tensorflow

Computation using data flow graphs for scalable machine learning

Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402Star half bd79095782ee4930099175e5ce7f4c89fa3ddabcd56fffcc7c74f6f2a2d46b27 (3 ratings)
Rated 4.5 out of 5
Subscribe to updates I use tensorflow


Statistics on tensorflow

Number of watchers on Github 92201
Number of open issues 1461
Average time to close an issue about 12 hours
Main language C++
Average time to merge a PR about 21 hours
Open pull requests 593+
Closed pull requests 477+
Last commit 4 months ago
Repo Created over 2 years ago
Repo Last Updated 4 months ago
Size 152 MB
Homepage https://tensorflo...
Organization / Authortensorflow
Latest Releasev1.6.0
Contributors54
Page Updated
Do you use tensorflow? Leave a review!
View open issues (1461)
View tensorflow activity
View on github
Latest Open Source Launches
Trendy new open source projects in your inbox! View examples

Subscribe to our mailing list

Evaluating tensorflow for your project? Score Explanation
Commits Score (?)
Issues & PR Score (?)
What people are saying about tensorflow Leave a review
Promises to be the best. Documentation and functionality catching up.
standardizes deep network models



Documentation Linux CPU Linux GPU Mac OS CPU Windows CPU Android
Documentation Build Status Build Status Build Status Build Status Build Status Download

TensorFlow is an open source software library for numerical computation using data flow graphs. The graph nodes represent mathematical operations, while the graph edges represent the multidimensional data arrays (tensors) that flow between them. This flexible architecture lets you deploy computation to one or more CPUs or GPUs in a desktop, server, or mobile device without rewriting code. TensorFlow also includes TensorBoard, a data visualization toolkit.

TensorFlow was originally developed by researchers and engineers working on the Google Brain team within Google's Machine Intelligence Research organization for the purposes of conducting machine learning and deep neural networks research. The system is general enough to be applicable in a wide variety of other domains, as well.

Keep up to date with release announcements and security updates by subscribing to announce@tensorflow.org.

Installation

See Installing TensorFlow for instructions on how to install our release binaries or how to build from source.

People who are a little more adventurous can also try our nightly binaries:

Nightly pip packages

  • We are pleased to announce that TensorFlow now offers nightly pip packages under the tf-nightly and tf-nightly-gpu project on pypi. Simply run pip install tf-nightly or pip install tf-nightly-gpu in a clean environment to install the nightly TensorFlow build. We support CPU and GPU packages on Linux, Mac, and Windows.

Individual whl files

Try your first TensorFlow program

$ python
>>> import tensorflow as tf
>>> hello = tf.constant('Hello, TensorFlow!')
>>> sess = tf.Session()
>>> sess.run(hello)
'Hello, TensorFlow!'
>>> a = tf.constant(10)
>>> b = tf.constant(32)
>>> sess.run(a + b)
42
>>> sess.close()

Contribution guidelines

If you want to contribute to TensorFlow, be sure to review the contribution guidelines. This project adheres to TensorFlow's code of conduct. By participating, you are expected to uphold this code.

We use GitHub issues for tracking requests and bugs. So please see TensorFlow Discuss for general questions and discussion, and please direct specific questions to Stack Overflow.

The TensorFlow project strives to abide by generally accepted best practices in open-source software development:

CII Best Practices

For more information

Learn more about the TensorFlow community at the community page of tensorflow.org for a few ways to participate.

