ops for real signals to `tf.spectral`. 2461* Added a `tf.bincount` function. 2462* Added Keras 2 API to contrib. 2463* Added a new lightweight queue-like object - `RecordInput`. 2464* Added `tf.contrib.image.compose_transforms` function. 2465* Bring `tf.estimator.*` into the API. Non-deprecated functionality from `tf.contrib.learn.Estimator` is moved to `tf.estimator.Estimator` with cosmetic changes. 2466* Docker images: TF images on gcr.io and Docker Hub are upgraded to ubuntu:16.04. 2467* Added the following features to TensorFlow Debugger (tfdbg): 2468 * Ability to inspect Python source file against TF ops and tensors (command `print_source` / `ps`) 2469 * New navigation bar in Curses-based UI 2470 * 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. 2471* Initial release of installation guides for Java, C, and Go. 2472* Added Text Dashboard to TensorBoard. 2473 2474## Deprecations 2475 2476* 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. 2477 2478## Changes to contrib APIs 2479* The behavior of RNNCells is now stricter due to the transition towards making RNNCells act more like Keras layers. 2480 * If an RNNCell is used twice in two different variable scopes, an error is raised describing how to avoid this behavior. 2481 * 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`. 2482* Deprecated contrib/distributions `pmf`, `pdf`, `log_pmf`, `log_pdf`. 2483* Moved `bayesflow.special_math` to distributions. 2484* `tf.contrib.tensor_forest.python.tensor_forest.RandomForestDeviceAssigner` removed. 2485* Changed some MVN classes and parameters: 2486 * `tf.contrib.distributions.MultivariateNormalFull` replaced by `tf.contrib.distributions.MultivariateNormalTriL`. 2487 * `tf.contrib.distributions.MultivariateNormalCholesky` replaced by `tf.contrib.distributions.MultivariateNormalTriL` 2488 * `tf.contrib.distributions.MultivariateNormalDiagWithSoftplusStDev` replaced 2489 by `tf.contrib.distributions.MultivariateNormalDiagWithSoftplusScale` 2490 * `tf.contrib.distributions.MultivariateNormalDiag` arguments changed from `mu`, `diag_stddev` to `log`, `scale_diag`. 2491 * `tf.contrib.distributions.MultivariateNormalDiagPlusVDVT` removed. 2492 * `tf.contrib.distributions.MultivariateNormalDiagPlusLowRank` added. 2493 2494## Bug Fixes and Other Changes 2495* Java: Support for loading models exported using the SavedModel API (courtesy @EronWright). 2496* Go: Added support for incremental graph execution. 2497* Fix a bug in the WALS solver when single-threaded. 2498* Added support for integer sparse feature values in `tf.contrib.layers.sparse_column_with_keys`. 2499* Fixed `tf.set_random_seed(0)` to be deterministic for all ops. 2500* Stability improvements for the GCS file system support. 2501* Improved TensorForest performance. 2502* Added support for multiple filename globs in `tf.matching_files`. 2503* `LogMessage` now includes a timestamp as beginning of a message. 2504* Added MultiBox person detector example standalone binary. 2505* Android demo: Makefile build functionality added to build.gradle to fully support building TensorFlow demo in Android on Windows. 2506* Android demo: read MultiBox priors from txt file rather than protobuf. 2507* Added colocation constraints to `StagingArea`. 2508* `sparse_matmul_op` reenabled for Android builds. 2509* Restrict weights rank to be the same as the broadcast target, to avoid ambiguity on broadcast rules. 2510* Upgraded libxsmm to 1.7.1 and applied other changes for performance and memory usage. 2511* Fixed bfloat16 integration of LIBXSMM sparse mat-mul. 2512* Improved performance and reduce memory usage by allowing ops to forward input buffers to output buffers and perform computations in-place. 2513* Improved the performance of CPU assignment for strings. 2514* Speed up matrix * vector multiplication and matrix * matrix with unknown shapes. 2515* C API: Graph imports now support input remapping, control dependencies, and returning imported nodes (see `TF_GraphImportGraphDefWithReturnOutputs()`) 2516* Multiple C++ API updates. 2517* Multiple TensorBoard updates including: 2518 * Users can now view image summaries at various sampled steps (instead of just the last step). 2519 * Bugs involving switching runs as well as the image dashboard are fixed. 2520 * Removed data download links from TensorBoard. 2521 * TensorBoard uses a relative data directory, for easier embedding. 2522 * TensorBoard automatically ignores outliers for domain calculation, and formats proportional values consistently. 2523* Multiple tfdbg bug fixes: 2524 * Fixed Windows compatibility issues. 2525 * Command history now persists across runs. 2526 * Bug fix in graph validation related to `tf.while_loops`. 2527* Java Maven fixes for bugs with Windows installation. 2528* Backport fixes and improvements from external keras. 