TensorFlow開發峰會2019主題演講 (TensorFlow Dev Summit 2019 Keynote)

字幕列表影片播放

♪ (upbeat music) ♪
Good morning everyone, I'm Alina,
program manager for TensorFlow.
(applause)
Thank you.
Welcome to the 2019 TensorFlow Developer Summit.
This is our third annual and largest developer summit to date,
and I'm so happy to have all of you here
both right here in the room and on the Livestream.
Welcome.
So I'm just curious by a show of hands,
who traveled a little bit further, maybe, to get here?
Europe?
Asia?
Africa?
As far as Australia?
Woo, awesome.
Welcome to all of you.
We have a lot of great talks ahead, some exciting announcements
and cool demos, so let's get going.
We are living in a formative moment of history right now,
where machine learning is experiencing an unprecedented revolution.
The way we fundamentally think about and interact with computer systems
has inherently changed
due to the breakthroughs in the field of AI.
and this is due to three major factors.
First, we have lots more compute specially designed ML accelerator
like these TPUs,
let you train models faster than ever before.
Secondly, we have breakthroughs in the field of machine learning.
Their novel algorithms created every month
like a BERT and an innovative approach to natural language processing
which lets anyone around the world
train their own state-of-the-art question-answering system.
And finally, we have lots and lots of data.
We're seeing new waves of data sets come from all kinds of disciplines.
For example, the new open images extended data set.
This is a collection of over 478,000 images
that volunteers have added
with the pursuit of inclusivity and diversity.
So, all three of these are basically changing
how we solve challenging, real-world problems,
and it's really cool to see that TensorFlow is the platform
that's powering this machine learning revolution.
It's allowing developers, businesses and researchers around the world
to benefit from intelligent applications.
And we've been really amazed
by what the community has built with TensorFlow.
Developers have been using TensorFlow to solve problems
in in their local communities.
So I don't know if any of you were in the Bay Area
during the tragic Paradise fire,
but one of the consequences was
that air quality was really bad.
It was in the high to mid to 200s on the Air Quality Index.
And as difficult as this was for us in Delhi, India during winter.
The air quality can get up to about the 400s on Air Quality Index,
and this is considered very dangerous.
So pollution sensors can help gauge air quality
but they're very expensive to deploy at scale.
So a group of students in Delhi built image classifiers in TensorFlow
and use those to build an app called Air Cognizer,
and what it does is basically just by using the images on a smartphone
it gives an accurate estimation of the air quality.
Businesses are also fundamentally
improving their products and services built with TensorFlow,
for example, Twitter strives to keep its global users informed
with relevant and healthy content.
But this can be hard,
when the users follow hundreds or even thousands of people,
so to solve this, Twitter launched ranked timeline, an ML power feed
which has the most relevant tweets at the top of the time timeline,
ensuring users never missed their best and most relevant content.
And by using TensorFlow's ecosystem of tools
like TensorFlow Hub, TensorBoard
and TensorFlow Model Analysis,
Twitter was able to reduce both training and model iteration time
as well as increase the timeline quality and engagement for users.
Specific industries are also being very much transformed by ML.
GE Healthcare, for example
is using TensorFlow to improve MRI imaging.
These TensorFlow models, they're real-time on MRI scanners
and can actually detect the orientation of the patient inside the scanner.
And this is really great
because not only does this help the diagnosis,
but also lowers the errors and exam time.
But also, what's really cool is it basically expands this technology
to many many more people around the world.
TensorFlow also powers bleeding-edge research.
A team of scientists, researchers and engineers
at nurse Oak Ridge National Laboratory at VIDYA
recently won the Gordon Bell Prize
for applying deeplearning
to study the effects of extreme weather patterns
using high-performance computing.
They built and scaled a neural network using TensorFlow,
of course, to a run on Summit, the world's fastest supercomputer.
