What's new in TensorFlow.js

[Yannick Assogba]

Hello, We wanted to share some updates of what has been going on recently with the TensorFlow.js!

What’s new in TensorFlow.js

tfjs v0.13 has landed! Check out the release notes here, a few highlights include:

• MobileNet has been sped up by 2x due to texture packing.

• Complex64 dtype added which paves the way for things like our new speech commands model!

• Added tf.profile() to profile memory characteristics of functions.

• Precision and recall metrics added to layers API.

We have also been working on a whole host of features over the past several months including:

• Operations now accept JavaScript arrays as well as tensors. You can now do tf.square([1, 2, 3])

• Precision improvements for iOS and desktop Safari.

• tfjs-node now has GPU support on Windows

• New Ops including:  tf.image.CropAndResize, tf.image.nonMaxSupression, tf.where, tf.spaceToBatchND, tf.batchToSpaceND, tf.topK and support for TensorArray* in the converter.

We’ve also added a gallery to collect demos, projects and tutorials that are being created by the TensorFlow.js community.

New Visualization Tools

We have just released tfjs-vis. A tool for visualization of browser based TensorFlow.js machine learning. tfjs-vis provides:

• A hideable drawer called the visor that provides a place to render visualizations in a way that doesn’t interfere with the app you are developing.

• A collection of charts for flexible rendering of data from machine learning models.

• A collection of high level methods to visualize tensorflow.js concepts such as models and layers.

• Utility functions for evaluating machine learning models such as per-class-accuracy and confusion matrices.

Check it out! We also have a few demos to show what you can do with it currently:

• https://storage.googleapis.com/tfjs-vis/mnist/dist/index.html

• https://storage.googleapis.com/tfjs-vis/mnist_internals/dist/index.html

New Models

Coco-ssd is an object detection model that identifies multiple objects in a single image.

The speech-commands model enables recognition of spoken commands of isolated English words from a small vocabulary. The default vocabulary includes the following words: the ten digits from "zero" to "nine", "up", "down", "left", "right", "go", "stop", "yes", "no", as well as the additional categories of "unknown word" and "background noise". It is easy to extend this model to custom speech commands using transfer learning.

New Guides

We have a few new guides on our tutorials and guides page including: How to get started with X: a guide for TensorFlow.js Users and How to define a custom WebGL operation.

Talks

• This SFHTML talk by Nick Kreeger further introduces using TensorFlow.js on Node

• Ping Yu talks about implementing Audio Models in Tensorflow.js

• Still haven’t tried out TensorFlow.js? This overview talk from Google I/O will give you a sense of what you can do with Tensorflow.js