1# Release 1.13.0 2 3## Major Features and Improvements 4 5* 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`. 6* TensorFlow GPU binaries are now built against CUDA 10 and TensorRT 5.0. 7* Support for Python3.7 on all operating systems. 8* Moved NCCL to core. 9 10## Behavioral changes 11 12* Disallow conversion of python floating types to uint32/64 (matching behavior of other integer types) in `tf.constant`. 13* 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). 14 15## Bug Fixes and Other Changes 16 17* Documentation 18 * Update the doc with the details about the rounding mode used in quantize_and_dequantize_v2. 19 * Clarify that tensorflow::port::InitMain() _should_ be called before using the TensorFlow library. Programs failing to do this are not portable to all platforms. 20* Deprecations and Symbol renames. 21 * Removing deprecations for the following endpoints: `tf.acos`, `tf.acosh`, `tf.add`, `tf.as_string`, `tf.asin`, `tf.asinh`, `tf.atan`, `tf.atan2`, `tf.atanh`, `tf.cos`, `tf.cosh`, `tf.equal`, `tf.exp`, `tf.floor`, `tf.greater`, `tf.greater_equal`, `tf.less`, `tf.less_equal`, `tf.log`, `tf.logp1`, `tf.logical_and`, `tf.logical_not`, `tf.logical_or`, `tf.maximum`, `tf.minimum`, `tf.not_equal`, `tf.sin`, `tf.sinh`, `tf.tan` 22 * Deprecate `tf.data.Dataset.shard`. 23 * Deprecate `saved_model.loader.load` which is replaced by `saved_model.load` and `saved_model.main_op`, which will be replaced by `saved_model.main_op` in V2. 24 * Deprecate tf.QUANTIZED_DTYPES. The official new symbol is tf.dtypes.QUANTIZED_DTYPES. 25 * Update sklearn imports for deprecated packages. 26 * Deprecate `Variable.count_up_to` and `tf.count_up_to` in favor of `Dataset.range`. 27 * Export `confusion_matrix` op as `tf.math.confusion_matrix` instead of `tf.train.confusion_matrix`. 28 * Add `tf.dtypes.` endpoint for every constant in dtypes.py; moving endpoints in versions.py to corresponding endpoints in `tf.sysconfig.` and `tf.version.`; moving all constants under `tf.saved_model` submodules to `tf.saved_model` module. New endpoints are added in V1 and V2 but existing endpoint removals are only applied in V2. 29 * Deprecates behavior where device assignment overrides collocation constraints inside a collocation context manager. 30* Keras & Python API 31 * Add to Keras functionality analogous to `tf.register_tensor_conversion_function`. 32 * Subclassed Keras models can now be saved through `tf.contrib.saved_model.save_keras_model`. 33 * `LinearOperator.matmul` now returns a new `LinearOperator`. 34* New ops and improved op functionality 35 * Add a Nearest Neighbor Resize op. 36 * Add an `ignore_unknown` argument to `parse_values` which suppresses ValueError for unknown hyperparameter types. Such * Add `tf.linalg.matvec` convenience function. 37 * `tf.einsum()`raises `ValueError` for unsupported equations like `"ii->"`. 38 * Add DCT-I and IDCT-I in `tf.signal.dct` and `tf.signal.idct`. 39 * Add LU decomposition op. 40 * Add quantile loss to gradient boosted trees in estimator. 41 * Add `round_mode` to `QuantizeAndDequantizeV2` op to select rounding algorithm. 42 * Add `unicode_encode`, `unicode_decode`, `unicode_decode_with_offsets`, `unicode_split`, `unicode_split_with_offset`, and `unicode_transcode` ops. Amongst other things, this Op adds the ability to encode, decode, and transcode a variety of input text encoding formats into the main Unicode encodings (UTF-8, UTF-16-BE, UTF-32-BE) 43 * Add "unit" attribute to the substr op, which allows obtaining the substring of a string containing unicode characters. 44 * Broadcasting support for Ragged Tensors. 45 * `SpaceToDepth` supports uint8 data type. 46 * Support multi-label quantile regression in estimator. 47 * We now use "div" as the default partition_strategy in `tf.nn.safe_embedding_lookup_sparse`, `tf.nn.sampled_softmax` and `tf.nn.nce_loss`. 48 hyperparameter are ignored. 49* Performance 50 * Improve performance of GPU cumsum/cumprod by up to 300x. 51 * Added support for weight decay in most TPU embedding optimizers, including AdamW and MomentumW. 52* TensorFlow 2.0 Development 53 * Add a command line tool to convert to TF2.0, tf_upgrade_v2 54 * Merge `tf.spectral` into `tf.signal` for TensorFlow 2.0. 55 * Change the default recurrent activation function for LSTM from 'hard_sigmoid' to 'sigmoid' in 2.0. Historically recurrent activation is 'hard_sigmoid' since it is fast than 'sigmoid'. With new unified backend between CPU and GPU mode, since the CuDNN kernel is using sigmoid, we change the default for CPU mode to sigmoid as well. With that, the default LSTM will be compatible with both CPU and GPU kernel. This will enable user with GPU to use CuDNN kernel by default and get a 10x performance boost in training. Note that this is checkpoint breaking change. If user want to use their 1.x pre-trained checkpoint, please construct the layer with LSTM(recurrent_activation='hard_sigmoid') to fallback to 1.x behavior. 56* TensorFlow Lite 57 * Move from `tensorflow/contrib/lite` to `tensorflow/lite`. 58 * Add experimental Java API for injecting TensorFlow Lite delegates 59 * Add support for strings in TensorFlow Lite Java API. 60* `tf.contrib`: 61 * Add Apache Ignite Filesystem plugin to support accessing Apache IGFS. 62 * Dropout now takes `rate` argument, `keep_prob` is deprecated. 63 * Estimator occurrences references `tf.contrib.estimator` were changed to `tf.estimator`: 64 * `tf.contrib.estimator.BaselineEstimator` with `tf.estimator.BaselineEstimator` 65 * `tf.contrib.estimator.DNNLinearCombinedEstimator` with `tf.estimator.DNNLinearCombinedEstimator` 66 * `tf.contrib.estimator.DNNEstimator` with `tf.estimator.DNNEstimator` 67 * `tf.contrib.estimator.LinearEstimator` with `tf.estimator.LinearEstimator` 68 * `tf.contrib.estimator.InMemoryEvaluatorHook` and tf.estimator.experimental.InMemoryEvaluatorHook`. 69 * `tf.contrib.estimator.make_stop_at_checkpoint_step_hook` with `tf.estimator.experimental.make_stop_at_checkpoint_step_hook`. 70 * Expose `tf.distribute.Strategy as the new name for tf.contrib.distribute.DistributionStrategy. 71 * Migrate linear optimizer from contrib to core. 72 * Move `tf.contrib.signal` to `tf.signal` (preserving aliases in tf.contrib.signal). 73 * Users of `tf.contrib.estimator.export_all_saved_models` and related should switch to `tf.estimator.Estimator.experimental_export_all_saved_models`. 74* tf.data: 75 * Add `tf.data.experimental.StatsOptions()`, to configure options to collect statistics from `tf.data.Dataset` pipeline using `StatsAggregator`. Add nested option, `experimental_stats` (which takes a `tf.data.experimen tal.StatsOptions` object), to `tf.data.Options`. Deprecates `tf.data.experimental.set_stats_agregator`. 76 * Performance optimizations: 77 * Add `tf.data.experimental.OptimizationOptions()`, to configure options to enable `tf.data` performance optimizations. Add nested option, `experimental_optimization` (which takes a `tf.data.experimental.OptimizationOptions` object), to `tf.data.Options`. Remove performance optimization options from `tf.data.Options`, and add them under `tf.data.experimental.OptimizationOptions` instead. 78 * Enable `map_and_batch_fusion` and `noop_elimination` optimizations by default. They can be disabled by configuring `tf.data.experimental.OptimizationOptions` to set `map_and_batch = False` or `noop_elimination = False` respectively. To disable all default optimizations, set `apply_default_optimizations = False`. 79 * Support parallel map in `map_and_filter_fusion`. 80 * Disable static optimizations for input pipelines that use non-resource `tf.Variable`s. 81 * Add NUMA-aware MapAndBatch dataset. 82 * Deprecate `tf.data.Dataset.make_one_shot_iterator()` in V1, removed it from V2, and added tf.compat.v1.data.make_one_shot_iterator()`. 83 * Deprecate `tf.data.Dataset.make_initializable_iterator()` in V1, removed it from V2, and added `tf.compat.v1.data.make_initializable_iterator()`. 84 * Enable nested dataset support in core `tf.data` transformations. 85 * For `tf.data.Dataset` implementers: Added `tf.data.Dataset._element_structured property` to replace `Dataset.output_{types,shapes,classes}`. 86 * Make `num_parallel_calls` of `tf.data.Dataset.interleave` and `tf.data.Dataset.map` work in Eager mode. 87* Toolchains 88 * Fixed OpenSSL compatibility by avoiding `EVP_MD_CTX_destroy`. 89 * Added bounds checking to printing deprecation warnings. 90 * Upgraded CUDA dependency to 10.0 91 * To build with Android NDK r14b, add "#include <linux/compiler.h>" to android-ndk-r14b/platforms/android-14/arch-*/usr/include/linux/futex.h 92 * Removed `:android_tensorflow_lib_selective_registration*` targets, use `:android_tensorflow_lib_lite*` targets instead. 93* XLA 94 * Move `RoundToEven` function to xla/client/lib/math.h. 95 * A new environment variable `TF_XLA_DEBUG_OPTIONS_PASSTHROUGH` set to "1" or "true" allows the debug options passed within an XRTCompile op to be passed directly to the XLA compilation backend. If such variable is not set (service side), only a restricted set will be passed through. 96 * Allow the XRTCompile op to return the ProgramShape resulted form the XLA compilation as a second return argument. 97 * XLA HLO graphs can now be rendered as SVG/HTML. 98* Estimator 99 * Replace all occurences of `tf.contrib.estimator.BaselineEstimator` with `tf.estimator.BaselineEstimator` 100 * Replace all occurences of `tf.contrib.estimator.DNNLinearCombinedEstimator` with `tf.estimator.DNNLinearCombinedEstimator` 101 * Replace all occurrences of `tf.contrib.estimator.DNNEstimator` with `tf.estimator.DNNEstimator` 102 * Replace all occurrences of `tf.contrib.estimator.LinearEstimator` with `tf.estimator.LinearEstimator` 103 * Users of `tf.contrib.estimator.export_all_saved_models` and related should switch to `tf.estimator.Estimator.experimental_export_all_saved_models`. 104 * Update `regression_head` to the new Head API for Canned Estimator V2. 105 * Switch `multi_class_head` to Head API for Canned Estimator V2. 106 * Replace all occurences of `tf.contrib.estimator.InMemoryEvaluatorHook` and `tf.contrib.estimator.make_stop_at_checkpoint_step_hook` with `tf.estimator.experimental.InMemoryEvaluatorHook` and `tf.estimator.experimental.make_stop_at_checkpoint_step_hook` 107 * Migrate linear optimizer from contrib to core. 108 109 110## Thanks to our Contributors 111 112This release contains contributions from many people at Google, as well as: 113 114Abhinav 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 115 116# Release 1.12.0 117 118## Major Features and Improvements 119 120* Keras models can now be directly exported to the SavedModel 121 format(`tf.contrib.saved_model.save_keras_model()`) and used with Tensorflow 122 Serving. 123* Keras models now support evaluating with a `tf.data.Dataset`. 124* TensorFlow binaries are built with XLA support linked in by default. 125* Ignite Dataset added to contrib/ignite that allows to work with Apache 126 Ignite. 127 128## Bug Fixes and Other Changes 129 130* tf.data: 131 * tf.data users can now represent, get, and set options of TensorFlow 132 input pipelines using `tf.data.Options()`, `tf.data.Dataset.options()`, 133 and `tf.data.Dataset.with_options()` respectively. 134 * New `tf.data.Dataset.reduce()` API allows users to reduce a finite 135 dataset to a single element using a user-provided reduce function. 136 * New `tf.data.Dataset.window()` API allows users to create finite windows 137 of input dataset; when combined with the `tf.data.Dataset.reduce()` API, 138 this allows users to implement customized batching. 139 * All C++ code moves to the `tensorflow::data` namespace. 140 * Add support for `num_parallel_calls` to `tf.data.Dataset.interleave`. 141* `tf.contrib`: 142 * Remove `tf.contrib.linalg`. `tf.linalg` should be used instead. 143 * Replace any calls to `tf.contrib.get_signature_def_by_key(metagraph_def, 144 signature_def_key)` with 145 `meta_graph_def.signature_def[signature_def_key]`. Catching a ValueError 146 exception thrown by `tf.contrib.get_signature_def_by_key` should be 147 replaced by catching a KeyError exception. 148* `tf.contrib.data` 149 * Deprecate, and replace by tf.data.experimental. 150* Other: 151 * Instead of jemalloc, revert back to using system malloc since it 152 simplifies build and has comparable performance. 153 * Remove integer types from `tf.nn.softplus` and `tf.nn.softsign` OpDefs. 154 This is a bugfix; these ops were never meant to support integers. 155 * Allow subslicing Tensors with a single dimension. 156 * Add option to calculate string length in Unicode characters 157 * Add functionality to SubSlice a tensor. 158 * Add searchsorted (ie lower/upper_bound) op. 159 * Add model explainability to Boosted Trees. 160 * Support negative positions for tf.substr 161 * There was previously a bug in the bijector_impl where the 162 _reduce_jacobian_det_over_event does not handle scalar ILDJ 163 implementations properly. 164 * In tf eager execution, allow re-entering a GradientTape context 165 * Add tf_api_version flag. If --define=tf_api_version=2 flag is passed in, 166 then bazel will build TensorFlow API version 2.0. Note that TensorFlow 167 2.0 is under active development and has no guarantees at this point. 168 * Add additional compression options to TfRecordWriter 169 * Performance improvements for regex full match operations. 170 * Replace tf.GraphKeys.VARIABLES with `tf.GraphKeys.GLOBAL_VARIABLES` 171 * Remove unused dynamic learning rate support. 172 173## Thanks to our Contributors 174 175This release contains contributions from many people at Google, as well as: 176 177(David) Siu-Kei Muk, Ag Ramesh, Anton Dmitriev, Artem Sobolev, Avijit-Nervana, 178Bairen Yi, Bruno Goncalves, By Shen, candy.dc, Cheng Chen, Clayne Robison, 179coder3101, Dao Zhang, Elms, Fei Hu, feiquan, Geoffrey Irving, Guozhong Zhuang, 180hellcom, Hoeseong Kim, imsheridan, Jason Furmanek, Jason Zaman, Jenny Sahng, 181jiefangxuanyan, Johannes Bannhofer, Jonathan Homer, Koan-Sin Tan, kouml, Loo 182Rong Jie, Lukas Geiger, manipopopo, Ming Li, Moritz KröGer, Naurril, Niranjan 183Hasabnis, Pan Daoxin, Peng Yu, pengwa, rasmi, Roger Xin, Roland Fernandez, Sami 184Kama, Samuel Matzek, Sangjung Woo, Sergei Lebedev, Sergii Khomenko, shaohua, 185Shaohua Zhang, Shujian2015, Sunitha Kambhampati, tomguluson92, ViníCius Camargo, 186wangsiyu, weidankong, Wen-Heng (Jack) Chung, William D. Irons, Xin Jin, Yan 187Facai (颜发才), Yanbo Liang, Yash Katariya, Yong Tang, 在原佐为 188 189# Release 1.11.0 190 191## Major Features and Improvements 192 193* Nvidia GPU: 194 * Prebuilt binaries are now (as of TensorFlow 1.11) built against cuDNN 7.2 and TensorRT 4. See updated install guides: [Installing TensorFlow on Ubuntu](https://www.tensorflow.org/install/install_linux#tensorflow_gpu_support) 195* Google Cloud TPU: 196 * Experimental tf.data integration for Keras on Google Cloud TPUs. 197 * Experimental / preview support for eager execution on Google Cloud TPUs. 198* DistributionStrategy: 199 * Add multi-GPU DistributionStrategy support in tf.keras. Users can now use `fit`, `evaluate` and `predict` to distribute their model on multiple GPUs. 200 * Add multi-worker DistributionStrategy and standalone client support in Estimator. See [README] (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/distribute) for more details. 201* Add C, C++, and Python functions for querying kernels 202 203## Breaking Changes 204 205* Keras: 206 * The default values for tf.keras `RandomUniform`, `RandomNormal`, and `TruncatedNormal` initializers have been changed to match those in external Keras. 207 * 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. 