Open source China global exclusive benefits, computing conference ticket rebate! >>>  

TensorFlow 1.3.0 has been released with a number of exciting new features. Here’s a look at the major new features and improvements:

  • From cuDNN 5.1 to cuDNN 6. CuDNN 7 is expected to be used in the upcoming TensorFlow 1.4 release

  • The tf.contrib.data.Dataset class has received some important updates

  • Advanced API functions and statistical distribution. A new addition is the multiple statistical distribution. Use a class to represent a statistical distribution and initialize it with parameters that define the distribution

  • Some changes have also been made to existing functions. The Tf. Gather function, which is used to select variables from tensors, now adds axis parameters for more flexibility in data collection

  • Import TensorFlow runs faster

  • New algorithms and Python bundles have been introduced for Cloud TPU

  • To be symmetric with TensorFlow-Android, tensorFlow-ios CocoPod was added

  • The basic implementation of ClusterResolvers was introduced

  • Changed the reference to LIBXSMM to version 1.8.1

  • Added tf.contrib.signal, a primitive library for signal processing

  • Add tF.contrib. resampler, which contains differentiable resampling of CPU and GPU graphics.

Major API changes:

  • Tf.rewriterconfig has been removed from the Python API since the 1.2 final beta release. Graph rewriting is still available, but not like tf.rewriterConfig. Instead, an explicit import method has been added.

  • One important change to tf.contrib.data.Dataset is the nested structure. List objects have been modified to tf.tensor, and you might need to change the use of lists in your existing code to tuples. In addition, dictionary objects are now supported as nested structures

Known issues:

  • Tensorflow_gpu fails to compile using Bazel 0.5.3

Click here for the full release notes

Download address:

  • Source code (zip)

  • Source code (tar.gz)