Cloud Native Computing Foundation KubeCon + CloudNativeCon China 2018 (Shanghai), 26 Nov 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Keynote: A Tale of Using Kubeflow to Make the Electricity Smarter in China – Julia Han & Xin Zhang

Description

Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io

Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects

Keynote: A Tale of Using Kubeflow to Make the Electricity Smarter in China – Julia Han, COO & Xin Zhang, CEO, Caicloud

Artificial Intelligence (AI) has the potential to permeate and intertwine with every industry in our society; however, even though open-source AI frameworks such as TensorFlow have lowered the barrier to experiment new algorithms, without sufficient computing power all the advanced deep neural networks would grind to a halt. Kubeflow aims to tailor Kubernetes to support data analytical workloads and to provide an elastic AI pipeline for model developers to improve their model development velocity, solve their devops chores, and build robust and efficient production training and serving systems. We share a journey of using Kubeflow to empower engineers from State Grid of China, which exemplifies a traditional corporate, to develop OCR, vision, and time series prediction models, enabling practical innovation for their finance department and core business.

To learn more click here: https://sched.co/FzGn

A

(Upbeat electronic music).

A

(Julia han speaking in foreign language), [interpreter] hi, good morning., welcome to our venue, and I will be.

A

The first presenter today.

A

and I'm translator, and my name is han jiayao,, and this is my colleague xin zhang, and basically our presentation,.

A

We will talk about all ways that people know.

A

To make the national boards smile.

A

And in his company and how will we talk about.

A

A more traditional size industry that is the power generation, and is our power utility being controlled.

A

But actually, in china, we have a lot of changes..

A

We have a lot of traditional industry,, traditional soe and traditional enterprises,, but they are all embarked on their particular journey, to move more online, and onto the clouds.

A

When they are doing this,, it is because of the market's competition in china, with american's market competition.. If they don't embark on the digital journey,.

A

Then they will be surpassed by other nation's commerce,.

B

For instance,, like the internet, start ups,.

A

And therefore, they need to embark on the digital transition,, and this means we will have new applications narrows.

A

And therefore, you attach the data, so, for the gpu market, after the competitors at 2017,, it endures a growth rate of almost 230%,.

A

And therefore look how to deloitte, and as a forbes, and as a mckinsey, direct,.

B

And also been a survey, actually,.

A

They all find that in the future we will have more ai and also more machine learning.

A

And relevant devices.. So what I'll say is that ai and also machine learning.

A

Generally, now has permeated into other aspects of our lives.

A

And our national state grid also at this cross-road,.

A

Cross-Section, so in terms of the power generation.

B

And a power transformation, and as a power transmission,.

A

And as in the end, the end-users and adjoining power journey,, there will be massive data in a way, and ai, and also machine learning can play a huge role. Here.

A

And these are some examples.

A

for instance, in power generation, and also in the pan-trans matter, control,, and we can use ai and also ml to do a lot of things.

A

For instance,, we can use a monitoring surveillance camera to do the managing., and also this is the end-user's behavior,.

A

For instance, we can forecast what is the trend.

A

Of the future power utilization.

A

And this is the power generation, for instance,.

A

That the thermal power station,.

A

And where you can use ai to monitor the exhaust gas temperature.

A

And also in hydropowered stations,, we can monitor the water level in the most modern way.

A

To prevent any flood., and also we can use the computer vision.

A

To detect the status of equipment, devices,.

A

And also the whole system.

A

And also for some non-core departments., the physical department today can use mostly our computer vision.

A

To automatically see the invoice.

A

And also to forecast the future of financial growth.

A

So, we have talked about many applications, models,.

A

And they include a lot of amounts in china's.

A

To operate more smartly, and what is the traditional way of doing that??

A

They usually will hold a third party.

A

To translate all their data and also information to the third party,, and it will, this vendor, and they will provide,.

A

The vendor will provide an api or a black box.

A

For algorithm to the state grant and also other enterprises.

A

And if the applications narrows it too much.

A

Or the data is too much, and for the black box algorithm,, it is very expensive.

A

To buy and how to tweak it actually, it is a black box,. It is very hard to adjust.

A

Or to tweak by ourselves., and also for many enterprises, they don't want to just pay their way,. They are core information,.

A

Core data, and also they are core innovation to the.

A

Third party vendor,, and they don't, simply don't want to do that.

A

because for them it is the most important asset.

A

And also core components in their enterprise.

A

In this data is information and also elevation, so for the enterprises. Actually, they want to have more control ability to do the development.

A

And also to fully utilize the new generation of ai.

A

And also machine learning and many people say that.

A

Now we have a lot of open source and there's an algorithm and, for instance, we have pytorch.

A

And caffe and tensorflow they are open,, they are all open source, and it's like.

A

A lowering the threshold., but we have a lot of framework, a lot of architecture.

A

Which one do I choose and how can I compare between them?.

A

So for many enterprise they're the one choosing which one to use, it takes a lot of worked in points.

