With just 15 days from TensorFlow’s major release of 1.7.0, the TensorFlow Community is highly energetic and enthusiast to distribute its 1.8.0 in coming days. The team has announced the beta release 1.8.0-rc0 with numerous exciting features and bug fixes. This newer version has paid more attention towards supporting GPU memory, running on multiple GPUs and cloud performance.
Major features and improvements in TensorFlow 1.8.0-rc0:
- Adds Gradient Boosted Trees for pre-made Estimators with BoostedTreesClassifier and BoostedTreesRegressor.
- Adds 3rd generation pipeline config for Cloud TPUs for performance improvement and usability.
- Support for running Estimator model on multiple GPUs by passing tf.contrib.distribute.MirroredStrategy() to tf.estimator.RunConfig().
- Support for prefetching GPU memory using tf.contrib.data.prefetch_to_device().
- Moving Bayesian computation tf.contrib.bayesflow to its own dedicated repository.
- Allows generic proto parsing and RPC communication with tf.contrib.{proto,rpc}.
Bug Fixes in TensorFlow 1.8.0-rc0:
- Enabled support for prefetching of dataset elements to GPU memory with tf.contrib.data.prefetch_to_device.
- Allows automatic tuning of prefetch buffer sizes with tf.contrib.data.AUTOTUNE.
- Added support for building datasets of CSV files with tf.contrib.data.make_csv_dataset.
- Provision of creating iterators with Both Dataset.__iter__() and Dataset.make_one_shot_iterator() in eager execution mode.
- Enabled automatic device placement.
- tf.GradientTape voluntarily moved out of future contributions to the library.
- Added new data preprocessing functions and fashion mnist dataset to tf.keras.
- Accelerated Linear Algebra (XLA) with lexicographical feature.
- Allows exclusion of nodes in tensor-filter operations.
- Fixed spurious background colors in text terminals.
- Fixed batch dimensions reshaping with BatchReshape.
- Support for explicit gradient checkpointing on TPU with tf.contrib.layers.recompute_grad works.
For the complete list of bug fixes and improvements, you can read TensorFlow’s Github page.
Miscellaneous changes:
- Easy calling to TensorFlow C API.
- Description of shapes and pointer noted in tf.distributions.Distribution tutorial notebook.
- Scatter operations extended and updated with tf.scatter_min and tf.scatter_max
- cuDNN RNN operations moved to TensorFlow codebase.
- Added float64 support for Conv2d, Conv2dBackpropInput, Conv2dBackpropFilter and AvgPool/AvgPoolGrad.
- Localised graph name scope thread for multi-threaded environments.
- Updated nsync synchronization library for avoiding slow primitives on Linux.
- Non-uniformity of orthogonal matrices fixed.
- Multi-image Estimator eval summaries displays correctly.
You can download the source code to access all the exciting features of TensorFlow 1.8.0-rc0.