.Net Foundation On .NET, 5 Feb 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Learning Machine Learning with .NET, PyTorch and the ONNX Runtime

Description

ONNX is a open format to represent deep learning models that is supported by various frameworks and tools. This format makes it easier to interoperate between frameworks and to maximize the reach of your hardware optimization investments

In this episode, Seth Juarez (@sethjuarez) sits with Rich to show us how we can use the ONNX runtime inside of our .NET applications. He gives us a quick introduction to training a model with PyTorch, and also explains some foundational concepts around prediction accuracy.

[T01:01] - Brief overview of training a machine learning model
[T02:54] - PyTorch training in Python
[T07:43] - Understanding accuracy and loss
[T10:50] - What is ONNX ?
[T11:53] - Consuming model with ONNX in .NET
[T25:38] - ML .NET and ONNX
[T26:58] - Industry usage

Useful Links

ONNX on GitHub
https://github.com/microsoft/onnxruntime

ONNX NuGet
https://www.nuget.org/packages/Microsoft.ML.OnnxRuntime?WT.mc_id=ondotnet-channel9-cephilli

ONNX Runtime is now open source
https://azure.microsoft.com/en-us/blog/onnx-runtime-is-now-open-source?WT.mc_id=ondotnet-channel9-cephilli

PyTorch
https://pytorch.org/

Seth’s Pytorch demo on GitHub
https://github.com/sethjuarez/DeepLearningWithPyTorch

A

Today, on.net, we're going to be working with seth juarez,.

A

To find out magic numbers using machine, learning. check it out.

A

Welcome to another episode of on.net., this is actually our first show that we're shooting in 2019.

A

What! yeah. so, we've got seth juarez who's going to be telling us about machine learning.

B

A

Py torch, and stuff like that. yes,, but mostly .net., but mostly .net. so,. I actually think this is a little bit weird. I'm interviewing you. why? yeah, cause well you're. The pro interviewer, here at microsoft.

B

Who is this guy?? I mean you know, I'm saying?

B

yeah. so,, I'm going to try my best today.

B

okay, and if you disappoint me, you know that we're going to have that thing that I said.

A

Yeah. okay, for sure.

A

so,. What is our our topic?.

B

Okay. so, today we're talking about the onyx runtime, and how to use it inside of .net,, and for many of you this might be a little confusing right,.

B

But when it comes to machine learning,, the cool thing about it is that there are two phases when you're doing machine learning, specifically with deep learning. okay.. The first phase is a data scientist, or programmer,, or even you, because I think anyone can do this.

B

If you train a thing called a model,.

B

And you think of a model as a function, or an asset, or a jar file or an assembly or whatever it's something you call.

B

So, once you train that model, the question becomes,.

B

Well,, now that it's done training, how do I actually use it?.

B

Totally. so,: that's what we're going to be focused on today., hopefully I'll get a little into the training..

B

If you have any questions, I'd love to answer any questions about how these things are trained., what it's doing,, but once you have that model file, all you got to think about is bring the file into my project.

B

Put some stuff in, and get the answer.

B

right. so,. It is a little bit like.

A

A gymnast trains for four years does all of these contortions.

A

Yeah. then somehow they go to a different city where.

A

The olympics are being held and they actually like basically do their model, and they end up with a gold medal?.

B

Yeah, one thing. yeah. you're, absolutely right.. There is a lot of contortions. yeah., that's basically all of machine learning,, literally the secret here. but, yeah,. It's a little bit different in the sense that,.

B

Imagine you have a,, let me see if I can come up..

B

Imagine you'd like set together your sprinkler system, and for those who don't know,, we have automatic sprinkler systems, in some part of us and you turn it on, and then water goes through the pipes. but,. Imagine being able to turn off enough valves, so that you can water all of your lawn at one time the best way possible.

B

right. so,. What happens is a model is basically like the pipes where you turn the knobs, is basically what we're doing.

B

we're trying out a bunch of settings in the knobs so that we can cover all of the things in this case..

B

In this case we're going to recognize numbers. okay.. So that's is it. so,! You can see here if we can go to my screen,.

B

Basically, this is just py torch.

B

You can get this is a github repo., we'll put a link on it., the deep learning with py torch, and under my github account, which is seth juarez..

B

Basically, these are the pipes I set up..

B

I set up this thing called a convolutional neural network.. These are the pipes., then what we do is on the interior there's, basically a bunch of.