2529* Keras config file handling fix. 2530 2531## Thanks to our Contributors 2532 2533This release contains contributions from many people at Google, as well as: 2534 2535A. Besir Kurtulmus, Adal Chiriliuc, @akash, Alec-Desouza, Alex Rothberg, Alex 2536Sergeev, Alexander Heinecke, Allen Guo, Andreas Madsen, Ankesh Anand, Anton 2537Loss, @Aravind, @Arie, Ashutosh Das, AuréLien Geron, Bairen Yi, @bakunyo, Ben 2538Visser, Brady Zhou, Calpa Liu, Changming Sun, Chih Cheng Liang, Christopher 2539Berner, Clark Zinzow, @Conchylicultor, Dan Ellis, Dan J, Dan Jarvis, Daniel 2540Ylitalo, Darren Garvey, David Norman, David Truong, @DavidNorman, Dimitar 2541Pavlov, Dmitry Persiyanov, @Eddie, @elirex, Erfan Noury, Eron Wright, Evgeny 2542Mazovetskiy, Fabrizio (Misto) Milo, @fanlu, Fisher Coder, Florian Courtial, 2543Franck Dernoncourt, Gagan Goel, Gao, Xiang, @Gautam, Gefu Tang, @guilherme, 2544@guschmue, Hannah Provenza, Hans Pabst, @hartb, Hsiao Yi, Huazuo Gao, Igor 2545ChorążEwicz, Ivan Smirnov, Jakub Kolodziejczyk, Jason Gavris, Jason Morton, Jay 2546Young, Jayaram Bobba, Jeremy Sawruk, Jiaming Liu, Jihun Choi, @jiqiu, Joan Thibault, 2547John C F, Jojy George Varghese, Jon Malmaud, Julian Berman, Julian Niedermeier, 2548Junpeng Lao, Kai Sasaki, @Kankroc, Karl Lessard, Kyle Bostelmann, @Lezcano, Li 2549Yi, Luo Yun, @lurker, Mahmoud-Abuzaina, Mandeep Singh, Marek Kolodziej, Mark 2550Szepieniec, Martial Hue, Medhat Omr, Memo Akten, Michael Gharbi, MichaëL Defferrard, 2551Milan Straka, @MircoT, @mlucool, Muammar Ibn Faisal, Nayana Thorat, @nghiattran, 2552Nicholas Connor, Nikolaas Steenbergen, Niraj Patel, Niranjan Hasabnis, @Panmari, 2553Pavel Bulanov, Philip Pries Henningsen, Philipp Jund, @polonez, Prayag Verma, Rahul 2554Kavi, Raphael Gontijo Lopes, @rasbt, Raven Iqqe, Reid Pryzant, Richard Shin, Rizwan 2555Asif, Russell Kaplan, Ryo Asakura, RüDiger Busche, Saisai Shao, Sam Abrahams, @sanosay, 2556Sean Papay, @seaotterman, @selay01, Shaurya Sharma, Sriram Narayanamoorthy, Stefano 2557Probst, @taknevski, @tbonza, @teldridge11, Tim Anglade, Tomas Reimers, Tomer Gafner, 2558Valentin Iovene, Vamsi Sripathi, Viktor Malyi, Vit Stepanovs, Vivek Rane, Vlad Firoiu, 2559@wangg12, @will, Xiaoyu Tao, Yaroslav Bulatov, Yi Liu, Yuan (Terry) Tang, @Yufeng, 2560Yuming Wang, Yuxin Wu, Zafar Takhirov, Ziming Dong 2561 2562We are also grateful to all who filed issues or helped resolve them, asked and 2563answered questions, and were part of inspiring discussions. 2564 2565 2566# Release 1.0.1 2567 2568## Bug Fixes and Other Changes 2569* Change GraphConstructor to not increase the version when importing, but instead take the min of all versions. 2570* Google Cloud Storage fixes. 2571* 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. 2572 2573# Release 1.0.0 2574 2575## Major Features and Improvements 2576* XLA (experimental): initial release of [XLA](https://www.tensorflow.org/versions/master/experimental/xla/), a domain-specific compiler for TensorFlow graphs, that targets CPUs and GPUs. 2577* TensorFlow Debugger (tfdbg): command-line interface and API. 2578* New python 3 docker images added. 2579* Made pip packages pypi compliant. TensorFlow can now be installed by `pip 2580 install tensorflow` command. 2581* Several python API calls have been changed to resemble NumPy more closely. 2582* Android: person detection + tracking demo implementing Scalable Object 2583 Detection using Deep Neural Networks. 2584* New (experimental) [Java API](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/java). 2585* Add new Android image stylization demo based on "A Learned Representation For Artistic Style", and add YOLO object detector support. 2586 2587## Breaking Changes to the API 2588To help you upgrade your existing TensorFlow Python code to match the API changes below, we have prepared a [conversion script](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/tools/compatibility). 2589* TensorFlow/models have been moved to a separate github repository. 2590* Division and modulus operators (/, //, %) now match Python (flooring) 2591 semantics. This applies to `tf.div` and `tf.mod` as well. To obtain forced 2592 integer truncation based behaviors you can use `tf.truncatediv` 2593 and `tf.truncatemod`. 2594* `tf.divide()` is now the recommended division function. `tf.div()` will 2595 remain, but its semantics do not respond to Python 3 or `from future` 2596 mechanisms. 2597* tf.reverse() now takes indices of axes to be reversed. E.g. 2598 `tf.reverse(a, [True, False, True])` must now be written as 2599 `tf.reverse(a, [0, 2])`. `tf.reverse_v2()` will remain until 1.0 final. 2600* `tf.mul`, `tf.sub` and `tf.neg` are deprecated in favor of `tf.multiply`, 2601 `tf.subtract` and `tf.negative`. 2602* `tf.pack` and `tf.unpack` are deprecated in favor of `tf.stack` and 2603 `tf.unstack`. 2604* `TensorArray.pack` and `TensorArray.unpack` are getting deprecated in favor of 2605 `TensorArray.stack` and `TensorArray.unstack`. 2606* The following Python functions have had their arguments changed to use `axis` 2607 when referring to specific dimensions. We have kept the old keyword arguments 2608 for compatibility currently, but we will be removing them well before the 2609 final 1.0. 2610 * `tf.argmax`: `dimension` becomes `axis` 2611 * `tf.argmin`: `dimension` becomes `axis` 2612 * `tf.count_nonzero`: `reduction_indices` becomes `axis` 2613 * `tf.expand_dims`: `dim` becomes `axis` 2614 * `tf.reduce_all`: `reduction_indices` becomes `axis` 2615 * `tf.reduce_any`: `reduction_indices` becomes `axis` 2616 * `tf.reduce_join`: `reduction_indices` becomes `axis` 2617 * `tf.reduce_logsumexp`: `reduction_indices` becomes `axis` 2618 * `tf.reduce_max`: `reduction_indices` becomes `axis` 2619 * `tf.reduce_mean`: `reduction_indices` becomes `axis` 2620 * `tf.reduce_min`: `reduction_indices` becomes `axis` 2621 * `tf.reduce_prod`: `reduction_indices` becomes `axis` 2622 * `tf.reduce_sum`: `reduction_indices` becomes `axis` 2623 * `tf.reverse_sequence`: `batch_dim` becomes `batch_axis`, `seq_dim` becomes `seq_axis` 2624 * `tf.sparse_concat`: `concat_dim` becomes `axis` 2625 * `tf.sparse_reduce_sum`: `reduction_axes` becomes `axis` 2626 * `tf.sparse_reduce_sum_sparse`: `reduction_axes` becomes `axis` 2627 * `tf.sparse_split`: `split_dim` becomes `axis` 2628* `tf.listdiff` has been renamed to `tf.setdiff1d` to match NumPy naming. 