They achieved a peak and sustained throughput
of 1.13 exaFLOPS and FPC-16
which is equivalent to more than a quantalian computations per second.
I think I need to pause for a second because that is ridiculously fast.
Right?
In addition to these awesome examples,
there are thousands and thousands of people all over the world
doing amazing work using TensorFlow,
and the power and impact of TensorFlow would not be what it is
without all of you, thank you.
It's with your help and interest
that TensorFlow has become the most widely adopted ML framework in the world.
And right here, I'd like to show the latest map of GitHub stars
who self-identified their location.
I'm sure many of the dots on this map are right here
in the room and on the Livestream,
so I just want to say thank you one more time.
And this growth has been absolutely amazing.
TensorFlow has been downloaded over 41 million times,
and has over 1800 contributors worldwide.
Last November, we celebrated TensorFlow's third birthday
by taking a look back at the different components
that we've added throughout the years.
But today, we'd like to talk
about how TensorFlow has matured as a platform
to become an entire end-to-end ecosystem.
And TensorFlow 2.0 is the start of a new era,
and we're committed and focused on making it
the best ML platform for all our users.
To talk more about TensorFlow 2.0
I'd like to introduce Rajat Monga,
Engineering Director of TensorFlow on stage.
Thank you.
(applause)
Thank You, Alina.
Hello, everyone, I'm Rajat.
I am an engineer at TensorFlow
and have been involved with this since the very beginning.
It's been great to see what we've been up to
over the last few years.
All the amazing growth
and all the amazing things that you've done with it.
It's also been great to hear from you.
You told us what you like about TensorFlow
and equally importantly, what you would like to see improved
in TensorFlow.
Your feedback has been loud and clear.
You asked for simpler, more intuitive APIs in developer experiences.
You pointed out areas of redundancy and complexity,
and you asked for better documentation and examples,
and this is exactly what we've been focusing on with TensorFlow 2.0.
To make it easy, we focused on Keras
for a single set of API's
and combine it with Eager Execution for the simplicity of Python.
With flexibility to try the craziest ideas
and ability to go beyond an exaFLOP
TensorFlow is more powerful than ever.
With the same robustness and performance
you expect in production, battle-tested in Google.
Let's start with the overall architecture for TensorFlow.
You may be familiar with this high-level architecture.
There have been lots of components and features
we've added throughout the years
to help support workloads to go from training to deployment.
With TensorFlow 2.0, we're really making sure
that these components work better together.
Here's how these powerful API components fit together
for the entire training workflow.
With tf.data for data ingestion and transformation,
keras and premade estimators from model building,
training with eager execution and graphs,
and finally packaging for deployment with SavedModel.
Let's take a look at some examples.
The first thing you need is data.
Often, you may want to validate results or test your new ideas
in common public data set.
TensorFlow data sets includes
a large and rapidly growing collection of public data sets
that you can get started with very easily.
And combined with tf.data,
it is simple to wrap your own data too.
Here is a small sample of the data sets that are available,
and all of these and many more are included there.
Then with keras, you can express the model with layers,
just as you are used to thinking about it.
Standard training and evaluation is packaged as well,
with model.fit and .evaluate.
Since deep learning models are often computationally expensive,
you may want to try scaling this across more than one device.
TensorFlow comes pre-built with MirroredStrategy
that works with small additions to your code.
Starting from a pre-trained model or component
also works well to reduce some of this computational cost.
To make it easy,
TensorFlow Hub provides a large collection of pre-trained components
that you can include in your model
and even fine tune for your specific data set.
Keras and .estimator offers high-level building blocks
for an easy-to-use package.
They come with everything you might need for typical training jobs.
But, sometimes you need a bit more control.
For example, when you're exploring new kinds of algorithms.
Let's say, you wanted to build a custom encoder for machine translation.
Here's how you might do this by subclassing a model.
Here, you can focus
on implementing the computational algorithm,
and let the framework take care of the rest.