B

Numbers that when another digit that we drew comes in,.

B

We multiply, we passed down., we do some things. We multiply passed down.. We need to find what those numbers are., as you can see right now, that's basically what it's doing,.

B

Because I ran it before we started., you can see that it's just going through and trying to figure.

B

Out what are the best numbers to multiply this digit? We drew by,.

B

In order to get the answer right,- and this can take awhile.- there's some models that I've run before that have taken days. some take weeks.

B

but basically, once you're done with the training,.

B

You have this file with a bunch of numbers.

B

right. so,. You have to say,.

A

Hey boss, I can't do any work for the next two weeks.

B

Basically., I'm training this model. or you could use our cloud. yeah. we rent by the hour I believe.

B

yeah,, something like that. so,. We actually should give a pitch to vs code.

A

Yeah. so, is vs code, a good ide or editor for python?.

B

Absolutely,, I absolutely love it. the reason why it's not just because python is sort of built in, and I can do some crazy things like for example,.

B

What's this "module" thing. nice., I can actually go to the definition and in python like it will bring it up.

B

Right,, like not just like here's, where the assembly is because it's literally this is the code.

B

So, a lot of the times when I'm looking at how this stuff actually works, I can go, and see exactly how it works.

A

Right. so,, you can understand it.

B

Yeah. absolutely. but again,. This is python stuff, and it may not be interesting to you as a .net dev.

B

The only thing I wanted to point out is that we have this loop right here..

B

You can see this loop here going throughout,.

B

Is that we go through all of it.

B

so,? We have 60 thousand hand-drawn digits that are 28 pixels by 28 pixels.

A

Okay,This is your test data?.

B

This is the data we use to train. yeah. so,. What we do is we give this thing all of this test data. It's like the water going through the pipes..

B

This loop right now is going through and adjusting all the knobs.

B

Right, trying them in different ways.

B

once it's figured out the best way to do it,.

B

Imagine pulling that out of the dirt with the knobs intact.

B

Then we're going to put it somewhere and just turn the water off.

B

right.. That's what we're doing. so here, right, now,, I'm running through this takes about 15 epochs, but it's already about 98 percent accuracy.

B

right., but basically the output that you get right. Here.

B

Let me go to reveal an explorer, is.

B

This thing called "model.onyx" and if you zoom in,.

B

This is a really cool. This is called netron..

B

It's a really cool program for looking at, like neural networks.

B

You can see that their input is this thing that they labeled "zero"..

B

Then it basically goes through the "w". What the size of it,.

B

And that's just a bunch of numbers, the "b" that we said we multiply, and add with this..

B

We add this on. Then we make it smaller.. Then we run a function.

B

Then we do this all the way to get an output, that's of size, ten.

B

You can see that one by ten here.. Basically, we want to see if that is between zero and nine,, and so there's ten.

B

so, in the 0th index. If you get a one there, that means that it's zero., if it's not sure you just pick the "max index", and that's the right answer.

A

The reason that there's ten of these elements is because.

A

All of these images are only go from zero to nine..

A

That's right. so,. If the images were like zero to 99,.

A

Then we would actually have a hundred elements to choose. From?.

B

We could, you could do that, but that might be a harder problem.

B

You could probably create two machine learning algorithms, one to detect a box around a single digit,.

B

And then pass that to this model. I see.. Do you see what I'm saying?.

A

I see. so,, you could do both like I said,.

B

Because the most important part of data science is like the science bit,, where science is literally guess and check, but more rigorous and we write stuff down.

A

I see. so,, the smart person would do the thing you said: let's make this problem simpler., let's take apart these two numbers,.

A

But then still deal with them as a single digit.

A

maybe,, I guess it depends.

B

On how people write numbers the same way.

B

but,, let's just say it was handwriting recognition..

B

The script was russian. yeah.. Russian script is notoriously difficult to understand because everything kind of runs, together.

B

maybe,. It wouldn't work that way. maybe,, it's better to learn, words, or maybe it's better to learn something else., so it depends? it depends. yeah., I'm never going to say it. so,. Now. This thing, I think, should be about it's.

B

A one more epoch where you can see that it's getting 98 percent accuracy on the test, data,, because what happens is once we trained on the data.

B

We leave some other data out, and we see how well are you doing on this data that you've never seen.

B

exactly. and that's the test data set.. The other thing that's important is this loss. Function., this loss function basically is measuring. How bad are we at it.

