What Is New In TensorFlow 2.0
Machine Learning is one of the most in-demand technologies and one of the fastest-growing frameworks for machine learning is TensorFlow. TensorFlow is Google’s open-source machine learning platform that is most widely adopted by machine learning developers worldwide.
Today, at the TensorFlow Developer Summit, the TensorFlow team announced the updates and roadmap of the product that includes availability of Tensor 2.0 alpha, TensorFlow.js 1.0, TensorFlow Lite and TensorFlow Extended. The TensorFlow 2.0 release is now available as a developer preview.
TensorFlow 2.0 Alpha is available now.
According to the presentation, the TensorFlow has been downloaded more than 41 million times, 50,000 commits, 9,900 pull requests, and has 1,800 contributors.
TensorFlow 2.0 is easy, more powerful, and scalable. There are easier APIs with better code examples and documentation.
TensorFlow can be deployed anywhere including servers, desktops, in the cloud, mobile, edge devices, and in browsers with the help of three versions of TensorFlow, TensorFlow Extended, TensorFlow Lite, and TensorFlow.js.
TensorFlow Extended is an integrated end to end platform for managing every stage of the machine learning life cycle.
TensorFlow Lite 1.0 is the solution for running models on mobile and IoT devices.
TensorFlow.js is a JavaScript library for training and deploying machine learning models in Web browsers, Node, and other platforms. TensorFlow.js already has been downloaded 300,000 times.
Watch the following video to learn more about TensorFlow 2.0 and its upcoming new products and features.
According to the official announcements, the TensorFlow 2.0 Roadmap includes the following :
APIs
- Mixed precision training API in Keras.
- Premade estimators for boosted trees, random forest, approximate nearest-neighbor search, k-means clustering, and more.
- tf.distribute.Strategy support for Keras subclass models, TPUs, and multi-node training
- Improved support for model saving and loading SavedModel, and conversion of existing 1.x TF-Hub modules
Reference models
- Updated model repository with TF 2.0 best-practice reference models and research model implementations. These will include ResNet, MobileNet, DenseNet, Mask-RCNN, NMT, NCF, Transformer, and many other models
- Collection of TF Hub modules, loadable in TensorFlow 2.0.
- More performance optimizations.
TensorBoard
- General bug fixes and enhancements to make TensorBoard great with TensorFlow 2.0.
- Improvements to hyperparameter tuning capabilities and workflow.
- Enable hosted TensorBoard to make sharing dashboards easy and search/compare experiments.
- Create a better plugin process for testing and adding TensorBoard plugins.
- Enable plugins to use different frontend technologies (like React).
TensorFlow Lite
- Increase coverage of supported ops in TensorFlow Lite.
- Easier conversion of TensorFlow 2.0 models to use in TensorFlow Lite.
- Extended support for Edge TPUs, TPU AIY boards.
- More documentation and tutorials.
TensorFlow.js
- Continued browser performance improvements.
- Implement prototype using compute shaders or WebGPU.
- Improve CPU performance, implement SIMD+ Web Assembly support (when available).
- More off-the-shelf models for image, audio, text models, and more.
- Improve parity in features and performance on Node with core TensorFlow.
- Support for TensorBoard visualization in Node training with TensorFlow.js.
- Integration with more JavaScript platforms.
- More documentation and getting started content.
TensorFlow with Swift
- Focus on researchers for the first half of 2019.
- Significant new feature development, such as support for transfer learning.
- Polish existing features, such as control flow support by the AutoDiff system.
- Deeper integrations with the TensorFlow ecosystem, such as TensorBoard.
- End-to-end tutorials for getting started on Swift for TensorFlow in Colab and additional technical documentation.
- Collaboration with fast.ai
TensorFlow Extended (TFX)
- Integration of all TFX components with TensorFlow 2.0.
- Fully orchestrated end-to-end workflows with common configuration and metadata tracking.
- Advanced training features, such as warm-starting.
- Notebook embedded visualizations and experiment tracking