TensorFlow Dev Summit 2020 Rewind
As the 4 annual TensorFlow Developer, Summit 2020 happened in the month of March. Thus this article is an effort to make all the engineers, developers, computer scientists, and data scientists of this wonderful AI community aware of the recent updates and announcements which were made in that summit.
“We choose to go to the Moon”
Let me take you back to September 12, 1962, to the 35th president of the United States John F. Kennedy “We choose to go to the Moon ” speech at rice university. Kennedy said.“We choose to go to the Moon in this decade and do the other things, not because they are easy, but because they are hard.”
Well, I think the TensorFlow engineering team took this speech at their heart, and after receiving lots of feedback from across the community, they started working on the hard problem which was to make TensorFlow as easy as possible to everyone in the world who basically knows how a computer works or to those who can leverage the superpowers of TensorFlow like platform and make the world a better place to live in.
Well, in that front TensorFlow Team achieved a major milestone last year by launching TensorFlow 2.0 which allowed them to move an inch more closer towards their goal of making an easy to use end to end “One in all” platform.
Building Momentum:
TensorFlow 2.0 was launched with features such as :
- Simplifies API’s
- Easy Model Building with Keras and Eager execution by default.
- Enabling Model Deployment to more devices(such as Raspberry Pi).
- Powerful experimentation for research(such as Tensorboard.dev).
Therefore by building on the momentum from 2.0 version released last year, the TensorFlow team announced the latest version TensorFlow 2.2 with enhancements such as :
- Improved Performance from the last version.
- Compatibility with the rest of the TF Ecosystem.
Thus, all those users who have faced trouble migrating to 2.0, worry not because the TF team has now made the entire system compatible with your favorite libraries. In brief what works with 1.x works with 2.x too.
3. Lastly, any software is judged on the bases of scalability and stability. Apparently, the TF team committed the core library which confirms that they won’t be making any major changes anytime soon. Thus More stability is given to the core library.
TensorFlow Ecosystem
Ever since Alexnet with 60 million parameters popularized deep convolutional neural networks by winning the ImageNet challenge ILSVRC 2012, the general trend has been to make deeper and more complicated networks in order to achieve higher accuracy.
Similarly, in order to improve the accuracy of our understanding of TensorFlow 2.x, we need to dig deeper too.
A. TensorFlow for Research
There are plenty of libraries and extensions that exist in TensorFlow such as JAX, TF Probability, and Tensor Flow Quantum, which is strengthening the research in the ML world. The researchers prefer TensorFlow over other platforms due to its new extensibility of allowing eager execution at the core and keeping it similar to NumPy, very pythonic, and integrates easily with the rest of the Python ecosystem.
However, in this section, I would briefly talk about one feature which I liked the most was the introduction of a new tool called Performance Profiler.
This new tool in TensorBoard will help researchers in two ways:
- It will consistently monitor the TF model performance on the host (CPU), the device (GPU), or on a combination of both the host and device(s) with respect to the various TensorFlow operations (ops) in the TF model.
- It will provide in-depth debugging guidance to researchers which ultimately will help in resolving performance bottlenecks.
End results: it will make the model execute faster.[win-win for everybody]
B. TensorFlow for Production
TensorFlow has not only been the first choice among the research communities, but it also has become the undisputed choice among the various industry sectors like healthcare and automobile. As TensorFlow provides researchers the ability to test some new ideas as fast as possible similarly it has enabled industrialists or production engineers to take their ideas and create a real-world Impact by making the training to deployment phase hassle-free.
So, let’s learn more about the TensorFlow 2.x updates and libraries to fulfill our end-to-end ML goals.
Long Live Marriage of TensorFlow and Keras
As the entire AI fraternity including me loves to work with Keras because it’s easy to use and it even allows it’s users to build and train custom models pretty quickly that’s why the TensorFlow team with keeping our needs into consideration have committed to keeping tf.keras as a default High-Level API.
TensorFlow Hub: A tool for Transfer Learning Fans
If in your day to day job you work on various computer vision-related applications and you want to make faster progress on it you more or less end up using Transfer learning.
(For Non-AI Audience)Well, Transfer learning is nothing but an approach in which you can download open-source weights from publicly available datasets (MS-COCO, ImageNet) that took someone else many weeks, months to figure out and use that as a very good initialization[instead of random intilization] to your own Neural Network to transfer knowledge from some of these large public datasets to your own problem.
So, now as per the recent announcement, there are more than 1000 models available in TF Hub with documentation, code snippets, demos, and interactive Google Colab all waiting to be used.
TensorFlow Extended (TFX): An End-to-End ML Platform
As per the recent updates, Once you are ready to move your model into production. TensorFlow Extended now will allow you to build a production-ready[end-to-end] ML pipeline for your scalable, high-performance machine learning tasks.
TensorFlow 2.1 supports Cloud TPU’s
This update is especially for TPU’s lovers. Now you can train and deploy your models and pipeline on custom hardware specifically designed for AI workflow such as TPU’s. In the latest version, TensorFlow has been optimized for Cloud TPU using Keras which means the same API you started with, will help you to scale to petaflops of TPU compute which would be enough to transform your business or create the next research breakthrough.
The End
Well, folks now we have reached the end of this article. I know through this article I was only able to scratch the surface related to announcements and updates made at TensorFlow Dev Summit 2020.
Therefore do visit this link to watch all the sessions at TensorFlow Dev Summit 2020.
Thank you for your attention
You using your time to read my work means the world to me. I fully mean that.
Also, follow me on Medium, LinkedIn, or Twitter if you want to! I would love that.
