Compile | AI technology base (public ID: rgznai100)

Participate in | zhang

TensorFlow 1.7.0 has been released for the first time.

 

Major features and improvements

  • Remove the Eager mode from contrib, now use tf.enable_EAGer_Execution ()

  • Graph has rewritten the emulated fixed-point quantizer and is now compatible with TensorFlow Lite, supported by the new tF.contrib.quantize package

  • Using TF.custom_gradient, you can easily customize the gradient calculation

  • The TensorBoard debugger plug-in, the GraphicalUser Interface (GUI) for the TensorFlow Debugger (TFDBG), is now available in alpha

  • Using new tf. Contrib. Data SqlDataset, support a sqlite database as a Dataset object to read

  • Tf. Contrib. Framework. The CriticalSection add Mutex/CriticalSection distributed

  • Tf.regex_replace better supports text processing

  • Contrib.data.bucket_by_sequence_length supports simple and efficient sequence input

 

Bug fixes and other improvements

  • Accelerated Linear Algebra (XLA) :

    • XLA adds support for MaxPoolGradGrad

    • XLA disallows CSE pass from TensorFlow

  • Tf. Data:

    • tf.data.Dataset

  1. The C++ Dataset op kernel can be built as an external library by using tf.load_op_library() mechanism

  2. Dataset. List_files () performs random shuffling of output by default

  3. Dataset.shuffle(… , seed=tf.constant(0, dtype=tf.int64)) now and Dataset. Shuffle (… ,seed=0) returns the same element sequence

  • Tf.data. TFRecordDataset adds the num_parallel_reads parameter

  • Tf. Contrib:

    • Tf. Contrib. Bayesflow. Halton_sequence now support the randomization

    • Tf.contrib.all_reduce adds support for scalars

    • Tf. Contrib. Bayesflow. Add effective_sample_size McMc_diagnostics

    • Tf. Contrib. Bayesflow. Add potential_scale_reduction McMc_diagnostics

    • Added BatchNormalization, Kumaraswamy bijectors

    • Tf.contrib. learn is no longer supported in the future. See the instructions in contrib/learn/ readme.md to convert the existing code

    • tf.contrib.data

    1. Remove no longer support classes, including tf. Contrib. Data. The Dataset, tf, contrib. Data. The Iterator, tf, contrib. Data. FixedLengthRecordDataset, Tf. Contrib. Data TextLineDataset, and tf contrib. Data. TFRecordDataset

    2. Add bucket_by_sequence_length, sliding_WINDOW_batch, and make_batched_features_dataset

    • Remove tf.contrib.ndlstm that is no longer maintained. You can find it from the external web site: https://github.com/tmbarchive/tfndlstm

    • Migrate most of the content of tF.contrib.BayesFlow to its own repository: TFP

  • other

    • If an exception is thrown, tF.py_func will now print out the complete stack trace

    • TPUClusterResolver is integrated with GKE to support Cloud TPU

    • Add a library of sampler statistical tests

    • For Cloud TPU, help functions for streaming data are added from the GCE VM

    • ClusterResolvers and TPUEstimator are integrated

    • Unified interfaces for Metropolis_Hastings and the HMC kernel

    • Migrate LIBXSMM convolution to a separate — defineFlag, so this operation is now disabled by default

    • Repair the MomentumOptimizerlambda

    • Reduce the tfP. layers boilerplate code with programmable docstrings

    • Adding auc_with_confidence_intervals, the method can calculate AUC values and confidence intervals with linear time complexity

    • Regression_head now accepts custom connection functions as input, allowing users to define their own connection functions in cases where array_ops.identity does not support them

    • Fixed initialized_value and Initial_value behavior when creating ResourceVariables objects from VariableDef

    • TensorSpec was added as the explanation document for Tensors

    • Deterministic constant folding operations

    • Tf.linalg.* Supports the dType of Float16

    • Add tf) estimator) export. TensorServingInputReceiver allow tf. The estimator. The estimator. Export_savedmodel incoming raw tensors to simulate the function

    The author | yifeif

    The original link

    https://github.com/tensorflow/tensorflow/releases/tag/v1.7.0-rc0?from=groupmessage&isappinstalled=0