2629* `tf.inv` has been renamed to be `tf.reciprocal` (component-wise reciprocal) 2630 to avoid confusion with `np.inv` which is matrix inversion 2631* tf.round now uses banker's rounding (round to even) semantics to match NumPy. 2632* `tf.split` now takes arguments in a reversed order and with different 2633 keywords. In particular, we now match NumPy order as 2634 `tf.split(value, num_or_size_splits, axis)`. 2635* `tf.sparse_split` now takes arguments in reversed order and with different 2636 keywords. In particular we now match NumPy order as 2637 `tf.sparse_split(sp_input, num_split, axis)`. NOTE: we have temporarily 2638 made `tf.sparse_split` require keyword arguments. 2639* `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)`. 2640* `tf.image.decode_jpeg` by default uses the faster DCT method, sacrificing 2641 a little fidelity for improved speed. One can revert to the old 2642 behavior by specifying the attribute `dct_method='INTEGER_ACCURATE'`. 2643* `tf.complex_abs` has been removed from the Python interface. `tf.abs` 2644 supports complex tensors and should be used instead. 2645* In the C++ API (in tensorflow/cc), Input, Output, etc. have moved 2646 from the tensorflow::ops namespace to tensorflow. 2647* Template.`var_scope` property renamed to `.variable_scope` 2648* SyncReplicasOptimizer is removed and SyncReplicasOptimizerV2 renamed to SyncReplicasOptimizer. 2649* `tf.zeros_initializer()` and `tf.ones_initializer()` now return a callable 2650 that must be called with initializer arguments, in your code replace 2651 `tf.zeros_initializer` with `tf.zeros_initializer()`. 2652* `SparseTensor.shape` has been renamed to `SparseTensor.dense_shape`. Same for 2653 `SparseTensorValue.shape`. 2654* 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. 2655* Replace tf.train.SummaryWriter and tf.train.SummaryWriterCache with tf.summary.FileWriter and tf.summary.FileWriterCache. 2656* Removes RegisterShape from public API. Use C++ shape function registration 2657 instead. 2658* Deprecated `_ref` dtypes from the python API. 2659* In the C++ API (in tensorflow/cc), Input, Output, etc. have moved 2660 from the tensorflow::ops namespace to tensorflow. 2661* Change arg order for `{softmax,sparse_softmax,sigmoid}_cross_entropy_with_logits` to be (labels, predictions), and force use of named args. 2662* 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. 2663* `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. 2664* The shape keyword argument of the `SparseTensor` constructor changes its name to `dense_shape` between Tensorflow 0.12 and Tensorflow 1.0. 2665 2666## Bug Fixes and Other Changes 2667* Numerous C++ API updates. 2668* New op: `parallel_stack`. 2669* Introducing common tf io compression options constants for 2670 RecordReader/RecordWriter. 2671* Add `sparse_column_with_vocabulary_file`, to specify a feature column that 2672 transform string features to IDs, where the mapping is defined by a vocabulary 2673 file. 2674* Added `index_to_string_table` which returns a lookup table that maps indices to 2675 strings. 2676* Add `string_to_index_table`, which returns a lookup table that matches strings 2677 to indices. 2678* Add a `ParallelForWithWorkerId` function. 2679* Add `string_to_index_table`, which returns a lookup table that matches strings 2680 to indices. 2681* Support restore session from checkpoint files in v2 in `contrib/session_bundle`. 2682* Added a tf.contrib.image.rotate function for arbitrary angles. 2683* Added `tf.contrib.framework.filter_variables` as a convenience function to 2684 filter lists of variables based on regular expressions. 2685* `make_template()` takes an optional `custom_getter_ param`. 2686* Added comment about how existing directories are handled by 2687 `recursive_create_dir`. 2688* Added an op for QR factorizations. 2689* Divides and mods in Python API now use flooring (Python) semantics. 2690* Android: pre-built libs are now built nightly. 2691* Android: cmake/gradle build for TensorFlow Inference library under 2692 `contrib/android/cmake` 2693* Android: Much more robust Session initialization code. 2694* Android: TF stats now exposed directly in demo and log when debug mode is 2695 active 2696* Android: new/better README.md documentation 2697* saved_model is available as `tf.saved_model`. 2698* Empty op is now stateful. 2699* Improve speed of scatter_update on the cpu for ASSIGN operations. 2700* Change `reduce_join` to treat `reduction_indices` in the same way as other `reduce_` ops. 2701* Move `TensorForestEstimator` to `contrib/tensor_forest`. 2702* Enable compiler optimizations by default and allow configuration in configure. 2703* `tf.divide` now honors the name field. 2704* Make metrics weight broadcasting more strict. 2705* Add new queue-like `StagingArea` and new ops: `stage` and `unstage`. 2706* Enable inplace update ops for strings on CPU. Speed up string concat. 