TensorFlow version 1.11.0 is now available.

This article covers some of the major changes, major features and improvements, Bug fixes, and other changes.

Main features and improvements

Nvidia GPU:

  • Now (according to TensorFlow 1.11) the pre-built binaries are built for cuDNN 7.2 and TensorRT 4. Please see the installation guide: Installing TensorFlow on Ubuntu (https://www.tensorflow.org/install/install_linux#tensorflow_gpu_support)

Google Cloud TPU:

  • Experimental TF.data integration for Keras on Google Cloud TPU

  • Pilot/preview support for Eager Execution on Google Cloud TPU

Distributed strategy:

  • Added multi-GPU distributed policy support for TF.keras. Fit, Evaluate, and Predict support for distribution

  • Add distributed policy support and independent client support for multi-worker in Estimator. See [README] (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/distribute) for more details

Add C, C++, and Python functions to query the kernel

Major change

Keras:

  • The default values of RandomUniform, RandomNormal, and TruncatedNormal initializers for TF.keras have been changed to ensure that they match the default values in the external KERAS

  • Major change: Running the model.get_config() method for the Sequential model now returns the Config Dictionary (in line with other model instances) instead of the base layer configuration list

Bug fixes and other changes

C + + :

  • Change the SessionFactory: : NewSession signature, so it can give detailed information to the failure

Tf. Data:

  • Contrib.data.make_csv_dataset () removes the num_parallel_parser_calls parameter

  • If the parameter does not match a file, tf.data.dataset. List_files () will raise an exception during initialization

  • For clarity, rename the BigTable category to BigtableTable

  • Record Cloud Bigtable API usage

  • Add tf.contrib.data.reduce_dataset, which can be used to reduce a dataset to a single element

  • Generalization tf. Contrib. Data. Sliding_window_batch

INC:

  • The trig solution is improved in operation

Tf. Contrib:

  • For tf. Keras. The layers. LocallyConnected2D and tf keras. The layers. The LocallyConnected1D implementation for parameter was added. The new mode (implementation = 2) performs forward propagation as a product of a single dense matrix, which results in significant acceleration in some scenarios (but may result in performance degradation in some scenarios – see DocString). This option also allows the use of padding = same

  • New documentation has been added to explain the difference between tf.fill and tf.constant

  • IndexedDatasets were experimentally added

  • Add selective registration targets using the Lite Proto runtime

  • Add simple Tensor and DataType categories for TensorFlow Lite Java

  • Added support for bitcasting of uint32 and uint64

  • An Estimator subclass was added, which can be created from SavedModel (SavedModelEstimator)

  • Add leaf index mode as parameter

  • From the tf. Contrib. Image. The transform of the input is allowed in the different output shape

  • Change the state_size order of the StackedRNNCell to the natural order. To preserve the existing behavior, users can add reverse_STATE_ORDER = True when constructing StackedRNNCell

  • Deprecate self.test_session() in favor of self.session() or self.cached_session()

  • Direct import tensor.proto.h (pass import will be removed from tensor.h soon)

  • Estimator.train() now supports out-of-the-box tF.contrib. summary. * Each call to.train() now creates a separate TFEvents file and does not reuse the shared file

  • Fixed shrinkage performance of FTRL optimizer L2: Gradient of L2 shrinkage term should not terminate in accumulator

  • Fix toCO compilation/execution bug on Windows

  • The GoogleZoneProvider category was added to detect where parts of the TensorFlow program run in the Google Cloud Engine

  • It is now safe to call the TF_Delete * function of any C API on NullPTR

  • Log some error messages on Android

  • The FakeQuant numbers in TFLite were matched to improve the accuracy of TFLite quantitative reasoning model

  • Check optional bucket locations for GCS file systems

  • StringSplitOp and StringSplitV2Op have been enhanced

  • The performance of regular expression replacement operations is improved

  • If if.write() fails, TFRecordWriter will now report an error

  • The TPU: TPUClusterResolvers cluster parser will provide more detailed and useful error messages

  • It is not recommended to use the SavedModelBuilder method to add the MetaGraphs legacY_init_op parameter. Use the equivalent main_op parameter instead. We now explicitly check for a single main_OP or legacy_init_op when we build SavedModel, whereas previously main_op was checked only at load time

  • The protocol for Estimator training can now be configured in RunConfig

  • The performance of solving trigonometric numbers is improved

  • The API of RNN unit of TF and Keras is unified. A new get_initial_state() method was added for the RNN cells of Keras and TF, which will replace the existing Zero_state () method in the future

  • Updated initialization of variables in Keras

  • Update “constrained_optimization” in tensorflow/contrib

  • Lift tree algorithm: Add pruning mode

  • Tf.train.Checkpoint By default, old checkpoints are not deleted

  • TFDBG: When debugging, the Tensor data cached will take up to 100 GB of disk space. Allow the environment variable TFDBG_DISK_BYTES_LIMIT to be added to adjust this upper limit

Thanks to our contributors

This launch was made possible by the contributions of many people at Google:

Aapeli, adoda, Ag Ramesh, Amogh Mannekote, Andrew Gibiansky, Andy Craze, Anirudh Koul, Aurelien Geron, Avijit, Avijit-Nervana, Ben, Benjamin H. Myara, Bhack, Brett Koonce, Tianqi Zhang, Xiaofei Zhang……

Finally, please click “Read the original article” to see the full version of the description.