B

If the answer is zero, then we're zero bad at it, which is great.

B

If the answer is 50 million or 50 million bad about it, which is not so great.

B

so, as this single loops down, what I'm looking for is increased accuracy,, decreased loss, and loss, just measures the difference between what we predict, and what the right answer: is., right, and so,. It sounds like there's.

A

Kind of a real relationship between those two numbers, but not exactly because they're determined in two very different ways.

A

yes. so,, the difference between the loss, function,.

B

And the accuracy is the accuracy you think of it as an absolute like it either..

B

The max index is either right or it's wrong.

B

but, for the loss function. It's going to be filling numbers out through the whole thing.

B

so,. When you subtract those continuous values, you're going to get a continuous answer,.

B

Right, and for accuracy it's either right or wrong, and you count that up. I see.yeah. so,. This should be the last epoch.

B

What it's going to do is it's going to output that "model.onyx".

B

so,. I better close this before I get like a "file open, error".

B

Because then we have to start again.

A

≫ yeah, because we have to start all over again,.

B

But, like I said it's pretty simple. just think of this. As a ai, you have to pick a model shape apriori.

B

Then you pick some hyper parameters, like the learning rate within learning rate is like how fast do we learn this stuff and another batch size and epochs how many, and then you run it..

B

Then it's either good or it's not..

B

If it's not good, you switch to different model or you switch. Hyper parameter is wherever.

A

Right., as soon as you get into this,,.

A

You either have to teach yourself, those sorts of things so that you can direct your efforts.

A

Or it would help to have someone like seth on your team.

A

Maybe. so, that they could give you tips on what to do. Next.

B

So I am like this school train data science, and I had been making demo where, for the most of my career.

B

so,, we have other people on our team like francesca lazzeri,,.

B

Who is a bonafide data scientists,?

B

They have a good horse sense about these things. Right, like they look at the learning rate and they look at the loss function..

B

If the loss function goes, up, they'll automatically be like the learning rate is too high..

B

So he's like a model whisperer?.

B

yeah,, she is. she's, pretty amazing.

B

for example,. If your batch size is too large,.

B

It might not be working, or for example, if you notice that it's not learning very fast or the difference.

B

Between your training error or your test error is either too high or too low,.

B

You start to get a sense for what is actually going on, and once you know what it's optimizing, it makes it a little bit more sense., but once you pick the model function, once you pick the right, optimization method,, once you pick the right hype-parameters,, then the output is this onnx file.

B

right. so,? How do we use it?? That's a great question. so,. Let me describe what onnx actually is. notice. It got about 98.4 percent.

B

so,. Are you happy with that? As the?

B

no,? There are some that do better,, but I wrote this myself the other day, and so maybe I should tweak it a little bit. Better. right?.

A

What would make you happy would be like we're done, and this was good.

B

I think a 100 percent accuracy feels wrong to me because it over-learned.

B

yeah. there's something called over fitting.. But if you get 99.9, I'd be okay with that.

A

B

So basically,, what we're going to do now is we have this onnx file that it just built, and what I need to do is I need to use it inside of a .net application.. Now let me explain what onyx is onyx is the open neural network exchange format.

B

Think of it as like the pdf for neural networks.

B

so,, anyone can build anything anywhere,, but if you can export your model, which is like those pilots with.

B

All the gauges, pre-turned correctly,.

B

Then you can use that almost anywhere and we happen to have something called the onnx runtime that you can literally just pull in with the nougat package,.

B

Just pulling in the nougat package for the onnx runtime,, and I don't know where it is- that we're a system., it's like system.runtime, there. It is the onnx runtime.

B

Yeah. that makes more sense. the perimeters for that and lots of pixelation,.

B

And all this, oh, is that that's pretty you like that?

B

yeah.. I want everyone to see,. I could zoom in even more for you.

A

No. that's good.

B

So, there's two things that happened once.. Once we do that,, we basically need to load the file and wow. This is this feels like it's really small. so,. Let me,.

A

It's a bit to do this. Guy down, here.

A

enbigen this to 200. okay.

A

B

So,, what I'm going to do is the first thing is I'm going to load the file. right? because it said onnx file., I'm using wind forms just because I wanted everyone to know like this is just regular .net. I think it's a .net core thing.

B

So,, the first thing I do is I create this inference. Session.

B

What I do is I keep that it globally as a global variable inside of the class, because I might do more than one inference,.