License

Apache License 2.0

tensorflow open issues Ask a question     (View All Issues)
  • over 1 year print_selective_registration_header
  • over 1 year GPU-enabled Mac build of TensorFlow version 0.11.0rc2-py2 was compiled against cuDNN v5.1
  • over 1 year Unable to install GPU enabled TensorFlow
  • over 1 year Running own TensorFlow model on Android gives native inference error: “Session was not created with a graph before Run()!”
  • over 1 year [docs] python3 bytes -> string in example in README
  • over 1 year Inconsistent behavior for variables_collections and outputs_collections parameters (contrib/layers)
  • over 1 year unable to generate tensorflow/go/op
  • over 1 year Restore and predict tensorflow contrib.learn.DNNRegressor/DNNClassifier without running at least one step of training (which was a missing feature).
  • over 1 year Use of Inception v5 model to extract characters from video frames
  • over 1 year lack of up-to-date files after upgrading to tensorflow 0.11
  • over 1 year Constant folding doesn't remove control edges
  • over 1 year tensorflow/examples/learn/text_classification_builtin_rnn_model.py references removed TensorFlowRNNClassifier
  • over 1 year help with finetune process
  • over 1 year Sticky "No OpKernel was registered to support Op 'Conv2D' with these attrs."
  • over 1 year Compiling error when adding -c dbg parameter
  • over 1 year Disable display “successfully opened CUDA library ****” when import tensorflow
  • over 1 year no modules in tensorflow/python ...
  • over 1 year How to allow GPU memory growth under TF slim?
  • over 1 year Inaccuracies in tf.reduce_sum
  • over 1 year XLA's Dot should follow broadcast semantics from np.matmul, not np.dot
  • over 1 year Extracting Bazel installation
  • over 1 year Do not require execution of ./configure when changing branches
  • over 1 year Fine tune inception without TFRecords.
  • over 1 year Op type not registered 'SqrtGrad'
  • over 1 year TensorFlow is 1.3-7 times slower than Theano
  • over 1 year Supervisor should_stop not working in TF distributed
  • over 1 year Simplify Android integration
  • over 1 year Re-entering an existing scope with variable_scope produces inconsistent results
  • over 1 year Build a static library
  • over 1 year Android arm64-v8a (armv8) native libs build error
tensorflow open pull requests (View All Pulls)
  • Fix doc to test GPU kernels using docker
  • Use standard types and add missing returns to fix compitable with MSVC
  • Add Tan, Asin, Acos, Atan trigonometric functions
  • use string::append to catenate strings
  • remove some signed/unsigned integer comparison warnings
  • Enable GPU reduction for doubles
  • Added a link to some tutorial examples.
  • add distributed, non-blocking thread pool implementation
  • Improved 1_notmnist
  • Adding instructions for wheel on python3.5
  • Update CROSSTOOL with Toolchain for ARM
  • added //tensorflow/cc:cc_ops to //tensorflow:libtensorflow.so
  • Fix training of decoder embeddings.
  • rnn_cell.py: Allow user to explicitly set forget bias in LSTMCell
  • Non-blocking queuing and running on thread pool
  • Adding basic CMake support
  • Enable building with CUDA support on Mac OS X
  • optimize slice_input_producer
  • [WIP] Add Adadelta optimizer
  • Added GPU implementation for resize nearest neighbor.
  • [doc] Add anchor links to headers in api_doc files
  • Recursively copying elements from one graph to another
  • gradient of diag operator
  • ci_build - debian jessie
  • Fixing python3 issues
  • DO NOT MERGE: Replaced python configuration by a Skylark Remote Repository
  • Fix types image ops tests
  • #1293 Implement Matrix Trace
  • [WIP] Added GPU implementation for resize nearest neighbor grad.
  • Mean of cross entropy vs sum in summaries tutorial
  • Retraining Example with Notebook
  • Fixing basic Ubuntu CMake Build
  • Use external repository for protobuf dependency.
  • Rint Operator
  • Go bindings
  • Add ability to specify which gcc to use
  • ci_build: support tensorflow as submodule
  • add reference for several concepts in tutorial
  • Update tensorflow to use new version of protobuf library
  • Go API
  • [WIP] CuDNN Batch Normalization Op
  • Add getting started doc for skflow
  • Reduction of CPU test time for cwise_ops_test and rnn_test.
  • Added implementation for GridRNN
  • Initial iOS Support and Example
  • Add Tan, Asin, Acos, Atan trigonometric functions. See #1108
  • 549 unsorted segment max
  • some learning decays from Stanford CS231n Karpathy lecture 6
  • fixes #1423 - add check for tests to see if tensorflow was built with gpu
  • update protobuf to 3.0.0b2.post2
  • Ref #2063: Adding validation summary writer in ValidationMonitor
  • internal to upstream 2016/4/26
  • build doc enhancements
  • only update batch norm averages during training
  • Update os_setup.md
  • Some Sampling Ops
  • [skflow] High-level Autoencoder
  • Attention Mask Ops
  • solve compatibility issue for ubuntu 16.04
  • FIFOBucketedQueue
  • Updated docstrings to link to descriptions of padding algorithms.
  • LSTM*BlockOp: Monolithic kernels for the LSTM, approx 40-50% faster.
  • [WIP] Add further changes for complex128 support
  • Fix two typos in API docs for ExponentialMovingAverage.
  • Add polygamma and zeta function to tensorflow
  • [skflow] Added support for different parameters per layer in dnn
  • [skflow] Added verbosity support to Validation monitor
  • Enables passwords for Jupyter in a Docker image
  • Match conv2 weights stddev with cuda-convnet's layer def.
  • Fix typos in API docs and Python docstring for ExponentialMovingAverage object.
  • Enable GPU for SoftmaxCrossEntropyWithLogits float64