208 209## Bug Fixes and Other Changes 210 211* C++: 212 * Changed the signature of SessionFactory::NewSession so that it can 213 return a meaningful error message on failure. 214* tf.data: 215 * Remove `num_parallel_parser_calls` argument from 216 `tf.contrib.data.make_csv_dataset()`. [tf.data] Remove 217 `num_parallel_parser_calls` argument from 218 `tf.contrib.data.make_csv_dataset()`. 219 * `tf.data.Dataset.list_files()` raises an exception at initialization 220 time if the argument matches no files. 221 * Renamed BigTable class to BigtableTable for clarity 222 * Document use of the Cloud Bigtable API 223 * Add `tf.contrib.data.reduce_dataset` which can be used to reduce a 224 dataset to a single element. 225 * Generalization of `tf.contrib.data.sliding_window_batch`. 226* INC: 227 * Runtime improvements to triangular solve. 228* `tf.contrib`: 229 * Add an `implementation` argument to `tf.keras.layers.LocallyConnected2D` 230 and `tf.keras.layers.LocallyConnected1D`. The new mode 231 (`implementation=2`) performs forward pass as a single dense matrix 232 multiplication, allowing dramatic speedups in certain scenarios (but 233 worse performance in others - see docstring). The option also allows to 234 use `padding=same`. 235 * Add documentation clarifying the differences between tf.fill and 236 tf.constant. 237 * Add experimental IndexedDatasets. 238 * Add selective registration target using the lite proto runtime. 239 * Add simple Tensor and DataType classes to TensorFlow Lite Java 240 * Add support for bitcasting to/from uint32 and uint64. 241 * Added a subclass of Estimator that can be created from a SavedModel 242 (SavedModelEstimator). 243 * Adds leaf index modes as an argument. 244 * Allow a different output shape from the input in 245 tf.contrib.image.transform. 246 * Change the state_size order of the StackedRNNCell to be natural order. 247 To keep the existing behavior, user can add reverse_state_order=True 248 when constructing the StackedRNNCells. 249 * Deprecate self.test_session() in favor of self.session() or 250 self.cached_session(). 251 * Directly import tensor.proto.h (the transitive import will be removed 252 from tensor.h soon) 253 * Estimator.train() now supports tf.contrib.summary.\* summaries out of 254 the box; each call to .train() will now create a separate tfevents file 255 rather than re-using a shared one. 256 * Fix FTRL L2-shrinkage behavior: the gradient from the L2 shrinkage term 257 should not end up in the accumulator. 258 * Fix toco compilation/execution on Windows 259 * GoogleZoneProvider class added to detect which Google Cloud Engine zone 260 tensorflow is running in. 261 * It is now safe to call any of the C API's TF_Delete\* functions on 262 nullptr 263 * Log some errors on Android to logcat 264 * Match FakeQuant numerics in TFLite to improve accuracy of TFLite 265 quantized inference models. 266 * Optional bucket location check for the GCS Filesystem. 267 * Performance enhancements for StringSplitOp & StringSplitV2Op. 268 * Performance improvements for regex replace operations. 269 * TFRecordWriter now raises an error if .write() fails. 270 * TPU: More helpful error messages in TPUClusterResolvers. 271 * The legacy_init_op argument to SavedModelBuilder methods for adding 272 MetaGraphs has been deprecated. Please use the equivalent main_op 273 argument instead. As part of this, we now explicitly check for a single 274 main_op or legacy_init_op at the time of SavedModel building, whereas 275 the check on main_op was previously only done at load time. 276 * The protocol used for Estimator training is now configurable in 277 RunConfig. 278 * Triangular solve performance improvements. 279 * Unify RNN cell interface between TF and Keras. Add new 280 get_initial_state() to Keras and TF RNN cell, which will use to replace 281 the existing zero_state() method. 282 * Update initialization of variables in Keras. 283 * Updates to "constrained_optimization" in tensorflow/contrib. 284 * boosted trees: adding pruning mode 285 * tf.train.Checkpoint does not delete old checkpoints by default. 286 * tfdbg: Limit the total disk space occupied by dumped tensor data to 100 287 GBytes. Add environment variable `TFDBG_DISK_BYTES_LIMIT` to allow 288 adjustment of this upper limit. 289 290## Thanks to our Contributors 291 292This release contains contributions from many people at Google, as well as: 293 294Aapeli, 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 295Ricardo Perez-Lopez, 张天启, 张晓飞 296 297 298# Release 1.10.1 299## Bug Fixes and Other Changes 300 301* `tf.keras`: 302 * Fixing keras on Cloud TPUs. No new binaries will be built for Windows. 303 304 305# Release 1.10.0 306 307## Major Features And Improvements 308 309* The `tf.lite` runtime now supports `complex64`. 310* Initial [Google Cloud Bigtable integration](https://github.com/tensorflow/tensorflow/tree/r1.10/tensorflow/contrib/bigtable) for `tf.data`. 311* Improved local run behavior in `tf.estimator.train_and_evaluate` which does not reload checkpoints for evaluation. 312* `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`. 313* 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. 314* 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) 315 316## Breaking Changes 317 318* 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). 319* Starting from TensorFlow 1.11, Windows builds will use Bazel. Therefore, we will drop official support for cmake. 320 321## Bug Fixes and Other Changes 322 323* `tf.data`: 324 * `tf.contrib.data.group_by_reducer()` is now available via the public API. 325 * `tf.contrib.data.choose_from_datasets()` is now available via the public API. 326 * 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()`. 327* `tf.estimator`: 328 * `Estimator`s now use custom savers included in `EstimatorSpec` scaffolds for saving SavedModels during export. 329 * `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. 330 * Support sparse_combiner in canned Linear Estimators. 331 * Added batch normalization to `DNNClassifier`, `DNNRegressor`, and `DNNEstimator`. 332 * Adding ranking support for boosted trees. 333 * Adding center bias option for boosted trees. 334* Add `synchronization` and `aggregation` args to get_variable(). These args will be used for distributed variables. 335* Add `synchronization` and `aggregation` args to the layer `add_weight()` API. These args will be used for distributed variables. 336* `tf.losses.*` do not add to the global collection when executing eagerly (to avoid leaking memory). 337* Support different summary and checkpoint directories in `tf.train.MonitoredTrainingSession()`. 338* Added IndRNN, IndyGRU, and IndyLSTM cells to `tf.contrib.rnn`. 339* Add safe static factory functions for SparseTensor and convert all CHECKs to DCHECKs. Using the constructor directly is unsafe and deprecated. 340* Make the Bigtable client connection pool configurable & increase the default # of connections for performance. 341* Added derivative of `tf.random_gamma` with respect to the alpha parameter. 342* Added derivative of `tf.igamma(a, x)` and `tf.igammac(a, x)` with respect to a. 343* Modified Bessel functions of order zero and one. 344* Add FillTriangular Bijector to create triangular matrices. 345* Added support for Type III DCT, and `tf.spectral.idct(type=2|3)`. 346* Correctly handle CuDNN RNN weight loaded when nest in `TimeDistributed`. 347* Adding per-element weight support for `WALSComputePartialLhsAndRhsOp`. 348* ZerosLike and OnesLike ops treated as constants by Graph Transform Tool. 349* Gamma distribution and the derived distributions (Beta, Dirichlet, Student's t, inverse Gamma) now fully reparameterized. 350* Java: Experimental wrapper classes to make graph generation easier. Thanks @karllessard and @kbsriram 351* Build & link in secure gRPC components (switch from the insecure grpc dependency to secure grpc dependency). 352* 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: 353 * New endpoints in `tf.image` namespace: `tf.image.extract_image_patches` 354 * New endpoints in `tf.debugging` namespace: `tf.debugging.check_numerics`, `tf.debugging.is_finite`, `tf.debugging.is_inf`, `tf.debugging.is_nan`. 355 * New endpoints in `tf.dtypes` namespace: `tf.dtypes.as_string`. 356 * 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`. 357 * 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`). 358 * 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` 359 * 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`. 360 * 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`. 361 * 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`). 362 363 364## Thanks to our Contributors 365 366This release contains contributions from many people at Google, as well as: 367 368Ag 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 369 370# Release 1.9.0 371 372## Major Features And Improvements 373* Updated docs for `tf.keras`: New Keras-based [get started](http://tensorflow.org/versions/r1.9/get_started), 374 and [programmers guide page](http://tensorflow.org/versions/r1.9/programmers_guide/keras). 375* Update `tf.keras` to the Keras 2.1.6 API. 376* 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). 377* 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/boosted_trees). 378* The [python interface](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/lite) 379 for the [TFLite Optimizing Converter](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/toco/README.md) 380 has been expanded, and the command line interface (AKA: `toco`, `tflite_convert`) is once again 381 included in the standard `pip` installation. 382* Improved data-loading and text processing with: 383 * [`tf.decode_compressed`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/decode_compressed) 384 * [`tf.string_strip`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/string_strip) 385 * [`tf.strings.regex_full_match`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/strings/regex_full_match) 386* Added experimental support for new pre-made Estimators: 387 * [`tf.contrib.estimator.BaselineEstimator`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/estimator/BaselineEstimator) 388 * [`tf.contrib.estimator.RNNClassifier`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/estimator/RNNEstimator) 389 * [`tf.contrib.estimator.RNNEstimator`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/estimator/RNNClassifier) 390* The [distributions.Bijector](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/distributions/bijectors/Bijector) 391 API supports broadcasting for Bijectors with new API changes. 392 393## Breaking Changes 394 * If you're opening empty variable scopes; replace `variable_scope('', ...)` by 395 `variable_scope(tf.get_variable_scope(), ...)`. 396 * Headers used for building custom ops have been moved from site-packages/external into site-packages/tensorflow/include/external. 397 398## Bug Fixes and Other Changes 399 400* `tfe.Network` is deprecated. Please inherit from `tf.keras.Model`. 401* Layered variable names have changed in the following conditions: 402 * Using `tf.keras.layers` with custom variable scopes. 403 * Using `tf.layers` in a subclassed `tf.keras.Model` class. See 404 [here](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/layers) 405 for more details 406* `tf.data`: 407 * `Dataset.from_generator()` now accepts an `args` list, in order to 408 create nested generators. 409 * `Dataset.list_files()` now produces deterministic results when 410 `shuffle=False` or a `seed` is passed. 411 * `tf.contrib.data.sample_from_datasets()` and 412 `tf.contrib.data.choose_from_datasets()` make it easier to sample or 413 deterministically choose elements from multiple datasets. 414 * `tf.contrib.data.make_csv_dataset()` now supports line breaks in quoted 415 strings, and two infrequently used arguments removed. 416 * (C++) `DatasetBase::DebugString()` is now `const`. 417 * (C++) `DatasetBase::MakeIterator()` has been renamed to 418 `DatasetBase::MakeIteratorInternal()`. 419 * (C++) `IteratorBase::Initialize()` method was added to support raising 420 errors during iterator construction. 421* Eager Execution: 422 * Added the ability to pause recording operations for gradient computation 423 via `tf.GradientTape.stop_recording`. 424 * Updated documentation, introductory notebooks. 425* `tf.keras`: 426 * Move Keras code out of _impl folder and remove API files. 427 * `tf.keras.Model.save_weights` now saves in TensorFlow format by default. 428 * Enable dataset iterators to be passed to `tf.keras.Model` training/eval 429 methods. 430* TensorFlow Debugger (tfdbg) CLI: fix an issue in which the TensorBoard 431 Debugger Plugin could not handle total source file size exceeding gRPC 432 message size limit (4 MB). 433* `tf.contrib`: 434 * `tf.contrib.framework.zero_initializer` supports ResourceVariable. 435 * Adding "constrained_optimization" to tensorflow/contrib. 436* Other: 437 * Add GCS Configuration Ops. 438 * Changing signature of `MakeIterator` to enable propagating error status. 439 * KL divergence for two Dirichlet distributions. 440 * More consistent GcsFileSystem behavior for certain reads past EOF. 441 * Update benchmark for tf.scan to match ranges across eager and graph 442 modes. 443 * Fixed bug in `tf.reduce_prod gradient` for complex dtypes. 444 * Allow the use of '.' in variables (e.g. "hparams.parse('a.b=1.0')"), 445 which would previously raise an error. This will correspond to an 446 attribute name with an embedded '.' symbol (e.g. 'a.b'), which can only 447 be accessed indirectly (e.g. through getattr and setattr). To set this 448 up the user will first need to explicitly add the variable to the hparam 449 object (e.g. "hparams.add_hparam(name='a.b', value=0.0)"). 450 * Benchmark for tf.scan in graph and eager modes. 451 * Added complex128 support to FFT, FFT2D, FFT3D, IFFT, IFFT2D, and IFFT3D. 452 * Making ids unique in `nn.embedding_lookup_sparse`. This helps to reduce 453 RPC calls for looking up the embeddings when there are repeated ids in 454 the batch. 455 * Support indicator column in boosted trees. 456 * Prevent `tf.gradients()` from backpropagating through integer tensors. 457 * LinearOperator[1D,2D,3D]Circulant added to `tensorflow.linalg`. 458 * Conv3D, Conv3DBackpropInput, Conv3DBackpropFilter now supports 459 arbitrary. 460 * Added `tf.train.Checkpoint` for reading/writing object-based 461 checkpoints. 462 * Added LinearOperatorKronecker, a dense-free implementation of the 463 Kronecker Product. 464 * Allow LinearOperator to broadcast. 465 * SavedModelBuilder will now deduplicate asset names that point to files 466 with the same basename and the same contents. Note that this may result 467 in new asset files included in SavedModels in cases where assets with 468 the same name but different contents were previously overwriting each 469 other. 470 471## Thanks to our Contributors 472 473This release contains contributions from many people at Google, as well as: 474 475Abdullah 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 476 477# Release 1.8.0 478 479## Major Features And Improvements 480* Can now pass `tf.contrib.distribute.MirroredStrategy()` to `tf.estimator.RunConfig()` to run an Estimator model on multiple GPUs on one machine. 481* Add `tf.contrib.data.prefetch_to_device()`, which supports prefetching to GPU memory. 482* Added Gradient Boosted Trees as pre-made Estimators: BoostedTreesClassifier, BoostedTreesRegressor. 483* Add 3rd generation pipeline config for Cloud TPUs which improves performance and usability. 484* `tf.contrib.bayesflow` is moving out to it's own repo. 485* Added `tf.contrib.{proto,rpc}` to allow generic proto parsing and RPC communication<sup>[1](#rpc-issue)</sup>. 