B

And I don't want to have to load this file up.. I know it's not very big. How big is it? Let's see, here.

B

It's like 99 k.

B

so, and that's not going to kill you.

A

It's not going to kill you,, but there are some models.

B

Are like 250, megs., okay. so,? It varies a lot.

A

Yeah, it depends.

B

This is a really small model, with not a lot of numbers., so basically, once I loaded up,, I'm saying well, I'm going to.

A

Can we just pause for a moment? the reason is,, because your domain is small.

A

Yes., because you said the images were like 28 by 28 or something.

B

28 by 28 pixels.

A

Yeah,, so that's already a very small,.

B

And it's basically black and white.

A

It is black and white grayscale. yeah. so,, just the amount.

A

Of possibilities that can exist in that space is very.

A

Small, as opposed to a model that is capable.

A

Of reading feature-length movies.

A

And giving you some sort of feedback on it.

B

Yeah. so, for example,, if you want it to look at a feature length movie and.

B

Predict whether there was a I don't know, a dog in it.

B

That's difficult right because now we just have a single vector that 784, that represents the pixels itself.

B

But now we have a 40 tensor for the height, width, length, color.

B

So, rgb or maybe even alpha channel and finally time.

B

So,, it's a 5d, huge thing, and we got to pass this in, and generally, the more stuff you pass it in, like you've, got to have a better predictive model,, and these things tend to get pretty deep.

B

yeah. so,. That would be bigger. yeah. That would be a much bigger or even simpler. One would be like looking at a picture and predicting whether it's a dog or a cat. for example,. The picture is like, let's just say it's 500 by 500 pixels picture..

B

You have 500 by 500 by.

B

Three, which is the pixel depending on the channel as a channel, goes in the front.

B

so,, it's three by 500 by 500, that one goes into the network and then it does multiplication.

B

But you have to a-priori- and this is the bit.

B

That's cool, is you have to actually pick the model function.

B

I pick this one, for this. It's pretty small., other ones are much bigger,, and this is again we're speaking strictly about computer vision..

B

Nlp natural language processing uses networks which are considerably different, but similar.

B

okay., yeah., okay. So once it's, did I answer your question?

B

yes., all right. so,. This is basically coming from the you can see.

B

The inference session comes from the the onnx runtime,.

B

And I load it up from the file and that's it.

B

then,. What I do now is when I want to actually do the prediction,.

B

I have to get the picture. so,. What I'm going to do is, I think I set strategically setup breakpoints and I'm going to run this,.

B

And the first thing I'm going to do by the way: it's all on github, we'll put a link below you, can totally use it.

B

somebody will load model, and I'm going get this.. This is the output.

B

right,. Does the thing that you just did.

A

We literally just built it. so,, I'm going to hit this and look,.

B

See it's using a local model here,, and now I'm just going to with my finger, draw the number three.

A

Okay and it's not,, I mean to be honest: it's not even a very well-written three.

B

Yes,- and I knew you were going to bring that up, because I did that on purpose..

B

It might not do a good job because maybe it didn't have enough digits- or maybe it never saw something like this, and maybe it didn't reset space of the optimization.

B

so,, I'm going to hit recognize, and the first thing we need to do is we need to convert that into a 28 by 28, pixel image.

B

Right, because that's all.

A

Because that's all it knows. yeah exactly. so, notice that I'm doing that. I'm using like,.

B

I think I used inkcanvas, inkcanvas on a winforms application inside of like a,.

B

I don't know what they call them, where they can put his xaml inside of winforms, but I did that..

B

You can see. I resize it to the appropriate size which is 28 by 28,.

B

And then a create a single float array, that's size by size,.

B

And then for each one of them, I take the average of all of the colors and subtract 255.

A

Makes sense because the bitmap is color oriented.

B

A

Yeah. this one is not.

B

This one is not. now,. I return this thing. so,. If I go to run the cursor here, you'll see that this is basically just a vector or an array of floats.

A

It's just an array.

B

Yeah. all right. so,. Now, let's go to the other one.

B

Let's see if I put a breakpoint there because that's, yes, I did. so, I'll hit at five. notice that it gets the digit in there, which is a float array and now we're using standard .net types.

B

The tensor of t is a data type that was introduced, probably maybe a couple a year, ago.

B

I think it was in dynacore, 2.1.

B

so, tensor of t is a way of nicely indexing into memory, as if it was an array,.