  • Enable GPU for L2Loss float64
  • fix shuffle_batch zero_division bug
  • A couple of bugfixes
  • Set -headerpad_max_install_names on Darwin for targets that need binary header padding to rename paths
  • Basic implementation of immediate execution mode for TensorFlow
  • Enable GPU for MatMul float64
  • python 3.5 compatibility and remove unused imports
  • python 3.5 compatibility and better display effect
  • Update seq2seq
  • Set workspace name
  • Additional instructions on how to fix missing submodules
  • Confusion with QueueRunner doc
  • Fixed Tensorboard minimap not being drawn on Safari (OS X, iOS)
  • Add parameter to specify inner activations for rnn cells
  • Added comments to clarify how/why the embeddings are being minimized.
  • Markdown links using `[This link](http://example.net/)` syntax
  • Add complex128 support
  • Fixes for protobuf changes
  • Updated API docs to sync with supported data types in csiwe_op_*.
  • Fix init value in inner_product
  • Initial proof-of-concept NodeJS binding.
  • add check for read only filesystem
  • Adding validation_steps and batch_size to VaidationMonitor
  • Fix convert_to_tensor error in tf.one_hot
  • Add genrule_strategy=standalone to bazelrc
  • Learning rates based on Karpathy CS231n lecture 6
  • [tf.learn] Fix logistic_regression() summary name conflict.
  • Experimental makefile support
  • [WIP] TensorFlow Studio prototype
  • Update iOS Eigen version
  • add support for nesterov momentum
  • Logistic regression weights: Using a `tf.Tensor` as a Python `bool` is not allowed
  • Immediate execution mode with graph caching.
  • Updating the setup web page for GPU to use primary with nvidia-docker then docker_run_gpu as alternative
  • Dynamically shaped image support for `tf.image.resize_image_with_crop_or_pad`
  • MNIST downloader doesn't work as expected
  • Enable whole batch whitening.
  • [WIP] Add scan_sum and scan_prod ops
  • Update Cmake build files to reflect recent dependency change
  • Gated Feedback LSTM implementation
  • Enable tf.size() for SparseTensor
  • [learn] add an lstm regression example
  • Implement bidirectional_dynamic_rnn (#1779)
  • Update RNN PTB example to use state_is_tuple
  • Support GIF image decode ops by FreeImage library
  • Correct resnet.py and hdf5_classification.py
  • Templatize dtype in RGBToHSV and HSVToRGB
  • Adds conv3d_transpose operation (3D "deconvolution")
  • Add sparse_transpose wrapper
  • Support GIF image decode ops by giflib library
  • Added complex type support to ops for GPU
  • Update Eigen to new version with cumsum/cumprod fixes
  • Enable tf.tanh() for SparseTensor
  • Added Xcode version check and logging utilities for kernel errors
  • fix word2vec_test's tmp path
  • fix dim mismatch bug in resnet example.
  • Fix bugs in Residual Network implementation
  • [WIP] Gradient reversal op
  • [learn] update TensorFlowEstimator's functions (fit, partial_fit, predict)
  • kmeans estimator fixes
  • Update Eigen to version that includes scan op fix
  • [WIP] Add cuda_configure repository rule to autodetect cuda.
  • Extend #3020 to decode animated GIF
  • fix mispellings and add minor corrections
  • Remove support for CuDNN v2 runtime/compile from TensorFlow.
  • Change cudnn version checking to check Major version.
  • Fix issue 3186: buffer alignment needs to be 8-byte for uint64 vals
  • docker exec -it to start the interactive terminal
  • Add Float64 support to CTC Decoder Ops
  • Explanation of blank label in ctc_loss
  • bug fix: the benchmark only run FLAGS.num_batches-1 times
  • Restore a working CMake build on Linux
  • Added esn cell
  • Changed visibility of protos_cc to public
  • Use tf.contrib.layers.conv2d instead of learn.ops.conv2d which is dep…
  • Moved eightbit graph trimming to before output_nodes definition
  • Hard-code libcuda version number to "1". Fixes #2865.
  • word2vec_basic.py : to prevent possible issue with libpng
  • Extended adjust_hue and adjust_saturation to work on 4D tensors
  • Fix go build errors
  • Correction in Dequantize comments
  • fix sync update in distributed mnist demo
  • Prevent evaluating stale output node.
  • Speed up retraining times approximately 2x
  • Update the quantization how-to
  • Remove read_analogies() from word2vec class initialization
  • time_major parameter for ctc_loss op
  • Add some new features of mnist_replica
  • docs(slim): fixes typos and capitalizes Tensorflow
  • Fix reduce_prod gradient for scalar reduction indices params
  • Add a new target including the C api.
  • Allows running "-m tensorflow.tensorboard" (alternative solution)
  • #3767 tanhgrad gradient and sigmoidgrad gradient
  • seq2seq.py: Fixed the documentation to be consistent with the code
  • Fix tensorboard CSV download in Python 3
  • Reduce memory usage and increase performance for convolution on iOS
  • add int32 support for param seq_length of op reverse_sequence
  • extract element from list when py_func's output type is a single tensorflow type
  • Explanation of blank label in ctc_loss
  • Fix TypeError in gradient clipping ops
  • Add layer_norm op to contrib.layers.
  • Fix analogy response of word2vec interactive when the response is unicode.
  • Cross-import KMeans and GMM in tf.learn.estimators
  • Starting on a tf.contrib.learn notebook
  • Remove some signed/unsigned integer comparison warnings
  • Fixed inconsistency of outputs collection
  • remove XCode 7.2 instructions
  • use 'typedef' instead of 'using' to compile on gcc4.8.2
  • improve error message when a category is empty
  • set attr num_sampled <= range_max
  • update reader._file_to_word_ids
  • Remove unused cache files from Docker image