2707 2708## Thanks to our Contributors 2709 2710This release contains contributions from many people at Google, as well as: 2711 2712Aaron Hu, Abhishek Aggarwal, Adam Michael, Adriano Carmezim, @AfirSraftGarrier, 2713Alexander Novikov, Alexander Rosenberg Johansen, Andrew Gibiansky, Andrew Hundt, 2714Anish Shah, Anton Loss, @b0noI, @BoyuanJiang, Carl Thomé, Chad Kennedy, Comic 2715Chang, Connor Braa, Daniel N. Lang, Daniel Trebbien, 2716@danielgordon10, Darcy Liu, Darren Garvey, Dmitri Lapin, Eron Wright, Evan 2717Cofer, Fabrizio Milo, Finbarr Timbers, Franck Dernoncourt, Garrett Smith, 2718@guschmue, Hao Wei, Henrik Holst, Huazuo Gao, @Ian, @Issac, Jacob Israel, 2719Jangsoo Park, Jin Kim, Jingtian Peng, John Pope, Kye Bostelmann, Liangliang He, 2720Ling Zhang, Luheng He, Luke Iwanski, @lvli, Michael Basilyan, Mihir Patel, 2721Mikalai Drabovich, Morten Just, @newge, Nick Butlin, Nishant Shukla, 2722Pengfei Ni, Przemyslaw Tredak, @rasbt, @Ronny, Rudolf Rosa, @RustingSword, 2723Sam Abrahams, Sam Putnam, @SeongAhJo, Shi Jiaxin, @skavulya, Steffen MüLler, 2724@TheUSER123, @tiriplicamihai, @vhasanov, Victor Costan, Vit Stepanovs, 2725Wangda Tan, Wenjian Huang, Xingdong Zuo, Yaroslav Bulatov, Yota Toyama, 2726Yuan (Terry) Tang, Yuxin Wu 2727 2728We are also grateful to all who filed issues or helped resolve them, asked and 2729answered questions, and were part of inspiring discussions. 2730 2731 2732# Release 0.12.0 2733 2734## Major Features and Improvements 2735 2736* TensorFlow now builds and runs on Microsoft Windows (tested on Windows 10, 2737 Windows 7, and Windows Server 2016). Supported languages include Python (via a 2738 pip package) and C++. CUDA 8.0 and cuDNN 5.1 are supported for GPU 2739 acceleration. Known limitations include: It is not currently possible to load 2740 a custom op library. The GCS and HDFS file systems are not currently 2741 supported. The following ops are not currently implemented: 2742 Dequantize, QuantizeAndDequantize, QuantizedAvgPool, 2743 QuantizedBatchNomWithGlobalNormalization, QuantizedBiasAdd, QuantizedConcat, 2744 QuantizedConv2D, QuantizedMatmul, QuantizedMaxPool, 2745 QuantizeDownAndShrinkRange, QuantizedRelu, QuantizedRelu6, QuantizedReshape, 2746 QuantizeV2, RequantizationRange, and Requantize. 2747* Go: Experimental API in Go to create and execute graphs 2748 (https://godoc.org/github.com/tensorflow/tensorflow/tensorflow/go) 2749* New checkpoint format becomes the default in `tf.train.Saver`. Old V1 2750 checkpoints continue to be readable; controlled by the `write_version` 2751 argument, `tf.train.Saver` now by default writes out in the new V2 2752 format. It significantly reduces the peak memory required and latency 2753 incurred during restore. 2754* Added a new library for library of matrix-free (iterative) solvers for linear 2755 equations, linear least-squares, eigenvalues and singular values in 2756 tensorflow/contrib/solvers. Initial version has lanczos bidiagonalization, 2757 conjugate gradients and CGLS. 2758* Added gradients for `matrix_solve_ls` and `self_adjoint_eig`. 2759* Large cleanup to add second order gradient for ops with C++ gradients and 2760 improve existing gradients such that most ops can now be differentiated 2761 multiple times. 2762* Added a solver for ordinary differential equations, 2763 `tf.contrib.integrate.odeint`. 2764* New contrib module for tensors with named axes, `tf.contrib.labeled_tensor`. 2765* Visualization of embeddings in TensorBoard. 2766 2767## Breaking Changes to the API 2768 2769* `BusAdjacency` enum replaced with a protocol buffer `DeviceLocality`. PCI bus 2770 indexing now starts from 1 instead of 0, and `bus_id==0` is used where 2771 previously `BUS_ANY` was used. 2772* `Env::FileExists` and `FileSystem::FileExists` now return a tensorflow::Status 2773 instead of a bool. Any callers to this function can be converted to a bool 2774 by adding .ok() to the call. 2775* The C API type `TF_SessionWithGraph` has been renamed to `TF_Session`, 2776 indicating its preferred use in language bindings for TensorFlow. 2777 What was previously `TF_Session` has been renamed to `TF_DeprecatedSession`. 2778* Renamed `TF_Port` to `TF_Output` in the C API. 2779* Removes RegisterShape from public API. Use C++ shape function registration instead. 2780 indexing now starts from 1 instead of 0, and `bus_id==0` is used where 2781 previously `BUS_ANY` was used. 2782* Most RNN cells and RNN functions now use different variable scopes to be 2783 consistent with layers (`tf.contrib.layers`). This means old checkpoints 2784 written using this code will not load after this change without providing 2785 `Saver` a list of variable renames. Examples of variable scope changes 2786 include `RNN` -> `rnn` in `tf.nn.rnn`, `tf.nn.dynamic_rnn` and moving from 2787 `Linear/Matrix` -> `weights` and `Linear/Bias` -> `biases` in most RNN cells. 2788* Deprecated tf.select op. tf.where should be used instead. 2789* `SparseTensor.shape` has been renamed to `SparseTensor.dense_shape`. Same for 2790 `SparseTensorValue.shape`. 2791* `Env::FileExists` and `FileSystem::FileExists` now return a 2792 `tensorflow::Status` instead of a bool. Any callers to this function can be 2793 converted to a bool by adding `.ok()` to the call. 2794* C API: Type `TF_SessionWithGraph` has been renamed to `TF_Session`, indicating 2795 its preferred use in language bindings for TensorFlow. What was previously 2796 `TF_Session` has been renamed to `TF_DeprecatedSession`. 