First impressions of TensorFlow Dev Summit, 2019
The 2019 edition of the TensorFlow dev summit got off to a great start on a rather cold and rainy morning in Sunnyvale, CA. This time around, the TensorFlow team has made certain visible changes from previous year’s edition:
- Summit is now a 2-day event with the first day full of talks and demos whereas the second day focussing on hands-on sessions where attendees can code along with the TensorFlow team members and engage them on the real-life problems they are solving.
- The event has a much larger invitee list and this is visible from their venue (Google Event Center)
- Talks are more technical with a lot of code snippets and live demos using Google Colab.
TensorFlow is our machine learning library of choice at IdeaChain where we use it to build sophisticated language models that understand and interpret your ideas.
Let’s go over the 15 key takeaways from a full day of talks and the hands-on sessions. To view all the talks, TensorFlow team has been fairly quick this time to put up the videos on their youtube channel.
15 Key Takeaways
- TensorFlow 2.0 is clearly the theme for this edition of the summit. All the talks from TensorFlow.js to TensorFlow lite emphasized their support for the new version.
- The 2.0 version of this ever powerful tool is a major release focussing on three user-centric areas: Usability, Clarity, and Flexibility
- With a focus on tf.keras as the preferred high-level API and eager execution now coming as default, the TensorFlow team has tried making their offering more accessible to users who were not familiar with sessions and graphs that formed the core of TensorFlow 1.x
- Graphs still exist but with eager execution becoming the default, sess.run is deprecated and will be replaced by the tf.function capability.
- The TensorFlow team has promised more intuitive syntax, removal of duplicate functionality and full lower-level API support to make the 2.0 version more clear and flexible for the user.
- For all those people looking to upgrade to TensorFlow 2.0 (at least for inference), your saved_models generated using version 1.x will load and work in 2.0. For upgrade script from TF 1.x to 2.0, refer here.
8. The announcement of TensorFlow World, a conference where engineers, innovators, executives, and product managers can come and discuss their product/service offering that has been powered by TensorFlow.
9. One of my favorite contribution to the TensorFlow ecosystem is TensorFlow Extended (TFX) and the TFX team has certainly delivered what it promised in the 2018 edition. All the various components in TFX (DataValidator, Trainer, ModelValidator, Pusher) now work together for an end-to-end ML offering. Bonus: TFX now integrates with open source orchestrators such as AirFlow and KubeFlow.
10. TensorFlow has stepped into the hardware space with the launch of the Raspberry pi style Coral DevBoard powered by the edge TPU ML accelerator. Priced at 150$, it is more expensive than a Raspberry Pi but cheaper than an Nvidia Jetson.
11. TensorFlow lite team focussed their talk on speaking about their expanding list of use-cases, from Google assistant to YouDao’s on-device translation service. Keeping up with the usability theme of TensorFlow 2.0, TensorFlow lite has focussed on reducing the footprint of the models and making inference faster. Documentation for TensorFlow lite has also been improved.
12. TensorFlow is now supported by the Julia programming language (tensorflow.jl). To give you a hint of why you might prefer Julia over Python for your next project, have a look at the code syntax and runtime for the below piece of code.
13. Ever considered doing machine learning on decentralized data, checkout TensorFlow Federated.
14. Sonnet, a high-level library built by DeepMind on top of TensorFlow, announced it’s support for TF2.0.
15. Are you developing models that are too big to fit on an off-the-shelf cloud instance? Need model parallelism? Checkout Mesh-TensorFlow.
I am personally excited to try TensorFlow 2.0 and see how it improves our machine learning performance at IdeaChain. With tf.Keras, I can expect our codebase to become more readable and have more flexibility in debugging using the eager execution mode.
There were a lot more updates and examples of how TensorFlow is being used in the industry and research. The above key takeaways capture just a glimpse of what was covered in the summit. To go over all the content, refer to TensorFlow’s YouTube channel. Consider applying to TensorFlow Dev Summit 2020 if you would like to meet some really smart people applying machine learning to real-life use-cases.
About the author: Gaurav is a data science manager at EY’s Innovation Advisory in Dublin, Ireland and Chief Data Scientist at Ideachain. Find more at gauravkaila.com