Cloud Native Computing Foundation Kubernetes AI Day EU 2022, 19 May 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Computer Vision Dog Breed Classification with Convol... Konstantinos Andriopoulos, Dorothea Kalliora

Description

Computer Vision Dog Breed Classification with Convolutional Neural Networks, TensorFlow and Kubeflow - Konstantinos Andriopoulos, Dorothea Kalliora, Arrikto

Sick of strangers at the dog park constantly commenting on how good looking your pup is, but at a loss when they ask, “What breed is it?” Me too! Why not use AI to answer the question for you? For data scientists looking for an open source and scalable way to tackle these sorts of problems, Kubernetes and Kubeflow make analyzing content in images and video much easier than trying to build everything from scratch and run it on bare metal or VMs. In this talk we’ll work through the development of a Notebook that leverages the combined powers of TensorFlow, ResNet-50 models, convolutional neural networks, VGG16, and transfer learning to see how accurately these algorithms can predict the breed of my dog. Spoiler alert! I have the genealogy results, so there will be a big reveal with DNA pitted against a variety of algorithms.

A

Said uh hello, everyone, as we said this session, is about computer vision, dog breath classification with convolutional neural networks, tensorflow and kubeflow, I'm dorothea, I'm a software engineer at rikto and yeah. So what you'll see in this session uh today, we will build a jupiter notebook that leverages the combined powers of kubeflow, tensorflow, cnns and transfer learning. In order to see how accurately we can predict the breed of a dog, then we will use gail to run this jupiter notebook as a keyflow pipeline.

A

uh Keyflow is the machine learning platform we are going to use for those who are not familiar with it. Kubeflow is an open source project that makes the deployments of machine learning workflows on kubernetes, simple, portable and scalable. Let's see how kubeflow is used in the machine learning process.

A

This is a high level overview of the machine learning process. A machine learning engineer would follow.

A

The process begins with experimenting with machine learning tools and frameworks such as spiders or tensorflow that we're going to use today to perform machine learning tasks.

A

Then kubeflow encapsulates applications and services to run those machine learning tools on top of kubernetes clusters. Lastly, kubernetes runs on-prem or any other cloud.

A

Now, let's have a closer look of the machine. Learning workflow as a data scientist would experience it. Data science begins with identifying the problem and collecting and analyzing the data. Then the data scientist has to choose a machine learning algorithm and code their models.

A

Subsequently, they experiment with data and model training.

A

After they build a good enough model, they can optimize it with hyper parameter tuning and, of course, in the very end they can serve the best model uh that produce the best result. As we see, kubeflop provides various components that implement all the aforementioned stages.

A

In this talk, we will focus on jupiter, notebooks, kail and kevlar pipelines.

A

So can how can someone interact with kubeflow? Let's have a closer look in some key components for this session.

A

Here you can see the keyflow central dashboard. It is a graphical user interface to manage notebooks volumes, snapshots and other various components. Qflow also has its own apis and sdks that you can interact with. One of them is a pipelines sdk, which we will see in the next slides.

A

We also saw that qflow intergrates with jupyter notebooks.

A

These notebooks run as containers inside the kubernetes pod and enable users to run web-based development environments inside the kubernetes cluster. They also expose a jupiter lab id to the user's browser used for interactive data science computations.

A

Lastly, they have persistent volumes attached to them, which store installed libraries and data plus enable snapshotting and reproducibility.

A

Another basic component of kubeflow is keyflow pipelines, as we said here is how the queue flow pipelines interface looks like keyflow pipelines is a platform for building and deploying portable scalable ml workflows and because- and it consists of a ui, as you see in this slide- for managing training, experiments, jobs and runs a stable sdk for creating pipelines and components.

A

And, lastly, an engine for scheduling, multi-step, machine learning, workflows.

A

We also saw that keyflow integrates with k, so what is scale kale simplifies the use of kubeflow, giving data scientists the tools they need to orchestrate and to end machine learning. Workflows. Let's talk more about its main features, offers a ui in the form of a jupiter lab extension.

A

As you see in this slide with scale, you can enable it and annotate cells with within jupiter notebooks in order to define the pipeline steps perform hyper parameter, tuning use a gpu and track metrics.

A

Kl also provides apis that enable users to create pipeline components and kfp dsl resolve the dependencies. Inject data objects into each step, deploy the data science pipeline and also serve models.

