Cloud Native Computing Foundation Kubernetes Forum Seoul 2019, 10 Dec 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Keynote: Kubeflow Endoscopy - Hong Seok Hwan, CEO, Dudaji, Inc.

Description

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: Kubeflow Endoscopy - Hong Seok Hwan, CEO, Dudaji, Inc.

They are a medical AI team. It consists of doctor and S/W engineer. They are actively working on AI projects, including publishing the topic ""Reading gastric endoscopic gastric cancer"" in the journal Endoscopy. This group is interested in doing AI research efficiently. The group is doing AI research with Kubeflow and Katib this year and seeing great results. They share their experiences in this session and cover the following:

(1) Basic knowledge of Hyper parameter tuning
(2) Kubeflow and Katib basic knowledge
(3) Pros and Cons of Hyperparameter Tuning Library
(4) Kubernetes and Kubeflow at the Medical AI Center
(5) Kubeflow and Katib use cases

In this session, the audience will learn how Kubeflow and Katib have been applied to real-world research. If an institute that is doing AI projects inefficiently is looking for an efficient platform, it can be the best starting point.

A

Ah Yes, Hello Nice to meet you.

A

Today I came to the 2nd round to tell you about the Kuplo I experienced. Uh I'm, a little nervous. Yes I'll, see you at 10 o'clock.

A

So, I do a lot of activities related to it. Well, naturally,. There were a lot of dg that I wanted to improve my workflow.

A

In, the meantime, we discovered who flow, and while using it, I, felt that this is a really good tool, and today, I would like to tell you about AccuFlo,, which we experienced in the project of diagnosing cancer and then Amira lesions in gastroscopic, images.

A

So we proceeded with an ai project that reads: gastric cancer from gastroscopic images. In, 2018, yes,. We conducted that research. We call this Plate One.

A

And then, in the second half of 2019,, let's expand the scope a little more and do a study on not only cancer but also lesions. I.

A

Used Flow, but I didn't use it in the first project.

A

There is a very interesting research, timeline. Well, data preparation and then ML code. Writing.-.

A

Oh, we were able to reduce time in a very meaningful way.

A

Looking at it a little more schematically,. We.

A

Were able to progress almost 10 times faster by shortening the optimization period to 2 weeks infinitely for the optimization that took a total of 5 months, and the serving part was also our first In, the second project,. It takes 1.5 months every other year,, but this could be shortened to 2 weeks.

A

And, since this second fire was very fast,, we and the hospital also proceeded with the full project. More aggressively., To see, I recently signed a mou, and this time I am starting a project to actually make a product. So. If this project is successful,, the software that helps doctors when performing gastroscopy in a real hospital is finally completed. Yesterday.

A

I'm, looking forward to being able to put in the product. Going.

A

Back to the point where the first project was finished,, looking back,, we were very dissatisfied with the results. After the first project was over. I.

A

Heard a lot of doubts about.

A

Whether it was, so I thought about whether we could improve the plot, more. There.

A

Were also a lot of things,. So if you look at three major things,, we were manually provisioning resources such as uh gpu,, and we wrote some scripts and used live tuning for Hypo 8 Mart,, but we still did it manually, and especially like the workflow. In, the case,, when necessary, I just had to write a sherry or python script and ran it,. But after I progressed a little with the application, I proceeded with the solution in a very unmanageable workflow, state.

A

Yes, then we thought about how we could improve this. After doing it,. We found out about the 9 Plus,, so we did a little research, and we were worried about.

A

Something. In, the project, we used 9%. The definition of 9% itself is ml toolkit tools. So,. There are a lot of components in the flow after,, but among them, we will explain more about the three components that we are very satisfied with and used.

A

The first is It's a Jupyter notebook. If. There are a lot of engineers. Here, I think it's probably a tool that everyone knows. Well.

A

It is easy and it is possible to put the work done on the created notebook back into Docker and put it into the process of automation in the workflow, or run it from another source. It's.

A

Been strengthened so it's more usable,, it's gotten better. It's car tv. It's. The component. We received the most help from while optimizing. The.

A

Biggest advantage is that you can do other things that you can do without modifying the core ml code.

A

How, can this be possible? If? You sell it, it's possible because you do more Lord. So, actually,. We also did a cares code called Tensor Plug at that time,, but both of them modified only one or two lines of the format related to pain, and licked and sucked it using a funny. Tip. Jun was able to do it very easily.

A

Is a component that simplifies the workflow and orchestrates it. As I said before, after we finished our first project, there were a lot of shell scripts. After.

A

Arranging the workflow,, the workflow becomes visible in the form of a graph,, so there is an advantage that we can execute the desired workflow by triggering when we want. In our case,. We prepare data in the second project. The process of tuning the parameters and the process of serving. The workflow of Yose Dalji was largely defined and used, and I was very satisfied with it.

A

I got 0 out of 7 a few weeks. Ago. The part that does is the part called cape serving. Uh,. This is kat on top of yeast, and kf, serving on top of it. The.

A

Definition itself is ml, 2 kits,, but I feel like it is gradually getting into the shape of a platform as I go toward 1.0. While.

A

Preparing for this presentation, I asked acquaintances around me, and I.

A

Heard feedback that they both have something in common, that they are both picky. In the case of a great white shark,. They say that you have to move even when you sleep, so that you don't die. It's a little sad that engineers have to keep moving,, but it seems like a good fate.. It seems to be a.

A

Very good tool for ai research., A research, lab or a team that has concerns about whether we can do a better project on the project., A research lab or a team shoulder line. If you get a chance,. It would be nice to try using blue after a while. Yes, thank you.