B

But it holds it in contiguous memories., I'm not sure the internals,, but I think it uses span of t in memory of t in order to make sure you could do that.

B

in fact,. You can use tensors and marshal memory into like c applications.

B

Yeah,, which was a great in the office program. then when they write the c program, writes into it because you're using memory of t,, it's writing directly into the memory space of the .net application. right. so,, no copying. yeah, which is really good.

B

so, for example,. If I wanted to write my own linear, algebra package,, I could literally call into kubla law pack. For example, that's written all in fortran c,.

B

And it will write into that memory for me., so now,. We can build some really efficient algorithms that use all the stuff that we it used. For.

B

so,, that's one of the cool things that I like about this. but basically,. The reason why we use tensor of t is because that's what the onnx runtime expects for the input.

B

right.. The next thing I'm going to do is I normalize the image, because that's how it's trained., now, you're, probably running wondering what that means.

B

Let me run the cursor, there, you go. basically,. What happens is instead of training on zero to 255,.

B

I divide all the numbers by 255. so,, because it's a number between zero and one.

B

right.. That's how we trained it. so,, that's how we have to pass it. In.

B

now, you're, probably wondering what is this bit?

B

well,. This is to describe the input, and I'm creating it from this tensor.

B

I'm creating a tensor from this named onnx value and like,.

B

Where does that come from? well,? Remember when I said we looked at this model and.

B

There is an input that was named zero. That's the thing that's going on.!

B

I don't know why it called it zero.

B

when it exploited it just did that. it doesn't really matter., it doesn't matter,, but that's the name of it. so,. If you don't know what it is, you're like, "oh shoot. What is it?, you can just look in using this or you can go into the session itself..

B

Did I right-click it here?? Let me see. lets go to "quick watch" right here, and you can see that in the session, we have this input, data and zero.

B

Yeah. so,, that's how I knew. that's the one to use., okay. so,. Now that we've done that,, we do the named onnx value then- and I always put this in a try-block just in case something goes horribly wrong..

B

I do a session run. But.

A

Yeah. so,, here's a good example. You were talking about,.

A

This is just running on dynaframework.

A

yeah.. I can tell from those version numbers.

B

I am going to circle this, so everyone can see your phantom finger that you will not see on the screen.

B

yes. rookie mistake. yeah, totally. okay. so,. Now I'm going do the prediction: here. so, I'll do f10..

B

I do a run and I just get the first answer. Out,.

B

And I go to tensor to array and you can see that now it's basically a vector, that has a bunch of shoot..

B

Let me just do a quick wide, so it doesn't disappear. Here.

B

so, I always, there. We go, quick watch., it's literally a vector of size, 10,.

B

And that's going to tell us what number it thinks it is.

B

Right, and we pick the max index,, so what did it guess??

B

I'm guessing it's this guy right, here.

A

Yeah., I will point, I'll use.

B

Sorry. rookie mistake again: there. yeah. gosh! this one right: here, right., yeah., that's.,.

A

Because the other numbers are very, very small.

B

Very small., even though they have a number.

A

Before the decimal place, they're- minus 14.

B

Yeah. so. > >. Maybe you should fix that like put like a little bar chart or something.

A

Something like that. right? so,. It's extremely confident that it is three.

A

That's right.. What I do is when I return it,.

B

I basically return like the,.

B

That it's not an empty prediction. that the prediction is the max index.

B

so. array.indexof,, I'm bringing pretty fast to loose with this, but I just wanted to show..

B

Then I select all the scores. Then I make a list, because that's what it expect to do. now,, I don't have to do that.. I could just return the array, but I had to set it up this way,, because this also will call into rest endpoints to do the same kind of inference., so, I'll. Do f5 and you'll see that it thinks it's three.

A

Right. we have that, those scores shown in a bit more of a human-friendly form.

B

Yeah., that's right.

A

Now,, the 250 seconds, that was more about how long it took you to get them out of those.

B

Break. yeah. yeah.: let's do it again: here. so, we'll do recognize and you-. Oh shit.,.

A

Yeah., I think you'd have to turn off all the breakpoints just.

B

Turn off all the breakpoints.

A

Just saying. just saying.

B

So, recognize, right? okay. now, we're well under a second.

A

B

Well. under a second. yeah,, because that's like.

A

Yeah. seven one thousandths of a second.

A

Let's do an eight here, right?.

B

Now, do something like do a letter. Instead.

A

Not only that., let me just start to screw it up. okay.

