TensorFlow is a popular and leading open-source framework for developing machine learning and deep learning applications. Developed and pioneered by Google, TensorFlow is a flexible and ever-changing framework favored by deep learning industry professionals and experts.
Each year, the team behind TensorFlow hosts a dev summit event that comprises two days filled with technical updates from the TensorFlow team and presentations from users showcasing amazing applications they’ve built using TensorFlow. There are also hacker-rooms, breakout sessions, and workshops.
This year, the TensorFlow Dev Summit 2020 had a different flavor to it. With the outbreak of the coronavirus in many countries, the TensorFlow team decided to prioritize the health and safety of their attendees (and rightly so!). They changed the dev summit to the first-ever live stream on YouTube and provided recordings for those who missed it.
I am thrilled to present the top highlights from the TensorFlow Dev Summit 2020 in this article. I have included the video for each talk as well so you can watch it in its entirety!
You can also check out Analytics Vidhya’s TensorFlow tutorials here:
- Introduction to Implementing Neural Networks using TensorFlow
- TensorFlow Serving: Deploying Deep Learning Models Just Got Easier!
- Build a Machine Learning Model in your Browser using TensorFlow.js and Python
Alright – let’s dive into the top sessions!
1. TensorFlow Keynote
Megan Kacholia, technical program manager for TensorFlow, kicked-off the summit with the story of Dr. Erwin, a self-proclaimed AI enthusiast and a radiologist in the Philippines. He has built a deep learning application using Tensorflow.js that can classify bone fracture images.
Since then, Dr. Erwin has been giving talks about how TensorFlow could potentially change the medical industry and regularly invites enthusiasts to come and build such systems.
Here are a few other projects built using TensorFlow:
- Disaster Watch: A crisis mapping platform that aggregates data and predicts physical constraints caused by a natural disaster
- Deep Pavlov: NLP library for a dialogue system
Megan then introduced a new update – Tensorflow 2.2. Below are few of its features:
- This update is a new baseline for measuring performance in a structured way, i.e improving speed and performance
- TensorFlow Ecosystem (main talking point of this TensorFlow Dev Summit 2020): For those who had trouble migrating to TensorFlow 2.0, this ecosystem provides them with their favorite libraries and model to work with 2.x
- Improved stability
Key takeaways and major features of TensorFlow from this keynote:
- tensorflow datasets: The TensorFlow datasets are out which are compiled by 2019 Google Summer of Code students
- tensorflow.dev: New toolkit to share your TensorBoard results with a sharable link
- Addons and extensions to the TensorFlow ecosystem
- Google Cloud AI Pipeline: Released for building end-to-end production pipeline with Kubeflow and Tensorflow Extended
- Performance profiling tool in TensorFlow 2.x
- TensorFlow 2.1 supports Cloud TPUs
- TensorFlow Runtime (TFRT): We won’t be exposed toTFRT as a developer or a researcher but it will be working under the hood to give the best performance possible for your deep learning model
Next in the TensorFlow Dev Summit 2020 keynote, we had a talk by Mansi Joshi, Engineering Director in the TensorFlow team where she introduced the TensorFlow system for Responsible AI. She stated what Responsible AI is and how the TensorFlow ecosystem can help build such systems:
Finally, Kemal El Moujahid, product director in the TensorFlow team, took the stage. He introduced and explained various resources and opportunities to connect more with the TensorFlow team using TF user groups, TF SIGs (Significant Interest Groups) and other options:
You can watch the full keynote of the TensorFlow Dev Summit 2020 here:
2. TensorFlow Hub: Making Model Discovery Easy
Speaker: Sandeep Gupta
TensorFlow Hub is the place to easily find the latest ready-to-use deep learning TensorFlow models with documentation, code snippets and much more. TensorFlow Hub’s rich repository of models covers a wide range of deep learning tasks, like:
- For Image: Models for image classification, object detection, image augmentation, etc.
- For Text: Models for text classification, state-of-the-art models like BERT
- For Video: Models for action recognition, gestures
- For Audio: Models for pitch detection
More than 1000 models are available with documentation and code snippets. You’ll find interactive Google colab notebooks on this link.
What’s New in TensorFlow Hub:
- Improved search and discovery
- Expanded support for TF formats including in tensorflow.js and tensflow.lite
- New improved state-of-art models for various streams
Watch the full talk on TensorFlow Hub here:
3. Collaborative ML with TensorBoard.dev
Speaker: Gal Oshri
TensorBoard is a TensorFlow visualization toolkit that is commonly used by deep learning researchers and engineers to understand their experiment results. TensorBoard lets us track metrics, visualize our deep learning model, and explore parameters among other things.