486 487## Bug Fixes and Other Changes 488* `tf.data`: 489 * Add `tf.contrib.data.prefetch_to_device`, which enables prefetching dataset elements to GPU memory. 490 * Add `tf.contrib.data.AUTOTUNE`, which allows the tf.data runtime to automatically tune the prefetch buffer sizes based on your system and environment. 491 * Add `tf.contrib.data.make_csv_dataset` for building datasets of CSV files. 492* Eager Execution: 493 * 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. 494 * 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) 495 * `tf.GradientTape` has moved out of contrib. 496* `tf.keras`: 497 * Added the fashion mnist dataset. 498 * New data preprocessing functions: `image/random_brightness`, `sequence/TimeseriesGenerator`, and `text/hashing_trick`. 499* Accelerated Linear Algebra (XLA): 500 * Select and scatter in reference util and evaluator now use lexicographical order to break ties. 501* TensorFlow Debugger (tfdbg) CLI: 502 * During tensor-filter operations, allow exclusion of nodes by regular expressions. 503 * Fix spurious background colors in some text terminals. 504* `tf.contrib`: 505 * Add meta-distribution BatchReshape which reshapes batch dimensions. 506 * `tf.contrib.layers.recompute_grad` works for explicit gradient checkpointing on TPU. 507 * Add `tf.contrib.framework.argsort`. 508 * Allow `DNNBoostedTreeCombinedEstimator` to work with core versions of feature columns and losses. 509 * Add non-linear image warping ops: `tf.contrib.image.sparse_image_warp`, `tf.contrib.image.dense_image_warp`, and `tf.contrib.image.interpolate_spline`. 510 * Fix bug in `tf.contrib.opt.MultitaskOptimizerWrapper` where types of tensors were mismatched. 511* Other: 512 * 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. 513 * Add description of shapes and a pointer to tutorial notebook in `tf.distributions.Distribution`. 514 * Update scatter operations: 515 * Add `tf.scatter_min` and `tf.scatter_max` 516 * Extend scatter operations to work with a scalar update parameter. 517 * Move cuDNN RNN ops to core for use in TensorFlow codebase only. 518 * Add `float64` support for `Conv2d`, `Conv2dBackpropInput`, and `Conv2dBackpropFilter`. 519 * Add `float64` support for `AvgPool`/`AvgPoolGrad`. 520 * Make graph name scope thread local so that they work correctly in multi-threaded environments. 521 * Update nsync synchronization library to avoid slow primitives on Linux. 522 * Removed need to put nsync/public on C include path when building custom ops. 523 * Add `tf.image.psnr`, `tf.image.ssim`, `tf.image.ssim_multiscale`, `tf.image.image_gradients`, `tf.image.sobel_edges`. 524 * Add links to https://js.tensorflow.org. 525 * Fix non-uniformity of orthogonal matrices. 526 * Fix bug where multi-image Estimator eval summaries were not displayed correctly. 527 528<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. 529 530## Thanks to our Contributors 531 532This release contains contributions from many people at Google, as well as: 533 5344d55397500, 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 535 536# Release 1.7.0 537 538## Major Features And Improvements 539* Eager mode is moving out of contrib, try `tf.enable_eager_execution()`. 540* Graph rewrites emulating fixed-point quantization compatible with TensorFlow Lite, supported by new `tf.contrib.quantize` package. 541* Easily customize gradient computation with `tf.custom_gradient`. 542* [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. 543* Experimental support for reading a sqlite database as a `Dataset` with new `tf.contrib.data.SqlDataset`. 544* Distributed Mutex / CriticalSection added to `tf.contrib.framework.CriticalSection`. 545* Better text processing with `tf.regex_replace`. 546* Easy, efficient sequence input with `tf.contrib.data.bucket_by_sequence_length` 547* Initial support for `tf.contrib.tensorrt` that enables native TensorRT in 548 TensorFlow. 549 550## Bug Fixes and Other Changes 551* Accelerated Linear Algebra (XLA): 552 * Add `MaxPoolGradGrad` support for XLA 553 * CSE pass from Tensorflow is now disabled in XLA. 554* `tf.data`: 555 * `tf.data.Dataset` 556 * Add support for building C++ Dataset op kernels as external libraries, using the `tf.load_op_library()` mechanism. 557 * `Dataset.list_files()` now shuffles its output by default. 558 * `Dataset.shuffle(..., seed=tf.constant(0, dtype=tf.int64))` now yields the same sequence of elements as `Dataset.shuffle(..., seed=0)`. 559 * Add `num_parallel_reads` argument to `tf.data.TFRecordDataset`. 560* `tf.contrib`: 561 * `tf.contrib.bayesflow.halton_sequence` now supports randomization. 562 * Add support for scalars in `tf.contrib.all_reduce`. 563 * Add `effective_sample_size` to `tf.contrib.bayesflow.mcmc_diagnostics`. 564 * Add `potential_scale_reduction` to `tf.contrib.bayesflow.mcmc_diagnostics`. 565 * Add `BatchNormalization`, `Kumaraswamy` bijectors. 566 * Deprecate `tf.contrib.learn`. Please check contrib/learn/README.md for instructions on how to convert existing code. 567 * `tf.contrib.data` 568 * Remove deprecated `tf.contrib.data.Dataset`, `tf.contrib.data.Iterator`, `tf.contrib.data.FixedLengthRecordDataset`, `tf.contrib.data.TextLineDataset`, and `tf.contrib.data.TFRecordDataset` classes. 569 * Added `bucket_by_sequence_length`, `sliding_window_batch`, and `make_batched_features_dataset` 570 * Remove unmaintained `tf.contrib.ndlstm`. You can find it externally at https://github.com/tmbarchive/tfndlstm. 571 * Moved most of `tf.contrib.bayesflow` to its own repo: `tfp` 572* Other: 573 * tf.py_func now reports the full stack trace if an exception occurs. 574 * Integrate `TPUClusterResolver` with GKE's integration for Cloud TPUs. 575 * Add a library for statistical testing of samplers. 576 * Add Helpers to stream data from the GCE VM to a Cloud TPU. 577 * Integrate ClusterResolvers with TPUEstimator. 578 * Unify metropolis_hastings interface with HMC kernel. 579 * Move LIBXSMM convolutions to a separate --define flag so that they are disabled by default. 580 * Fix `MomentumOptimizer` lambda. 581 * Reduce `tfp.layers` boilerplate via programmable docstrings. 582 * Add `auc_with_confidence_intervals`, a method for computing the AUC and confidence interval with linearithmic time complexity. 583 * `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. 584 * Fix `initialized_value` and `initial_value` behaviors for `ResourceVariables` created from `VariableDef` protos. 585 * Add TensorSpec to represent the specification of Tensors. 586 * Constant folding pass is now deterministic. 587 * Support `float16` `dtype` in `tf.linalg.*`. 588 * Add `tf.estimator.export.TensorServingInputReceiver` that allows `tf.estimator.Estimator.export_savedmodel` to pass raw tensors to model functions. 589 590## Deprecations 591 592* TensorFlow 1.7 may be the last time we support Cuda versions below 8.0. 593 Starting with TensorFlow 1.8 release, 8.0 will be the minimum supported 594 version. 595* TensorFlow 1.7 may be the last time we support cuDNN versions below 6.0. 596 Starting with TensorFlow 1.8 release, 6.0 will be the minimum supported 597 version. 598 599## Thanks to our Contributors 600 601This release contains contributions from many people at Google, as well as: 602 6034d55397500, 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 604 605 606 607# Release 1.6.0 608 609## Breaking Changes 610* Prebuilt binaries are now built against CUDA 9.0 and cuDNN 7. 611* Prebuilt binaries will use AVX instructions. This may break TF on older CPUs. 612 613## Major Features And Improvements 614* New Optimizer internal API for non-slot variables. Descendants of AdamOptimizer that access _beta[12]_power will need to be updated. 615* `tf.estimator.{FinalExporter,LatestExporter}` now export stripped SavedModels. This improves forward compatibility of the SavedModel. 616* FFT support added to XLA CPU/GPU. 617 618## Bug Fixes and Other Changes 619* Documentation updates: 620 * Added a second version of Getting Started, which is aimed at ML 621newcomers. 622 * Clarified documentation on `resize_images.align_corners` parameter. 623 * Additional documentation for TPUs. 624* Google Cloud Storage (GCS): 625 * Add client-side throttle. 626 * Add a `FlushCaches()` method to the FileSystem interface, with an implementation for GcsFileSystem. 627* Other: 628 * Add `tf.contrib.distributions.Kumaraswamy`. 629 * `RetryingFileSystem::FlushCaches()` calls the base FileSystem's `FlushCaches()`. 630 * Add `auto_correlation` to distributions. 631 * Add `tf.contrib.distributions.Autoregressive`. 632 * Add SeparableConv1D layer. 633 * Add convolutional Flipout layers. 634 * When both inputs of `tf.matmul` are bfloat16, it returns bfloat16, instead of float32. 635 * Added `tf.contrib.image.connected_components`. 636 * Add `tf.contrib.framework.CriticalSection` that allows atomic variable access. 637 * Output variance over trees predictions for classifications tasks. 638 * For `pt` and `eval` commands, allow writing tensor values to filesystem as numpy files. 639 * gRPC: Propagate truncated errors (instead of returning gRPC internal error). 640 * Augment `parallel_interleave` to support 2 kinds of prefetching. 641 * Improved XLA support for C64-related ops log, pow, atan2, tanh. 642 * Add probabilistic convolutional layers. 643 644## API Changes 645* Introducing `prepare_variance` boolean with default setting to False for backward compatibility. 646* Move `layers_dense_variational_impl.py` to `layers_dense_variational.py`. 647 648## Known Bugs 649* Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or 650 `CUDA_ILLEGAL_ADDRESS` failures. 651 652 Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA 9 653 and CUDA 9.1 sometimes does not properly compute the carry bit when 654 decomposing 64-bit address calculations with large offsets (e.g. `load [x + 655 large_constant]`) into 32-bit arithmetic in SASS. 656 657 As a result, these versions of `ptxas` miscompile most XLA programs which use 658 more than 4GB of temp memory. This results in garbage results and/or 659 `CUDA_ERROR_ILLEGAL_ADDRESS` failures. 660 661 A fix in CUDA 9.1.121 is expected in late February 2018. We do not expect a 662 fix for CUDA 9.0.x. Until the fix is available, the only workaround is to 663 [downgrade](https://developer.nvidia.com/cuda-toolkit-archive) to CUDA 8.0.x 664 or disable XLA:GPU. 665 666 TensorFlow will print a warning if you use XLA:GPU with a known-bad version of 667 CUDA; see e00ba24c4038e7644da417ddc639169b6ea59122. 668 669## Thanks to our Contributors 670 671This release contains contributions from many people at Google, as well as: 672 6734d55397500, Ag Ramesh, Aiden Scandella, Akimasa Kimura, Alex Rothberg, Allen Goodman, 674amilioto, Andrei Costinescu, Andrei Nigmatulin, Anjum Sayed, Anthony Platanios, 675Anush Elangovan, Armando Fandango, Ashish Kumar Ram, Ashwini Shukla, Ben, Bhavani Subramanian, 676Brett Koonce, Carl Thomé, cclauss, Cesc, Changming Sun, Christoph Boeddeker, Clayne Robison, 677Clemens Schulz, Clint (Woonhyuk Baek), codrut3, Cole Gerdemann, Colin Raffel, Daniel Trebbien, 678Daniel Ylitalo, Daniel Zhang, Daniyar, Darjan Salaj, Dave Maclachlan, David Norman, Dong--Jian, 679dongsamb, dssgsra, Edward H, eladweiss, elilienstein, Eric Lilienstein, error.d, Eunji Jeong, fanlu, 680Florian Courtial, fo40225, Fred, Gregg Helt, Guozhong Zhuang, Hanchen Li, hsm207, hyunyoung2, 681ImSheridan, Ishant Mrinal Haloi, Jacky Ko, Jay Young, Jean Flaherty, Jerome, JerrikEph, Jesse 682Kinkead, jfaath, Jian Lin, jinghuangintel, Jiongyan Zhang, Joel Hestness, Joel Shor, Johnny Chan, 683Julian Niedermeier, Julian Wolff, JxKing, K-W-W, Karl Lessard, Kasper Marstal, Keiji Ariyama, 684Koan-Sin Tan, Loki Der Quaeler, Loo Rong Jie, Luke Schaefer, Lynn Jackson, ManHyuk, Matt Basta, 685Matt Smith, Matthew Schulkind, Michael, michaelkhan3, Miguel Piedrafita, Mikalai Drabovich, 686Mike Knapp, mjwen, mktozk, Mohamed Aly, Mohammad Ashraf Bhuiyan, Myungjoo Ham, Naman Bhalla, 687Namrata-Ibm, Nathan Luehr, nathansilberman, Netzeband, Niranjan Hasabnis, Omar Aflak, Ozge 688Yalcinkaya, Parth P Panchal, patrickzzy, Patryk Chrabaszcz, Paul Van Eck, Paweł Kapica, Peng Yu, 689Philip Yang, Pierre Blondeau, Po-Hsien Chu, powderluv, Puyu Wang, Rajendra Arora, Rasmus, Renat 690Idrisov, resec, Robin Richtsfeld, Ronald Eddy Jr, Sahil Singh, Sam Matzek, Sami Kama, sandipmgiri, 691Santiago Castro, Sayed Hadi Hashemi, Scott Tseng, Sergii Khomenko, Shahid, Shengpeng Liu, Shreyash 692Sharma, Shrinidhi Kl, Simone Cirillo, simsicon, Stanislav Levental, starsblinking, Stephen Lumenta, 693Steven Hickson, Su Tang, Taehoon Lee, Takuya Wakisaka, Ted Chang, Ted Ying, Tijmen Verhulsdonck, 694Timofey Kondrashov, vade, vaibhav, Valentin Khrulkov, vchigrin, Victor Costan, Viraj Navkal, 695Vivek Rane, wagonhelm, Yan Facai (颜发才), Yanbo Liang, Yaroslav Bulatov, yegord, Yong Tang, 696Yoni Tsafir, yordun, Yuan (Terry) Tang, Yuxin Wu, zhengdi, Zhengsheng Wei, 田传武 697 698# Release 1.5.0 699 700## Breaking Changes 701* Prebuilt binaries are now built against CUDA 9.0 and cuDNN 7. 702* Starting from 1.6 release, our prebuilt binaries will use AVX instructions. 703 This may break TF on older CPUs. 704 705## Major Features And Improvements 706* [Eager execution](https://github.com/tensorflow/tensorflow/tree/r1.5/tensorflow/contrib/eager) 707 preview version is now available. 708* [TensorFlow Lite](https://github.com/tensorflow/tensorflow/tree/r1.5/tensorflow/lite) 709 dev preview is now available. 710* CUDA 9.0 and cuDNN 7 support. 711* Accelerated Linear Algebra (XLA): 712 * Add `complex64` support to XLA compiler. 713 * `bfloat` support is now added to XLA infrastructure. 714 * Make `ClusterSpec` propagation work with XLA devices. 715 * Use a deterministic executor to generate XLA graph. 716* `tf.contrib`: 717 * `tf.contrib.distributions`: 718 * Add `tf.contrib.distributions.Autoregressive`. 719 * Make `tf.contrib.distributions` QuadratureCompound classes support batch 720 * Infer `tf.contrib.distributions.RelaxedOneHotCategorical` `dtype` from arguments. 721 * Make `tf.contrib.distributions` quadrature family parameterized by 722 `quadrature_grid_and_prob` vs `quadrature_degree`. 723 * `auto_correlation` added to `tf.contrib.distributions` 724 * Add `tf.contrib.bayesflow.layers`, a collection of probabilistic (neural) layers. 725 * Add `tf.contrib.bayesflow.halton_sequence`. 726 * Add `tf.contrib.data.make_saveable_from_iterator.` 727 * Add `tf.contrib.data.shuffle_and_repeat`. 728 * Add new custom transformation: `tf.contrib.data.scan()`. 729 * `tf.contrib.distributions.bijectors`: 730 * Add `tf.contrib.distributions.bijectors.MaskedAutoregressiveFlow`. 731 * Add `tf.contrib.distributions.bijectors.Permute`. 732 * Add `tf.contrib.distributions.bijectors.Gumbel`. 733 * Add `tf.contrib.distributions.bijectors.Reshape`. 734 * Support shape inference (i.e., shapes containing -1) in the Reshape bijector. 735* Add `streaming_precision_recall_at_equal_thresholds,` a method for computing 736 streaming precision and recall with `O(num_thresholds + size of predictions)` 737 time and space complexity. 738* Change `RunConfig` default behavior to not set a random seed, making random 739 behavior independently random on distributed workers. We expect this to 740 generally improve training performance. Models that do rely on determinism 741 should set a random seed explicitly. 742* Replaced the implementation of `tf.flags` with `absl.flags`. 743* Add support for `CUBLAS_TENSOR_OP_MATH` in fp16 GEMM 744* Add support for CUDA on NVIDIA Tegra devices 745 746## Bug Fixes and Other Changes 747* Documentation updates: 748 * Clarified that you can only install TensorFlow on 64-bit machines. 749 * Added a short doc explaining how `Estimator`s save checkpoints. 