A

And also for many large, now, enterprises, it says that well, I can set up a algorithm team.

A

Or a architecture team by myself,, but this does not mean.

A

That they will not encounter any pinpoints,.

A

Cause they're concern is efficiency, for instance,.

A

This is our high-interest (mumbles), for the (mumbles) eleven. They will acquire a new system.

A

And also for starbucks, they require that before.

A

The eleventh day they need to have this new framework, new architecture..

A

That means there will be an increase with, of,.

A

Lots of data, and how shall we develop in.

A

A very short time, the architecture and also the algorithm,.

A

And put it into practice?

A

and artificial intelligence requires a neural network.

A

Divided and we have a lot of improvements in terms of.

A

Algorithm, but actually the normal is very complicated,.

A

And it require a lot of computing capacity,.

A

This means we need to have more time to train the framework.

A

And the last challenge also comes from one of our.

A

Client in power generation industry.- and this is a comfort pace,, and I think this.

A

System itself, shift, right, shift.

A

And you can see it coloring,.

A

So what does the report mean,.

A

Actually, it has for this enterprise,.

A

This enterprise has 400 gpu, and how to assign.

A

To the 20 model in the power plan.

A

And here is one nine,: if you can see it, clearly,.

B

It says that in afternoon of six two,.

A

The gpu was assigned to the models, and it will be.

A

Finished by the afternoon of the 20th.

A

And I think many of you come to kubecon,.

A

And we have a lot of case sharing and best practice sharing.

A

In the past one, two days., so one marriage, the gpu source, msljava resources,,.

A

Or china, where have come into an era of cloud native,.

A

But here, when we are trying to manage the gpu,.

A

And concerning ai and also machine learning,.

A

It's like we are still at the preface to our cloud era.

A

So this is a big issue for all of us,.

A

And it is a comfort place provided by one of our clients,.

A

Well, we have a lot of pinpoints and, as our challenge is ahead of us, and as we hurry our algorithm's pinpoints,.

A

And it says the workbench in 2018, they will issue.

A

The enterprise report- and we all know that the.

A

Classical software of development.

A

Basically, there's a two-dimensional issue., we know the properties of the software and then we know the architecture of the software, the landscape. It is really simple. We just engineer it.

A

To implement it,, so we have x axis and some y axis,.

A

But the new generation software development, a part particular to for laboratory, but also we need to think about data processing in that the model compares them and the parameter adjustment.

A

And also the relevant neural network.

A

So it is more than four-dimensional,, even five-dimension,,.

A

And that means we need to think about the time,.

A

We need to think about the time cost,, the cost of time.

B

And it will be dozens of times.

A

Compared with the present era, so type out, what are our solutions?.

A

So I will invite my colleague zhang xin, to talk about it.

A

(Zhang xin speaking echoing english).

B

(Foreign language).

A

So, as I was saying in instant chart, right, now,.

A

We talk about machine learning, is about.

B

Four dimensional,, so in the middle we have.

A

A machine learning code,, and apart from it, we still need to think about the data processing, and also when we're faced with massive data,.

A

How can we allocate our resources, and also, how can we shorten the lifecycle.

B

And all of that we all need to think about.

A

So machine learning is not only about algorithm.

A

And mapping from last year we began to realize the.

A

Pin point, and ruthead, and also google and also tycloud.

A

And together we come up with a new solution,.

A

That is kubeflow. and kubeflow is based on the cubal values and also.

A

Its relevant cloud made up trying to solve all.

A

Of the pinpoints it will be ai, develop,.

A

An ai iteration, and also the ai online production management., all of our path of kubeflow, are still in its infancy. State,.

A

And currently the kubeflow features and functions are as follows,. First of all, we have the distributed computing, structure,.

A

In order to shorten the dimensional model, iteration cycle.

A

And also we have a fifth row plank model,.

B

That project, which means errors now come back.

B

And to have an end-to-end machine learning, flow,.

A

And apart from training, apart from this flow,.

A

And when we implement the project, we also found.

A

That for traditional enterprises, we need to provide them with more features than other manufacturers.

A

In order to facilitate their digital transformation,.

A

And, first of all, what we learn.

A

The most important thing is data.

A

data,, data, databank. and for ai development. The most important asset is always data,, for instance, how to connect data and also how to unify the format.

A

And also how to do data washing and actually data clarification. But if we do not have a very valid data relative.

A

Impact on data management platform for traditional enterprises,, they are still using the traditional way, like a man-made search manual equal to their data and to manage the data manually.

A

And if we are continue to use, and if we misaligned data.

B

And we still need to think about how to train data.

B

And I think we still need to put in a lot of effort. Here.

A

We need to do the error and the trial, and to experiment with the data,.

A

And also, we need to tweak the data when.

A

The result is not satisfactory.

A

So during the way, and in the scenario.

A

Of software development, and when we face difficulties.

A

Usually, we need to have a second pair of eyes.

B

To find a new perspective.

