TensorFlow – the robust, open-source machine learning Software library, is used for numerical computation using data flow graphs that consist of nodes, which in turn represent operation related to mathematics and on the contrary, the graph edges represent Multidimensional arrays called Tensors that are communicated between them.

The open-source deep learning library comes with a flexible architecture that allows its users to deploy computation to one or more GPUs or CPUs in a desktop, mobile device or server, with a single API(Application Programming Interface).

TensorFlow is licensed under the Apache open source software license, That simply means that you can use it, modify it, and even redistribute the modified version of it. Therefore, To use TensorFlow, You don't have to pay any licensing fee.

The use of TensorFlow has also helped researchers, engineers, artists, students, and many others to make progress with everything, may it be from language translation to early detection of skin cancer or even preventing blindness in diabetics for that matter.

A few of the main features that make it the only one among so many options to the developers are:

• Robustness

• Ease Of Use

• Flexible and Thoughtful Architecture.

• Open-Source

There have also been many releases in Tensorflow a few of which include: Tensorflow 1.0, Tensorflow 1.2, Tensorflow 1.2 in the list.

To add To it again, Here we have the release Tensorflow 1.4.0!

This very new Version of Tensorflow- (Tensorflow 1.4.0) comes with some major features and improvements a few of which are mentioned below:

• tf.keras, tf.data(API now subject to backward compatibility guarantees)- now a part of core Tensorflow API.

• Usability Improvement: The frequently used tensor filters are now added to hooks and Session wrappers by default(has_inf_or_nan). That results in no need for clients to call .add_tensor_filter(tf_debug.has_inf_or_nan) anymore.

• Major new features include: Addition of Dataset.apply() method that is used for applying custom transformation functions and Dataset.from_generator() that is used for building an input pipeline from a Python generator.

• Addition of several custom transformation functions like -tf.contrib.data.sloppy_interleave()

• Extending the (tf.contrib.distributions.bijector) API for the handling of a few non-injective transforms.

• Java: Generics for improved type safety, Support Multidimensional String Tensors.

A few among the other changes and bug fixes in the new version also include:

• Resolved race condition in TensorForest TreePredictionsV4Op.

• Google Amazon S3 file system, Cloud Storage file system, and Hadoop file system support are from now default build options.

• Change RunConfig default behavior to not set a random seed, in turn, making random behavior independently random on distributed workers. Thus, returning an improvement in training performance.

• Models should set a random seed explicitly(that do rely on determinism).

For More Information: GitHub