750 * Add documentation for ops supported by the `tf2xla` bridge. 751 * Fix minor typos in the doc of `SpaceToDepth` and `DepthToSpace`. 752 * Updated documentation comments in `mfcc_mel_filterbank.h` and `mfcc.h` to 753 clarify that the input domain is squared magnitude spectra and the weighting 754 is done on linear magnitude spectra (sqrt of inputs). 755 * Change `tf.contrib.distributions` docstring examples to use `tfd` alias 756 rather than `ds`, `bs`. 757 * Fix docstring typos in `tf.distributions.bijectors.Bijector`. 758 * `tf.assert_equal` no longer raises `ValueError.` It now raises 759 `InvalidArgumentError,` as documented. 760 * Update Getting Started docs and API intro. 761* Google Cloud Storage (GCS): 762 * Add userspace DNS caching for the GCS client. 763 * Customize request timeouts for the GCS filesystem. 764 * Improve GCS filesystem caching. 765* Bug Fixes: 766 * Fix bug where partitioned integer variables got their wrong shapes. Before 767 * Fix correctness bug in CPU and GPU implementations of Adadelta. 768 * Fix a bug in `import_meta_graph`'s handling of partitioned variables when 769 importing into a scope. WARNING: This may break loading checkpoints of 770 graphs with partitioned variables saved after using `import_meta_graph` with 771 a non-empty `import_scope` argument. 772 * Fix bug in offline debugger which prevented viewing events. 773 * Added the `WorkerService.DeleteWorkerSession` method to the gRPC interface, 774 to fix a memory leak. Ensure that your master and worker servers are running 775 the same version of TensorFlow to avoid compatibility issues. 776 * Fix bug in peephole implementation of BlockLSTM cell. 777 * Fix bug by casting dtype of `log_det_jacobian` to match `log_prob` in 778 `TransformedDistribution`. 779 * Fix a bug in `import_meta_graph`'s handling of partitioned variables when 780 * Ensure `tf.distributions.Multinomial` doesn't underflow in `log_prob`. 781 Before this change, all partitions of an integer variable were initialized 782 with the shape of the unpartitioned variable; after this change they are 783 initialized correctly. 784* Other: 785 * Add necessary shape util support for bfloat16. 786 * Add a way to run ops using a step function to MonitoredSession. 787 * Add `DenseFlipout` probabilistic layer. 788 * A new flag `ignore_live_threads` is available on train. If set to `True`, it 789 will ignore threads that remain running when tearing down infrastructure 790 after successfully completing training, instead of throwing a RuntimeError. 791 * Restandardize `DenseVariational` as simpler template for other probabilistic 792 layers. 793 * `tf.data` now supports `tf.SparseTensor` components in dataset elements. 794 * It is now possible to iterate over `Tensor`s. 795 * Allow `SparseSegmentReduction` ops to have missing segment IDs. 796 * Modify custom export strategy to account for multidimensional sparse float 797 splits. 798 * `Conv2D`, `Conv2DBackpropInput`, `Conv2DBackpropFilter` now supports arbitrary 799 dilations with GPU and cuDNNv6 support. 800 * `Estimator` now supports `Dataset`: `input_fn` can return a `Dataset` 801 instead of `Tensor`s. 802 * Add `RevBlock`, a memory-efficient implementation of reversible residual layers. 803 * Reduce BFCAllocator internal fragmentation. 804 * Add `cross_entropy` and `kl_divergence` to `tf.distributions.Distribution`. 805 * Add `tf.nn.softmax_cross_entropy_with_logits_v2` which enables backprop 806 w.r.t. the labels. 807 * GPU back-end now uses `ptxas` to compile generated PTX. 808 * `BufferAssignment`'s protocol buffer dump is now deterministic. 809 * Change embedding op to use parallel version of `DynamicStitch`. 810 * Add support for sparse multidimensional feature columns. 811 * Speed up the case for sparse float columns that have only 1 value. 812 * Allow sparse float splits to support multivalent feature columns. 813 * Add `quantile` to `tf.distributions.TransformedDistribution`. 814 * Add `NCHW_VECT_C` support for `tf.depth_to_space` on GPU. 815 * Add `NCHW_VECT_C` support for `tf.space_to_depth` on GPU. 816 817## API Changes 818* Rename `SqueezeDims` attribute to `Axis` in C++ API for Squeeze op. 819* `Stream::BlockHostUntilDone` now returns Status rather than bool. 820* Minor refactor: move stats files from `stochastic` to `common` and remove 821 `stochastic`. 822 823## Known Bugs 824* Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or 825 `CUDA_ILLEGAL_ADDRESS` failures. 826 827 Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA 9 828 and CUDA 9.1 sometimes does not properly compute the carry bit when 829 decomposing 64-bit address calculations with large offsets (e.g. `load [x + 830 large_constant]`) into 32-bit arithmetic in SASS. 831 832 As a result, these versions of `ptxas` miscompile most XLA programs which use 833 more than 4GB of temp memory. This results in garbage results and/or 834 `CUDA_ERROR_ILLEGAL_ADDRESS` failures. 835 836 A fix in CUDA 9.1.121 is expected in late February 2018. We do not expect a 837 fix for CUDA 9.0.x. Until the fix is available, the only workaround is to 838 [downgrade](https://developer.nvidia.com/cuda-toolkit-archive) to CUDA 8.0.x 839 or disable XLA:GPU. 840 841 TensorFlow will print a warning if you use XLA:GPU with a known-bad version of 842 CUDA; see e00ba24c4038e7644da417ddc639169b6ea59122. 843 844## Thanks to our Contributors 845 846This release contains contributions from many people at Google, as well as: 847 848Adam Zahran, Ag Ramesh, Alan Lee, Alan Yee, Alex Sergeev, Alexander, Amir H. Jadidinejad, 849Amy, Anastasios Doumoulakis, Andrei Costinescu, Andrei Nigmatulin, Anthony Platanios, 850Anush Elangovan, arixlin, Armen Donigian, ArtëM Sobolev, Atlas7, Ben Barsdell, Bill Prin, 851Bo Wang, Brett Koonce, Cameron Thomas, Carl Thomé, Cem Eteke, cglewis, Changming Sun, 852Charles Shenton, Chi-Hung, Chris Donahue, Chris Filo Gorgolewski, Chris Hoyean Song, 853Chris Tava, Christian Grail, Christoph Boeddeker, cinqS, Clayne Robison, codrut3, concerttttt, 854CQY, Dan Becker, Dan Jarvis, Daniel Zhang, David Norman, dmaclach, Dmitry Trifonov, 855Donggeon Lim, dongpilYu, Dr. Kashif Rasul, Edd Wilder-James, Eric Lv, fcharras, Felix Abecassis, 856FirefoxMetzger, formath, FredZhang, Gaojin Cao, Gary Deer, Guenther Schmuelling, Hanchen Li, 857Hanmin Qin, hannesa2, hyunyoung2, Ilya Edrenkin, Jackson Kontny, Jan, Javier Luraschi, 858Jay Young, Jayaram Bobba, Jeff, Jeff Carpenter, Jeremy Sharpe, Jeroen BéDorf, Jimmy Jia, 859Jinze Bai, Jiongyan Zhang, Joe Castagneri, Johan Ju, Josh Varty, Julian Niedermeier, 860JxKing, Karl Lessard, Kb Sriram, Keven Wang, Koan-Sin Tan, Kyle Mills, lanhin, LevineHuang, 861Loki Der Quaeler, Loo Rong Jie, Luke Iwanski, LáSzló Csomor, Mahdi Abavisani, Mahmoud Abuzaina, 862ManHyuk, Marek ŠUppa, MathSquared, Mats Linander, Matt Wytock, Matthew Daley, Maximilian Bachl, 863mdymczyk, melvyniandrag, Michael Case, Mike Traynor, miqlas, Namrata-Ibm, Nathan Luehr, 864Nathan Van Doorn, Noa Ezra, Nolan Liu, Oleg Zabluda, opensourcemattress, Ouwen Huang, 865Paul Van Eck, peisong, Peng Yu, PinkySan, pks, powderluv, Qiao Hai-Jun, Qiao Longfei, 866Rajendra Arora, Ralph Tang, resec, Robin Richtsfeld, Rohan Varma, Ryohei Kuroki, SaintNazaire, 867Samuel He, Sandeep Dcunha, sandipmgiri, Sang Han, scott, Scott Mudge, Se-Won Kim, Simon Perkins, 868Simone Cirillo, Steffen Schmitz, Suvojit Manna, Sylvus, Taehoon Lee, Ted Chang, Thomas Deegan, 869Till Hoffmann, Tim, Toni Kunic, Toon Verstraelen, Tristan Rice, Urs KöSter, Utkarsh Upadhyay, 870Vish (Ishaya) Abrams, Winnie Tsang, Yan Chen, Yan Facai (颜发才), Yi Yang, Yong Tang, 871Youssef Hesham, Yuan (Terry) Tang, Zhengsheng Wei, zxcqwe4906, 张志豪, 田传武 872 873We are also grateful to all who filed issues or helped resolve them, asked and 874answered questions, and were part of inspiring discussions. 875 876# Release 1.4.1 877 878## Bug Fixes and Other Changes 879* `LinearClassifier` fix. 880 881# Release 1.4.0 882 883## Major Features And Improvements 884* `tf.keras` is now part of the core TensorFlow API. 885* [`tf.data`](http://tensorflow.org/guide/datasets) is now part of 886 the core TensorFlow API. 887 * The API is now subject to backwards compatibility guarantees. 888 * For a guide to migrating from the `tf.contrib.data` API, see the 889 [README](https://github.com/tensorflow/tensorflow/blob/r1.4/tensorflow/contrib/data/README.md). 890 * Major new features include `Dataset.from_generator()` (for building an input 891 pipeline from a Python generator), and the `Dataset.apply()` method for 892 applying custom transformation functions. 893 * Several custom transformation functions have been added, including 894 `tf.contrib.data.batch_and_drop_remainder()` and 895 `tf.contrib.data.sloppy_interleave()`. 896* Add `train_and_evaluate` for simple distributed `Estimator` training. 897* Add `tf.spectral.dct` for computing the DCT-II. 898* Add Mel-Frequency Cepstral Coefficient support to `tf.contrib.signal` 899 (with GPU and gradient support). 900* Add a self-check on `import tensorflow` for Windows DLL issues. 901* Add NCHW support to `tf.depth_to_space` on GPU. 902* TensorFlow Debugger (tfdbg): 903 * Add `eval` command to allow evaluation of arbitrary Python/numpy expressions 904 in tfdbg command-line interface. See 905 [Debugging TensorFlow Programs](https://www.tensorflow.org/guide/debugger) 906 for more details. 907 * Usability improvement: The frequently used tensor filter `has_inf_or_nan` is 908 now added to `Session` wrappers and hooks by default. So there is no need 909 for clients to call `.add_tensor_filter(tf_debug.has_inf_or_nan)` anymore. 910* SinhArcsinh (scalar) distribution added to `contrib.distributions`. 911* Make `GANEstimator` opensource. 912* `Estimator.export_savedmodel()` now includes all valid serving signatures 913 that can be constructed from the Serving Input Receiver and all available 914 ExportOutputs. For instance, a classifier may provide regression- and 915 prediction-flavored outputs, in addition to the classification-flavored one. 916 Building signatures from these allows TF Serving to honor requests using the 917 different APIs (Classify, Regress, and Predict). Furthermore, 918 `serving_input_receiver_fn()` may now specify alternative subsets of nodes 919 that may act as inputs. This allows, for instance, producing a prediction 920 signature for a classifier that accepts raw `Tensors` instead of a serialized 921 `tf.Example`. 922* Add `tf.contrib.bayesflow.hmc`. 923* Add `tf.contrib.distributions.MixtureSameFamily`. 924* Make `Dataset.shuffle()` always reshuffles after each iteration by default. 925* Add `tf.contrib.bayesflow.metropolis_hastings`. 926* Add `log_rate` parameter to `tf.contrib.distributions.Poisson`. 927* Extend `tf.contrib.distributions.bijector` API to handle some non-injective 928 transforms. 929* Java: 930 * Generics (e.g., `Tensor<Integer>`) for improved type-safety 931 (courtesy @andrewcmyers). 932 * Support for multi-dimensional string tensors. 933 * Support loading of custom operations (e.g. many in `tf.contrib`) on Linux 934 and OS X 935* All our prebuilt binaries have been built with CUDA 8 and cuDNN 6. 936 We anticipate releasing TensorFlow 1.5 with CUDA 9 and cuDNN 7. 937 938## Bug Fixes and Other Changes 939* `tf.nn.rnn_cell.DropoutWrapper` is now more careful about dropping out LSTM 940 states. Specifically, it no longer ever drops the `c` (memory) state of an 941 `LSTMStateTuple`. The new behavior leads to proper dropout behavior 942 for LSTMs and stacked LSTMs. This bug fix follows recommendations from 943 published literature, but is a behavioral change. State dropout behavior 944 may be customized via the new `dropout_state_filter_visitor` argument. 945* Removed `tf.contrib.training.python_input`. The same behavior, in a more 946 flexible and reproducible package, is available via the new 947 `tf.contrib.data.Dataset.from_generator` method! 948* Fix `tf.contrib.distributions.Affine` incorrectly computing log-det-jacobian. 949* Fix `tf.random_gamma` incorrectly handling non-batch, scalar draws. 950* Resolved a race condition in TensorForest TreePredictionsV4Op. 951* Google Cloud Storage file system, Amazon S3 file system, and Hadoop file 952 system support are now default build options. 953* Custom op libraries must link against libtensorflow_framework.so 954 (installed at `tf.sysconfig.get_lib()`). 955* Change `RunConfig` default behavior to not set a random seed, making random 956 behavior independently random on distributed workers. We expect this to 957 generally improve training performance. Models that do rely on determinism 958 should set a random seed explicitly. 959 960## Breaking Changes to the API 961* The signature of the `tf.contrib.data.rejection_resample()` function has been 962 changed. It now returns a function that can be used as an argument to 963 `Dataset.apply()`. 964* Remove `tf.contrib.data.Iterator.from_dataset()` method. Use 965 `Dataset.make_initializable_iterator()` instead. 966* Remove seldom used and unnecessary `tf.contrib.data.Iterator.dispose_op()`. 967* Reorder some TF-GAN loss functions in a non-backwards compatible way. 968 969## Known Issues 970* In Python 3, `Dataset.from_generator()` does not support Unicode strings. 971 You must convert any strings to bytes objects before yielding them from 972 the generator. 973 974## Thanks to our Contributors 975 976This release contains contributions from many people at Google, as well as: 977 9784d55397500, Abdullah Alrasheed, abenmao, Adam Salvail, Aditya Dhulipala, Ag Ramesh, 979Akimasa Kimura, Alan Du, Alan Yee, Alexander, Amit Kushwaha, Amy, Andrei Costinescu, 980Andrei Nigmatulin, Andrew Erlichson, Andrew Myers, Andrew Stepanov, Androbin, AngryPowman, 981Anish Shah, Anton Daitche, Artsiom Chapialiou, asdf2014, Aseem Raj Baranwal, Ash Hall, 982Bart Kiers, Batchu Venkat Vishal, ben, Ben Barsdell, Bill Piel, Carl Thomé, Catalin Voss, 983Changming Sun, Chengzhi Chen, Chi Zeng, Chris Antaki, Chris Donahue, Chris Oelmueller, 984Chris Tava, Clayne Robison, Codrut, Courtial Florian, Dalmo Cirne, Dan J, Darren Garvey, 985David Kristoffersson, David Norman, David RöThlisberger, DavidNorman, Dhruv, DimanNe, 986Dorokhov, Duncan Mac-Vicar P, EdwardDixon, EMCP, error.d, FAIJUL, Fan Xia, 987Francois Xavier, Fred Reiss, Freedom" Koan-Sin Tan, Fritz Obermeyer, Gao, Xiang, 988Guenther Schmuelling, Guo Yejun (郭叶军), Hans Gaiser, HectorSVC, Hyungsuk Yoon, 989James Pruegsanusak, Jay Young, Jean Wanka, Jeff Carpenter, Jeremy Rutman, Jeroen BéDorf, 990Jett Jones, Jimmy Jia, jinghuangintel, jinze1994, JKurland, Joel Hestness, joetoth, 991John B Nelson, John Impallomeni, John Lawson, Jonas, Jonathan Dekhtiar, joshkyh, Jun Luan, 992Jun Mei, Kai Sasaki, Karl Lessard, karl@kubx.ca, Kb Sriram, Kenichi Ueno, Kevin Slagle, 993Kongsea, Lakshay Garg, lhlmgr, Lin Min, liu.guangcong, Loki Der Quaeler, Louie Helm, 994lucasmoura, Luke Iwanski, Lyndon White, Mahmoud Abuzaina, Marcel Puyat, Mark Aaron Shirley, 995Michele Colombo, MtDersvan, Namrata-Ibm, Nathan Luehr, Naurril, Nayana Thorat, Nicolas Lopez, 996Niranjan Hasabnis, Nolan Liu, Nouce, Oliver Hennigh, osdamv, Patrik Erdes, 997Patryk Chrabaszcz, Pavel Christof, Penghao Cen, postBG, Qingqing Cao, Qingying Chen, qjivy, 998Raphael, Rasmi, raymondxyang, Renze Yu, resec, Roffel, Ruben Vereecken, Ryohei Kuroki, 999sandipmgiri, Santiago Castro, Scott Kirkland, Sean Vig, Sebastian Raschka, Sebastian Weiss, 1000Sergey Kolesnikov, Sergii Khomenko, Shahid, Shivam Kotwalia, Stuart Berg, Sumit Gouthaman, 1001superzerg, Sven Mayer, tetris, Ti Zhou, Tiago Freitas Pereira, Tian Jin, Tomoaki Oiki, 1002Vaibhav Sood, vfdev, Vivek Rane, Vladimir Moskva, wangqr, Weber Xie, Will Frey, 1003Yan Facai (颜发才), yanivbl6, Yaroslav Bulatov, Yixing Lao, Yong Tang, youkaichao, 1004Yuan (Terry) Tang, Yue Zhang, Yuxin Wu, Ziming Dong, ZxYuan, 黄璞 1005 1006We are also grateful to all who filed issues or helped resolve them, asked and 1007answered questions, and were part of inspiring discussions. 