And you could even customize the training loop
to get full control over the gradients and the optimization process.
While training models,
whether packaged with keras or more complex ones,
it's often valuable to understand the progress,
and even analyze the model in detail.
TensorBoard provides a lot of visualization to help with this,
and now it comes full integration
with intercollab and other Jupiter notebooks,
allowing you to see the same visualizations
right from within your notebook.
All of these features are available in TensorFlow 2.0,
and I'm really excited to announce that our alpha release is available
for you as of today.
(applause)
Many of you in the room and across the world
really helped with lots of work in testing to make this possible.
I really like to take this moment to thank you all.
Please give yourself a round of applause.
We really couldn't have done this without you.
(applause)
In addition to all the great improvements we talked about,
this release comes with a Conversion Script
to help you upgrade from 1.X.,
and a compatibility module to give you access to 1.X. APIs
for easy transition,
and we are working towards the full release over the next quarter.
There's a lot of work going on to make TensorFlow 2.0
really work well for you.
You can track the progress and provide feedback
on the TensorFlow GitHub projects page.
You asked for better documentation
and we worked to streamline our docks for APIs, guides and tutorials.
All of this material will be available today
on the newly redesigned TensorFlow.org website.
Where you'll find more examples, documentation and tools to get started.
We're really very excited
about these changes and what's to come.
And to tell you more about improvements in TensorFlow
for research and production.
I'd like to welcome Megan Kacholia on stage, thank you.
(applause)
Thanks, Rajat.
TensorFlow has always been a platform for research to production.
We just saw how TensorFlow, as high-level APIs,
make it easy to get started and build your models.
Now, let's talk about how it improves powerful experimentation for researchers,
and let's you take models from research and prototyping
all the way through to production.
Researchers have been using TensorFlow for state-of-the-art research.
We can see it in paper publications,
which are shown over the past few years in this chart.
But powerful experimentation begins and really needs flexibility.
This begins with Eager Execution with TensorFlow.
In TensorFlow 2.0 by default, every Python command is immediately executed.
This means you can write your code in the style you're used to
without having to use Session.Run.
This also makes a big difference in the realm of debugging.
As you iterate through with Eager Mode,
you'll eventually want to distribute your code onto GPUs, TPUs
and other hardware or accelerators.
For this, we've provided tf.function
which turns your eager code into a graph, function by function.
You get all of the familiar tools
like Python, control-flow, asserts, even print
but can convert to a graph anytime you need to,
including when you're ready to move your model into production.
And even with this, you'll continue to get great debugging.
Debug ability is great, not just in Eager,
but we've made huge improvements in tf.function and graphs as well.
In this example shown here,
we're splitting a tensor using tf.function which creates a graph,
but because of the mismatched inputs, you get an error.
As you can see, we now give users the information
about the file and the line number where the error occurred in the model
to help you more quickly track things down,
so you can continue iterating.
We've made the error messages concise, easy to understand and actionable.
We hope you enjoy these changes and they make it much easier for you
to quickly iterate and progress with your models.
Performance is another area we know that researchers
as well as all users for that matter, care about,
and we've continued improving core performance in TensorFlow.
Since last year,
we've sped up training on eight NVIDIA Tesla V100 by almost double.
Using a Google Cloud TPU V2,
we've boosted performance by 1.6x.
And with Intel MKL Acceleration
we've got an inference speed up by almost three times.
Performance will continue to be a big focus of TensorFlow 2.0
and a core part of our progress to final release.
TensorFlow also provides flexibility to enable researchers,
and this is with many add-on libraries
that extend and expand TensorFlow
in new and useful ways.
Some of these add-on libraries or extensions
to make certain problems easier,
like TF.Text with Unicode
and the new ragged Tensor type.
In other cases, it lets us explore
how we can make machine learning models fairer and safer
by a TF Privacy.
You'll also hear new announcements on TF-Agents for reinforcement learning,
and tomorrow, we'll be discussing the new TF federated library
for federated learning.