  • lstm code example clarification
  • add Rint operation
  • Issue3963
  • DO NOT MERGE -- Use canonical cuda include directory path in CROSSTOOL.
  • modify scatter_update to hack variable initialization check
  • close opened file descriptors properly
  • 549 unsorted segment max
  • Fuse resize and mirror padding ops into convolutions
  • Register missing gradient for tf.nn.max_pool_with_argmax
  • extend tf.select to broadcast a scalar condition #3945
  • time_major parameter for ctc_loss op
  • Update(mnist): fix the confusing cluster setter
  • Fixing comment, wrong data type in argument description.
  • check_version: fails if native.bazel_version is undefined
  • Correct height/width ordering to mnist docs
  • Adding missing module
  • change syntax of curl & tar
  • Documentation for creating language bindings and shape inference
  • fix spelling errors in comments
  • Fix typo in comment.
  • Fail configure script immediately on error.
  • Enable to pass feed_fn to fit method
  • Merge release branch back into master.
  • dynamic_rnn_decoder initial commit
  • Fix skflow resnet example to work on v0.10.0
  • Generate additional cudart linkopts in cuda_configure.
  • GridRNN cell uses tuples for output and states
  • Support older versions of git
  • [issue4339]Add adjust_gamma for image
  • Develop device framework
  • Adding support for Linux s390x
  • List of 2Ds -> 3D Tensor, seq2seq_loss
  • DO NOT MERGE: Change GPU test to use cuda8. Update docker ubuntu to 16.04.
  • Add Substr Op
  • WIP Add Python kernel and tests for decode_image
  • [wip] 4d image ops
  • Minor Python code health improvements
  • Fixed error with NaN elements in Wide & Deep Learning Tutorial
  • cherry-picks for 0.11.0rc1
  • Fixes and improvements for Windows platform code
  • Add basic WebDriver functional tests of TensorBoard.
  • Add a resource container for sync tokens in SyncReplicasOptimizerV2.
  • Fix typo in xla_prerelease.md
  • Branch 136382518
  • Remove duplicated arg description in doc string
  • Matrix Triangular Solve GPU op
  • Use an OrderedDict in convert_collection_to_dict
  • Update the version of re2 to support building with g++ 6.2.
  • Add support for dict Input to DataFeeder, StreamingDataFeeder, Estimator.fit() and Estimator.evaluate()
  • Deprecated the `is_training` argument in vgg.py.
  • Add data_format option to convolution2d_transpose
  • Issue 4746
  • Branch 136231860
  • Verify file_io.create_dir and file_io.list_directory in GCS test .
  • Fix too long sentence
  • Update load_csv > load_csv_with_header for r 0.11
  • Added list functionality to registry object
  • Make TensorFlow pip package build on Windows
  • Fixed fused_batch_norm gradients
  • Added support for new fused Winograd algorithms in cudnn 5.1
  • Issue 1702
  • Fixed the wrong comments of `tf.scan`.
  • Replace all calls to contrib.learn.datasets.base.load_csv with calls …
  • Fixed leak `_EventLoggerThread` threads
  • Add PIL formats(gif,png...) support for retrain.py
  • updates to einsum
  • add required positional argument: 'features_dtype'
  • move type2index into EIGEN_HAS_INDEX_LIST
  • Revert "extend tf.select to broadcast a scalar condition #3945"
  • Fix merging of summaries to match the tutorial "TensorFlow Mechanics 101"
  • Cherrypicks and rc2.
  • for completeness that if the tarball from Alex's website is already i…
  • dynamic_rnn time_major=False transpose fixed for inputs of rank>3
  • os_setup.md references "Installing from sources" consistently
  • Made mnist tutorial data storage folder consistent
  • doc: fix possible copy&paste error in cumprod docs
  • Updating README adjusting to GPU support
  • add support for fetch and feed session.run conversion functions
  • Do not set testonly to //tensorflow/python:construction_fails_go
  • replaced inception file with v3
  • Little change in tf.contrib.learn Quickstart tutorial
  • Use fast IDCT for JPEG decoding by default
  • Issue 1702
  • Fix a typo in export meta graph sample code.
  • Remove a premature checkpoint existence test in Saver.restore().
  • Add ci scripts for cmake build and test on windows.
  • Updates download_progress_hook comment
  • Issue5147
  • fix failed copy for op without outputs but with control outputs
  • Adding support for Linux s390x
  • GridRNN cell uses tuples for output and states
  • Update 1_notmnist.ipynb
  • fix windows gpu build
  • Gradle change in 2.2 and sdk update
  • Add synthetic datasets
  • Add atrous_conv2d_transpose python function
  • Fix confusing CTC documentation
  • Orthogonal initializer
  • [bug]Fix a issue (#4419)
  • Fix CMake config if used as a subproject
  • Allow deprecation decorator to specify owner
  • Checks that quantized Tensor's elements have type tuple.
  • Deprecated array and embedding ops in contrib.learn
  • Add support for decoding RGBA images in tfexample_decoder.
  • Fix syntax error in UnchangedShape example code
  • Branch 138943977
  • enable avgpooling for windows gpu builds
  • Branch 138939121
  • Add string tensor tags to new summary interface
  • make RestoreV2() work for windows
  • optional libjpeg-turbo support
  • Add support for dict Input to DataFeeder, StreamingDataFeeder, Estimator.fit() and Estimator.evaluate()
  • word2vec: Create directories for saving summaries, if necessary.
  • [Windows/CMake] Add support for tf.contrib.
  • Add maxout op to tf.contrib.layers
  • Added Student T CDF as per 5413
  • Removed unnecessary classproperty decorator
  • Remove deprecated copy of checkpoints lib
  • Refactor cuda_configure to better support for different CUDA installations
  • Updating CMake requirements #5487
  • Force precision to be float for the MNIST fully connected example.