1008 1009# Release 1.3.0 1010 1011See also [TensorBoard 0.1.4](https://github.com/tensorflow/tensorboard/releases/tag/0.1.4) release notes. 1012 1013## Major Features and Improvements 1014* Added canned estimators to Tensorflow library. List of added estimators: 1015 * `DNNClassifier` 1016 * `DNNRegressor` 1017 * `LinearClassifier` 1018 * `LinearRegressor` 1019 * `DNNLinearCombinedClassifier` 1020 * `DNNLinearCombinedRegressor`. 1021* All our prebuilt binaries have been built with cuDNN 6. We anticipate releasing TensorFlow 1.4 with cuDNN 7. 1022* `import tensorflow` now goes much faster. 1023* 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. 1024* Added an axis parameter to `tf.gather`. 1025* Added a `constant_values` keyword argument to `tf.pad`. 1026* Adds `Dataset.interleave` transformation. 1027* Add `ConcatenateDataset` to concatenate two datasets. 1028* Added Mobilenet support to TensorFlow for Poets training script. 1029* Adds a block cache to the GCS filesystem with configurable block size and count. 1030* SinhArcSinh bijector added. 1031* Added `Dataset.list_files` API. 1032* Introduces new operations and Python bindings for the Cloud TPU. 1033* Adding TensorFlow-iOS CocoaPod for symmetry with tensorflow-android. 1034* Introduces base implementations of ClusterResolvers. 1035* Unify memory representations of TensorShape and PartialTensorShape. As a consequence, tensors now have a maximum of 254 dimensions, not 255. 1036* Changed references to LIBXSMM to use version 1.8.1. 1037* TensorFlow Debugger (tfdbg): 1038 * Display summaries of numeric tensor values with the `-s` flag to command `print_tensor` or `pt`. 1039 * Display feed values with the `print_feed` or `pf` command and clickable links in the curses UI. 1040 * Runtime profiler at the op level and the Python source line level with the `run -p` command. 1041* Initial release of the statistical distribution library `tf.distributions`. 1042* GPU kernels and speed improvements for unary `tf.where` and `tf.nn.top_k`. 1043* Monotonic Attention wrappers added to `tf.contrib.seq2seq`. 1044* Added `tf.contrib.signal`, a library for signal processing primitives. 1045* Added `tf.contrib.resampler`, containing CPU and GPU ops for differentiable resampling of images. 1046 1047## Breaking Changes to the API 1048* `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. 1049* 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. 1050 1051## Changes to contrib APIs 1052* Adds tf.contrib.nn.rank_sampled_softmax_loss, a sampled-softmax variant that can improve rank loss. 1053* `tf.contrib.metrics`.{streaming_covariance,streaming_pearson_correlation} modified to return nan when they have seen less or equal to 1 unit of weight. 1054* Adds time series models to contrib. See contrib/timeseries/README.md for details. 1055* Adds FULLY_CONNECTED Op to tensorflow/lite/schema.fbs 1056 1057## Known Issues 1058* Tensorflow_gpu compilation fails with Bazel 0.5.3. 1059 1060## Bug Fixes and Other Changes 1061* Fixes `strides` and `begin` dtype mismatch when slicing using int64 Tensor index in python. 1062* Improved convolution padding documentation. 1063* Add a tag constant, gpu, to present graph with GPU support. 1064* `saved_model.utils` now support SparseTensors transparently. 1065* A more efficient implementation of non-max suppression. 1066* Add support for the shrinkage-type L2 to FtrlOptimizer in addition to the online L2 it already supports. 1067* Fix negative variance in moments calculation. 1068* Expand UniqueOp Benchmark Tests to cover more collision cases. 1069* Improves stability of GCS filesystem on Mac. 1070* Add time estimation to HloCostAnalysis. 1071* 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. 1072* Added None check for save_path in `saver.restore`. 1073* Register devices under their legacy names in device_mgr to ease the transition to clusterspec-propagated configurations. 1074* VectorExponential added to distributions. 1075* Add a bitwise module with bitwise_and, bitwise_or, bitwise_xor, and invert functions. 1076* Add fixed-grid ODE integration routines. 1077* Allow passing bounds to ScipyOptimizerInterface. 1078* Correctness fixes for fft_length parameter to `tf.spectral.rfft` & `tf.spectral.irfft`. 1079* 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. 1080* Add in-memory caching to the Dataset API. 1081* Set default end_of_sequence variable in datasets iterators to false. 1082* [Performance] Increase performance of `tf.layers.conv2d` when setting use_bias=True by 2x by using nn.bias_add. 1083* Update iOS examples to use CocoaPods, and moved to tensorflow/examples/ios. 1084* Adds a family= attribute in `tf.summary` ops to allow controlling the tab name used in Tensorboard for organizing summaries. 1085* When GPU is configured, do not require --config=cuda, instead, automatically build for GPU if this is requested in the configure script. 1086* Fix incorrect sampling of small probabilities in CPU/GPU multinomial. 1087* 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. 1088* Allow uses of over-parameterized separable convolution. 1089* TensorForest multi-regression bug fix. 1090* Framework now supports armv7, cocoapods.org now displays correct page. 1091* Script to create iOS framework for CocoaPods. 1092* Android releases of TensorFlow are now pushed to jcenter for easier integration into apps. See https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/android/README.md for more details. 1093* TensorFlow Debugger (tfdbg): 1094 * Fixed a bug that prevented tfdbg from functioning with multi-GPU setups. 1095 * Fixed a bug that prevented tfdbg from working with `tf.Session.make_callable`. 1096 1097## Thanks to our Contributors 1098 1099This release contains contributions from many people at Google, as well as: 1100 11014F2E4A2E, Adriano Carmezim, Adrià Arrufat, Alan Yee, Alex Lattas, Alex Rothberg, 1102Alexandr Baranezky, Ali Siddiqui, Andreas Solleder, Andrei Costinescu, Andrew Hundt, 1103Androbin, Andy Kernahan, Anish Shah, Anthony Platanios, Arvinds-Ds, b1rd, Baptiste 1104Arnaud, Ben Mabey, Benedikt Linse, Beomsu Kim, Bo Wang, Boyuan Deng, Brett Koonce, 1105Bruno Rosa, Carl Thomé, Changming Sun, Chase Roberts, Chirag Bhatia, Chris Antaki, 1106Chris Hoyean Song, Chris Tava, Christos Nikolaou, Croath Liu, cxx, Czxck001, Daniel 1107Ylitalo, Danny Goodman, Darren Garvey, David Brailovsky, David Norman, DavidNorman, 1108davidpham87, ddurham2, Dhruv, DimanNe, Drew Hintz, Dustin Tran, Earthson Lu, ethiraj, 1109Fabian Winnen, Fei Sun, Freedom" Koan-Sin Tan, Fritz Obermeyer, Gao, Xiang, Gautam, 1110Guenther Schmuelling, Gyu-Ho Lee, Hauke Brammer, horance, Humanity123, J Alammar, 1111Jayeol Chun, Jeroen BéDorf, Jianfei Wang, jiefangxuanyan, Jing Jun Yin, Joan Puigcerver, 1112Joel Hestness, Johannes Mayer, John Lawson, Johnson145, Jon Malmaud, Jonathan Alvarez-Gutierrez, 1113Juang, Yi-Lin, Julian Viereck, Kaarthik Sivashanmugam, Karl Lessard, karl@kubx.ca, Kevin 1114Carbone, Kevin Van Der Burgt, Kongsea, ksellesk, lanhin, Lef Ioannidis, Liangliang He, 1115Louis Tiao, Luke Iwanski, LáSzló Csomor, magixsno, Mahmoud Abuzaina, Marcel Hlopko, Mark 1116Neumann, Maxwell Paul Brickner, mdfaijul, MichaëL Defferrard, Michał JastrzęBski, Michele 1117Colombo, Mike Brodie, Mosnoi Ion, mouradmourafiq, myPrecious, Nayana Thorat, 1118Neeraj Kashyap, Nelson Liu, Niranjan Hasabnis, Olivier Moindrot, orome, Pankaj Gupta, Paul 1119Van Eck, peeyush18, Peng Yu, Pierre, preciousdp11, qjivy, Raingo, raoqiyu, ribx, Richard S. 1120Imaoka, Rishabh Patel, Robert Walecki, Rockford Wei, Ryan Kung, Sahil Dua, Sandip Giri, Sayed 1121Hadi Hashemi, sgt101, Shitian Ni, Shuolongbj, Siim PõDer, Simon Perkins, sj6077, SOLARIS, 1122Spotlight0xff, Steffen Eberbach, Stephen Fox, superryanguo, Sven Mayer, Tapan Prakash, 1123Tiago Morais Morgado, Till Hoffmann, Tj Rana, Vadim Markovtsev, vhasanov, Wei Wu, 1124windead, Yan (Asta) Li, Yan Chen, Yann Henon, Yi Wang, Yong Tang, yorkie, Yuan (Terry) 1125Tang, Yuxin Wu, zhengjiajin, zhongzyd, 黄璞 1126 1127We are also grateful to all who filed issues or helped resolve them, asked and 1128answered questions, and were part of inspiring discussions. 1129 1130# Release 1.2.1 1131 1132## Bug Fixes and Other Changes 1133* Updating markdown version required to >= 2.6.8. 1134* Support tensors as dropout rates again, by removing the min(max(..)) 1135 1136# Release 1.2.0 1137 1138## Major Features and Improvements 1139* Python 3.6 support on Windows. 1140* Added `tf.layers.conv3d_transpose` layer for spatio temporal deconvolution. 1141* Added `tf.Session.make_callable()`, which provides a lower overhead means of running a similar step multiple times. 1142* Added libverbs-based RDMA support to contrib (courtesy @junshi15 from Yahoo). 1143* Bring `tf.feature_column.*` into the API. Non-deprecated functionality from `tf.contrib.layers.*` is moved to `tf.feature_column.*` with cosmetic changes. 1144* `RNNCell` objects now subclass `tf.layers.Layer`. The strictness described 1145 in the TensorFlow 1.1 release is gone: The first time an RNNCell is used, 1146 it caches its scope. All future uses of the RNNCell will reuse variables from 1147 that same scope. This is a breaking change from the behavior of RNNCells 1148 in TensorFlow versions <= 1.0.1. TensorFlow 1.1 had checks in place to 1149 ensure old code works correctly with the new semantics; this version 1150 allows more flexible uses of RNNCell but can lead to subtle errors if 1151 using code meant for TensorFlow <= 1.0.1. For example, writing: 1152 `MultiRNNCell([lstm] * 5)` will now build a 5-layer LSTM stack where each 1153 layer shares the **same** parameters. To get 5 layers each with their own 1154 parameters, write: `MultiRNNCell([LSTMCell(...) for _ in range(5)])`. 1155 If at all unsure, first test your code with TF 1.1; ensure it raises no 1156 errors, and then upgrade to TF 1.2. 1157* RNNCells' variable names have been renamed for consistency with Keras layers. 1158 Specifically, the previous variable names "weights" and "biases" have 1159 been changed to "kernel" and "bias", respectively. 1160 This may cause backward incompatibility with regard to your old 1161 checkpoints containing such RNN cells, in which case you can use the tool 1162 [checkpoint_convert script](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/rnn/python/tools/checkpoint_convert.py) 1163 to convert the variable names in your old checkpoints. 1164* Many of the RNN functions and classes that were in the `tf.nn` namespace 1165 before the 1.0 release and which were moved to `tf.contrib.rnn` have now 1166 been moved back to the core namespace. This includes 1167 `RNNCell`, `LSTMCell`, `GRUCell`, and a number of other cells. These 1168 now reside in `tf.nn.rnn_cell` (with aliases in `tf.contrib.rnn` for backwards 1169 compatibility). The original `tf.nn.rnn` function is now `tf.nn.static_rnn`, 1170 and the bidirectional static and state saving static rnn functions are also 1171 now back in the `tf.nn` namespace. 1172 1173 Notable exceptions are the `EmbeddingWrapper`, `InputProjectionWrapper` and 1174 `OutputProjectionWrapper`, which will slowly be moved to deprecation 1175 in `tf.contrib.rnn`. These are inefficient wrappers that should often 1176 be replaced by calling `embedding_lookup` or `layers.dense` as pre- or post- 1177 processing of the rnn. For RNN decoding, this functionality has been replaced 1178 with an alternative API in `tf.contrib.seq2seq`. 1179* Intel MKL Integration (https://software.intel.com/en-us/articles/tensorflow-optimizations-on-modern-intel-architecture). Intel developed a number of 1180 optimized deep learning primitives: In addition to matrix multiplication and 1181 convolution, these building blocks include: 1182 Direct batched convolution 1183 Pooling: maximum, minimum, average 1184 Normalization: LRN, batch normalization 1185 Activation: rectified linear unit (ReLU) 1186 Data manipulation: multi-dimensional transposition (conversion), split, 1187 concat, sum and scale. 1188* TensorForest Estimator now supports SavedModel export for serving. 1189* Support client-provided ClusterSpec's and propagate them to all workers to enable the creation of dynamic TensorFlow clusters. 1190* TensorFlow C library now available for Windows. 1191* We released a new open-source version of TensorBoard. 1192* [`SavedModel CLI`](https://www.tensorflow.org/versions/master/guide/saved_model_cli) tool available to inspect and execute MetaGraph in SavedModel 1193* Android releases of TensorFlow are now pushed to jcenter for easier 1194 integration into apps. See 1195 https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/android/README.md 1196 for more details. 1197 1198## Deprecations 1199 1200* TensorFlow 1.2 may be the last time we build with cuDNN 5.1. Starting with 1201 TensorFlow 1.3, we will try to build all our prebuilt binaries with cuDNN 6.0. 1202 While we will try to keep our source code compatible with cuDNN 5.1, it will 1203 be best effort. 1204 1205## Breaking Changes to the API 1206* `org.tensorflow.contrib.android.TensorFlowInferenceInterface` now throws exceptions where possible and has simplified method signatures. 1207 1208## Changes to contrib APIs 1209* Added `tf.contrib.util.create_example`. 1210* Added bilinear interpolation to `tf.contrib.image`. 1211* Add `tf.contrib.stateless` for random ops with custom seed control. 1212* MultivariateNormalFullCovariance added to contrib/distributions/ 1213* tensorflow/contrib/rnn undergoes RNN cell variable renaming for 1214 consistency with Keras layers. Specifically, the previous variable names 1215 "weights" and "biases" are changed to "kernel" and "bias", respectively. 1216 This may cause backward incompatibility with regard to your old 1217 checkpoints containing such RNN cells, in which case you can use the 1218 [checkpoint_convert script](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/rnn/python/tools/checkpoint_convert.py) 1219 to convert the variable names in your old checkpoints. 1220* Added `tf.contrib.kernel_methods` module with Ops and estimators for primal 1221 (explicit) kernel methods in TensorFlow. 1222 1223## Bug Fixes and Other Changes 1224* In python, `Operation.get_attr` on type attributes returns the Python DType 1225 version of the type to match expected get_attr documentation rather than the 1226 protobuf enum. 1227* tensorflow/contrib/rnn undergoes RNN cell variable renaming for 1228 consistency with Keras layers. Specifically, the previous variable names 1229 "weights" and "biases" are changed to "kernel" and "bias", respectively. 1230* Changed MIN_SDK version to 8.0 when building iOS libraries. 1231* Fixed LIBXSMM integration. 1232* Make decode_jpeg/decode_png/decode_gif handle all formats, since users frequently try to decode an image as the wrong type. 1233* Improve implicit broadcasting lowering. 1234* Improving stability of GCS/BigQuery clients by a faster retrying of stale transmissions. 1235* Remove OpKernelConstruction::op_def() as part of minimizing proto dependencies. 1236* VectorLaplaceDiag distribution added. 1237* Android demo no longer requires libtensorflow_demo.so to run (libtensorflow_inference.so still required) 1238* Added `categorical_column_with_vocabulary_file`. 1239* Introduce ops for batching/unbatching tensors across Session::Run() calls. 1240* Add tf.log_sigmoid(x) = tf.log(tf.sigmoid(x)) = -tf.nn.softplus(-x). 