A

We see how kale converts jupiter notebooks into pipelines, but we haven't yet addressed the importance of pipelines in the machine learning workflow.

A

Well, if you really think about it, when working in a jupiter notebook, you already follow a workflow that looks like a pipeline cells can be viewed as separate steps that depend on each other pipelines follow the same logic. They consist of clearly defined steps that can even run in parallel as isolated code execution units.

A

This enables hyper-parameter tuning of a notebook and its steps. Also, it becomes very easy when running pipelines in the keep flow pipelines to apply data versioning and reproducibility to these independent steps.

A

The various cells of your notebook that become in tibetan steps might have different requirements. For example, the training of a deep learning model might need a gpu, whereas another cell may not all the above are feasible with kale and rock.

A

So, let's talk about how kale and rock can dramatically simplify the workflow of building a kefla pipeline. Here is the machine learning development process. Without these components, as you can see, you need to follow various steps. First, you need to write your machine learning code then create docker images, write the dsl kfp code, compile it then upload the kf3 pile the pipeline to kfb and, lastly, run the pipeline.

A

So what happens? If you make a mistake, and you need to amend your machine learning code, you would have to start from creating docker images again with your new code and follow the next steps. As you see here, every time you make a mistake, when you introduce scale and rock this procedure is simplified to three steps, you will write your machine learning code.

A

You will tag your notebook cells with scale and run the pipeline with a click of a button. If you make a mistake this time, you just edit your notebook, so this reduces the iteration time by seventy percent.

A

uh Today, uh since we have introduced all the basic components, we will talk about keyfloss service. We will use quiplus a service to run our dog breed demo.

A

Kiplus a service is the fastest way to deploy kubeflow.

A

Here you can see the central dashboard you can get started with q flow in minutes with just a click of a button like I'm going to show you now so um give me a second.

A

A

A

You can help me go to the script.

A

Okay thanks uh so we're going to open a new window and type this url qflow.director.com.

A

You can sign up here with your credentials and create a new account. I have already an account, so is this email.

A

Let me type my password.

A

Oh, it's, okay! I can change it later.

A

A

Okay, so here's the central dashboard that you saw earlier, uh I have already at employment, but if you click here, you can create a new queue flow deployment. Just by clicking this button, this procedure would take under 30 minutes. We can click create.

A

You can also see a video of how you can use cube flow as you're waiting.

A

So this is creating uh and we can view the ready deployment, here's the credentials. We have a username and the password. We can copy our password and go to this url.

A

We will use user as the username and this password in order to login, and here you can see the keyflow central dashboard that we saw before with the various components. For example, we can go to pipelines and you can see some ready pipelines that we have prepared.

A

We can go over to the slides.

A

And now I'm going to hand it over to constantinos to show you the demo.

B

uh One two one: two yeah, so hello, everybody, I'm constantine sandrioplus, I'm a software engineer director along with artheia, and you can call me costis. You know later in the questions it's much easier, so um the rossell showed how we can easily create a kubeflow as a service instance and access cube flow really easily. So in this part of the presentation we are going to run a demo for you. We have prepared the jupyter notebook.

B

We are going to convert it into a pipeline and you'll see how we will use cubicle as a service to do that. So let's jump right in and talk specifics about the example itself.

B

So, as the title says today we're going to talk about the dog breed classification problem, it's based on a really famous udacity project. Some of you may know it, and the purpose, as the name suggests, is to classify the images of dogs according to their breed. So it accepts any user defined input image and you can easily you know, go ahead and take a picture of your dog upload it and use our trained models and get a prediction of your dog read back.

B

So for the purpose of the demo, we are going to use cnns, convolutional, neural networks and transfer learning. So let's jump right in and talk a bit about this part of machine learning.

B

So, as I said, convolutional neural networks uh is the framework, the type of deep networks that we are going to use, and they are a huge class of artificial neural networks and they have infinite applications in image, classification and image analysis.

B

So an interesting feature about cnns is that uh their architecture consists of, let's say, windows that are in a way sliding across the image and they detect different types of interesting features in that image such as corners such as edges.

B

So that leads them to understand the contents of the image like items that are contained in the image or dogs in our case and the framework that we are going to use in order to implement all this in. This demo is tensorflow, I'm sure you know tensorflow.

B

We chose it because well, of course, it's open source. It has an infinite amount of applications. um It's excellent for training and inference of deep neural networks, and it is also natively supported in kubeflow.