  • enabling VocabularyProcessor.transform to also return document length.
  • Swap dictionary and tensor for readability
  • Update mnist_softmax.py
  • New "TimeMillis" method has been added.
  • change validation_metrics to MetricSpec
  • Use passthrough args rather than wrapper functions
  • Fix projector_api_test, ensuring it runs.
  • Simplify CMake thirdparty library target names
  • Fix: Create a new script for gpu-mac tests.
  • Fix: Do not use perl in configure, use only standard bash tools.
  • [Windows] Don't try to read numa node from SysFS
  • fix broken link in docs
  • Translate.py - Added the option to translate a file
  • More similar typo to #5973
  • Fix #5946.
  • fixed the argscope example
  • seperable -> separable
  • Run cc tests on Windows with Bazel
  • [TF-Slim README] A few typos and latest APIs
  • Enable sandboxed builds for tensorflow.
  • fixed run_path_pairs AttributeError
  • Revert "Removed unnecessary classproperty decorator"
  • Make placement of constants follow consumers if they are all on the same device
  • Add load method to tf.Variable
  • [CMake] Add TensorBoard dependencies to PIP package.
  • better documentation of return value of boolean_mask
  • Fix a bug which repeated variable creation in optimizer when using bucketing
  • A theano.clone() equivalent for Tensorflow
  • Propagate seed in parallel_read to readers. Fixes #6735
  • temporarily add /sbin to PATH for ldconfig call
  • update Defun doc to clarify that definitions are frozen at first .run call
  • GridRNN cell uses tuples for output and states
  • ./configure: A better way to exclude fetching //tensorflow/examples/android/...
  • add virtual START&END symbol to CRF and implement viterbi decoding
  • GPU impl. of hue adjustment op
  • Fix wrong order for op_name/summary_name in range_input_producer
  • Update LSTMBlockCell to use LSTMStateTuple as state
  • Enhance cuda kernel helper.
  • Adding support for Big Endian in decode_raw_op_test
  • Max pool grad grad
  • change tensorboard html data request path to relative
  • Allow for different input signatures for different modes and additionally updating zlib fix
  • Add adaptive softmax implemtation in nn_impl.py
  • Make placement of constants follow consumers if they are all on the same device
  • Added absolute path expansion of parent directory check in saver.py, …
  • update license to 2017
  • [Java] Bulk data transfer for Tensor class
  • Fix comment on update_op of streaming_concat()
  • training.slot_creator: support for non-fully-defined shaped vars
  • Add sequence_loss and sequence_loss_by_example working with dynamic_rnn without having to change outputs shape
  • Sparsemax
  • docs: add a troubleshooting section to the faq
  • made contrib flatten layer accept tensors with dynamic shapes
  • accelerate crf_log_norm
  • Adding fast layer normalization GPU kernels and layer functions to tf.contrib.layers
  • Fix Python 3 dict object to list, described in issue #5488
  • slim.learning.train and slim.evaluation.evaluation to handle OutOfRange gracefully
  • Initialise DNNClassifier with normalizer_fn
  • Fix possible flake sources in saver_test and supervisor_test.
  • Bump up size for estimator_test
  • Remove local_run in Experiment
  • Update tf_upgrade.py to handle more upgrade cases
  • Move default for dnn_activation_fn to _dnn_linear_combined_model_fn for consistency
  • Catch dnn_hidden_units to avoid obsture error message
  • Update proto library reference and make curl forward/backward compatible
  • [WIP] Add standlone Poplar plugin scaffolding
  • Adding checks for broken bottleneck files
  • Update iris.py
  • cross_validation is deprecated
  • Upgraded to libxsmm 1.7
  • Restore deprecation warnings on {histogram,scalar}_summary in logging_ops
  • CudaRoot() returns the configured CUDA toolkit path.
  • Replace all gate_gradients=1 with gate_gradients=GATE_OP in documenta…
  • FreeBSD compatibility
  • improved and fix bug of getting next batch in mnist
  • Add support for dict generator input_fn in learn_io
  • Remove errorneous squeeze in `tf.contrib.losses.sparse_softmax_cross_entropy_loss`
  • Add gradient for `placeholder_with_default`
  • Add note to new op and source build doc (#6473)
  • TF-549 Adds unsorted segment max Op
  • Make XLA command-line option to mnist_softmax_xla.py actually work.
  • Added Intel MKL graph optimization code and the ability to enable MKL when running configure.
  • Feature: log-log scale
  • Add support for different feature signatures
  • Add ProfilerHook for capturing CPU/GPU profiling information with MonitoredSession
  • GraphDef is deprecated.
  • During makefile CI build, adjust ownership of the files docker writes…
  • Fixed off-by-one error in L115-116 in wordvec_basic.py
  • Fix regression in TensorBoard
  • Fix: make sure to join all threads to avoid flakes in sync_replicas_optimizer_test.
  • Change mismatched ref input error message
  • Pass required input_length to batch_sequences_with_states() in example
  • Remove leading slash from data directory
  • Expose param_regularizers in python batch_norm layer
  • Update shape checking logic in einsum
  • resize_image_with_crop_or_pad can work with batch of images
  • Add BUILD file and test case for "fact" user op
  • Fixing example strided_slice
  • Adding MKL matmul op
  • Fixed a bug where inception layers had 2 3x3 conv kernels instead of …
  • fesetround is not part of the C++ std library
  • Add HomeBrew instructions
  • Make CNN input float32
  • change tf.contrib.learn.datasets.load_dataset('boston') to sklearn.da…
  • resolve #6762 on ldconfig only available on root PATH
  • fix docs formatting of tf.while_loop