Deep learning research is also being applied to real-world applications
using TensorFlow.
Here are a few examples from researchers at Google
where we see them applying it to areas like our data centers.
We're making them more efficient with AI control system
that delivers energy savings.
Our apps like Google Maps, the one shown in the middle,
which has a new navigation feature called global localization.
It combines visual positioning service, street view, and machine learning
to more accurately identify position and orientation.
And devices like the Google Pixel
that use machine learning to improve depth estimation
to create better portrait mode photos
like the one shown here.
In order to make these real-world applications a reality,
you must be able to take models from research and prototyping.
all the way through to launching and production.
This has always been a core strength and focus for TensorFlow.
Using TensorFlow, you can deploy your models
on a number of platforms like shown here.
And models end up in a lot of places,
so we want to make sure TensorFlow works well
across all of these on servers and in cloud,
on mobile and other Edge devices,
in browser and JavaScript platforms.
We have products for each of these:
TensorFlow Extended, TensorFlow Lite and TensorFlow.js
which I'll briefly talk through.
TensorFlow Extended is our end-to-end platform
for managing every stage of the machine learning lifecycle.
This spans all the way from ingesting and transforming your data
to deploying your machine learning models at scale.
In orange shown here, you can see the libraries of open-sourced so far.
What this slide alludes to is
that we're now taking a step further
and providing components built from these libraries
that make up an end-to-end platform.
And note these are the same components
that are used internally
in thousands of production systems,
powering Google's most important products.
The components are only part of the story.
2019 is the year we're putting it all together,
and providing you with an integrated end-to-end platform.
First, you can bring your own orchestrator.
Here, we're showing airflow or kubeflow, even raw kubernetes,
whatever you want.
No matter what orchestrator you choose,
the TensorFlow extending components integrate with a meta-data store.
This store keeps track of all the component runs,
the artifacts that went into them
and the artifacts that were also produced.
This enables advanced features like experiments, experimentation
and experiment tracking, model comparison
and things along those lines
that I'm sure you'll be excited about
and will help you as you iterate through
and work with your production systems.
We have an end-to-end talk coming up later today from Clemens and his team
in which they'll take you on a complete tour
of using TensorFlow Extended to solve a real problem.
Moving on, TensorFlow Lite is our solution for running models
on mobile and IoT hardware.
it uses a custom streamline file format and a stripped-down runtime,
so you can deploy TensorFlow models everywhere your users are.
On-device models can be more responsive to input than cloud backends,
and they keep user data on device for privacy
which is very important, especially in this day and age.
Google and our partners like IGE in China use TF Lite
for all kinds of tools,
including predictive text generation, video segmentation
and things like Edge detection.
But under the hood, TensorFlow Lite is about performance.
You can deploy models to CPU, GPU and even Edge TPUs,
and expect fast performance,
and we've been refining since we launched TensorFlow Lite last year.
By using the latest quantization techniques on CPU,
adding support for OpenGL 3.1 and Metal on GPUs,
and tuning our performance on Edge TPUs,
we're constantly pushing the limits of what is possible on device,
and you should can expect even greater enhancements in the year ahead.
We'll hear details from Raziel and his colleagues coming up
in a little bit this morning.
Javascript is the number one programming language in the world
and until recently hasn't necessarily benefited
from all the machine learning development and tools.
Last year we released TensorFlow.js,
a library for training and deploying machine learning models
in the browser and on Node.js.
Since then we've seen huge adoption in the JavaScript community
with more than 300,000 downloads and 100 contributors,
but we're just at the beginning
given how big the JavaScript and web ecosystem is.
Today we're excited to announce TensorFlow.js version 1.0.
This comes with many improvements and new features.
We have a library of off-the-shelf models for common machine learning problems
that run both in the browser and on node.