1241* Changed hooks lists to immutable tuples, and now allow any iterable for the associated arguments. 1242* Introduce TFDecorator. 1243* Added an Mfcc op for speech feature generation. 1244* 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. 1245* Added unreduced NONE, and reduced MEAN options for losses. Removed "WEIGHTED_" prefix from other Reduction constants. 1246* assertAllClose now handles dicts. 1247* Added Gmock matcher for HloInstructions. 1248* Add var name to errors on variable restore. 1249* Added an AudioSpectrogram op for audio feature generation. 1250* Added `reduction` arg to losses. 1251* `tf.placeholder` can represent scalar shapes and partially known. 1252* Remove estimator_spec(mode) argument. 1253* Added an AudioSpectrogram op for audio feature generation. 1254* TensorBoard disables all runs by default if there are more than 40 runs. 1255* Removed old doc generator code. 1256* GCS file system integration now supports domain buckets, e.g gs://bucket.domain.com/path. 1257* Add `tf.summary.text` for outputting text to TensorBoard. 1258* The "run" command of tfdbg's command-line interface now supports filtering of tensors by node name, op type and tensor dtype. 1259* `tf.string_to_number` now supports int64 and float64 outputs. 1260 1261## Thanks to our Contributors 1262 1263This release contains contributions from many people at Google, as well as: 1264 12654F2E4A2E, Aaron Schumacher, Abhi Agg, admcrae, Adriano Carmezim, Adrià Arrufat, 1266agramesh1, Akimitsu Seo, Alan Mosca, Alex Egg, Alex Rothberg, Alexander Heinecke, 1267Alexander Matyasko, Alexandr Baranezky, Alexandre Caulier, Ali Siddiqui, Anand Venkat, 1268Andrew Hundt, Androbin, Anmol Sharma, Arie, Arno Leist, Arron Cao, AuréLien Geron, Bairen Yi, 1269Beomsu Kim, Carl Thomé, cfperez, Changming Sun, Corey Wharton, critiqjo, Dalei Li, Daniel 1270Rasmussen, Daniel Trebbien, DaríO Hereñú, David Eng, David Norman, David Y. Zhang, Davy Song, ddurham2, 1271Deepak Subburam, Dmytro Kyrychuk, Dominic Rossi, Dominik SchlöSser, Dustin Tran, 1272Eduardo Pinho, Egil Martinsson, Elliot Saba, Eric Bigelow, Erik Smistad, Evan Klitzke, 1273Fabrizio Milo, Falcon Dai, Fei Gao, FloopCZ, Fung Lam, Gautam, GBLin5566, Greg Peatfield, 1274Gu Wang, Guenther Schmuelling, Hans Pabst, Harun Gunaydin, Huaizheng, Ido Shamay, Ikaro 1275Silva, Ilya Edrenkin, Immexxx, James Mishra, Jamie Cooke, Jay Young, Jayaram Bobba, 1276Jianfei Wang, jinghua2, Joey Meyer, John Maidens, Jonghoon Jin, Julian Villella, 1277Jun Kim, Jun Shi, Junwei Pan, jyegerlehner, Karan Desai, Karel Van De Plassche, 1278Kb Sriram, KhabarlakKonstantin, Koan-Sin Tan, krivard, Kwotsin, Leandro Gracia Gil, 1279Li Chen, Liangliang He, Louie Helm, lspvic, Luiz Henrique Soares, LáSzló Csomor, 1280Mark Wong, Mathew Wicks, Matthew Rahtz, Maxwell Paul Brickner, Michael Hofmann, Miguel 1281Flores Ruiz De Eguino, MikeTam1021, Mortada Mehyar, Mycosynth, Namnamseo, 1282Nate Harada, Neven Miculinic, Nghia Tran, Nick Lyu, Niranjan Hasabnis, Nishidha, Oleksii 1283Kuchaiev, Oyesh Mann Singh, Panmari, Patrick, Paul Van Eck, Piyush Chaudhary, Quim Llimona, 1284Raingo, Richard Davies, Ruben Vereecken, Sahit Chintalapudi, Sam Abrahams, Santiago Castro, 1285Scott Sievert, Sean O'Keefe, Sebastian Schlecht, Shane, Shubhankar Deshpande, Spencer Schaber, 1286Sunyeop Lee, t13m, td2014, Thomas H. P. Andersen, Toby Petty, Umang Mehta, 1287Vadim Markovtsev, Valentin Iovene, Vincent Zhao, Vit Stepanovs, Vivek Rane, Vu Pham, wannabesrevenge, 1288weipingpku, wuhaixutab, wydwww, Xiang Gao, Xiaolin Lin, xiaoyaozhuzi, Yaroslav Bulatov, Yi Liu, 1289Yoshihiro Sugi, Yuan (Terry) Tang, Yuming Wang, Yuxin Wu, Zader Zheng, Zhaojun Zhang, zhengjiajin, 1290ZhipengShen, Ziming Dong, zjj2wry 1291 1292We are also grateful to all who filed issues or helped resolve them, asked and 1293answered questions, and were part of inspiring discussions. 1294 1295# Release 1.1.0 1296 1297## Major Features and Improvements 1298* Added Java API support for Windows. 1299* Added `tf.spectral` module. Moved existing FFT ops to `tf.spectral` while 1300 keeping an alias in the old location (`tf.*`). 1301* Added 1D, 2D and 3D Fourier transform ops for real signals to `tf.spectral`. 1302* Added a `tf.bincount` function. 1303* Added Keras 2 API to contrib. 1304* Added a new lightweight queue-like object - `RecordInput`. 1305* Added `tf.contrib.image.compose_transforms` function. 1306* Bring `tf.estimator.*` into the API. Non-deprecated functionality from `tf.contrib.learn.Estimator` is moved to `tf.estimator.Estimator` with cosmetic changes. 1307* Docker images: TF images on gcr.io and Docker Hub are upgraded to ubuntu:16.04. 1308* Added the following features to TensorFlow Debugger (tfdbg): 1309 * Ability to inspect Python source file against TF ops and tensors (command `print_source` / `ps`) 1310 * New navigation bar in Curses-based UI 1311 * 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. 1312* Initial release of installation guides for Java, C, and Go. 1313* Added Text Dashboard to TensorBoard. 1314 1315## Deprecations 1316 1317* 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. 1318 1319## Changes to contrib APIs 1320* The behavior of RNNCells is now stricter due to the transition towards making RNNCells act more like Keras layers. 1321 * If an RNNCell is used twice in two different variable scopes, an error is raised describing how to avoid this behavior. 1322 * 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`. 1323* Deprecated contrib/distributions `pmf`, `pdf`, `log_pmf`, `log_pdf`. 1324* Moved `bayesflow.special_math` to distributions. 1325* `tf.contrib.tensor_forest.python.tensor_forest.RandomForestDeviceAssigner` removed. 1326* Changed some MVN classes and parameters: 1327 * `tf.contrib.distributions.MultivariateNormalFull` replaced by `tf.contrib.distributions.MultivariateNormalTriL`. 1328 * `tf.contrib.distributions.MultivariateNormalCholesky` replaced by `tf.contrib.distributions.MultivariateNormalTriL` 1329 * `tf.contrib.distributions.MultivariateNormalDiagWithSoftplusStDev` replaced 1330 by `tf.contrib.distributions.MultivariateNormalDiagWithSoftplusScale` 1331 * `tf.contrib.distributions.MultivariateNormalDiag` arguments changed from `mu`, `diag_stddev` to `log`, `scale_diag`. 1332 * `tf.contrib.distributions.MultivariateNormalDiagPlusVDVT` removed. 1333 * `tf.contrib.distributions.MultivariateNormalDiagPlusLowRank` added. 1334 1335## Bug Fixes and Other Changes 1336* Java: Support for loading models exported using the SavedModel API (courtesy @EronWright). 1337* Go: Added support for incremental graph execution. 1338* Fix a bug in the WALS solver when single-threaded. 1339* Added support for integer sparse feature values in `tf.contrib.layers.sparse_column_with_keys`. 1340* Fixed `tf.set_random_seed(0)` to be deterministic for all ops. 1341* Stability improvements for the GCS file system support. 1342* Improved TensorForest performance. 1343* Added support for multiple filename globs in `tf.matching_files`. 1344* `LogMessage` now includes a timestamp as beginning of a message. 1345* Added MultiBox person detector example standalone binary. 1346* Android demo: Makefile build functionality added to build.gradle to fully support building TensorFlow demo in Android on Windows. 1347* Android demo: read MultiBox priors from txt file rather than protobuf. 1348* Added colocation constraints to `StagingArea`. 1349* `sparse_matmul_op` reenabled for Android builds. 1350* Restrict weights rank to be the same as the broadcast target, to avoid ambiguity on broadcast rules. 1351* Upgraded libxsmm to 1.7.1 and applied other changes for performance and memory usage. 1352* Fixed bfloat16 integration of LIBXSMM sparse mat-mul. 1353* Improved performance and reduce memory usage by allowing ops to forward input buffers to output buffers and perform computations in-place. 1354* Improved the performance of CPU assignment for strings. 1355* Speed up matrix * vector multiplication and matrix * matrix with unknown shapes. 1356* C API: Graph imports now support input remapping, control dependencies, and returning imported nodes (see `TF_GraphImportGraphDefWithReturnOutputs()`) 1357* Multiple C++ API updates. 1358* Multiple TensorBoard updates including: 1359 * Users can now view image summaries at various sampled steps (instead of just the last step). 1360 * Bugs involving switching runs as well as the image dashboard are fixed. 1361 * Removed data download links from TensorBoard. 1362 * TensorBoard uses a relative data directory, for easier embedding. 1363 * TensorBoard automatically ignores outliers for domain calculation, and formats proportional values consistently. 1364* Multiple tfdbg bug fixes: 1365 * Fixed Windows compatibility issues. 1366 * Command history now persists across runs. 1367 * Bug fix in graph validation related to `tf.while_loops`. 1368* Java Maven fixes for bugs with Windows installation. 1369* Backport fixes and improvements from external keras. 1370* Keras config file handling fix. 1371 1372## Thanks to our Contributors 1373 1374This release contains contributions from many people at Google, as well as: 1375 1376A. Besir Kurtulmus, Adal Chiriliuc, @akash, Alec-Desouza, Alex Rothberg, Alex 1377Sergeev, Alexander Heinecke, Allen Guo, Andreas Madsen, Ankesh Anand, Anton 1378Loss, @Aravind, @Arie, Ashutosh Das, AuréLien Geron, Bairen Yi, @bakunyo, Ben 1379Visser, Brady Zhou, Calpa Liu, Changming Sun, Chih Cheng Liang, Christopher 1380Berner, Clark Zinzow, @Conchylicultor, Dan Ellis, Dan J, Dan Jarvis, Daniel 1381Ylitalo, Darren Garvey, David Norman, David Truong, @DavidNorman, Dimitar 1382Pavlov, Dmitry Persiyanov, @Eddie, @elirex, Erfan Noury, Eron Wright, Evgeny 1383Mazovetskiy, Fabrizio (Misto) Milo, @fanlu, Fisher Coder, Florian Courtial, 1384Franck Dernoncourt, Gagan Goel, Gao, Xiang, @Gautam, Gefu Tang, @guilherme, 1385@guschmue, Hannah Provenza, Hans Pabst, @hartb, Hsiao Yi, Huazuo Gao, Igor 1386ChorążEwicz, Ivan Smirnov, Jakub Kolodziejczyk, Jason Gavris, Jason Morton, Jay 1387Young, Jayaram Bobba, Jeremy Sawruk, Jiaming Liu, Jihun Choi, @jiqiu, Joan Thibault, 1388John C F, Jojy George Varghese, Jon Malmaud, Julian Berman, Julian Niedermeier, 1389Junpeng Lao, Kai Sasaki, @Kankroc, Karl Lessard, Kyle Bostelmann, @Lezcano, Li 1390Yi, Luo Yun, @lurker, Mahmoud-Abuzaina, Mandeep Singh, Marek Kolodziej, Mark 1391Szepieniec, Martial Hue, Medhat Omr, Memo Akten, Michael Gharbi, MichaëL Defferrard, 1392Milan Straka, @MircoT, @mlucool, Muammar Ibn Faisal, Nayana Thorat, @nghiattran, 1393Nicholas Connor, Nikolaas Steenbergen, Niraj Patel, Niranjan Hasabnis, @Panmari, 1394Pavel Bulanov, Philip Pries Henningsen, Philipp Jund, @polonez, Prayag Verma, Rahul 1395Kavi, Raphael Gontijo Lopes, @rasbt, Raven Iqqe, Reid Pryzant, Richard Shin, Rizwan 1396Asif, Russell Kaplan, Ryo Asakura, RüDiger Busche, Saisai Shao, Sam Abrahams, @sanosay, 1397Sean Papay, @seaotterman, @selay01, Shaurya Sharma, Sriram Narayanamoorthy, Stefano 1398Probst, @taknevski, @tbonza, @teldridge11, Tim Anglade, Tomas Reimers, Tomer Gafner, 1399Valentin Iovene, Vamsi Sripathi, Viktor Malyi, Vit Stepanovs, Vivek Rane, Vlad Firoiu, 1400@wangg12, @will, Xiaoyu Tao, Yaroslav Bulatov, Yi Liu, Yuan (Terry) Tang, @Yufeng, 1401Yuming Wang, Yuxin Wu, Zafar Takhirov, Ziming Dong 1402 1403We are also grateful to all who filed issues or helped resolve them, asked and 1404answered questions, and were part of inspiring discussions. 1405 1406 1407# Release 1.0.1 1408 1409## Bug Fixes and Other Changes 1410* Change GraphConstructor to not increase the version when importing, but instead take the min of all versions. 1411* Google Cloud Storage fixes. 1412* 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. 1413 1414# Release 1.0.0 1415 1416## Major Features and Improvements 1417* 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. 1418* TensorFlow Debugger (tfdbg): command-line interface and API. 1419* New python 3 docker images added. 1420* Made pip packages pypi compliant. TensorFlow can now be installed by `pip 1421 install tensorflow` command. 1422* Several python API calls have been changed to resemble NumPy more closely. 1423* Android: person detection + tracking demo implementing Scalable Object 1424 Detection using Deep Neural Networks. 1425* New (experimental) [Java API](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/java). 1426* Add new Android image stylization demo based on "A Learned Representation For Artistic Style", and add YOLO object detector support. 1427 1428## Breaking Changes to the API 1429To 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). 1430* TensorFlow/models have been moved to a separate github repository. 1431* Division and modulus operators (/, //, %) now match Python (flooring) 1432 semantics. This applies to `tf.div` and `tf.mod` as well. To obtain forced 1433 integer truncation based behaviors you can use `tf.truncatediv` 1434 and `tf.truncatemod`. 1435* `tf.divide()` is now the recommended division function. `tf.div()` will 1436 remain, but its semantics do not respond to Python 3 or `from future` 1437 mechanisms. 1438* tf.reverse() now takes indices of axes to be reversed. E.g. 1439 `tf.reverse(a, [True, False, True])` must now be written as 1440 `tf.reverse(a, [0, 2])`. `tf.reverse_v2()` will remain until 1.0 final. 1441* `tf.mul`, `tf.sub` and `tf.neg` are deprecated in favor of `tf.multiply`, 1442 `tf.subtract` and `tf.negative`. 1443* `tf.pack` and `tf.unpack` are deprecated in favor of `tf.stack` and 1444 `tf.unstack`. 1445* `TensorArray.pack` and `TensorArray.unpack` are getting deprecated in favor of 1446 `TensorArray.stack` and `TensorArray.unstack`. 1447* The following Python functions have had their arguments changed to use `axis` 1448 when referring to specific dimensions. We have kept the old keyword arguments 1449 for compatibility currently, but we will be removing them well before the 1450 final 1.0. 1451 * `tf.argmax`: `dimension` becomes `axis` 1452 * `tf.argmin`: `dimension` becomes `axis` 1453 * `tf.count_nonzero`: `reduction_indices` becomes `axis` 1454 * `tf.expand_dims`: `dim` becomes `axis` 1455 * `tf.reduce_all`: `reduction_indices` becomes `axis` 1456 * `tf.reduce_any`: `reduction_indices` becomes `axis` 1457 * `tf.reduce_join`: `reduction_indices` becomes `axis` 1458 * `tf.reduce_logsumexp`: `reduction_indices` becomes `axis` 1459 * `tf.reduce_max`: `reduction_indices` becomes `axis` 1460 * `tf.reduce_mean`: `reduction_indices` becomes `axis` 1461 * `tf.reduce_min`: `reduction_indices` becomes `axis` 1462 * `tf.reduce_prod`: `reduction_indices` becomes `axis` 1463 * `tf.reduce_sum`: `reduction_indices` becomes `axis` 1464 * `tf.reverse_sequence`: `batch_dim` becomes `batch_axis`, `seq_dim` becomes `seq_axis` 1465 * `tf.sparse_concat`: `concat_dim` becomes `axis` 1466 * `tf.sparse_reduce_sum`: `reduction_axes` becomes `axis` 1467 * `tf.sparse_reduce_sum_sparse`: `reduction_axes` becomes `axis` 1468 * `tf.sparse_split`: `split_dim` becomes `axis` 1469* `tf.listdiff` has been renamed to `tf.setdiff1d` to match NumPy naming. 1470* `tf.inv` has been renamed to be `tf.reciprocal` (component-wise reciprocal) 1471 to avoid confusion with `np.inv` which is matrix inversion 1472* tf.round now uses banker's rounding (round to even) semantics to match NumPy. 1473* `tf.split` now takes arguments in a reversed order and with different 1474 keywords. In particular, we now match NumPy order as 1475 `tf.split(value, num_or_size_splits, axis)`. 1476* `tf.sparse_split` now takes arguments in reversed order and with different 1477 keywords. In particular we now match NumPy order as 1478 `tf.sparse_split(sp_input, num_split, axis)`. NOTE: we have temporarily 1479 made `tf.sparse_split` require keyword arguments. 