  • Fix build error, where nccl requires -lrt link option
  • Don't use __has_builtin with version of apple clang that doesn't supp…
  • [WIP] Add standlone Poplar plugin scaffolding
  • Register variable's proto function with key 'LOCAL_VARIABLES'
  • verbs: Change IB/RoCE path MTU to 1K (instead of 4K)
  • Allows direct builds from a Visual Studio solution & Enables Debug builds
  • Fix error of version_info.cc not being generated on windows
  • Including batch_norm as the normalizer function by default, as mentioned in function description
  • Link to gcc_s and gcc if compiler is GCC version 5. Fixes issue #9593
  • Making the quantizedMatmul for FC layers multi-thread
  • Added options for linking to libraries that are present on system
  • Add a simple BMP decoder and enable it on Android
  • Delay Compensated Asynchronous Stochastic Gradient Descent
  • Add support of Sparse Tensor Slice
  • Unveil type check part in _VerifyGeneratedGradients that was not actu…
  • Dynamic ksize and strides with MaxPool and AvgPool
  • small change to allow multiple wrappings with EmbeddingWrapper
  • Add 3D operations for layers: conv3d, avg_pool3d and max_pool3d
  • Fix https://github.com/tensorflow/tensorflow/issues/8207
  • add support for flat both inner and outer dims
  • Add scaffolding for running XLA python tests on a plugin backend.
  • Shuffle Op: Fixes #9369
  • Add Feature: Sparse matrix multiplications for Tensors with rank > 2
  • Store step stats in benchmark model and use python timeline to visual…
  • a python implmentation of label_image
  • A replacement for tf.cond
  • [OpenCL] Implementation improvements
  • Add output_shape parameter to transposed2d_convolution layer
  • Ops and kernels for partial_reduces
  • CPU kernels for FFT (WIP)
  • Fix KeyError in quantize_graph round, quantize opt
  • Add TensorFlow equivalent to np.repeat
  • Add gzip and zlib support for FixedLengthRecordReader
  • Fix typos
  • Undo modification to `x` dict by deleting `_TARGET_KEY` after features dequeue
  • Fixed a comment typo in GraphView:InitializeNode(), executor.cc.
  • Fixed tf.contrib.crf.crf_log_norm to handle zero sequence length.
  • Add a tip for tf.train.LoggingTensorHook
  • fix the return value of Tensor::flat_inner_outer_dims
  • create boston.ipynb
  • Branch 155393864
  • [XLA] Add a strides parameter to the XLA slice operation
  • Support `if_darwin` condition
  • Release branch commits
  • [XLA] Ensure constants conform to the specified index type
  • MKL_INSTALL_PATH should not be clear when given
  • Add is_closed() method to Queue
  • Register GPU RefExit kernel
  • [XLA] Move some useful Literal conversion code into LiteralUtil
  • RecordInput mini batches for dividing processing among multiple devices.
  • Add SMAPE(symmetric mean absolute percentage error) loss function.
  • cudnn: Fix symlink includes handling
  • Automatically convert inputs to tensors in Dataset.from_tensor_slices
  • Add support for sparse_reduce_max and sparse_reduce_max_sparse
  • add Cuda{2D,3D}LaunchConfig that maximizes occupancy
  • [XLA] HLO Executor based backend example
  • Update rnn.py
  • Add batch_size and ValidationMonitor
  • New reader for LMDB databases
  • Multiplicative Integration Recurrent Neural Networks
  • Quantize conv2d transpose
  • Change SummaryWriter --> FileWriter in TensorBoard
  • Fixed _linear(.) to use *batch* matrix multiplication.
  • Disable AWS S3 virtual addressing
  • Make raw_rnn accept scalar or TensorArray values for state.
  • raw_input() was removed in Python 3
  • long was removed in Python 3
  • xrange() was removed in Python 3
  • Update README.md
  • replace deprecated keep_dims with keepdims in keras.backend
  • Imported lstm1d and lstm2d in ndlstm __init__.py.
  • Fix an imperfect implementation of tf.losses.mean_pairwise_squared_error
  • Lite: Supporting Raspberry Pi.
  • Fix build errors in contrib/mpi introduced by commit 6042b5d267f
  • common global variable with constant.py
  • Correct a small typo
  • remove SRU num_units == x.shape[-1] restriction
  • Compare_and_bitpack function for bool for big endian
  • change from deprecated version to a new version
  • cmake gpu build improvement
  • Added early stopping and CheckpointSaverListeners to train and evaluate
  • Fix of issue #13164 (Merges #13382)
  • Increase tolerance in `losses_impl_test.py`. fixes #16238
  • Fixed documentation formatting
  • py_func convert unicode string results to bytes for python2
  • Add alternative paths for CUDA installation.
  • Allow step callback for scipy SLSQP
  • Add LINM (Loop Invariant Node Motion) optimization pass in GraphOptim…
  • Update README.md
  • Update README.md
  • Add tf.multi_one_hot that one-hot encodes multiple columns of Tensor
  • Add a rnn example on mnist dataset using tf library
  • Fix two small issues of XLA
  • [Windows] Copy NominalCPUFrequency from Abseil
  • [MSVC] Workaround MSVC template/lambda parsing bug
  • typo fix
  • Add reduction parameter to mean_pairwise_squared_error loss
  • Improve formatting of Tensor shapes in tf.losses
  • dataset_ops.py batch() checks type immediately
  • Added detailed discussion of non-strict semantics
  • Improve shape function of NonMaxSuppression
  • Implementation of the unpooling layer in tf.contrib.layers
  • Enable validate_args for all distributions based on a global parameter
  • Update bazel version in docker
  • Adding the CMAKE_GENERATOR line to all external cmake files.
  • Fix warning about keep_dims. keep_dims -> keepdims for tf.reduce_sum().
  • GLSTMCell fix
  • Tensorflow Lite demo app for Android: add support for floating point models as Inception-v3
  • [tflite] fixed label_image resize bilinear problems