2797* C API: Renamed `TF_Port` to `TF_Output`. 2798* C API: The caller retains ownership of `TF_Tensor` objects provided to 2799 `TF_Run`, `TF_SessionRun`, `TF_SetAttrTensor` etc. 2800* Renamed `tf.image.per_image_whitening()` to 2801 `tf.image.per_image_standardization()` 2802* Move Summary protobuf constructors to `tf.summary` submodule. 2803* Deprecate `histogram_summary`, `audio_summary`, `scalar_summary`, 2804 `image_summary`, `merge_summary`, and `merge_all_summaries`. 2805* Combined `batch_*` and regular version of linear algebra and FFT ops. The 2806 regular op now handles batches as well. All `batch_*` Python interfaces were 2807 removed. 2808* `tf.all_variables`, `tf.VARIABLES` and `tf.initialize_all_variables` renamed 2809 to `tf.global_variables`, `tf.GLOBAL_VARIABLES` and 2810 `tf.global_variables_initializer` respectively. 2811* `tf.zeros_initializer()` and `tf.ones_initializer()` now return a callable 2812 that must be called with initializer arguments, in your code replace 2813 `tf.zeros_initializer` with `tf.zeros_initializer()` 2814 2815## Bug Fixes and Other Changes 2816 2817* Use threadsafe version of `lgamma` function. 2818* Fix `tf.sqrt` handling of negative arguments. 2819* Fixed bug causing incorrect number of threads to be used for multi-threaded 2820 benchmarks. 2821* Performance optimizations for `batch_matmul` on multi-core CPUs. 2822* Improve trace, `matrix_set_diag`, `matrix_diag_part` and their gradients to 2823 work for rectangular matrices. 2824* Support for SVD of complex valued matrices. 2825 2826 2827## Thanks to our Contributors 2828 2829This release contains contributions from many people at Google, as well as: 2830 2831@a7744hsc, Abhi Agg, @admcrae, Adriano Carmezim, Aki Sukegawa, Alex Kendall, 2832Alexander Rosenberg Johansen, @amcrae, Amlan Kar, Andre Simpelo, Andreas Eberle, 2833Andrew Hundt, Arnaud Lenglet, @b0noI, Balachander Ramachandran, Ben Barsdell, 2834Ben Guidarelli, Benjamin Mularczyk, Burness Duan, @c0g, Changming Sun, 2835@chanis, Corey Wharton, Dan J, Daniel Trebbien, Darren Garvey, David Brailovsky, 2836David Jones, Di Zeng, @DjangoPeng, Dr. Kashif Rasul, @drag0, Fabrizio (Misto) 2837Milo, FabríCio Ceschin, @fp, @Ghedeon, @guschmue, Gökçen Eraslan, Haosdent 2838Huang, Haroen Viaene, Harold Cooper, Henrik Holst, @hoangmit, Ivan Ukhov, Javier 2839Dehesa, Jingtian Peng, Jithin Odattu, Joan Pastor, Johan Mathe, Johannes Mayer, 2840Jongwook Choi, Justus Schwabedal, Kai Wolf, Kamil Hryniewicz, Kamran Amini, 2841Karen Brems, Karl Lattimer, @kborer, Ken Shirriff, Kevin Rose, Larissa Laich, 2842Laurent Mazare, Leonard Lee, Liang-Chi Hsieh, Liangliang He, Luke Iwanski, 2843Marek Kolodziej, Moustafa Alzantot, @MrQianjinsi, @nagachika, Neil Han, Nick 2844Meehan, Niels Ole Salscheider, Nikhil Mishra, @nschuc, Ondrej Skopek, OndřEj 2845Filip, @OscarDPan, Pablo Moyano, Przemyslaw Tredak, @qitaishui, @Quarazy, 2846@raix852, Philipp Helo, Sam Abrahams, @SriramRamesh, Till Hoffmann, Tushar Soni, 2847@tvn, @tyfkda, Uwe Schmidt, Victor Villas, Vit Stepanovs, Vladislav Gubarev, 2848@wujingyue, Xuesong Yang, Yi Liu, Yilei Yang, @youyou3, Yuan (Terry) Tang, 2849Yuming Wang, Zafar Takhirov, @zhongyuk, Ziming Dong, @guotong1988 2850 2851We are also grateful to all who filed issues or helped resolve them, asked and 2852answered questions, and were part of inspiring discussions. 2853 2854# Release 0.11.0 2855 2856## Major Features and Improvements 2857 2858* CUDA 8 support. 2859* cuDNN 5 support. 2860* HDFS Support. 2861* Adds Fused LSTM support via cuDNN 5 in `tensorflow/contrib/cudnn_rnn`. 2862* Improved support for NumPy style basic slicing including non-1 strides, 2863 ellipses, newaxis, and negative indices. For example complicated expressions 2864 like `foo[1, 2:4, tf.newaxis, ..., :-3:-1, :]` are now supported. In addition 2865 we have preliminary (non-broadcasting) support for sliced assignment to 2866 variables. In particular one can write `var[1:3].assign([1,11,111])`. 2867* 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. 2868* Introducing `core/util/tensor_bundle` module: a module to efficiently 2869 serialize/deserialize tensors to disk. Will be used in TF's new checkpoint 2870 format. 2871* Added tf.svd for computing the singular value decomposition (SVD) of dense 2872 matrices or batches of matrices (CPU only). 2873* Added gradients for eigenvalues and eigenvectors computed using 2874 `self_adjoint_eig` or `self_adjoint_eigvals`. 2875* Eliminated `batch_*` methods for most linear algebra and FFT ops and promoted 2876 the non-batch version of the ops to handle batches of matrices. 2877* Tracing/timeline support for distributed runtime (no GPU profiler yet). 2878* C API gives access to inferred shapes with `TF_GraphGetTensorNumDims` and 2879 `TF_GraphGetTensorShape`. 2880* Shape functions for core ops have moved to C++ via 2881 `REGISTER_OP(...).SetShapeFn(...)`. Python shape inference RegisterShape calls 2882 use the C++ shape functions with `common_shapes.call_cpp_shape_fn`. A future 2883 release will remove `RegisterShape` from python. 2884 2885 2886## Bug Fixes and Other Changes 2887 2888* Documentation now includes operator overloads on Tensor and Variable. 2889* `tensorflow.__git_version__` now allows users to identify the version of the 2890 code that TensorFlow was compiled with. We also have 2891 `tensorflow.