Welcome back.

- I'm Paige Bailey, and this is... - Laurence Moroney,

and we are here to answer all of your Ask TensorFlow questions

at the TensorFlow Dev Summit.

So, if you have any questions, please submit them to social media

with the hashtag #AskTensorFlow.

And we'll answer as many of them as we can,

but any that we can't get to today we'll try to reach out to you later

- to answer them. - Absolutely.

So let's get started with the first question.

- Okay, so shall I try this one? - Yeah, absolutely.

So this one, I think, came on Twitter, from @yuehgoh, and it said:

"Once I installed tensorflow-gpu, import did not work for me.

I have been trying to use it," but has been unable to do so.

"It fails to load native tensorflow runtime."

I remember seeing this one.

Actually, it was on YouTube,

and it was in response to one of my videos

about pip-installed TensorFlow GPU in Colab,

because once upon a time in Colab,

you had to pip install TensorFlow GPU to be able to use it,

and now if you try to do that, you end up having some issues.

The reason for that is actually really good, and it's good news,

and it's because you don't need to do it anymore.

Excellent!

So, actually, if we switch to Colab for a second on my laptop,

I can show you.

This was the notebook that I was showing earlier on.

And all you have to do, if you want to use GPU in Colab,

is just change your runtime type,

pick GPU as the hardware accelerator,

and now you don't need to pip install TensorFlow GPU;

it actually does it for you under the hood, behind the scenes.

It's really, really cool,

and that's why earlier I was able to train this so quickly,

because I was actually using the GPU.

And, as you can see, there's no pip install GPU on here.

(Paige) Excellent.

Whenever we were initially testing TensorFlow 2.0,

we had a kind of similar issue, as well, with the GPU install,

in that you needed specific CUDA drivers.

But now, CUDA 10 is supported in Colab, as well.

So Colab is a great experience if you're using a GPU

or if you're using any other accelerated hardware.

(Laurence) Yeah, and a pro tip going forward, as well,

if you want to do GPU stuff,

because this was something that I ran into a number of times

when trying to use the GPU,

was that you always have to carefully take a look

at the version of CUDA and cuDNN that you're using.

Because I made the mistake

that I just went to the vendor's website,

I downloaded the latest versions, I installed them,

and then I saw TensorFlow was actually supporting

a slightly earlier version.

So if you do get an error when you're trying to use GPU,

just take a look at the version

of the driver that it's looking to support,

and then, from the vendor's website,

download that specific version.

- (Paige) Yeah, driver issues... - Driver issues.

(Paige) They're always a treat, right?

Exactly!

It's one of the things that makes our job interesting.

Absolutely.

Alright, so shall we take a look at the next one?

Yeah, let's go-- Oh! @adkumar!

So @adkumar had at least eight excellent questions on Twitter.

Maybe he wins the volume award.

He absolutely does!

And we'll get to all of them,

not in this sort of Ask TensorFlow segment,

but let's focus on just one for today

and then answer the rest offline.

Yeah, and I think a lot of them were really about file formats

and how do I use different file formats,

so shall we drill into that?

Absolutely. So the question is:

"Which is the preferred format for saving the model going forward,"

saved_model or something else?

And, if we look at the laptop, we can take a gander

at one of the slides from the keynote this morning,

really showing that Keras is a first-class citizen

in TensorFlow 2.0

and SavedModel is at the heart of every deployment.

So here you can see SavedModel being used for TensorFlow Serving,

for TensorFlow Lite, for TensorFlow.js,

and lots of other language bindings,

so really, we're pushing for the SavedModel.

(Laurence) And if you focus on SavedModel you can't go wrong.

(Paige) Yes, absolutely.

It's a lot easier to use

than some of the other deployment options that we'd seen before.

Yeah, so I think the guidance would be in the recommendation,

not just for AD, but for everybody else.

And, when you're thinking about saving out your models,

take a look at SavedModel, consider using SavedModel,

and, as a result, it's not only is the advantage of the file format,

but just how it's supported across all of these things.

And an excellent point of TensorFlow 2.0--

I'm just going to keep selling it-- is that we have a number

of code samples and tutorials available today

about how you can deploy your models with SavedModel.

Yeah, and I've personally found,

from playing with some of the TensorFlow Lite stuff in 2.0,

saving as a SavedModel and then going through the conversion

to the TF Lite process, it was a lot easier for me

than in previous iterations

where I had to use TocoConverter and all that kind of stuff.

So it's really being refined. We're really iterating on that.

- And I think it looks really cool! - Excellent!

So thanks for all of those questions, AD. There's some great stuff in there.

We'll try to answer some of the rest of them,

but understood that most of them are focused

around the file format and hopefully SavedModel will help you.

