TensorFlow 1.8.0 Release Candidate Announced!

April 17, 2018, 3:03 p.m. By: Kirti Bakshi

TensorFlow 1.8.0

TensorFlow, an open-source software library is used for the purpose of dataflow programming across a range of tasks. It is also known as a symbolic math library and is used for applications that are related to machine learning such as neural networks as well.

As of now, till date, there have been many releases of the open source software library and in order to add to it even further, here we have the announcement of TensorFlow 1.8.0-rc0 Release Candidate!

Major Features And Improvements:

  • tf.contrib.distribute.MirroredStrategy() to tf.estimator.RunConfig() can now be passed in order to run an Estimator model on multiple GPUs on one machine.

  • Addition of tf.contrib.data.prefetch_to_device(), which supports the process off prefetching to GPU memory.

  • As pre-made Estimators, an addition of Gradient Boosted Trees: BoostedTreesRegressor, BoostedTreesClassifier.

  • Addition of 3rd generation pipeline config for Cloud TPUs which as a result improves performance and usability.

  • The moving out of tf.contrib.bayesflow to its own repo.

  • In order to allow generic proto parsing as well as RPC communication, an addition of tf.contrib.{proto,rpc.

*tf.data:

  1. Addition of tf.contrib.data.prefetch_to_device, which as a result enables prefetching dataset elements to GPU memory.

  2. Addition of tf.contrib.data.AUTOTUNE, which based on your system and environment allows the tf.data runtime to automatically tune the prefetch buffer sizes.

  3. Addition of tf.contrib.data.make_csv_dataset for building datasets of CSV files.

* Eager Execution:

  1. With the help of eager execution, the Datasets can now be used as standard python iterators (for a batch in dataset:). When eager execution is enabled, Both Dataset.iter() and Dataset.make_one_shot_iterator() can now be used in order to create iterators.

  2. Automatic device placement has been enabled

  3. The movement of tf.GradientTape out of contrib.

* tf.keras:

  1. Addition of the fashion mnist dataset.

  2. New data preprocessing functions: image/random_brightness, sequence/TimeseriesGenerator, and text/hashing_trick.

* Accelerated Linear Algebra (XLA):

  1. Select and scatter in reference util and evaluator now to break ties make the use of lexicographical order.

*TensorFlow Debugger (tfdbg) CLI:

  1. During operations of tensor-filter, the allowance of exclusion of nodes by regular expressions.

  2. In some text terminals, the Fixing of spurious background colours.

*tf.contrib:

  1. Addition of meta-distribution BatchReshape which reshapes batch dimensions.

  2. tf.contrib.layers.recompute_grad works for explicit gradient checkpointing on TPU.

  3. Addition of tf.contrib.framework.argsort.

  4. Permitting DNNBoostedTreeCombinedEstimator to work with core versions of feature columns as well as losses.

  5. Fixing of the bug in tf.contrib.opt.MultitaskOptimizerWrapper where there was a mismatch in the types of tensors.

Some Other changes include:

  • Low-level graph construction now calls the TensorFlow C API. This change to most of the users should be invisible, 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.

  • Addition of description of shapes and a pointer to the tutorial notebook in tf.distributions.Distribution.

  • The Updation of scatter operations: Addition of tf.scatter_min and tf.scatter_max and Extending of scatter operations in order to work with a scalar update parameter.

  • For use in TensorFlow codebase only, Movement of cuDNN RNN ops to core.

  • Addition of float64 support for Conv2d, Conv2dBackpropInput, and Conv2dBackpropFilter.

  • Addition of float64 support for AvgPool/AvgPoolGrad.

  • The making of graph name scope thread local so that they work correctly in multi-threaded environments.

  • Update of the nsync synchronization library in order to avoid slow primitives on Linux.

  • Removed the need to put nsync/public on C include path when building custom ops.

  • Addition of tf.image.psnr, tf.image.ssim, tf.image.ssim_multiscale, tf.image.image_gradients, tf.image.sobel_edges.

  • Addition of more links to https://js.tensorflow.org.

  • Fixing of non-uniformity of orthogonal matrices.

  • Fixing of bug where multi-image Estimator eval summaries were not displayed correctly.

Source And For More Information: GitHub