1480* `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)`. 1481* `tf.image.decode_jpeg` by default uses the faster DCT method, sacrificing 1482 a little fidelity for improved speed. One can revert to the old 1483 behavior by specifying the attribute `dct_method='INTEGER_ACCURATE'`. 1484* `tf.complex_abs` has been removed from the Python interface. `tf.abs` 1485 supports complex tensors and should be used instead. 1486* In the C++ API (in tensorflow/cc), Input, Output, etc. have moved 1487 from the tensorflow::ops namespace to tensorflow. 1488* Template.`var_scope` property renamed to `.variable_scope` 1489* SyncReplicasOptimizer is removed and SyncReplicasOptimizerV2 renamed to SyncReplicasOptimizer. 1490* `tf.zeros_initializer()` and `tf.ones_initializer()` now return a callable 1491 that must be called with initializer arguments, in your code replace 1492 `tf.zeros_initializer` with `tf.zeros_initializer()`. 1493* `SparseTensor.shape` has been renamed to `SparseTensor.dense_shape`. Same for 1494 `SparseTensorValue.shape`. 1495* 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. 1496* Replace tf.train.SummaryWriter and tf.train.SummaryWriterCache with tf.summary.FileWriter and tf.summary.FileWriterCache. 1497* Removes RegisterShape from public API. Use C++ shape function registration 1498 instead. 1499* Deprecated `_ref` dtypes from the python API. 1500* In the C++ API (in tensorflow/cc), Input, Output, etc. have moved 1501 from the tensorflow::ops namespace to tensorflow. 1502* Change arg order for `{softmax,sparse_softmax,sigmoid}_cross_entropy_with_logits` to be (labels, predictions), and force use of named args. 1503* 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. 1504* `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. 1505* The shape keyword argument of the `SparseTensor` constructor changes its name to `dense_shape` between Tensorflow 0.12 and Tensorflow 1.0. 1506 1507## Bug Fixes and Other Changes 1508* Numerous C++ API updates. 1509* New op: `parallel_stack`. 1510* Introducing common tf io compression options constants for 1511 RecordReader/RecordWriter. 1512* Add `sparse_column_with_vocabulary_file`, to specify a feature column that 1513 transform string features to IDs, where the mapping is defined by a vocabulary 1514 file. 1515* Added `index_to_string_table` which returns a lookup table that maps indices to 1516 strings. 1517* Add `string_to_index_table`, which returns a lookup table that matches strings 1518 to indices. 1519* Add a `ParallelForWithWorkerId` function. 1520* Add `string_to_index_table`, which returns a lookup table that matches strings 1521 to indices. 1522* Support restore session from checkpoint files in v2 in `contrib/session_bundle`. 1523* Added a tf.contrib.image.rotate function for arbitrary angles. 1524* Added `tf.contrib.framework.filter_variables` as a convenience function to 1525 filter lists of variables based on regular expressions. 1526* `make_template()` takes an optional `custom_getter_ param`. 1527* Added comment about how existing directories are handled by 1528 `recursive_create_dir`. 1529* Added an op for QR factorizations. 1530* Divides and mods in Python API now use flooring (Python) semantics. 1531* Android: pre-built libs are now built nightly. 1532* Android: cmake/gradle build for TensorFlow Inference library under 1533 `contrib/android/cmake` 1534* Android: Much more robust Session initialization code. 1535* Android: TF stats now exposed directly in demo and log when debug mode is 1536 active 1537* Android: new/better README.md documentation 1538* saved_model is available as `tf.saved_model`. 1539* Empty op is now stateful. 1540* Improve speed of scatter_update on the cpu for ASSIGN operations. 1541* Change `reduce_join` to treat `reduction_indices` in the same way as other `reduce_` ops. 1542* Move `TensorForestEstimator` to `contrib/tensor_forest`. 1543* Enable compiler optimizations by default and allow configuration in configure. 1544* `tf.divide` now honors the name field. 1545* Make metrics weight broadcasting more strict. 1546* Add new queue-like `StagingArea` and new ops: `stage` and `unstage`. 1547* Enable inplace update ops for strings on CPU. Speed up string concat. 1548 1549## Thanks to our Contributors 1550 1551This release contains contributions from many people at Google, as well as: 1552 1553Aaron Hu, Abhishek Aggarwal, Adam Michael, Adriano Carmezim, @AfirSraftGarrier, 1554Alexander Novikov, Alexander Rosenberg Johansen, Andrew Gibiansky, Andrew Hundt, 1555Anish Shah, Anton Loss, @b0noI, @BoyuanJiang, Carl Thomé, Chad Kennedy, Comic 1556Chang, Connor Braa, Daniel N. Lang, Daniel Trebbien, 1557@danielgordon10, Darcy Liu, Darren Garvey, Dmitri Lapin, Eron Wright, Evan 1558Cofer, Fabrizio Milo, Finbarr Timbers, Franck Dernoncourt, Garrett Smith, 1559@guschmue, Hao Wei, Henrik Holst, Huazuo Gao, @Ian, @Issac, Jacob Israel, 1560Jangsoo Park, Jin Kim, Jingtian Peng, John Pope, Kye Bostelmann, Liangliang He, 1561Ling Zhang, Luheng He, Luke Iwanski, @lvli, Michael Basilyan, Mihir Patel, 1562Mikalai Drabovich, Morten Just, @newge, Nick Butlin, Nishant Shukla, 1563Pengfei Ni, Przemyslaw Tredak, @rasbt, @Ronny, Rudolf Rosa, @RustingSword, 1564Sam Abrahams, Sam Putnam, @SeongAhJo, Shi Jiaxin, @skavulya, Steffen MüLler, 1565@TheUSER123, @tiriplicamihai, @vhasanov, Victor Costan, Vit Stepanovs, 1566Wangda Tan, Wenjian Huang, Xingdong Zuo, Yaroslav Bulatov, Yota Toyama, 1567Yuan (Terry) Tang, Yuxin Wu 1568 1569We are also grateful to all who filed issues or helped resolve them, asked and 1570answered questions, and were part of inspiring discussions. 1571 1572 1573# Release 0.12.0 1574 1575## Major Features and Improvements 1576 1577* TensorFlow now builds and runs on Microsoft Windows (tested on Windows 10, 1578 Windows 7, and Windows Server 2016). Supported languages include Python (via a 1579 pip package) and C++. CUDA 8.0 and cuDNN 5.1 are supported for GPU 1580 acceleration. Known limitations include: It is not currently possible to load 1581 a custom op library. The GCS and HDFS file systems are not currently 1582 supported. The following ops are not currently implemented: 1583 Dequantize, QuantizeAndDequantize, QuantizedAvgPool, 1584 QuantizedBatchNomWithGlobalNormalization, QuantizedBiasAdd, QuantizedConcat, 1585 QuantizedConv2D, QuantizedMatmul, QuantizedMaxPool, 1586 QuantizeDownAndShrinkRange, QuantizedRelu, QuantizedRelu6, QuantizedReshape, 1587 QuantizeV2, RequantizationRange, and Requantize. 1588* Go: Experimental API in Go to create and execute graphs 1589 (https://godoc.org/github.com/tensorflow/tensorflow/tensorflow/go) 1590* New checkpoint format becomes the default in `tf.train.Saver`. Old V1 1591 checkpoints continue to be readable; controlled by the `write_version` 1592 argument, `tf.train.Saver` now by default writes out in the new V2 1593 format. It significantly reduces the peak memory required and latency 1594 incurred during restore. 1595* Added a new library for library of matrix-free (iterative) solvers for linear 1596 equations, linear least-squares, eigenvalues and singular values in 1597 tensorflow/contrib/solvers. Initial version has lanczos bidiagonalization, 1598 conjugate gradients and CGLS. 1599* Added gradients for `matrix_solve_ls` and `self_adjoint_eig`. 1600* Large cleanup to add second order gradient for ops with C++ gradients and 1601 improve existing gradients such that most ops can now be differentiated 1602 multiple times. 1603* Added a solver for ordinary differential equations, 1604 `tf.contrib.integrate.odeint`. 1605* New contrib module for tensors with named axes, `tf.contrib.labeled_tensor`. 1606* Visualization of embeddings in TensorBoard. 1607 1608## Breaking Changes to the API 1609 1610* `BusAdjacency` enum replaced with a protocol buffer `DeviceLocality`. PCI bus 1611 indexing now starts from 1 instead of 0, and `bus_id==0` is used where 1612 previously `BUS_ANY` was used. 1613* `Env::FileExists` and `FileSystem::FileExists` now return a tensorflow::Status 1614 instead of a bool. Any callers to this function can be converted to a bool 1615 by adding .ok() to the call. 1616* The C API type `TF_SessionWithGraph` has been renamed to `TF_Session`, 1617 indicating its preferred use in language bindings for TensorFlow. 1618 What was previously `TF_Session` has been renamed to `TF_DeprecatedSession`. 1619* Renamed `TF_Port` to `TF_Output` in the C API. 1620* Removes RegisterShape from public API. Use C++ shape function registration instead. 1621 indexing now starts from 1 instead of 0, and `bus_id==0` is used where 1622 previously `BUS_ANY` was used. 1623* Most RNN cells and RNN functions now use different variable scopes to be 1624 consistent with layers (`tf.contrib.layers`). This means old checkpoints 1625 written using this code will not load after this change without providing 1626 `Saver` a list of variable renames. Examples of variable scope changes 1627 include `RNN` -> `rnn` in `tf.nn.rnn`, `tf.nn.dynamic_rnn` and moving from 1628 `Linear/Matrix` -> `weights` and `Linear/Bias` -> `biases` in most RNN cells. 1629* Deprecated tf.select op. tf.where should be used instead. 1630* `SparseTensor.shape` has been renamed to `SparseTensor.dense_shape`. Same for 1631 `SparseTensorValue.shape`. 1632* `Env::FileExists` and `FileSystem::FileExists` now return a 1633 `tensorflow::Status` instead of a bool. Any callers to this function can be 1634 converted to a bool by adding `.ok()` to the call. 1635* C API: Type `TF_SessionWithGraph` has been renamed to `TF_Session`, indicating 1636 its preferred use in language bindings for TensorFlow. What was previously 1637 `TF_Session` has been renamed to `TF_DeprecatedSession`. 1638* C API: Renamed `TF_Port` to `TF_Output`. 1639* C API: The caller retains ownership of `TF_Tensor` objects provided to 1640 `TF_Run`, `TF_SessionRun`, `TF_SetAttrTensor` etc. 1641* Renamed `tf.image.per_image_whitening()` to 1642 `tf.image.per_image_standardization()` 1643* Move Summary protobuf constructors to `tf.summary` submodule. 1644* Deprecate `histogram_summary`, `audio_summary`, `scalar_summary`, 1645 `image_summary`, `merge_summary`, and `merge_all_summaries`. 1646* Combined `batch_*` and regular version of linear algebra and FFT ops. The 1647 regular op now handles batches as well. All `batch_*` Python interfaces were 1648 removed. 1649* `tf.all_variables`, `tf.VARIABLES` and `tf.initialize_all_variables` renamed 1650 to `tf.global_variables`, `tf.GLOBAL_VARIABLES` and 1651 `tf.global_variables_initializer` respectively. 1652* `tf.zeros_initializer()` and `tf.ones_initializer()` now return a callable 1653 that must be called with initializer arguments, in your code replace 1654 `tf.zeros_initializer` with `tf.zeros_initializer()` 1655 1656## Bug Fixes and Other Changes 1657 1658* Use threadsafe version of `lgamma` function. 1659* Fix `tf.sqrt` handling of negative arguments. 1660* Fixed bug causing incorrect number of threads to be used for multi-threaded 1661 benchmarks. 1662* Performance optimizations for `batch_matmul` on multi-core CPUs. 1663* Improve trace, `matrix_set_diag`, `matrix_diag_part` and their gradients to 1664 work for rectangular matrices. 1665* Support for SVD of complex valued matrices. 1666 1667 1668## Thanks to our Contributors 1669 1670This release contains contributions from many people at Google, as well as: 1671 1672@a7744hsc, Abhi Agg, @admcrae, Adriano Carmezim, Aki Sukegawa, Alex Kendall, 1673Alexander Rosenberg Johansen, @amcrae, Amlan Kar, Andre Simpelo, Andreas Eberle, 1674Andrew Hundt, Arnaud Lenglet, @b0noI, Balachander Ramachandran, Ben Barsdell, 1675Ben Guidarelli, Benjamin Mularczyk, Burness Duan, @c0g, Changming Sun, 1676@chanis, Corey Wharton, Dan J, Daniel Trebbien, Darren Garvey, David Brailovsky, 1677David Jones, Di Zeng, @DjangoPeng, Dr. Kashif Rasul, @drag0, Fabrizio (Misto) 1678Milo, FabríCio Ceschin, @fp, @Ghedeon, @guschmue, Gökçen Eraslan, Haosdent 1679Huang, Haroen Viaene, Harold Cooper, Henrik Holst, @hoangmit, Ivan Ukhov, Javier 1680Dehesa, Jingtian Peng, Jithin Odattu, Joan Pastor, Johan Mathe, Johannes Mayer, 1681Jongwook Choi, Justus Schwabedal, Kai Wolf, Kamil Hryniewicz, Kamran Amini, 1682Karen Brems, Karl Lattimer, @kborer, Ken Shirriff, Kevin Rose, Larissa Laich, 1683Laurent Mazare, Leonard Lee, Liang-Chi Hsieh, Liangliang He, Luke Iwanski, 1684Marek Kolodziej, Moustafa Alzantot, @MrQianjinsi, @nagachika, Neil Han, Nick 1685Meehan, Niels Ole Salscheider, Nikhil Mishra, @nschuc, Ondrej Skopek, OndřEj 1686Filip, @OscarDPan, Pablo Moyano, Przemyslaw Tredak, @qitaishui, @Quarazy, 1687@raix852, Philipp Helo, Sam Abrahams, @SriramRamesh, Till Hoffmann, Tushar Soni, 1688@tvn, @tyfkda, Uwe Schmidt, Victor Villas, Vit Stepanovs, Vladislav Gubarev, 1689@wujingyue, Xuesong Yang, Yi Liu, Yilei Yang, @youyou3, Yuan (Terry) Tang, 1690Yuming Wang, Zafar Takhirov, @zhongyuk, Ziming Dong, @guotong1988 1691 1692We are also grateful to all who filed issues or helped resolve them, asked and 1693answered questions, and were part of inspiring discussions. 1694 1695# Release 0.11.0 1696 1697## Major Features and Improvements 1698 1699* CUDA 8 support. 1700* cuDNN 5 support. 1701* HDFS Support. 1702* Adds Fused LSTM support via cuDNN 5 in `tensorflow/contrib/cudnn_rnn`. 1703* Improved support for NumPy style basic slicing including non-1 strides, 1704 ellipses, newaxis, and negative indices. For example complicated expressions 1705 like `foo[1, 2:4, tf.newaxis, ..., :-3:-1, :]` are now supported. In addition 1706 we have preliminary (non-broadcasting) support for sliced assignment to 1707 variables. In particular one can write `var[1:3].assign([1,11,111])`. 1708* 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. 1709* Introducing `core/util/tensor_bundle` module: a module to efficiently 1710 serialize/deserialize tensors to disk. Will be used in TF's new checkpoint 1711 format. 1712* Added tf.svd for computing the singular value decomposition (SVD) of dense 1713 matrices or batches of matrices (CPU only). 1714* Added gradients for eigenvalues and eigenvectors computed using 1715 `self_adjoint_eig` or `self_adjoint_eigvals`. 1716* Eliminated `batch_*` methods for most linear algebra and FFT ops and promoted 1717 the non-batch version of the ops to handle batches of matrices. 1718* Tracing/timeline support for distributed runtime (no GPU profiler yet). 1719* C API gives access to inferred shapes with `TF_GraphGetTensorNumDims` and 1720 `TF_GraphGetTensorShape`. 1721* Shape functions for core ops have moved to C++ via 1722 `REGISTER_OP(...).SetShapeFn(...)`. Python shape inference RegisterShape calls 1723 use the C++ shape functions with `common_shapes.call_cpp_shape_fn`. A future 1724 release will remove `RegisterShape` from python. 1725 1726 1727## Bug Fixes and Other Changes 1728 1729* Documentation now includes operator overloads on Tensor and Variable. 1730* `tensorflow.__git_version__` now allows users to identify the version of the 1731 code that TensorFlow was compiled with. We also have 1732 `tensorflow.__git_compiler__` which identifies the compiler used to compile 1733 TensorFlow's core. 1734* Improved multi-threaded performance of `batch_matmul`. 1735* LSTMCell, BasicLSTMCell, and MultiRNNCell constructors now default to 1736 `state_is_tuple=True`. For a quick fix while transitioning to the new 1737 default, simply pass the argument `state_is_tuple=False`. 1738* DeviceFactory's AddDevices and CreateDevices functions now return 1739 a Status instead of void. 1740* Int32 elements of list(type) arguments are no longer placed in host memory by 1741 default. If necessary, a list(type) argument to a kernel can be placed in host 1742 memory using a HostMemory annotation. 