  • Python3 support of docs generation
  • Add broadcast support for softmax_cross_entropy_with_logits
  • Allow passing Saver write_version to 'evaluation_once' and 'evaluatio…
  • Add NLSTM RNN cell and the unit tests
  • fix tf.GIT_VERSION always 'unknown' on windows cmake build
  • Add stream selection support for `tf.contrib.ffmpeg.decode_video`
  • Windows: Enable tensorflow/contrib in Bazel build
  • [XLA] Fix subcomputation unification not adjusting conditionals
  • Updated adding_an_op.md to reflect newer API and to fix typos
  • Fixes issues in tf.contrib.keras.utils.Progbar
  • [WIP] More generic configuration
  • Refactoring by extracting duplicate code into methods
  • Moving code using new to absl::make_unique. This should make cleaning…
  • Fix tfcompile module label.
  • Update version string to 1.6.0
  • Update nn.py
  • Change repository command to valid value
  • Support GCS URL by tf.estimator.LatestExporter
  • Error in variable_scope initialization, via estimator api, due to a panda.DataFrame without column headers
  • Clean up output formatting of saved_model_cli.py
  • Cherrypick: Don't assign device for the keras part of _saved_first_checkpoint. Fi…
  • Change unicode() --> six.text_type() for Python 3
  • C++ gradients for MaxPool3D, AvgPool and AvgPool3D
  • Update TrainingSpec and EvalSpec pydoc
  • Extract kernel, bias values of a layer
  • Add dict(features) instead of features in beginners guide
  • Implementation of Unpooling operator
  • MKL: Update mkl docs
  • Fix build issues when having packed git refs.
  • [Intel MKL-DNN]: added MKLDNN dilated convolution support
  • LoggingTensorHook to read from runconfig in Estimator
  • Update TF Lite android demo to Android Studio 3
  • Add broadcasting functionality for Div and Sub ops.
  • Added a way to accept list as input to get_variable_value in Estimat…
  • Implement the bilinear initializer op
  • Fix a bug in tf.multiply documentation
  • Add NumPy style warning when casting complex to float
  • Minor improvements to `estimator.predict()` docs
  • Add ability to use default values via environment variables
  • Add missing `override'
  • make the TfLiteCameraDemo.apk built with bazel work again
  • tf.Dimension raises TypeError for tf.DType
  • Use relative and correct paths to get flatc and schema file.
  • tf.tile gradient supports IndexedSlice
  • minor edit
  • make benchmark_model for TFLite build
  • Fix broken link in programmers_guide/faq and some minor format
  • Fix broken link in kernel method tutorial
  • Fix windows gpu cmake build
  • Small markup documentation fixes around closing blocks
  • [CMake]Exclude duplicate tests from ${tf_test_src_simple}
  • Fix inconsistency of eq for TensorShape
  • MKL DNN: fix the TF1.6 speed issue by fixing MKL DNN LRN taking the optimum path
  • Update fold_old_batch_norms.cc to accommodate 'NCHW' format.
  • Fix the broken link of tf-learn's iris tutorial also some format and typo
  • C++ gradient for Slice
  • Support other dtypes in BeamSearchDecoder initialization
  • get_variables_and_layers
  • Update RELEASE.md
  • contrib/lite: add missing include assert.h (spectrogram.cc)
  • Run selective registration tool even if it's just been built
  • Remove TF_NEED_KAFKA from configure.py as it is not needed anymore.
  • Fix the tpu related broken link especially for imagenet_to_gcs.py
  • Fix markdown error in layers tutorial.
  • Fix the messed up list format in using_tpu.md
  • C++ gradient for StridedSlice
  • SYCL with ComputeCpp: local_config_sycl has multiple matches
  • add reflexive method for Dimension
  • Register half in some ops which support all floating point types
  • contrib/lite: spelling/code tweaks
  • Add scan command to saved_model_cli to check for security sensitive ops.
  • Enhancement with deprecated_argument_lookup
  • Improvements for Android TV
  • Fix minor typo in saved_model.md
tensorflow questions on Stackoverflow (View All Questions)
  • How do I run a python script with Tensorflow running in a Docker on Windows?
  • How to Fine-tuning a Pretrained Network in Tensorflow?
  • Initializing variable with another variable using tensorflow
  • Restoring TensorFlow model
  • How to invoke tf.initialize_all_variables() in C++? tensorflow
  • tensorflow: segmentation error in multiple threads
  • TensorFlow with Docker
  • Parameterized model over output functions using tensorflow
  • How to set layer-wise learning rate in Tensorflow?
  • Error while importing Tensorflow in python2.7 in Ubuntu 12.04. 'GLIBC_2.17 not found'
  • Preprocess a Tensorflow tensor in Numpy
  • Renormalize weight matrix using TensorFlow
  • Unable to build TensorFlow from source with bazel. 22nd January 2016
  • Update a subset of weights in TensorFlow
  • How to safely terminate a tensorflow program running on mutiple GPUs
  • Tensorflow: share value for two different variables within same operation
  • Tensorflow implementation of loss of Q-network with slicing
  • How to expand a Tensorflow Variable
  • Tensorflow indexing with boolean tensor
  • Running a Tensorflow model on Android
  • Tensorflow network diverges if reading/preprocessing is done on cpu
  • TensorFlow installation results in ImportError: No module named tensorflow
  • Tensorflow: Passing a session to a python multiprocess
  • Cannot gather gradients for GradientDescentOptimizer in TensorFlow
  • TensorFlow apply_gradients remotely
  • Interpolated sampling of points in an image with TensorFlow
  • Tensorflow exception with matmul
  • TensorFlow with Eclipse
  • TensorFlow: best method for non-one-hot vectors?
  • Tensorflow: Diagonal Subtensor for 3D Convolutional NN
tensorflow latest release notes
v1.6.0 TensorFlow 1.6.0