B

and that's why we need to talk about.

A

Coordinated development, and in the level of.

B

Cooperation in the year and as an algorithm, we need to.

A

Have a way to share, we need to have a shared platform.

A

And we still lack this kind of platform to share.

A

And if we have the datamap and if we have the best.

B

And the most steep-able model, then also this model.

A

Need to be (unintelligible) into the cloud.

A

To be accepted by our clients.

B

So we have a few nras,. We have a few trainers,.

A

And it's really easy for us to release the software.

A

But for the engineers and also for software developers.

A

Actually, they know a lot about algorithm.

A

But one problem about organization and they talk about.

A

Engineering actually for our software developers.

A

They still like a lot of those skills in this end.

B

And even though we solve the issue of the deployment.

A

There also we need to think about. We should not only.

A

To guarantee the adaptation of the production scenarios.

A

We should only be worried about what we have under.

A

The development scenario we have a 99% accuracy.

A

But if we let the clients do with it actually.

A

Their accuracy rate will be lowered,, and then we don't use the computer to solve the issue,.

A

Then there are the accuracy rate is even lower.

A

And the, and the study is that, if we have.

A

All through the difficulties, and we do have the flow.

A

And we have new data coming in and ways the changes.

B

Of the data, and model need to be iterated as well.

A

And also the speed of iteration need to be even higher.

B

Than java, and when we have the new data and we need.

A

To have the new model,, all the pain we suffered.

A

Will need to do it again, and therefore we need to think about how to do the ci and also.

B

(Unintelligible) to do a very fast thing.

B

and when we have the unique software and can.

A

We need to monitor the reliability and also the consistency and also the resource consumption.

A

Of the model,, and also because the behavior.

A

Of the consumers is changing and also the data is changing.

A

We need to think about whether the model.

B

Is as accurate as we expect.

A

And the kubeflow provides us with very good experiment and starting point., but for traditional industry like electricity, generator.

A

And we can use kubeflow to establish a new architecture.

A

So this is a new concept,, ai os,. We do not want to.

A

Compete with microsoft,, but we need to ride along.

A

Opportunity, and we have an operating system.

A

That is for ai,, so we have the system operator,.

A

We have the data scientists, and also.

B

We have the data engineers.

A

So we have, we want to use the ai os to enable all.

B

These three different stuff.

A

And also, we have, for instance, data and.

A

Resource management and also streamlined and interactive.

A

Pipeline and also we have the distributed.

B

And orchestrated training and recycled model to business loop., and if you want, you can take a mobile phone.

A

To take a picture of this part,, so this is based on.

A

For the models,, otherwise we are ll the.

A

Open forms, and they use the kubenavis to solve our.

A

Pinpoints in resource managements.

A

so pv and pvc, and also to get (unintelligible).

A

To cause a different,, everything shows narrows, and that's how it can better synergize.

A

The different working groups. and also based on programming, we use with ard.

A

And other operators and to realize that that to support gaps and also the agility.

A

And also with the kubeflow,, and we can separate.

B

The pytorch and tensorflow, and also we have the.

B

Operation interface and therefore.

A

For the staff from the traditional enterprise, they.

B

Can use this interface to operate it.

A

And when we apply the ai os to the power generation, industry.

A

We provide them with the smart terminal,.

A

And in informatics we can store all of the data format.

A

Required by the power generations, including.

A

The transmission,, for example, financial, physical data,,.

A

And also operation, data, and just now julia talk about.

A

Different scenarios,, for instance, we can apply this to, for instance, the mami train.

A

And some tracing and the generator mentions.

A

In different scenarios. and lastly, please watch a video, so, for instance, this is a interface.

A

And for the operators we can.

B

Use a different operating space.

A

And the operators will have managers and (unintelligible). and through the online environment we will have the (unintelligible) and all of the ideas are processed in an automatic way.

B

[Male translator], with the data variable from drop of the user interface, we will be able to attack and push out.

B

Those data behind all of the user experience..

B

We have used a lot of open source technologies.

B

We also be open up some of those technologies.

B

To contribute to the society to invest in the future., and we can also modify from the headache to the module.

B

After a simple configuration from the lower hierarchy,.

B

It will be able to initiate a training task and you can also run the interface..

B

It will automatically adjust the allocated.

B

Resources among all the different tasks and jobs.

B

We also have a multidimensional dashboard, not only on the utilization rate of the resources, but also on the active math of the model..

B

Here we can, when you have a multiphase, complete model,.

B

Hopefully, turn into a system that can be launched.

B

We are familiar with the microsoft service max container and the grp and apis.

B

We want to put everything together.

B

And it will become a system.

B

already launched an ai,, that's it.

B

Enjoy your day,, thank you to all of our colleagues.

B

Who have been contributing to our program and the projects. also would like to say thank you that this is the first time I am giving a keynote presentation,.

B

I think the best advantage when giving the keynote is, that I have no questions for you., so anyway, thank you. (applause).