Learnings from TensorFlow Dev Summit 2019

I attended TensorFlow Dev Summit on March 6 & 7 at Google event center in Sunnyvale. It was an incredible experience full of exciting new announcements and learnings. And I met many people from the TensorFlow community around the world.

Day 0

The Machine Learning GDEs (Google Developer Experts) gathered for dinner the evening before the Summit. We have been helping with testing and providing feedback on TensorFlow 2.0 preview. This was the first time for many of us to meet in person.

Day 1

It was a rainy day in Sunnyvale but many people tweeted photos of rainbows. We were happily waiting in the rain for check-in, and were very excited about the Summit.

Even though I knew the recordings will be made available later, I still took notes during talks because it helps me to learn. I also love taking photos with people I met so that I can remember them. There were so many new announcements and people to meet at the Summit, I could barely keep up.

TensorFlow 2.0 Alpha

(link to video | link to documentation)

One of the biggest announcements from the Summit was TensorFlow 2.0 Alpha. Give it a try:

pip install -U --pre tensorflow

Recently I also shared my notes on TF 2.0 from trying out the preview.

TensorFlow 2.0 is great for both beginners and experts; for both researchers and developers. It focuses on usability, clarity and flexibility:

• tf.Keras as the high level API and eager execution as default
• declutters the duplicated APIs for consistency
• still offers flexibility for those who need low level APIs

During breaks I got to hang out with the TensorFlow team and community.

It was great honor to meet Francois Chollet, creator of Keras. I got a new Keras sticker from him and I enjoyed chatting with him about Keras.

TensorFlow Lite

(link to video | link to roadmap)

There are a lot of interests in on-device machine learning (Pete Warden explained in 2018 “Why The Future of ML is Tiny”). And TensorFlow Lite is super important in enabling that.

Saw a demo of teachable machine built with Coral dev board, first product featuring edge TPU. Pete Warden gave a demo using Sparkfun’s microcontroller.

Lots of changes planned for 2019 focusing on the four areas:

• Usability — get it working end to end.
• Performance — fast execution: edge TPU delegate, GPU delegate.
• Optimization — smaller/faster models with quantization & pruning et.
• Documentation — better docs, tutorials and examples.

TensorFlow Datasets

(link to video | link to documentation)

The ML partnership of data + training models and play nice with each other. Load datasets with a few lines of code. There are currently 30 datasets available and more are being added. Make your data famous by adding your own dataset.

Swift for TensorFlow

(link to video | link to repo)

I’m very excited about Swift for TensorFlow, the next generation ML framework. Keras makes training easier but Swift is taking it further:

• Deploying dynamic models
• Application integrated neural nets
• Flexible and extensible autodiff
• Improved developer workflows

We saw a walkthrough of a image classification code which looks a lot like Keras. Swift for TensorFlow has seamless Python Interoperability. The great integration between Swift for TensorFlow and Python for TensorFlow should enable easier transition to Swift for TensorFlow.

TensorFlow Extended (TFX)

(Link to videos: TFX overview & Pre-training Workflow | Post-training Workflow)

TFX is an end-to-end ML platform and it now integrates with AirFlow and Kubeflow.

TensorFlow.js 1.0

(link to video)

TensorFlow Javascript is a library for training and deploying ML models in the browser:

• Pretrained models — Image, Audio and Text classifications (NPM or hosted scripts)
• Convert existing python models: with a command line tool or through Saved Model, TFHub or Keras.
• Train in the browser and Node.js — with Layers API
• Deploy on Platform

New learning resources

New courses were announced at the Summit from Udacity, Coursera and Fast.ai. Before the Summit MIT also made a new Deep Learning course with TensorFlow 2.0.

• Intro to TensorFlow for Deep Learning on Udacity

• New specialization TensorFlow: From Basics to Master (by Deeplearning.ai on Coursera). Course 1 Intro to TensorFlow for AI, ML & Deep Learning (link) is now available. Taught by Laurence Moroney from Google Brain team.

• Fast.ai announced that they will embrace Swift for deep learning.

Other announcements

• TensorFlow world, a 4-day conference by O’Reilly in Santa Clara in October, with hands-on training and talks. CFP is open till 4/23.
• Powered by TensorFlow 2.0 Challenge hosted on DevPost, for people to showcase what they have created with TensorFlow 2.0.
• Google Summer of Code, students get to work on TensorFlow while getting paid over the summer. They will be mentored by experts from the TensorFlow team and community.

Day 2

I went to a few more talks on the second day of the Summit.

TensorFlow On-Device: Compressing, Quantizing, and Distributing

This was a popular session. The room was so crowded that I managed to find a corner to stand and even the overflow room was filled.

There were 2 talks: first one was “TensorFlow Lite on Android” — about how to get started with TensorFlow Lite. Second talk was “TensorFlow lite: Quantization”. After that there was Q&As with the TensorFlow Lite team.

Research and Future talks:

• Federated Learning —ML training on decentralized data. (link to video | link to documentation)
• Mesh TensorFlow — for large-scaled models. (link to video)
• Sonnet (by DeepMind) — for complex models by researchers. (link to video)

Global docs sprint planning

Before I headed over to the airport, Paige and I chatted about planning for a global docs sprint to help improve the documentation of TensorFlow.

You can watch all session recordings from the Summit here and see the photo album here.

I summarized the announcements and learnings from the Summit in sketchnotes. Trying to fit my notes from the Summit on a single piece of paper really helps me to visualize the key take-aways.