Many engineers and researchers from deep learning field must be familiar with TensorFlow platform. Google introduced TensorFlow, an open source system for machine learning in 2015 and yesterday (Feb 15, 2017) at Google’s inaugural TensorFlow Dev Summit in Mountain View, California, they announced a new release of TensorFlow, version 1.0. Before jumping to TensorFlow 1.0 release update, let’s just catch up with what is TensorFlow!
What is TensorFlow?
TensorFlow is an open source machine learning library which was initially developed by engineers and researchers at Google Brain Team. Google made it open source in 2015. TensorFlow library can ease the process of applying deep learning algorithms into various areas. Graphical representation of mathematical operations simplifies the deep learning algorithms implementation. It is based on data flow graphs where graph node represents mathematical computations, while the graph edges represent the multidimensional data arrays (which they call as tensors) communicated between them.
What do you need to know before getting started with TensorFlow?
- How to code in Python
- Mathematical operations at least about Arrays
- Some knowledge of machine learning
About TensorFlow 1.0
Google claimed that TensorFlow 1.0 is now more production ready than ever, faster than earlier, more flexible and with more stable Python APIs. Many tools have been added to the framework which makes TensorFlow much better.
Faster than earlier: XLA (Accelerated Linear Algebra) now lays a foundation to the optimized TensorFlow v1.0. XLA is domain-specific compiler to speed up TensorFlow computations. It is in an experimental phase (alpha) and one can play with it via Just-in-time (JIT) compilation or ahead-of-time (AOT) compilation. XLA not only speed up the computations but also improves memory usage. Check out XLA details here.
More flexibility: Addition of new APIs and modules like tf.layers, tf.metrics, and tf.losses modules makes TensorFlow more flexible. TensorFlow 1.0 now also compatible with Keras which is another neural network library written in Python.
Python API changes: Changes in major Python APIs which now resembles with NumPy (fundamental scientific Python package for computations) are not backword-compatible. Older programs might not run on TensorFlow 1.0 as expected so this migration guide may come handy to upgrade your programs to V1.0
Java and Go support: TensorFlow 1.0 now adds APIs for Java and Go programming languages (experimental as of now).
TensorFlow Debugger: (tfdbg): TensorFlow 1.0 introduces new command line debugger so develoeprs can now debug live TensorFlow programs.
New Android demos: New demos are created with TensorFlow 1.0 release for Android which showcase object detection and localization, camera-based image stylization.
Installation improvements: TensorFlow instllation is made easy with Python image docker and PyPI compliant pip packages. Check out TensorFlow instllation details here..
Find out more about TensorFlow 1.0 in release notes.