- Alright. - Perfect, so let's go to the next one.

So this next one comes from Elie Gakuba,

asking, "Is it possible to run tensorboard on colabs?"

- And I know this made Paige really happy! - Ah, yes!!

Oh, dude! You are going to be so delighted!

Because before TensorBoard was running on Colabs,

we were talking about it: "We really want it on Colabs!"

It was so painful.

And if you wanted to get it working

in a Colab notebook or in Jupyter,

you ended up using a tool like Ngrok,

and that was kind of not approved by our bosses, or in general.

But yes, the good news is that you can run TensorBoard in Colabs.

(Laurence) And when it was first announced internally in Google,

before it was publically announced,

we all got this email from Paige,

and it was full of all these smiley emojis and hearts.

(Paige laughs)

So, Elie, thank you for the question,

because I think you've made Paige's day.

(Paige) Excellent! And so you can see here in the Colab,

I'm running through and downloading some files

in the hope that we could play with it a little bit,

but here you can see it actually working.

You should be able to do different operations like smoothing,

changing some of the values,

and then also using the embedding visualizer

directly from your Colab notebook

in order to understand your accuracies

and to be able to do model performance debugging.

Another nice thing that the team has been working very, very hard on

is that you don't have to specify ports.

So you don't have to remember

if you wanted to have multiple TensorBoard instances running,

that you were using, what is it, 6006, or whatever, for another.

It just automatically selects one that would be a good candidate

and creates it for you.

So the team is phenomenal.

If you have any interest whatsoever in TensorBoard at all,

I suggest stalking their PRs, like I do,

because that's how I found out

that TensorBoard got added to Jupyter notebooks and also to Colab.

But, yeah, so excited.

And we'll have this link in the documentation for the video,

as well, the little notes underneath, for you to go and play with.

And I do have to say, it's so great to have a PR stalker in our group.

(laughter)

I get push notifications to my phone-- It's a problem.

But yeah, they've been doing such great work.

(Laurence) So the question is yes, TensorBoard is working in Colab.

- And also Project Jupyter notebooks! - Nice.

Use it wherever! TensorBoard everywhere!

TensorBoard everywhere; we all love TensorBoard.

I haven't really played with it that much,

but do a lot of the plugins also work?

So TensorBoard really is a collection

of these different visualizations,

so you can see scalers, like your accuracy;

you can see histograms;

you can see embedding visualizers,

which allows you to do clustering,

like that great MNIST example

from the Dev Summit a couple years ago.

- With all that stuff. - Moving it around.

- I play with that all day. - It's beautiful.

And then also Beholder, which was created by a community member.

The plugins... It's so funny you mention it.

A couple of our GSOC, our Google Summer of Code projects

this summer are focused on getting

additional visualization plugins added to TensorBoard.

(Laurence) Cool! Nice!

So in addition to the What-If Tool, in addition to Beholder,

you could make your own TensorBoard plugin.

While I could geek out about this all day, I think we should move on

to some other questions that some folks have had.

Can't we kick out about this all day?

Can we extend the stream for three or four hours?

So the next question that came in, from Amirhosein Herandy,

"How would you use feature_columns with Keras?"

I know you know the answer to this question.

But I know that it's a particular passion of yours

that you're working with Estimators in Keras.

So, for the folks who are watching,

feature columns are really part of Estimators.

They're a way of really getting your data

efficiently into Estimators.

And with people saying, "Hey, it's all there in Estimators.

What about Keras?"

You've been working on some great stuff around that, right?

I have, but I know that your YouTube channel

has a series from Karmel, who spoke earlier today.

Yeah, so Karmel is our engineering director

for high-level APIs, and she has this great series

around high-level APIs for TensorFlow,

really, really teaching you how to use the high-level APIs.

And Karmel is working actively, and her team are working actively,

for parity of things like feature columns in TensorFlow 2.0.

I'm not sure if it's fully there yet in the Alpha.

I haven't checked into it yet.

But, yeah, it's on the way, if it's not there already.

So you should be able to use them in Keras.

(Paige) Yes, and we're also in the process of building out--

If you wanted to migrate your models from using Estimators

to being more of a TensorFlow 2.0 format

with Keras,

I am currently in the process of building a migration guide.

So if you have any interest around that,

please feel free to reach out.

And we're excited to get that released pretty soon.

I'm really excited to see that, personally,

because the very first stuff I did in TensorFlow

was with Estimators, before I learned Keras,

and I really want to go back and start changing them to use Keras

without rewriting them.

Absolutely.

Keras is just friendlier to work with, my feel.

Yeah, so they're both great.

Estimators really gives you the power. Keras, great for beginners.

So hopefully we'll get the best of both worlds.