The first thing you need is data.

You may want to validate results or test ideas on a common public

dataset.

It includes a large and rapidly growing collection of datasets

you can get

started with easily and combined with tf.data it is simple to

wrap your own data too. Here is a small sample of the datasets

available and all of these and many nor are included there.

Then with Keras, you can express the

models just like you are used to thinking about it.

Standard package is fit with model fit and evaluate as well.

Since deep learning models are often

commutationally expensive you way want

to try scaling this across more than one device.

Starting from a pre-trained model or component also works

well to reduce some

of this computational cost. TensorFlow helps provide a large

collection of pretained components you

can include in your model and Feinstein

-- fine tune for your dataset.

Keras comes with everything you might need for a typical

training job.

Sometimes you need a bit more control. For example when you

are exploring new kinds of algorithms.

Let's say you wanted to build a custom

encoder for machine translation, here is

how you could do this by subclassing the model.

You can even customize it training loop

to get full control over the gradients

and optimization process.

While training models, whether packaged with Keras or more

complex ones, it is often valuable to understand the

progress and even analyze the muddle in detail.

TensorFlow board provides a lot of visualization to help with

chis and

comes full integration with Colab and

other Jupyter notebooks allowing you to

see the same visuals. All of these features are available in

TensorFlow 2.0 and I am really excited to announce

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 to make this

possible. I would really like to take this moment to thank you

you all. Please give yourself a round of applause.

We really couldn't have done this

without you.

In addition to all the great improvements we talked about,

this

release comes with a converter script

and compatibility module to give you

access to the 1.X APIs. We are working for a full release over

the next quarter.

There is a lot of work going on to make TensorFlow 2.0 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 have worked to streamline our docs

for APIs, guides and tutorials. All of this material will be

available

today on the newly redesigned TensorFlow.org website where you

will find examples, documentation and tools to get

started. We are very excited about these changes and what's

to come. To tell you more about improvements in TensorFlow for

research and production, I would like to welcome Megan Kacholia

on stage. Thank you.

Thank you. Thanks Rajat. TensorFlow has always been a

platform for research to production.

We just saw how TensorFlow high-level APIs make it easy at

a get started and build models and now let's talk about

how it improves experimentation for research and let's you take

models from

research and production all the way through. We can see this in

paper publications

which are shown over the past few years in this chart.

Powerful experimentation begins and

really needs flexibility and this begins with eager execution

and TensorFlow 2.0 every Python command is immediately executed.

This means you can write your code in

the style you are used it without having to use session

realm. This makes a big difference in the realm of

debugging.

As you iterate through, you will want

to distribute your code on to GPUs and

TPUs and we have provided tf. function turning your eager code

into a graph function-by-function. You get

Python control flow, asserts and even print but can convert to a

graph any time you need to, including when you are ready to

move your model into production. Even with this, you will

continue to get great debugging.

Debugability is great not just in eager

but we have made improves in tf. function and graph. Because of

the mismatch inputs you get an error.

As you can see, we give information to

user about the file and line number where the error occurs.

We have made the error messages concise, easy to understand and

actionable. We hope you enjoy the changes and they make it

easier to progress with the models. Performance is another

area we know researchers as well as all users for that matter

care about. We have continued improving core performance in

TensorFlow. Since last year, we have sped up

training on eight Nvidia TeslaV100s by LLGS double.

With Intel and MKL acceleration we have gotten inference speed

up by almost three times. Performance will continue to be

a focus of TensorFlow 2.

0 and a core part of our progress to final release.

TensorFlow also provides flexibility

and many add on libraries that expand and extend TensorFlow.

Some are extensions to make certain

problems easier like tf.text with Unicode.

It helps us explore how we can make machine learning model

safer by a tf.privacy. You will hear new announcements on

reinforcement learning and tomorrow we

will discuss the new tf. federated library.

It is being applied to real world applications as well.

Here are a few examples from researchers at Google where we

see it applied to areas like data centers and making them

more efficient. Our apps like Google Maps, the one in

the middle, which has a new feature

called global localization and combines street service.

And devices like the Google Pixel that use machine learning

to improve depth estimation to create better portrait mode

photos like the ones shown here. In order to make these real

world applications a reality, you must be able

to take models from research and prototyping to launch and

production. This has been a core strength and focus for

TensorFlow. Using TensorFlow you can deploy models on a number of

platforms shown here and models end up in a lot of places so we

want to make sure TensorFlow works across all these servers,

Cloud, mobile,

edge devices and Java and number of platforms.

We have products for these.

TensorFlow Extended is the end the end platform.

In orange, shown here, you can see the

libraries we have Open SourceSourced so far. We are

taking a step further and providing components built from

these libraries that make up an end-to-end platform. These are

the same components used internally in thousands of

production systems powering Google's most important

products. Components are only part of there story.

2019 is the year we are putting it together and providing you

with an integrated end-to-end platform.

You can bring your own orchestrator.

Here is airflow or raw Kubernetes even.

Not matter what orchestrate you

chose, the items integrate with the metadata store.

This enables experiments, experimentation, experiment

tracking and model comparison and things I am sure you will be

excited about and will help you as you iterate through. We have

an end-to-end talk coming up from Clemens and his team and

they will

take you on a complete tour of

TensorFlow Extended to solve a real problem. TensorFlow Lite is

our solution for

running models on a mobile and IOt hardware.

On device models can be

reore responsive and keep users on device for privacy.

Google and partners like

iqiyi provide all sorts of things. TensorFlow Lite is about

performance.

You can deploy models to CPU, GPU, and even EdgeTPU.

By using the latest techniques and adding support for OpenGL

and metal on

GPUs and tuning performance on EdgeTPUs we are constantly

pushing the limits of what is possible. You should expect

greater enhancements in the year ahead.

We will hear details from Raziel and colleaguess coming up later.

JavaScript is the number one

programming laj language in the world