But there was a limitation – these results could only be shared in a picture format for others to review or correct. The actual implementation where the errors could be found more easily could not be shared.
TensorBoard.dev makes this task easier. With tensorflow.dev, we can upload our TensorBoard results and get a link that we can share with everyone for free! Others can easily view and interact with our TensorBoard to compare their performance or rectify our mistakes. Here is the link to get started.
Here’s the talk on TensorBoard.dev at the TensorFlow Dev Summit 2020:
4. Performance Profiling in TensorFlow 2.x
Speaker: Qiumin Xu
Performance Profiling tool is finally released in Tensorflow 2.x! This tool helps to improve our deep learning model’s performance like a professional player. What performance profiling does is it produces automated performance guidance and suggestions for improving the model performance and thereby increasing the productivity of performance engineers.
8 new tools have been released, 4 of which are common to CPU, GPU, and TPU. Some of them include:
- Overview page: This tool provides an overview of the performance of workload running on our device
- I/O Input pipeline analyzer: A most important tool to determine if there is any bottleneck present in the model or not and provides suggestions and resources for removing this bottleneck
- TensorFlow Stats: Present statistics in a chart or tabular format
I’m sure you’ll be excited as I am to try performance profiling in TensorFlow 2.0! You can do that on this link.
And don’t forget to watch the full talk on Performance Profiling in TensorFlow 2.x:
5. MLIR: Accelerating TensorFlow with Compilers
Speaker: Jacques Pienaar
Machine Learning models each day are increasing in complexity and size. So this requires an increase in computational requirements in training these models. System and hardware must rapidly adapt to more complex deep learning algorithms while supporting a wide variety of deployment scenarios.
So here is MLIR (Multi-Level Intermediate Representation). It is a compiler framework and intermediate representation for TensorFlow. Below are its key features:
- State-of-the-art compiler technology
- Modular and extensible (can easily be modified for your own model and hardware)
- Not opinionated (create a solution for your own problem space)
- Fully customizable
To start contributing to MLIR, follow this link.
You can view the talk here:
6.TFRT: A New TensorFlow Runtime
Speaker: Mingsheng Hong
Runtime is a low-level component that orchestrates all model execution by calling into relevant kernels that implement machine learning primitives like matrix multiplication. TFRT is introduced to replace the existing TensorFlow runtime for faster and bigger models and enhance research innovations.
Below are its key features:
- Focus is on performance, extensibility, and unification
- Improved error reporting
- Improved performance and reduced CPU usage
- Unified Training across multiple hardware.
It is still under production and the TensorFlow team plans to integrate TFRT with the TensorFlow stack within a year. This will largely improve performance and reduce hardware usage simultaneously.
Below is the link to the full talk:
7. TensorFlow Lite: ML for Mobile and IoT Devices
Tensorflow Lite is a production-ready, cross-platform framework for deploying machine learning and deep learning models on mobile devices and embedded systems. Since it’s launch in 2017, TensorFlow lite is now on more than 4 billion mobile devices globally.
Below are some of the key points from the talk that introduced many features in the ever-emerging TensorFlow lite library:
- TF lite support library has extended just beyond processing and transforming data to enable easy on-device deployment
- More Image And Language API introduced
- Android Studio integration (by far the most important integration into this TF lite). This will ensure single drag and drop into Android studio and then automatically generate Java classes for TF lite model
- Extended Model Metadata: This tool will allow us to easily generate model metadata which provides additional information on what the model does, expected input format for model and what does the output signify
- TFLite Model Maker: A new Python library that lets us customize models for our dataset, without requiring machine learning
- Core ML delegate for Apple devices to accelerate floating-point operations on iOS devices through the Apple neural engine
If you still haven’t tried this amazing TensorFlow feature, you can get started here.
Here’s the full session:
8. TensorFlow.js: Machine Learning for the Web and Beyond
Speaker: Na Li
- More models including body segmentation, toxicity detection and sound recognition
- Converter Wizard, an interactive command-line tool, to convert deep learning models to the Tensorflow.js format effortlessly
Watch the full video below:
Analytics Vidhya’s Take on TensorFlow Dev Summit 2020
Here are my takeaways from TensorFlow Dev Summit 2020:
- There are a lot of exciting libraries introduced in TensorFlow, including tensorboard.dev and TFRT
- Many new features are introduced in existing libraries like TensorFlow Lite and TensorFlow.js
- Major focus is on improving the speed and performance of our model. Performance profiling will heavily increase the productivity of performance engineers across the globe
- A lot of work is being done on improving the scalability of deep learning models and implementing machine learning models in Android devices and Arduino with much less latency and network issues
Let me know your pick from TensorFlow Dev Summit 2020 in the comments section below!You can also read this article on our Mobile APP