The original link: mp.weixin.qq.com/s/MKUM26_zW…
Since its open source release in 2015, TensorFlow has become the most widely used machine learning framework in the world, meeting a wide range of user and use case needs. During this time, TensorFlow has improved with rapid advances in computing hardware, machine learning research, and commercial deployment.
To reflect these rapid changes, the team plans to release a preview version of TensorFlow 2.0 later this year.
TensorFlow 2.0 will be an important milestone, with an emphasis on ease of use. Here’s what users expect from TensorFlow 2.0:
Eager Execution will be a core feature of 2.0. It better fine-tunes user expectations of the programming model through the TensorFlow practice, and should make TensorFlow easier to learn and apply. Support for more platforms and languages through standardized exchange formats and API consistency, and improve compatibility and parity between these components. Remove deprecated apis and reduce the number of duplicates, which can cause confusion for users.
Expose the 2.0 design process
In the near future, the team will hold a series of public design reviews covering planned changes. This process clarifies the functionality that will be part of TensorFlow 2.0 and allows the community to propose changes and ask questions. If you’d like to see comments and updates about the process, go to [email protected]. The team hopes to gather user feedback on planned changes after a preview release later this year.
Compatibility and continuity
TensorFlow 2.0 is an opportunity to correct errors and make improvements that are prohibited under the semantic version.
To simplify the transition, we will create a conversion tool that will update Python code to use the TensorFlow 2.0-compliant API, or issue a warning if such a conversion cannot be done automatically. Similar tools have made great contributions in the transition to version 1.0.
Not all changes can be completely automated. For example, apis, some of which have no direct equivalent, will be deprecated. In this case, the team will provide a compatibility module (tensorflow.pat.v1) that contains the full TensorFlow 1.x API and will be maintained throughout the life of TensorFlow 2.x.
It is not expected that there will be any further feature development on TensorFlow 1.x once the final version of TensorFlow 2.0 is released. Security patches will continue to be issued for TensorFlow 1.x for one year from the release date of TensorFlow 2.0.
Disk Compatibility
We probably won’t make major changes to SavedModels or stored GraphDef (we plan to include all current kernels in 2.0). However, the changes in 2.0 will mean that variable names in the original checkpoint may have to be transformed before being compatible with the new model.
tf.contrib
TensorFlow’s contrib module goes beyond maintainable and supported modules in a single repository. Larger projects would be better maintained individually, while the team would incubate smaller extensions with the TensorFlow master code. Therefore, as part of the release of TensorFlow 2.0, publishing tf.contrib will be stopped. The team will work with the respective owners over the next few months to develop a detailed migration plan, including how to publish your TensorFlow extension on community pages and documentation. For each contrib module, the team will:
Integrate the project into TensorFlow and move it to a separate repository to remove it completely
This means that all TF.contrib projects will be deprecated in the future and that adding new tF.contrib projects will stop today. The team is looking for owners/maintainers of some projects currently in tf.contrib. If you are interested, please contact us.
The next step
For questions about development or migration to TensorFlow 2.0, please contact our team by email at [email protected]. To stay up-to-date on the latest details of 2.0 development, subscribe at [email protected] and participate in the relevant design review.
In addition to the above communication methods, we have created a special post “TensorFlow 2.0 is coming!” in the “Technical QUESTIONS” section of tensorflowers.cn. Please leave a comment in the comments below and we will help you communicate with the TensorFlow development team.
Welcome to TensorFlow on Google’s official wechat account!