Release 1.6.0

Breaking Changes

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

Major Features And Improvements

  • New Optimizer internal API for non-slot variables. Descendants of AdamOptimizer that access _beta[12]_power will need to be updated.
  • tf.estimator.{FinalExporter,LatestExporter} now export stripped SavedModels. This improves forward compatibility of the SavedModel.
  • FFT support added to XLA CPU/GPU.
  • Android TF can now be built with CUDA acceleration on compatible Tegra devices (see contrib/makefile/README.md for more information)

Bug Fixes and Other Changes

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

API Changes

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

Known Bugs

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

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

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

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

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

  • The tensorboard command or module may appear to be missing after certain upgrade flows. This is due to pip package conflicts as a result of changing the TensorBoard package name. See the TensorBoard 1.6.0 release notes for a fix.

Thanks to our Contributors

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

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

v1.6.0-rc1 TensorFlow 1.6.0-rc1

Release 1.6.0

Breaking Changes

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

Major Features And Improvements

  • New Optimizer internal API for non-slot variables. Descendants of AdamOptimizer that access _beta[12]_power will need to be updated.
  • tf.estimator.{FinalExporter,LatestExporter} now export stripped SavedModels. This improves forward compatibility of the SavedModel.
  • FFT support added to XLA CPU/GPU.
  • Android TF can now be built with CUDA acceleration on compatible Tegra devices (see contrib/makefile/README.md for more information)

Bug Fixes and Other Changes

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

API Changes

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

Known Bugs

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

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

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

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

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

  • The tensorboard command or module may appear to be missing after certain upgrade flows. This is due to pip package conflicts as a result of changing the TensorBoard package name. See the TensorBoard 1.6.0 release notes for a fix.

Thanks to our Contributors

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

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

v1.6.0-rc0 TensorFlow 1.6.0-rc0

Release 1.6.0

Breaking Changes

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

Major Features And Improvements

  • tf.estimator.{FinalExporter,LatestExporter} now export stripped SavedModels. This improves forward compatibility of the SavedModel.
  • FFT support added to XLA CPU/GPU.

Bug Fixes and Other Changes

  • Documentation updates:
    • Added a second version of Getting Started, which is aimed at ML newcomers.
    • Clarified documentation on resize_images.align_corners parameter.
    • Additional documentation for TPUs.
  • Google Cloud Storage (GCS):
    • Add client-side throttle.
    • Add a FlushCaches() method to the FileSystem interface, with an implementation for GcsFileSystem.
  • Other:
    • New Optimizer internal API for non-slot variables. Descendants of AdamOptimizer that access _beta[12]_power will need to be updated.
    • Add tf.contrib.distributions.Kumaraswamy.
    • RetryingFileSystem::FlushCaches() calls the base FileSystem's FlushCaches().
    • Add auto_correlation to distributions.
    • Add tf.contrib.distributions.Autoregressive.
    • Add SeparableConv1D layer.
    • Add convolutional Flipout layers.
    • When both inputs of tf.matmul are bfloat16, it returns bfloat16, instead of float32.
    • Added tf.contrib.image.connected_components.
    • Add tf.contrib.framework.CriticalSection that allows atomic variable access.
    • Output variance over trees predictions for classifications tasks.
    • For pt and eval commands, allow writing tensor values to filesystem as numpy files.
    • gRPC: Propagate truncated errors (instead of returning gRPC internal error).
    • Augment parallel_interleave to support 2 kinds of prefetching.
    • Improved XLA support for C64-related ops log, pow, atan2, tanh.
    • Add probabilistic convolutional layers.

API Changes

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

Known Bugs

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

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

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

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

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

Thanks to our Contributors

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

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

Other projects in C++