Since the release of TensorFlow as an open-source in 2015, the machine learning framework has become the world’s most widely adopted framework, that caters to a spectrum of users and use-cases. Over time, TensorFlow along with rapid developments has evolved in Machine Learning research, commercial deployment, and computing hardware.
In order to reflect these rapid changes, the work on the next major version of TensorFlow has started!
TensorFlow 2.0, with its main focus on ease of use, will be a major milestone. Given Below are some highlights of what can be expected with the new version:
A central feature of 2.0 will be Eager execution. It aligns the expectations of the users about the programming model better with TensorFlow practice and is expected to make TensorFlow easier to learn and apply.
Support for more platforms and languages improved compatibility and via standardization on exchange formats and alignment of APIs, parity between these components.
Remove deprecated APIs and reduce the amount of duplication, which has caused confusion for users.
Public 2.0 design process:
Shortly, a series of public design reviews that will cover the planned changes will be held. This process will allow the community to propose changes and voice concerns and will also clarify the features that will be part of the new version.
Compatibility and continuity:
This new version, TensorFlow 2.0 is an opportunity to correct mistakes and make improvements which otherwise under semantic versioning are forbidden. And in order to ease the transition, there will be a creation of a conversion tool which updates code in Python in order to use APIs that are compatible with TensorFlow 2.0, or warns in cases where such a conversion is not automatically possible. During the transition to 1.0, a similar tool had helped tremendously.
It is to note that not all changes can be made fully automatically and once a final version of TensorFlow 2.0 is released there is no further anticipation of any feature development on TensorFlow 1.x. Though security patches issue for the last TensorFlow 1.x release will be issued for one year after the release date TensorFlow 2.0.
On-disk compatibility:
There is no intention of making any breaking changes to SavedModels or stored GraphDefs and by this, it is meant that the plan is to include all current kernels in 2.0. However, the changes in TensorFlow 2.0 will mean that variable names in raw checkpoints before being compatible with new models might have to be converted.
tf.contrib:
TensorFlow’s contrib module has grown beyond what in a single repository can be maintained and supported. It is better, that the Larger projects be maintained separately, while smaller extensions will be incubated along with the main TensorFlow code. And as part of releasing TensorFlow 2.0, the distribution of tf.contrib will come to a stop. In the coming months, the team plans to work on detailed migration plans, with the respective owners. For each of the contrib modules they plan either:
The Integration The Project Into Tensorflow;
Moving Of It To A Separate Repository
Or The Removal Of It Entirely.
Nevertheless, this does mean that there will be a depreciation of all of the tf.contrib, and a stop will be put in the addition of new tf.contrib projects.
Moving to its release date, The team plans to release a preview version of TensorFlow 2.0 later this year!
0 Comments