1743* `uniform_unit_scaling_initializer()` no longer takes a `full_shape` arg, 1744 instead relying on the partition info passed to the initializer function when 1745 it's called. 1746* The NodeDef protocol message is now defined in its own file `node_def.proto` 1747 `instead of graph.proto`. 1748* `ops.NoGradient` was renamed `ops.NotDifferentiable`. `ops.NoGradient` will 1749 be removed soon. 1750* `dot.h` / DotGraph was removed (it was an early analysis tool prior 1751 to TensorBoard, no longer that useful). It remains in history 1752 should someone find the code useful. 1753* re2 / regexp.h was removed from being a public interface of TF. 1754 Should users need regular expressions, they should depend on the RE2 1755 library directly rather than via TensorFlow. 1756 1757## Thanks to our Contributors 1758 1759This release contains contributions from many people at Google, as well as: 1760 1761Abid K, @afshinrahimi, @AidanGG, Ajay Rao, Aki Sukegawa, Alex Rothberg, 1762Alexander Rosenberg Johansen, Andrew Gibiansky, Andrew Thomas, @Appleholic, 1763Bastiaan Quast, Ben Dilday, Bofu Chen, Brandon Amos, Bryon Gloden, Cissp®, 1764@chanis, Chenyang Liu, Corey Wharton, Daeyun Shin, Daniel Julius Lasiman, Daniel 1765Waterworth, Danijar Hafner, Darren Garvey, Denis Gorbachev, @DjangoPeng, 1766Egor-Krivov, Elia Palme, Eric Platon, Fabrizio Milo, Gaetan Semet, 1767Georg Nebehay, Gu Wang, Gustav Larsson, @haosdent, Harold Cooper, Hw-Zz, 1768@ichuang, Igor Babuschkin, Igor Macedo Quintanilha, Ilya Edrenkin, @ironhead, 1769Jakub Kolodziejczyk, Jennifer Guo, Jihun Choi, Jonas Rauber, Josh Bleecher 1770Snyder, @jpangburn, Jules Gagnon-Marchand, Karen Brems, @kborer, Kirill Bobyrev, 1771Laurent Mazare, Longqi Yang, Malith Yapa, Maniteja Nandana, Martin Englund, 1772Matthias Winkelmann, @mecab, Mu-Ik Jeon, Nand Dalal, Niels Ole Salscheider, 1773Nikhil Mishra, Park Jiin, Pieter De Rijk, @raix852, Ritwik Gupta, Sahil Sharma, 1774Sangheum Hwang, @SergejsRk, Shinichiro Hamaji, Simon Denel, @Steve, @suiyuan2009, 1775Tiago Jorge, Tijmen Tieleman, @tvn, @tyfkda, Wang Yang, Wei-Ting Kuo, Wenjian 1776Huang, Yan Chen, @YenChenLin, Yuan (Terry) Tang, Yuncheng Li, Yunfeng Wang, Zack 1777Polizzi, @zhongzyd, Ziming Dong, @perhapszzy 1778 1779We are also grateful to all who filed issues or helped resolve them, asked and 1780answered questions, and were part of inspiring discussions. 1781 1782# Release 0.10.0 1783 1784## Major Features and Improvements 1785 1786* Added support for C++ shape inference 1787* Added graph-construction C API 1788* Major revision to the graph-construction C++ API 1789* Support makefile build for iOS 1790* Added Mac GPU support 1791* Full version of TF-Slim available as `tf.contrib.slim` 1792* Added k-Means clustering and WALS matrix factorization 1793 1794## Bug Fixes and Other Changes 1795 1796* Allow gradient computation for scalar values. 1797* Performance improvements for gRPC 1798* Improved support for fp16 1799* New high-level ops in tf.contrib.{layers,metrics} 1800* New features for TensorBoard, such as shape display, exponential smoothing 1801* Faster and more stable Google Cloud Storage (GCS) filesystem support 1802* Support for zlib compression and decompression for TFRecordReader and TFRecordWriter 1803* Support for reading (animated) GIFs 1804* Improved support for SparseTensor 1805* Added support for more probability distributions (Dirichlet, Beta, Bernoulli, etc.) 1806* Added Python interfaces to reset resource containers. 1807* Many bugfixes and performance improvements 1808* Many documentation fixes 1809 1810## Thanks to our Contributors 1811 1812This release contains contributions from many people at Google, as well as: 1813 1814Alex 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 1815 1816We are also grateful to all who filed issues or helped resolve them, asked and 1817answered questions, and were part of inspiring discussions. 1818 1819# Release 0.9.0 1820 1821## Major Features and Improvements 1822 1823* Python 3.5 support and binaries 1824* Added iOS support 1825* Added support for processing on GPUs on MacOS 1826* Added makefile for better cross-platform build support (C API only) 1827* fp16 support and improved complex128 support for many ops 1828* Higher level functionality in contrib.{layers,losses,metrics,learn} 1829* More features to Tensorboard 1830* Improved support for string embedding and sparse features 1831* The RNN api is finally "official" (see, e.g., `tf.nn.dynamic_rnn`, 1832 `tf.nn.rnn`, and the classes in `tf.nn.rnn_cell`). 1833* TensorBoard now has an Audio Dashboard, with associated audio summaries. 1834 1835## Bug Fixes and Other Changes 1836 1837* Turned on CuDNN Autotune. 1838* Added support for using third-party Python optimization algorithms (contrib.opt). 1839* Google Cloud Storage filesystem support. 1840* HDF5 support 1841* Add support for 3d convolutions and pooling. 1842* Update gRPC release to 0.14. 1843* Eigen version upgrade. 1844* Switch to eigen thread pool 1845* `tf.nn.moments()` now accepts a `shift` argument. Shifting by a good estimate 1846 of the mean improves numerical stability. Also changes the behavior of the 1847 `shift` argument to `tf.nn.sufficient_statistics()`. 1848* Performance improvements 1849* Many bugfixes 1850* Many documentation fixes 1851* TensorBoard fixes: graphs with only one data point, Nan values, 1852 reload button and auto-reload, tooltips in scalar charts, run 1853 filtering, stable colors 1854* Tensorboard graph visualizer now supports run metadata. Clicking on nodes 1855 while viewing a stats for a particular run will show runtime statistics, such 1856 as memory or compute usage. Unused nodes will be faded out. 1857 1858## Thanks to our Contributors 1859 1860This release contains contributions from many people at Google, as well as: 1861 1862Aaron 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 1863 1864We are also grateful to all who filed issues or helped resolve them, asked and 1865answered questions, and were part of inspiring discussions. 1866 1867# Release 0.8.0 1868 1869## Major Features and Improvements 1870 1871* Added a distributed runtime using GRPC 1872* Move skflow to `contrib/learn` 1873* Better linear optimizer in `contrib/linear_optimizer` 1874* Random forest implementation in `contrib/tensor_forest` 1875* CTC loss and decoders in `contrib/ctc` 1876* Basic support for `half` data type 1877* Better support for loading user ops (see examples in `contrib/`) 1878* Allow use of (non-blocking) Eigen threadpool with `TENSORFLOW_USE_EIGEN_THREADPOOL` define 1879* Add an extension mechanism for adding network file system support 1880* TensorBoard displays metadata stats (running time, memory usage and device used) and tensor shapes 1881 1882## Bug Fixes and Other Changes 1883 1884* Utility for inspecting checkpoints 1885* Basic tracing and timeline support 1886* Allow building against cuDNN 5 (not incl. RNN/LSTM support) 1887* Added instructions and binaries for ProtoBuf library with fast serialization and without 64MB limit 1888* Added special functions 1889* `bool`-strictness: Tensors have to be explicitly compared to `None` 1890* Shape strictness: all fed values must have a shape that is compatible with the tensor they are replacing 1891* Exposed `tf.while_loop` (deprecated `control_flow_ops.While`) 1892* run() now takes RunOptions and RunMetadata, which enable timing stats 1893* Fixed lots of potential overflow problems in op kernels 1894* Various performance improvements, especially for RNNs and convolutions 1895* Many bugfixes 1896* Nightly builds, tutorial tests, many test improvements 1897* New examples: transfer learning and deepdream ipython notebook 1898* Added tutorials, many documentation fixes. 1899 1900## Thanks to our Contributors 1901 1902This release contains contributions from many people at Google, as well as: 1903 1904Abhinav 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. 1905 1906We are also grateful to all who filed issues or helped resolve them, asked and 1907answered questions, and were part of inspiring discussions. 1908 1909 1910# Release 0.7.1 1911 1912## Bug Fixes and Other Changes 1913 1914* Added gfile.Open and gfile.Copy, used by input_data.py. 1915* Fixed Saver bug when MakeDirs tried to create empty directory. 1916* GPU Pip wheels are built with cuda 7.5 and cudnn-v4, making them 1917 required for the binary releases. Lower versions of cuda/cudnn can 1918 be supported by installing from sources and setting the options 1919 during ./configure 1920* Fix dataset encoding example for Python3 (@danijar) 1921* Fix PIP installation by not packaging protobuf as part of wheel, 1922 require protobuf 3.0.0b2. 1923* Fix Mac pip installation of numpy by requiring pip >= 1.10.1. 1924* Improvements and fixes to Docker image. 1925 1926 1927# Release 0.7.0 1928 1929## Major Features and Improvements 1930 1931* Allow using any installed Cuda >= 7.0 and cuDNN >= R2, and add support 1932 for cuDNN R4 1933* Added a `contrib/` directory for unsupported or experimental features, 1934 including higher level `layers` module 1935* Added an easy way to add and dynamically load user-defined ops 1936* Built out a good suite of tests, things should break less! 1937* Added `MetaGraphDef` which makes it easier to save graphs with metadata 1938* Added assignments for "Deep Learning with TensorFlow" udacity course 1939 1940 1941## Bug Fixes and Other Changes 1942 1943* Added a versioning framework for `GraphDef`s to ensure compatibility 1944* Enforced Python 3 compatibility 1945* Internal changes now show up as sensibly separated commits 1946* Open-sourced the doc generator 1947* Un-fork Eigen 1948* Simplified the `BUILD` files and cleaned up C++ headers 1949* TensorFlow can now be used as a submodule in another bazel build 1950* New ops (e.g., `*fft`, `*_matrix_solve`) 1951* Support for more data types in many ops 1952* Performance improvements 1953* Various bugfixes 1954* Documentation fixes and improvements 1955 1956 1957## Breaking Changes to the API 1958 1959* `AdjustContrast` kernel deprecated, new kernel `AdjustContrastv2` takes and 1960 outputs float only. `adjust_contrast` now takes all data types. 1961* `adjust_brightness`'s `delta` argument is now always assumed to be in `[0,1]` 1962 (as is the norm for images in floating point formats), independent of the 1963 data type of the input image. 1964* The image processing ops do not take `min` and `max` inputs any more, casting 1965 safety is handled by `saturate_cast`, which makes sure over- and underflows 1966 are handled before casting to data types with smaller ranges. 1967* For C++ API users: `IsLegacyScalar` and `IsLegacyVector` are now gone from 1968 `TensorShapeUtils` since TensorFlow is scalar strict within Google (for 1969 example, the shape argument to `tf.reshape` can't be a scalar anymore). The 1970 open source release was already scalar strict, so outside Google `IsScalar` 1971 and `IsVector` are exact replacements. 1972* The following files are being removed from `tensorflow/core/public/`: 1973 * `env.h` -> `../platform/env.h` 1974 * `status.h` -> `../lib/core/status.h` 1975 * `tensor.h` -> `../framework/tensor.h` 1976 * `tensor_shape.h` -> `../framework/tensor_shape.h` 1977 * `partial_tensor_shape.h` -> `../framework/partial_tensor_shape.h` 1978 * `tensorflow_server.h` deleted 1979* For C++ API users: `TensorShape::ShortDebugString` has been renamed to 1980 `DebugString`, and the previous `DebugString` behavior is gone (it was 1981 needlessly verbose and produced a confusing empty string for scalars). 1982* `GraphOptions.skip_common_subexpression_elimination` has been removed. All 1983 graph optimizer options are now specified via 1984 `GraphOptions.OptimizerOptions`. 1985* `ASSERT_OK` / `EXPECT_OK` macros conflicted with external projects, so they 1986 were renamed `TF_ASSERT_OK`, `TF_EXPECT_OK`. The existing macros are 1987 currently maintained for short-term compatibility but will be removed. 1988* The non-public `nn.rnn` and the various `nn.seq2seq` methods now return 1989 just the final state instead of the list of all states. 1990* `tf.scatter_update` now no longer guarantees that lexicographically largest 1991 index be used for update when duplicate entries exist. 1992* `tf.image.random_crop(image, [height, width])` is now 1993 `tf.random_crop(image, [height, width, depth])`, and `tf.random_crop` works 1994 for any rank (not just 3-D images). The C++ `RandomCrop` op has been replaced 1995 with pure Python. 1996* Renamed `tf.test.GetTempDir` and `tf.test.IsBuiltWithCuda` to 1997 `tf.test.get_temp_dir` and `tf.test.is_built_with_cuda` for PEP-8 1998 compatibility. 1999* `parse_example`'s interface has changed, the old interface is accessible in 2000 `legacy_parse_example` (same for related functions). 2001* New `Variable`s are not added to the same collection several times even if 2002 a list with duplicates is passed to the constructor. 2003* The Python API will now properly set the `list` member of `AttrValue` in 2004 constructed `GraphDef` messages for empty lists. The serialization of some 2005 graphs will change, but the change is both forwards and backwards compatible. 2006 It will break tests that compare a generated `GraphDef` to a golden serialized 2007 `GraphDef` (which is discouraged). 2008 2009 2010## Thanks to our Contributors 2011 2012This release contains contributions from many people at Google, as well as: 2013 2014Akiomi Kamakura, Alex Vig, Alexander Rosenberg Johansen, Andre Cruz, Arun Ahuja, 2015Bart Coppens, Bernardo Pires, Carl Vondrick, Cesar Salgado, Chen Yu, 2016Christian Jauvin, Damien Aymeric, Dan Vanderkam, Denny Britz, Dongjoon Hyun, 2017Eren Güven, Erik Erwitt, Fabrizio Milo, G. Hussain Chinoy, Jim Fleming, 2018Joao Felipe Santos, Jonas Meinertz Hansen, Joshi Rekha, Julian Viereck, 2019Keiji Ariyama, Kenton Lee, Krishna Sankar, Kristina Chodorow, Linchao Zhu, 2020Lukas Krecan, Mark Borgerding, Mark Daoust, Moussa Taifi, 2021Nathan Howell, Naveen Sundar Govindarajulu, Nick Sweeting, Niklas Riekenbrauck, 2022Olivier Grisel, Patrick Christ, Povilas Liubauskas, Rainer Wasserfuhr, 2023Romain Thouvenin, Sagan Bolliger, Sam Abrahams, Taehoon Kim, Timothy J Laurent, 2024Vlad Zavidovych, Yangqing Jia, Yi-Lin Juang, Yuxin Wu, Zachary Lipton, 2025Zero Chen, Alan Wu, @brchiu, @emmjaykay, @jalammar, @Mandar-Shinde, 2026@nsipplswezey, @ninotoshi, @panmari, @prolearner and @rizzomichaelg. 2027 2028We are also grateful to all who filed issues or helped resolve them, asked and 2029answered questions, and were part of inspiring discussions. 2030 2031 2032# Release 0.6.0 2033 2034## Major Features and Improvements 2035 2036* Python 3.3+ support via changes to python codebase and ability 2037 to specify python version via ./configure. 2038 2039* Some improvements to GPU performance and memory usage: 2040 [convnet benchmarks](https://github.com/soumith/convnet-benchmarks/issues/66) 2041 roughly equivalent with native cudnn v2 performance. Improvements mostly due 2042 to moving to 32-bit indices, faster shuffling kernels. More improvements to 2043 come in later releases. 2044 2045 2046## Bug Fixes 2047 2048* Lots of fixes to documentation and tutorials, many contributed 2049 by the public. 2050 2051* 271 closed issues on github issues. 2052 2053## Backwards-Incompatible Changes 2054 2055* `tf.nn.fixed_unigram_candidate_sampler` changed its default 'distortion' 2056 attribute from 0.0 to 1.0. This was a bug in the original release 2057 that is now fixed. 2058 2059# Release 0.5.0 2060 2061Initial release of TensorFlow. 2062