__git_compiler__` which identifies the compiler used to compile 2892 TensorFlow's core. 2893* Improved multi-threaded performance of `batch_matmul`. 2894* LSTMCell, BasicLSTMCell, and MultiRNNCell constructors now default to 2895 `state_is_tuple=True`. For a quick fix while transitioning to the new 2896 default, simply pass the argument `state_is_tuple=False`. 2897* DeviceFactory's AddDevices and CreateDevices functions now return 2898 a Status instead of void. 2899* Int32 elements of list(type) arguments are no longer placed in host memory by 2900 default. If necessary, a list(type) argument to a kernel can be placed in host 2901 memory using a HostMemory annotation. 2902* `uniform_unit_scaling_initializer()` no longer takes a `full_shape` arg, 2903 instead relying on the partition info passed to the initializer function when 2904 it's called. 2905* The NodeDef protocol message is now defined in its own file `node_def.proto` 2906 `instead of graph.proto`. 2907* `ops.NoGradient` was renamed `ops.NotDifferentiable`. `ops.NoGradient` will 2908 be removed soon. 2909* `dot.h` / DotGraph was removed (it was an early analysis tool prior 2910 to TensorBoard, no longer that useful). It remains in history 2911 should someone find the code useful. 2912* re2 / regexp.h was removed from being a public interface of TF. 2913 Should users need regular expressions, they should depend on the RE2 2914 library directly rather than via TensorFlow. 2915 2916## Thanks to our Contributors 2917 2918This release contains contributions from many people at Google, as well as: 2919 2920Abid K, @afshinrahimi, @AidanGG, Ajay Rao, Aki Sukegawa, Alex Rothberg, 2921Alexander Rosenberg Johansen, Andrew Gibiansky, Andrew Thomas, @Appleholic, 2922Bastiaan Quast, Ben Dilday, Bofu Chen, Brandon Amos, Bryon Gloden, Cissp®, 2923@chanis, Chenyang Liu, Corey Wharton, Daeyun Shin, Daniel Julius Lasiman, Daniel 2924Waterworth, Danijar Hafner, Darren Garvey, Denis Gorbachev, @DjangoPeng, 2925Egor-Krivov, Elia Palme, Eric Platon, Fabrizio Milo, Gaetan Semet, 2926Georg Nebehay, Gu Wang, Gustav Larsson, @haosdent, Harold Cooper, Hw-Zz, 2927@ichuang, Igor Babuschkin, Igor Macedo Quintanilha, Ilya Edrenkin, @ironhead, 2928Jakub Kolodziejczyk, Jennifer Guo, Jihun Choi, Jonas Rauber, Josh Bleecher 2929Snyder, @jpangburn, Jules Gagnon-Marchand, Karen Brems, @kborer, Kirill Bobyrev, 2930Laurent Mazare, Longqi Yang, Malith Yapa, Maniteja Nandana, Martin Englund, 2931Matthias Winkelmann, @mecab, Mu-Ik Jeon, Nand Dalal, Niels Ole Salscheider, 2932Nikhil Mishra, Park Jiin, Pieter De Rijk, @raix852, Ritwik Gupta, Sahil Sharma, 2933Sangheum Hwang, @SergejsRk, Shinichiro Hamaji, Simon Denel, @Steve, @suiyuan2009, 2934Tiago Jorge, Tijmen Tieleman, @tvn, @tyfkda, Wang Yang, Wei-Ting Kuo, Wenjian 2935Huang, Yan Chen, @YenChenLin, Yuan (Terry) Tang, Yuncheng Li, Yunfeng Wang, Zack 2936Polizzi, @zhongzyd, Ziming Dong, @perhapszzy 2937 2938We are also grateful to all who filed issues or helped resolve them, asked and 2939answered questions, and were part of inspiring discussions. 2940 2941# Release 0.10.0 2942 2943## Major Features and Improvements 2944 2945* Added support for C++ shape inference 2946* Added graph-construction C API 2947* Major revision to the graph-construction C++ API 2948* Support makefile build for iOS 2949* Added Mac GPU support 2950* Full version of TF-Slim available as `tf.contrib.slim` 2951* Added k-Means clustering and WALS matrix factorization 2952 2953## Bug Fixes and Other Changes 2954 2955* Allow gradient computation for scalar values. 2956* Performance improvements for gRPC 2957* Improved support for fp16 2958* New high-level ops in tf.contrib.{layers,metrics} 2959* New features for TensorBoard, such as shape display, exponential smoothing 2960* Faster and more stable Google Cloud Storage (GCS) filesystem support 2961* Support for zlib compression and decompression for TFRecordReader and TFRecordWriter 2962* Support for reading (animated) GIFs 2963* Improved support for SparseTensor 2964* Added support for more probability distributions (Dirichlet, Beta, Bernoulli, etc.) 2965* Added Python interfaces to reset resource containers. 2966* Many bugfixes and performance improvements 2967* Many documentation fixes 2968 2969## Thanks to our Contributors 2970 2971This release contains contributions from many people at Google, as well as: 2972 2973Alex 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 2974 2975We are also grateful to all who filed issues or helped resolve them, asked and 2976answered questions, and were part of inspiring discussions. 2977 2978# Release 0.9.0 2979 2980## Major Features and Improvements 2981 2982* Python 3.5 support and binaries 2983* Added iOS support 2984* Added support for processing on GPUs on MacOS 2985* Added makefile for better cross-platform build support (C API only) 2986* fp16 support and improved complex128 support for many ops 2987* Higher level functionality in contrib.{layers,losses,metrics,learn} 2988* More features to Tensorboard 2989* Improved support for string embedding and sparse features 2990* The RNN api is finally "official" (see, e.g., `tf.nn.dynamic_rnn`, 2991 `tf.nn.rnn`, and the classes in `tf.nn.rnn_cell`). 2992* TensorBoard now has an Audio Dashboard, with associated audio summaries. 2993 2994## Bug Fixes and Other Changes 2995 2996* Turned on CuDNN Autotune. 2997* Added support for using third-party Python optimization algorithms (contrib.opt). 