TensorFlow 1.0 Released
Google recently announced TensorFlow version 1.0. Python API is now stable and experimental APIs for Java and Go have been added. XLA, the new domain-specific compiler, delivers 7.3x performance increase in 8 GPUs setup and 58x for 64 GPUs for Inception-v3 neural network model. A new high level API can help with constructing convoluted neural networks, compute Evaluation-related metrics and loss functions operations. Keras can also be integrated with TensorFlow using a build-in module. Keras is a high-level Python neural networks library aiming to abstract deep learning for fast experimentation.
Shortly after the announcement, Google also announced tf.transform, a library for data preprocessing with TensorFlow. Based on Apache Beam, tf.transform can help avoiding 'training-serving skew', the problem having data in production differ from data used to train the underlying model.
In addition to these improvements, a command line debugger has been added, Python 3 docker images and easier installation via pip package management. A side effect of these improvements is that there are some backwards-incompatible changes which can be addressed via the migration guide and the conversion script.
TensorFlow, in a little over a year, is already used in over 6000 open-source repositories in GitHub. More information is available in the videos from TensorFlow Developer Summit, covering recent updates and interesting use cases.