After the recent release to TensorFlow 1.10.0 release candidate family, rc-0, the new release candidate rc-1 is out and available.
Key highlights of this new version include major features and improvements to model training and evaluation, along with lots of bug fixes to the existing ecosystem.
What’s new in TensorFlow 1.10.0 RC1?
- The tf.lite runtime module now supports complex64 type
- Bigtable is a high-performance storage system which can help you store and serve training data. This new version will support the initial bigtable integration for tf.data
- With improved local run behavior in tf.estimator.train_and_evaluate function, there is no need to reload checkpoints for evaluation
- Now you can restrict the way workers and PS interact by setting device_filters in RunConfig class. Thus speeding up the training process and ensuring clean shutdowns in specific situations. However, if you want the workers and PS to communicate in order to complete the jobs, you will have to set customized session_options in RunConfig class.
Feature additions and improvements
- Now you can find Distributions and Bijectors in TensorFlow Probability, which was initially found at tf.contrib.distributions. By the end of 2018 tf.contrib.distributions will be removed.
- New endpoints are added for existing TensorFlow symbols. Going forward these new endpoints are expected to be the preferred endpoints and may replace some of the existing endpoints in the future. You can find the new symbols added to the following modules: tf.debugging, tf.dtypes, tf.image, tf.io, tf.linalg, tf.manip, tf.math, tf.quantization, tf.strings.
Breaking changes done to the ecosystem
- All the new prebuilt libraries are built against NCCL 2.2. They no longer include NCCL in the binary install. If you want to bring the complete usage of TensorFlow with multiple GPUs and NCCL you will need to upgrade it to NCCL 2.2. You can find the updated installation guide on Installing TensorFlow on Ubuntu and Install TensorFlow from Sources.
- From TensorFlow 1.11 release onwards, Windows builds will use Bazel. Hence this change will drop the official support for cmake.
To get full details on the features list and bug fixes done in this release candidate, you can check out Tensorflow’s official release page on Github.