TIM NOVIKOFF: Hi.

My name is Tim, and I'm the product manager for Colab.

For those of you who don't already know,

Colab allows you to write and execute arbitrary Python

code via the browser.

It's hosted Jupyter Notebooks by Google.

It's integrated with your Google account, which

means if you're signed into Chrome or signed into Gmail,

you're already signed into Colab.

And it's integrated with Google Drive, which

means that when you create a Jupyter Notebook in Colab,

it's automatically saved to your drive account.

Colab started in 2012 as an internal tool

for data analysts.

In 2014, we started using it for our internal machine learning

crash course.

This is actually where a lot of Googlers

first learned TensorFlow.

In 2017, we launched it publicly.

We wanted the whole world to have access

to the same great tool that Googlers were

using to get their work done.

Colab is easy to use.

When you connect with a virtual machine in a Colab notebook,

your Python environment is already set up for you.

The latest version of CUDA is already installed for you.

It just works.

So how do we do this?

Well, on the back end, we maintain

a pool of pre-warmed VMs.

These virtual machines have hundreds of popular Python

libraries pre-installed, including

TensorFlow, TensorBoard, TF privacy, TF data sets,

and many more.

There are resource limits in Colab.

And those are there to ensure sustainability and reduce

abuse.

Without any further ado, here are my top 10 Colab tricks

for TensorFlow users.

Number 10.

Always specify your TensorFlow version in Colab.

In Colab, when you import TensorFlow, by default today

it imports a version of TensorFlow 1.

But soon, it will by default import

a version of TensorFlow 2.

You can start using TensorFlow 2 in Colab

today by specifying your TensorFlow version

with the syntax shown on the screen.

If you have old Colab notebooks that use TensorFlow 1

and you want those to keep working well

after we switch to default TF 2, make

sure to specify TF 1.x in those notebooks.

That will futureproof those notebooks

so they'll keep working well.

The bottom line-- always specify your TensorFlow version

in Colab.

Number 9.

Use TensorBoard.

Colab output cells can render essentially arbitrary HTML.

And the TensorBoard team has taken advantage of this

to allow you to use TensorBoard right in line with Colab.

It works great, and I highly recommend it.

Number 8.

TFLite?

No problem.

Even though TFLite is for on-device machine learning,

you can still train your TFLite models in the cloud in Colab.

And in fact, that's a popular way to use TFLite.

Number 7.

Use TPUs.

It's free in Colab and very easy to do.

You just go to change runtime type

and select TPUs from the runtime--

from the dropdown.

If you want to feel the raw power of TPUs

under your fingertips, give them a whirl for free in Colab.

And with TensorFlow 2.1, TPUs have never been easier to use.

Number 6.

Use local runtimes if you want.

If you have your own workstation and your own GPUs perhaps

and you want to run your workload on your own hardware,

that's fine.

You can still actually use the Colab UI with a local runtime.

And that's easy to do as well.

Number 5.

The Colab scratchpad notebook.

If you find yourself with Colab notebooks that have names

like "untitled 17" and "untitled 18,"

the Cloud scratchpad notebook might be for you.

This is a special notebook available at the URL shown here

that's not automatically saved to your drive account.

And it's really great for doing your scratch work.

Number 4.

Copy your data into your Colab VM.

Rather than calling data on external storage or runtime

when you're, for example, training your models,

first copy all your data into the Colab VM.

This can result in speed-ups even if you're only

going to use the data once.

Number 3.

Mind your memory.

If you've ever run out of memory in Colab,

you may have noticed that perhaps later you were assigned

a special high RAM runtime.

Well, that may seem like a great thing.

But it also means that you're more

likely to run into the resource limits of Colab later.

So the best thing to do is to not run out of memory at all

and mind your memory when you're doing your work.

Number 2.

Close your tabs when you're done with Colab.

This will help you disconnect from the underlying virtual

machines sooner.

Again, conserving resources makes it less likely

that you'll run into the resource limits of Colab.

Finally, the number one trick, only use

GPUs when they're actually needed for your work.

Now, of course, when using TensorFlow,

often GPUs are needed for your work.

But when you're doing work that doesn't require a GPU,

just use a default runtime that only has a CPU.

Again, conserving resources helps

you get the most out of Colab by avoiding later

running into the resource limits of Colab.

Now, we do hear from users all the time things

like, "We want faster GPUs.

We want more reliable access to P100s.

We want longer runtimes.

We want more lenient idle time out periods.

And we want longer maximum VM lifetimes.

Or we want higher memory with more reliable access

to high RAM runtimes."

Well, we do hear you.

And that's why we recently launched Colab Pro.

Colab Pro comes with faster GPUs, longer runtimes, and more

memory for $10 a month.

We launched it in February, and it's been a big success.

It's only available in the United States for now,

but we're working as fast as we can

to bring it to more countries.

All right.

So what's next in Colab?

Well, first of all, the free version of Colab

is not going away.

The free version of Colab and Colab Pro

will continue to co-exist, and we're

continuing to work on improving both of them.

And as for feature development, well, that's

where I would like to hear from you.

Tweet at me.

Send feedback via the product.

Tweet at GoogleColab.

However you do it, let us know what

you want us to build next in Colab,

because it's user feedback from people like you

that drives our roadmap going forward.

So please let us know what you want next in Colab.

[MUSIC PLAYING]