We're also adding support for more platforms
where JavaScript runs such as electron desktop apps
or mobile native platforms.
and a huge focus in TensorFlow.js 1.0
is on performance improvements.
As an example, compared to last year,
MobileNet inference and browser is now nine times faster.
You'll learn more about these advances in our talk later in the day.
Another language that we're really excited about is swift.
Swift for TensorFlow is reexamining
what it means for performance and usability.
With a new stack built on top of TensorFlow's core
and a new programming model
that intends to bring further usability.
And today, we're announcing that Swift for TensorFlow is now at version 0.2.
It's ready for you to experiment with, to try out,
and we're really excited to be bringing this to the community.
In addition to telling you about version 0.2,
we're also excited to announce that Jeremy Howard, a fast.ai
is writing a new course in Swift for TensorFlow.
Chris and Brennan will tell you a lot more about this later today.
So to recap everything we've shown you so far.
TensorFlow has grown to a full ecosystem
from research to production,
from server to mobile with many languages.
This growth has been fueled by our community,
and honestly would not have been possible without the community.
To talk about what we're planning for you and with you in 2019,
I'll hand it over to Kemal.
(applause)
It's all you.
(Kemal) Thank you, Megan.
Hi, my name is Kemal
and I'm the Product Director for TensorFlow.
I'm really excited to be here today for this celebration,
and what we're celebrating is the most important part
of what we're building, and that's the community.
Personally, I love building developer platforms.
I used to be a developer as an entrepreneur,
and now I get to enable other developers
by building together a better platform.
When we started working on 2.0,
we turned to the community,
we started with the request for common process,
consulting with all of you on important product decisions.
We received valuable feedback
and we couldn't have built 2.0 without you.
And some of you wanted to get more involved
so we created special interest groups or sigs
like Networking or Tensor Board to name a few.
And sigs are really a great way for the community
to build the pieces of TensorFlow
that they care the most about.
We also wanted to hear more about what you were building,
so we launched a Powered By TensorFlow campaign.
And I am going to say we were amazed by the creativity of the project,
from biological image analysis, to custom wearables,
to chat BOTS.
So after three years, our community is really thriving.
There're almost 70 machine learning GDEs right now.
Around the world,
1800 contributors on core alone,
and countless more of you who are doing amazing work
to help make TensorFlow successful.
So on behalf of the whole TensorFlow team
we want to say a huge thank you.
(applause)
So we have big plans for 2019,
and I would like to make a few announcements.
First, as our community grows,
we welcome people who are new to machine learning
and it's really important to provide them
with the best educational material,
so we're excited to announce
two new online courses.
One is with deeplearning.ai
and it's published in the Coursera platform.
And the other's with Udacity.
The first batch of these lessons is available right now,
and they provide an awesome introduction to TensorFlow for developers.
They require no prior knowledge to machine learning,
so I highly encourage you to check them out.
Next, if your students for the very first time,
you can apply to the Google Summer of Code program
and get to work with the TensorFlow engineering team
to help build a part of TensorFlow.
I also talked about the Powered By TensorFlow campaign.
We're so excited with the creativity
that we decided to launch a 2.0 hackathon on DevPost
post to let you share your latest and greatest,
and win cool prizes.
So we're really excited to see what you're going to build.
Finally, as our ecosystem grows,
we're now having a second day at the summit,
but we really wanted to do something more.
We wanted a place where you can share
what you've been building on TensorFlow,
so we're excited to announce
TensorFlow World,
a week-long conference dedicated to open-source collaboration.
This conference will be co-presented by O'Reilly Media and TensorFlow,
and will be held in Santa Clara end of October.
Our vision is to bring together the awesome TensorFlow World
and give a place for folks to connect with each other.
So I'd like to invite on stage Gina Blaber
to say a few words about the conference.
(applause)
(Gina) Thank You, Kemal.
O'Reilly is a learning company
with a focus on technology and business.
We have strong ties with the open source community
as many of you know,
and we have a history of bringing big ideas to life.