2998* Google Cloud Storage filesystem support. 2999* HDF5 support 3000* Add support for 3d convolutions and pooling. 3001* Update gRPC release to 0.14. 3002* Eigen version upgrade. 3003* Switch to eigen thread pool 3004* `tf.nn.moments()` now accepts a `shift` argument. Shifting by a good estimate 3005 of the mean improves numerical stability. Also changes the behavior of the 3006 `shift` argument to `tf.nn.sufficient_statistics()`. 3007* Performance improvements 3008* Many bugfixes 3009* Many documentation fixes 3010* TensorBoard fixes: graphs with only one data point, Nan values, 3011 reload button and auto-reload, tooltips in scalar charts, run 3012 filtering, stable colors 3013* Tensorboard graph visualizer now supports run metadata. Clicking on nodes 3014 while viewing a stats for a particular run will show runtime statistics, such 3015 as memory or compute usage. Unused nodes will be faded out. 3016 3017## Thanks to our Contributors 3018 3019This release contains contributions from many people at Google, as well as: 3020 3021Aaron 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 3022 3023We are also grateful to all who filed issues or helped resolve them, asked and 3024answered questions, and were part of inspiring discussions. 3025 3026# Release 0.8.0 3027 3028## Major Features and Improvements 3029 3030* Added a distributed runtime using GRPC 3031* Move skflow to `contrib/learn` 3032* Better linear optimizer in `contrib/linear_optimizer` 3033* Random forest implementation in `contrib/tensor_forest` 3034* CTC loss and decoders in `contrib/ctc` 3035* Basic support for `half` data type 3036* Better support for loading user ops (see examples in `contrib/`) 3037* Allow use of (non-blocking) Eigen threadpool with `TENSORFLOW_USE_EIGEN_THREADPOOL` define 3038* Add an extension mechanism for adding network file system support 3039* TensorBoard displays metadata stats (running time, memory usage and device used) and tensor shapes 3040 3041## Bug Fixes and Other Changes 3042 3043* Utility for inspecting checkpoints 3044* Basic tracing and timeline support 3045* Allow building against cuDNN 5 (not incl. RNN/LSTM support) 3046* Added instructions and binaries for ProtoBuf library with fast serialization and without 64MB limit 3047* Added special functions 3048* `bool`-strictness: Tensors have to be explicitly compared to `None` 3049* Shape strictness: all fed values must have a shape that is compatible with the tensor they are replacing 3050* Exposed `tf.while_loop` (deprecated `control_flow_ops.While`) 3051* run() now takes RunOptions and RunMetadata, which enable timing stats 3052* Fixed lots of potential overflow problems in op kernels 3053* Various performance improvements, especially for RNNs and convolutions 3054* Many bugfixes 3055* Nightly builds, tutorial tests, many test improvements 3056* New examples: transfer learning and deepdream ipython notebook 3057* Added tutorials, many documentation fixes. 3058 3059## Thanks to our Contributors 3060 3061This release contains contributions from many people at Google, as well as: 3062 3063Abhinav 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. 3064 3065We are also grateful to all who filed issues or helped resolve them, asked and 3066answered questions, and were part of inspiring discussions. 3067 3068 3069# Release 0.7.1 3070 3071## Bug Fixes and Other Changes 3072 3073* Added gfile.Open and gfile.Copy, used by input_data.py. 3074* Fixed Saver bug when MakeDirs tried to create empty directory. 3075* GPU Pip wheels are built with cuda 7.5 and cudnn-v4, making them 3076 required for the binary releases. Lower versions of cuda/cudnn can 3077 be supported by installing from sources and setting the options 3078 during ./configure 3079* Fix dataset encoding example for Python3 (@danijar) 3080* Fix PIP installation by not packaging protobuf as part of wheel, 3081 require protobuf 3.0.0b2. 3082* Fix Mac pip installation of numpy by requiring pip >= 1.10.1. 3083* Improvements and fixes to Docker image. 3084 3085 3086# Release 0.7.0 3087 3088## Major Features and Improvements 3089 3090* Allow using any installed Cuda >= 7.0 and cuDNN >= R2, and add support 3091 for cuDNN R4 3092* Added a `contrib/` directory for unsupported or experimental features, 3093 including higher level `layers` module 3094* Added an easy way to add and dynamically load user-defined ops 3095* Built out a good suite of tests, things should break less! 3096* Added `MetaGraphDef` which makes it easier to save graphs with metadata 3097* Added assignments for "Deep Learning with TensorFlow" udacity course 3098 3099 3100## Bug Fixes and Other Changes 3101 3102* Added a versioning framework for `GraphDef`s to ensure compatibility 3103* Enforced Python 3 compatibility 3104* Internal changes now show up as sensibly separated commits 3105* Open-sourced the doc generator 3106* Un-fork Eigen 3107* Simplified the `BUILD` files and cleaned up C++ headers 3108* TensorFlow can now be used as a submodule in another bazel build 3109* New ops (e.g., `*fft`, `*_matrix_solve`) 3110* Support for more data types in many ops 3111* Performance improvements 3112* Various bugfixes 3113* Documentation fixes and improvements 3114 3115 3116## Breaking Changes to the API 3117 3118* `AdjustContrast` kernel deprecated, new kernel `AdjustContrastv2` takes and 3119 outputs float only. `adjust_contrast` now takes all data types. 