The 5 biggest announcements from TensorFlow Developer Summit 2018

The second TensorFlow Developer Summit was filled with exciting product announcements and technical talks from the TensorFlow team and guest speakers. Here are 5 major features extended to the TensorFlow machine learning framework, announced at the Summit.

TensorFlow.js: Machine Learning brought to your browsers

Using TensorFlow.js, developers can now define, train, and run machine learning models entirely in the browser. This open-source library can be run using Javascript and a high-level layers API.

What does this mean from a developer’s perspective?

TensorFlow.js allows importing of an existing, pre-trained model, say a TensorFlow or Keras model into the TensorFlow.js format.

Developers can use transfer learning to re-train an imported model, using only a small amount of data.

What does this mean from a user’s perspective?

No need to install any libraries or drivers. Just open a webpage, and your program is ready to run.

TensorFlow.js automatically supports WebGL, so it will accelerate your code when a GPU is available.

With TensorFlow.js, users may also open their webpage from a mobile device, where the model will take advantage of sensor data from the mobile’s gyroscope or an accelerometer.

All the data stays on the client, making TensorFlow.js useful for privacy preserving and low-latency inference.

You can see TensorFlow.js in action by trying out the Emoji Scavenger Hunt game from a browser on your mobile phone.

TensorFlow Hub: A library for reusable Machine Learning modules in TensorFlow

The next major announcement at the TensorFlow Developer summit was the TensorFlow Hub. This platform is an aggregator to publish, discover, and reuse parts of machine learning modules in TensorFlow. Module here refers to a self-contained piece of a TensorFlow graph, along with its weights, that can be reused across other similar tasks. Model reusing helps a developer train a model using a smaller dataset, improve generalization, or speed up training. TensorFlow Hub comes with two tools that help in finding potential issues in neural networks. The first is a graphical debugger for inspecting the artificial neurons of an AI. The other visualize how well the model as a whole analyzes large amounts of data.

TensorFlow Model Analysis

TFMA is an open-source library that combines the power of TensorFlow and Apache Beam to compute and visualize evaluation metrics. TFMA ensures that ML models meet specific quality thresholds and behaves as expected for all relevant slices of data.

TFMA uses Apache Beam to do a full pass over the specified evaluation dataset. This allows more accurate calculation of metrics and also scales up to massive evaluation datasets.

TFMA allows developers to visualize model metrics over time in a time series graph. It visualizes metrics computed for a single model over multiple versions of the exported SavedModel.

TFMA uses Slicing metrics to analyze the performance of a model on a more granular level.

TensorFlow is now available in more languages and platforms

TensorFlow Developer Summit also brought a good news for swift programmers. As of April 2018, TensorFlow for Swift will be open sourced. TensorFlow for Swift is more than just language binding for TensorFlow. It integrates first-class compiler and language support, providing the full power of graphs with the usability of eager execution.

TensorFlow Lite, TensorFlow’s cross-platform solution for deploying trained ML models on mobile, also has major updates. It will now feature full support for Raspberry Pi and increased support for ops/models (including custom ops). The TensorFlow Lite core interpreter is now only 75 KB in size (vs 1.1 MB for TensorFlow) with speedups of up to 3x when running quantized image classification models.

New applications and domains opened using TensorFlow

TensorFlow Developer Summit also made announcements pertaining to sectors beyond the core deep learning and neural network models.

The TensorFlow Probability API provides state-of-the-art methods for Bayesian analysis. This library contains building blocks like probability distributions, sampling methods, and new metrics and losses.

They’ve also released Nucleus, a library for reading, writing, and filtering common genomics file formats for use in TensorFlow. This is released along with DeepVariant, an open-source TensorFlow based tool for genome variant discovery. Both these tools intend to help spur new research and advances in genomics.

The TensorFlow Developer Summit also showcased a new blog, YouTube channel, and other community resources.