B

So the first pre-trained cnn that we are going to use is called resnet50 and, as the name suggests, it's 50 layers deep, it's trained on imagenet. It has over a million different images and about a thousand categories of different objects contained in these images, such as keyboard, mouse pencil and many animals such as dogs.

B

So the next one we are going to use is called vg16, as the name suggests again, it's 16 layers deep and it's a cnn with infinite applications in image analysis.

B

It's also used in the imagenet visual database project and um well, as you probably guessed, it's really really popular.

B

So, let's jump in the sections of the notebook that we are going to use for this demo uh in the beginning, as any data scientist we're going to do a couple of downloads imports define some pipeline parameters because this notebook is going to turn into a pipeline.

B

uh In the end, we are going to load that in memory and then we are going to create three different uh types of cnns, the first one. We will build with our own bare hands using tensorflow.

B

The other two are pre-trained, so we are going to use transfer learning on top of them in order to use them in our specific problem in dog reclassification, then we are going to test them and see a couple of visualizations in order to see where we're at okay, so we can now jump in the demo.

B

Let me change the screen here: let's go here, okay, so as you as dorothea showed you, we have a instance ready to use so we're going to go this way. We are going to copy our password and we are going to connect to it right. So let me move that a bit in order to be easier.

B

So, as you can see, we have a uh kubeflow dashboard here, uh the ones that we saw in the slides, let's go ahead and create a new notebook server. In order to run our examples, so I'm going to go ahead, hit your.

C

B

I'm going to give it a name.

C

B

To kubecon all right so as you've seen, I have lots of other cube instances. Here we are going to add a couple of more gigabytes of volume, let's say 15 and I think we're ready. We can click launch and uh essentially, what we are doing now is.

B

We are creating a notebook server that runs a jupyter judo lab image, but this image not only exposes a jupyter lab ui, but it also this jupyter lab ui is extended by chaos lab extension, so we are going to see how kale goes over the jupiter lab ui and gives you the opportunity to create a whole bunch of different annotations in order to translate cells into actual pipeline steps, so we're connecting to it.

B

You see the jupiter lab ui pop up yeah, so this is empty. We need to download our uh notebook. So let's get the terminal. Our notebook is hosted in our github organization. Page it's under the examples repository. So I'm going to go ahead and clone it without seeing actually, let's see how that goes,.

B

It goes okay, yeah,.

C

B

Oh yeah, I forgot clone. Thank you.

B

Thanks for the help also, so if we refresh here, we can see that it has a new examples. uh Folder the file viewer. Here we are going to go ahead and pick the dog breed one, and we need this this notebook. So um we have prepared this for you. Let me go ahead and download this so that we don't do that later on um and we need to install also a couple of requirements.

B

So what I downloaded there is the data sets and some of the pre-trained features about the models that we are going to to use uh to classify dogs, the resident 50 individual g161 that we said before. So, let's go ahead and take a look at the notebook, okay and see here that we have a new icon on the left.

B

It's called kale and we can hit enable- and you see the difference in the ui now kale gives you the opportunity, as I said before, to annotate uh notebook cells and define them as pipeline steps or pipeline parameters that you can tune each time you run a new pipeline and also you have import cells and you have a whole bunch of different cells. Actually, so let's go ahead and see now that the notebook is annotated, um let's go ahead and see how you know what each cell does so.

B

Essentially in this step we load our data set. This is not that important. We can go ahead and see what we do with our models.

B

So you can see here that we are annotating a cell as a cnn from scratch, I mean this: is the cell that defines our uh newly created um cnn model and you'll see that we are using the sequential api by tensorflow in order to create it? It's not a very deep model, so don't expect really fascinating results, so we go ahead then, and compile it, we train it, and then we do a couple of tests.

B

We report test accuracy and also we'll you know, create a couple of visualizations for you to see so in another step or in another bunch of steps that are all annotated in with the same annotation, you can see that we are using the vg16 model and essentially, what we are doing is that we are adding a uh another layer in the end of the model, and this layer uh is not a pre-trained layer as the rest of the model is.

B

We are going to use this layer in order to do classification on top of our specific data, set the dog bit data set, so we compile the model, we train it and same thing. We do a couple of tests um and a couple of visualizations and predictions the third model, and the last one is the resonant 50 model. We do exactly the same thing as we did with vg16 and well. I think that these are the important parts of the mode look. We can go back and see how the downloads are going.