3120* `adjust_brightness`'s `delta` argument is now always assumed to be in `[0,1]` 3121 (as is the norm for images in floating point formats), independent of the 3122 data type of the input image. 3123* The image processing ops do not take `min` and `max` inputs any more, casting 3124 safety is handled by `saturate_cast`, which makes sure over- and underflows 3125 are handled before casting to data types with smaller ranges. 3126* For C++ API users: `IsLegacyScalar` and `IsLegacyVector` are now gone from 3127 `TensorShapeUtils` since TensorFlow is scalar strict within Google (for 3128 example, the shape argument to `tf.reshape` can't be a scalar anymore). The 3129 open source release was already scalar strict, so outside Google `IsScalar` 3130 and `IsVector` are exact replacements. 3131* The following files are being removed from `tensorflow/core/public/`: 3132 * `env.h` -> `../platform/env.h` 3133 * `status.h` -> `../lib/core/status.h` 3134 * `tensor.h` -> `../framework/tensor.h` 3135 * `tensor_shape.h` -> `../framework/tensor_shape.h` 3136 * `partial_tensor_shape.h` -> `../framework/partial_tensor_shape.h` 3137 * `tensorflow_server.h` deleted 3138* For C++ API users: `TensorShape::ShortDebugString` has been renamed to 3139 `DebugString`, and the previous `DebugString` behavior is gone (it was 3140 needlessly verbose and produced a confusing empty string for scalars). 3141* `GraphOptions.skip_common_subexpression_elimination` has been removed. All 3142 graph optimizer options are now specified via 3143 `GraphOptions.OptimizerOptions`. 3144* `ASSERT_OK` / `EXPECT_OK` macros conflicted with external projects, so they 3145 were renamed `TF_ASSERT_OK`, `TF_EXPECT_OK`. The existing macros are 3146 currently maintained for short-term compatibility but will be removed. 3147* The non-public `nn.rnn` and the various `nn.seq2seq` methods now return 3148 just the final state instead of the list of all states. 3149* `tf.scatter_update` now no longer guarantees that lexicographically largest 3150 index be used for update when duplicate entries exist. 3151* `tf.image.random_crop(image, [height, width])` is now 3152 `tf.random_crop(image, [height, width, depth])`, and `tf.random_crop` works 3153 for any rank (not just 3-D images). The C++ `RandomCrop` op has been replaced 3154 with pure Python. 3155* Renamed `tf.test.GetTempDir` and `tf.test.IsBuiltWithCuda` to 3156 `tf.test.get_temp_dir` and `tf.test.is_built_with_cuda` for PEP-8 3157 compatibility. 3158* `parse_example`'s interface has changed, the old interface is accessible in 3159 `legacy_parse_example` (same for related functions). 3160* New `Variable`s are not added to the same collection several times even if 3161 a list with duplicates is passed to the constructor. 3162* The Python API will now properly set the `list` member of `AttrValue` in 3163 constructed `GraphDef` messages for empty lists. The serialization of some 3164 graphs will change, but the change is both forwards and backwards compatible. 3165 It will break tests that compare a generated `GraphDef` to a golden serialized 3166 `GraphDef` (which is discouraged). 3167 3168 3169## Thanks to our Contributors 3170 3171This release contains contributions from many people at Google, as well as: 3172 3173Akiomi Kamakura, Alex Vig, Alexander Rosenberg Johansen, Andre Cruz, Arun Ahuja, 3174Bart Coppens, Bernardo Pires, Carl Vondrick, Cesar Salgado, Chen Yu, 3175Christian Jauvin, Damien Aymeric, Dan Vanderkam, Denny Britz, Dongjoon Hyun, 3176Eren Güven, Erik Erwitt, Fabrizio Milo, G. Hussain Chinoy, Jim Fleming, 3177Joao Felipe Santos, Jonas Meinertz Hansen, Joshi Rekha, Julian Viereck, 3178Keiji Ariyama, Kenton Lee, Krishna Sankar, Kristina Chodorow, Linchao Zhu, 3179Lukas Krecan, Mark Borgerding, Mark Daoust, Moussa Taifi, 3180Nathan Howell, Naveen Sundar Govindarajulu, Nick Sweeting, Niklas Riekenbrauck, 3181Olivier Grisel, Patrick Christ, Povilas Liubauskas, Rainer Wasserfuhr, 3182Romain Thouvenin, Sagan Bolliger, Sam Abrahams, Taehoon Kim, Timothy J Laurent, 3183Vlad Zavidovych, Yangqing Jia, Yi-Lin Juang, Yuxin Wu, Zachary Lipton, 3184Zero Chen, Alan Wu, @brchiu, @emmjaykay, @jalammar, @Mandar-Shinde, 3185@nsipplswezey, @ninotoshi, @panmari, @prolearner and @rizzomichaelg. 3186 3187We are also grateful to all who filed issues or helped resolve them, asked and 3188answered questions, and were part of inspiring discussions. 3189 3190 3191# Release 0.6.0 3192 3193## Major Features and Improvements 3194 3195* Python 3.3+ support via changes to python codebase and ability 3196 to specify python version via ./configure. 3197 3198* Some improvements to GPU performance and memory usage: 3199 [convnet benchmarks](https://github.com/soumith/convnet-benchmarks/issues/66) 3200 roughly equivalent with native cudnn v2 performance. Improvements mostly due 3201 to moving to 32-bit indices, faster shuffling kernels. More improvements to 3202 come in later releases. 3203 3204 3205## Bug Fixes 3206 3207* Lots of fixes to documentation and tutorials, many contributed 3208 by the public. 3209 3210* 271 closed issues on github issues. 3211 3212## Backwards-Incompatible Changes 3213 3214* `tf.nn.fixed_unigram_candidate_sampler` changed its default 'distortion' 3215 attribute from 0.0 to 1.0. This was a bug in the original release 3216 that is now fixed. 3217 3218# Release 0.5.0 3219 3220Initial release of TensorFlow. 3221