GAL OSHRI: Hi, everyone.

My name is Gal Oshri, and I'm product manager

on the TensorFlow team.

I'm here to talk to you about collaborative machine learning

with TensorBoard.dev.

Machine learning often involves collaboration

where you exchange ideas with a team, post questions on Stack

Overflow, share your code on GitHub,

and publish your findings to the broader research community

through papers and conferences.

In all of these, there's something

you're expressing about what you've done,

potentially through visualizations.

For example, it is common to share experiment results

and papers through various charts,

but it is also common to include this type of information

when asking for troubleshooting help.

In this case, someone posted an issue on GitHub asking why

they're not achieving the expected results alongside

a screenshot of their TensorBoard to explain why

they're-- what's going wrong.

The problem is the pictures don't

convey all the information.

For example, you might want to show what other metrics are

being tracked, how many trials were run,

or let the reader explore the model structure in more depth.

Even when the code is released, it

can require quite a bit of setup before the results can

be viewed.

So how can we make this easier?

Well, we already have TensorBoard,

which is TensorFlow's visualization toolkit.

It is commonly used by researchers and engineers

to understand their experiment results.

It lets them track metrics, visualize their model,

explore model parameters, view their embeddings,

and a lot more.

And we're consistently adding new capabilities to it.

Last year, we launched the HParams Dashboard

to help visualize hyperparameter [INAUDIBLE] results.

This can help you identify which hyperparameters are most

promising for further exploration.

Given that many people are already using TensorBoard,

and some are even sharing screenshots of it,

we launched TensorBoard.dev to make all of this easier.

With TensorBoard.dev, you can upload your experiment results

and get a link that can be shared with everyone for free.

Others can view and interact with your TensorBoard

with no setup or installation.

We started with the Scalars dashboard,

but more dashboards will be added soon.

We're only at the beginning of this journey,

and we want to continue evolving this to help drive

state of the art research.

As inspiration, I want to share a few awesome examples with you

today.

The TensorFlow/models repository on GitHub

provides a collection of officially

supported TensorFlow models on cloud TPUs.

Many of them now provide a link to TensorBoard.dev.

This allows you to quickly view the training dynamics

and gives you a point of reference,

if you're training them all yourself,

in trying to understand if something is going wrong.

In this example, [? Shawn ?] is using

Twitter to share some work on large-scale ML experimentation.

TensorBoard.dev is used to help tell

that story more effectively.

This team of researchers at Google

recently published a paper on optimizer

grafting to help better understand

optimizer performance.

Along with the paper, they've also

upload results to TensorBoard.dev

to show 550 grafting curves that help

illustrate their technique.

This will be difficult to convey purely in a paper.

Another project from Google Research

introduces big transfer, a recipe

for scaling up image-based general visual representation

learning.

By transferring using a simple heuristic,

they achieve state of the art results across multiple vision

tasks.

They uploaded their results to TensorBoard.dev

to show how their model's performance reaches state

of the art and compares with the baseline across multiple data

sets.

This paper from NeurIPS 2019 discusses

binarized neural networks.

Their GitHub repository includes several links

to TensorBoard.dev to illustrate model accuracy.

But what's really cool is that, as part of the NeurIPS

Reproducibility Challenge, another group of researchers

published a report on this paper.

In this report, they included links to TensorBoard.dev

as well to provide a more complete picture

of the training process and include

other useful information to help others with debugging.

So all this sounds great, but how do you get started?

You continue using TensorBoard in the same way

that you use it today.

In this example, the TensorBoard Keras callback is shown,

but there are other APIs that can be used.

If you're not familiar with TensorBoard,

check out TensorFlow.org/TensorBoard

for some great [INAUDIBLE] tutorials.

You can upload these logs with the TensorBoard dev upload

command and provided the log directory.

Just last week, we enabled optionally adding a name

and description to your experiments

to provide more context to those that view it.

This can include links to your paper, GitHub repo,

or simply point out what is really

interesting about your results.

You'll get a link to your TensorBoard

that you can immediately view, even

if your experiment is still in progress.

Finally, you can share the experiment by copying the URL

or using the Share button in the top right.

As I mentioned, we'll be adding more

of TensorBoard's dashboards over time

as well as enabling more collaboration capabilities.

You can learn more at TensorBoard.dev

and follow a simple example to get started.

Thank you for your time, and next up

is Julia to tell you about Kaggle with TensorFlow 2.0.

[MUSIC PLAYING]