That's why we're excited about partnering with TensorFlow
to create this new event
that brings machine learning and AI to the community.
The event of TensorFlow happening on October 28 to 31
in Santa Clara.
And when I say community, I mean everyone.
We want to bring together the entire TensorFlow community
of individuals and teams,
and enterprises.
This is the place where you'll meet experts
from around the world,
the team that actually creates TensorFlow,
and the companies and enterprises that will help you deploy it.
We have an open CFP right now on the TensorFlow World site.
I invite you all to check that out and send in your proposal soon,
so your voice is heard at that event.
We look forward to seeing you at TensorFlow World in October.
Thank you.
(applause)
Thank you, Gina. This is going to be great.
Are you guys excited?
Woo!
So we have a few calls to action for you.
Take a course, submit a talk to TF World, start hacking in 2.0.
By the way, the grand prizes for a hackathon on DevPost,
will include free tickets to TensorFlow World.
You know one thing that I love is to hear about these amazing stories
of people building awesome stuff on top of TensorFlow.
And as a team, we really believe that AI advances faster
when people have access to our tools
and can then apply them to the problems that they care about
in ways that we never really dreamed of.
And when people can really do that, some special things happen.
And I'd like to share with you something really special.
(urban noise)
Looking at historical documents
and especially documents from the Middle Age period,
requires a lot of time and also a lot of patience.
♪ (gentle music) ♪
In the Vatican Archives, there are 85km of documents
more or less the length of the Panama Canal.
The scriptures written in the medieval handwriting
are different from the ones we know nowadays.
If one day someone ask me to transcribe and translate
all the documents of the Vatican Archive,
I would tell them that they are completely crazy.
(woman) Looking at this book page by page
and trying to decipher, read and transcribe whatever is there
takes an enormous amount of time.
It would require an army of paleographer.
♪ (upbeat music) ♪
(woman 2) What I am excited the most about machine learning is
that it enabled us to solve problems
that up to 10, 15 years ago we thought unsolvable.
(Paolo) "In Codice Ratio" was born from this idea of building a software
that can read and interpret what is inside those manuscripts.
When we started discussing the problem,
we realized that a solution based on neural networks
was absolutely necessary.
The choice of TensorFlow was a natural one.
(Elena) Before using any kind of machine learning module,
we needed to collect data first.
You have thousands of images of dogs and cats on the Internet,
but there's very little images of ancient manuscripts.
We build our own custom web application for crowd sourcing
and we involved high school students to collect the data.
I didn't know much about machine learning in general,
but I found it very easy to create a TensorFlow environment.
When we were trying to figure out which model worked best for us,
Keras was the best solution.
The production model runs on TensorFlow layers and estimator interface.
We experimented with binary classification
with fully connected networks,
and finally we move to convolutional neural network
and multi-class classification.
In a short time, we were able to develop and test the first solutions.
When it comes to recognizing single characters,
we can get 95% average accuracy.
(Marco) Being able to access an IT tool greatly shortens the timing.
Being able to solve certain abbreviations and to understand a text
in that cryptic writing
is something exceptional.
(Serena) This will have an enormous impact in a short period of time.
We will have a massive quantity of historical information available.
I just think solving problems is fun.
It's a game against myself,
and how good I can do.
(Marco) The study of history is extremely important
to understand our present
and to get a perspective on the future.
(applause)
This is such a great story.
I think about the scholars who wrote these manuscripts.
They couldn't have been imagined that centuries later,
people will be using computers to bring back to life their work.
So we're really lucky to have Elena with us today.
Elena, would you stand?
(applause)
Don't miss the talk where she will share her story today.
I really hope you have a great day.
We have some really awesome things.
The team and I will be around.
Please come and say hi, we want to hear from you,
and with that, I'm going to hand it over to Martin
who will talk about TensorFlow 2.0. Thank you.
♪ (upbeat music) ♪