B

Okay, yeah. We have a couple of uh installations painting this take like like a minute, so essentially what we're going to do now now that we have set up our environment, uh we are installing everything and we uh have downloaded our data set. We have loaded uh our models, etc.

B

Now we are going to hit, uh compile and run, and when we do that, kale will automatically validate the notebook, see if there's any missing dependencies there um between the steps and it will, it is going to take a snapshot of the volume, and this snapshot is essentially a picture of the volume. Let's say that we can reproduce whenever we want and kale is going to create a new pvc out of that snapshot, and this pvc will find its way in each and every step of the pipeline. That's going to be created.

B

So this way we persist all of our user environment to the pipelines that we are going to create. So it seems like it's working. We can go ahead and view it and you'll see the execution graph. That kfp exposes it's in an early stage. The first step is creating a volume so because this is going to take a bit of time, we can actually show you oh see next step.

B

uh We can actually show you a pipeline that we ran well yesterday and you can see here how these three steps are the ones that we use to create our model and train it, and you see how they run in parallel, but they somehow depend on the same step. On the previous step, you see, you know all the dependencies using the arrows, so let's go ahead and click you know in resnet and go in the visualizations tab and you'll see that we have this html window here.

B

So, as you can see, the test accuracy of this model is 82 percent. Over our test data set, and we have uh shown we have uh uh decided to show to show your visualization of a specific part of the data set this dog. This is a dalmatian. This is the ground truth, and this is what the model predicted. So we see really really good results early on. uh The next model that we are going to see is the vgc-16 again visualizations tab.

B

You can see whole bunch of visualizations that uh you can do not that good, not that good of an accuracy it's 72 percent, but again it predicted correctly this specific dalmatian dock. We can do the same thing for our cnn from scratch, and we have some uh things here too. If we go and scroll right back like this you'll see that the test address is not that good.

B

At this point I mean uh we expected that it's like a really really not that deep layer, so it cannot be compared to resnet and the other two layers um so yeah. This is it from the demo. We can actually jump in the slides, and you know I want to end this by talking a bit about uh our community. um I would like to conclude by saying that in arikto we are we're really really passionate open source contributors.

B

We have contributed to many too many components in kubeflow, we've uh contributed to the jupiter management ui um to flow pipelines to our own opinionated version of kubeflow mini kf up to the to the linux kernel itself.

B

So a bit more about cube flow qflo, as you may know, is a great community made up with uh many many big companies that are actively contributing and it is a community that is always always trying to find new members new contributors that are passionate users, so please feel free to get involved, and here you can find like a lot of pointers from our uh github organization to our slack community.

B

We are really really happy to support you and to onboard you. So without further ado. I would like to say thank you. Here's a couple pointers and I hope you enjoyed it.

D

Now do we have any questions.

D

I can't see any hands up. Anybody have any questions. I I have one question so I I found out earlier today that we have quite a few dog lovers in the room. Yeah can can these people some of these people might have dogs, then they need classifying. Can they drop an image into the slack channel? Will you classify for them? We can come back to that later on yeah.

B

Yeah, as we said, it's a public example, so if you just go ahead in our github account download it and run it, you can actually take a picture of your dog and see what the actual breed is.

D

See that later on, oh so, we do have a question: okay,.

C

uh Friend, congratulations on the presentation. uh From the our point of view, is it possible to predict a mixed breed, uh something like uh percentage of.

B

Yeah of a combination of two or three breeds you say well, it actually depends on the data set that we used. I think that the data set that we used, although I have not seen all of the targets, but I think it's all uh unique breeds.

B

So if you create a data set that has different targets, you know extends the one that we used and has a combination of others and also provide pictures of these, so that you can train your model, then sure, but I think this one doesn't did.

D

You have another question.

E

About the kale product, yes, it looks very cool, but it looks a kind of deprecated product. It's just twist, 500 stars in github, and the last comment is like for the october last year. Yes,.

B

Yes, so uh what we showed you here is actually our um our own private version of kale that comes with kubeflow as a service, so it isn't very in sync, with what you see in our upstream repo in github. It has a whole bunch of new features, and it's you know it's in a really really updated stage than the one that you see in the open source.

D

Any more questions jessica. Do you have any questions in the slack channel.

D

Okay, great stuff, okay, thank you very much.