B

Okay. so, I'm going to do this. I'm going to do this..

B

Then I'm going to like a line between here and.

B

See what happen. it's two things.

B

It's an eight but notice that there's uncertainty that starts to come in. now,. What I'm going to do is I'm going to try to do this, to try to make it like a seven two and see what happen.

B

Right., do you see how now the probabilities start to go a little bit, down, right?.

B

Because it's becoming less and less confident about these things.

B

yeah., it's interesting to know that it's.

A

Leaning more towards three than it is towards seven., but here's the thing.. I want.

B

To be very clear about this., it's not thinking., it's! Basically, when you pass those 740,.

B

700, a four numbers in,, the weights that are picked to multiply by and then function over and then multiply a function,, that's what it's returning at the output, but it's not thinking..

B

Ai has no feelings. for of all those who talk about like, "is ai, going to start thinking?" no,. I think people should start thinking.

B

yes., because you might as well ascribe.

A

That would be a good idea.. You might as well ascribe like.

B

This kind of intelligence to a rock and it would be the same thing. yeah., but we're basically doing an optimization algorithm to come up with something that will- that spit something out. notice that it's guessing three and I've screwed up..

B

I could fill the whole thing in and it will still guess something because again,, it's just multiplying matrices., oh look,, it thinks that's a two., then there's times where it will get things wrong. Too.

B

That's because it only train,, we trained it on specific digits..

B

If there's numbers away., I see.

A

-That you write differently,, it will not. so,. This is interesting.. This goes back to the non-thinking part..

A

This system doesn't really have a way of saying, "I, don't know.".

A

Like you or me, with our ai system or just I,,.

A

We can see this, last thing that you said, and said,.

A

"Well,, I'm not prepared to say what number I think that is because this is gobbledygook.", whereas this one basically just says "I'm always going to do best effort..

A

I will give you a statement of how accurate I think this is,.

A

But that is all I'm incapable of., I don't have the his is gobbledygook."..

A

I even hesitate to say like it's,.

B

This is how accurate I think I am. basically,. What I did is I forced a function at the endpoint,.

B

That forced them to look like probabilities. is called the softmax. so,. I basically forcing it to pretend to be confident with those numbers, right?

B

so,, it's not even like, like I could have just not done the softmax and just pick the max number, and you'll see that some numbers are higher,.

B

Some numbers are positive,, some are negative, right?, but in the end like, if you want to ascribe like the unsuredness,.

B

You would return a vector of 11, and give it examples of things. It's not sure about..

B

Oh I see. then,. It would return that. okay., but you didn't build that.. I didn't build that., that's the thing.! You cannot do anything, you did not explicitly train it. For,.

B

With the right network,, with generating the right, weights., okay makes sense.

A

All right. any questions about it?.

B

By the way,, this is all available right now: pytorch 1.0,.

B

Will natively export the right onnx format.

B

For you to use inside of .net applications.

A

Right oh. so, in this case,.

A

You use the onnx runtime directly.

A

What about ml.net, can it also used these onnx files?.

A

So, the thing about ml.net is: it's used for training.

B

Okay. we got to separate the process,, but you can also use ml.net to do the inferencing as well,.

B

But in your brain you should still separate the two processes.

B

Training takes a long time and is very tedious..

B

What was the word? You use, the gymnast used?, the contortion. contortion.

A

There's a lot of contortionist during the training.. I did that on purpose., but when you are running the model,, we call it inference.

B

It's a separate thing., it's pretty fixed.

B

okay., yeah., so, okay.. So with ml.net, you can both train the model and do inferencing.

B

ml.net, as far as I know, does only where it called shallow models.

B

It's not a bad thing.. Shallow models, work really well like decision trees or logistic regressions..

B

It does those kind of models,. This for onnx.

B

in particular,, is for neural networks.

A

I see. so, kind of apples and oranges.

B

It could be. it could be like yeah.

B

like,, remember how I- the example was that you put down the pipes and turn the knobs.

B

there's a fix class of this style of pipes you can put down and where knobs are.

B

I see. for deep learning,, it's their neural networks,. There stack. for other ones, there more shallow, but they get the job done. To.

A

So just in closing,, can you tell us a little bit of,? Are you working with microsoft customers today who are using the sort of thing that you just showed?, oh absolutely.,.

B

This exact example, no,, because this came out at connect.

B

yes, but more like.

A

Other industries, where you're seeing this sort of thing being picked, up?.

B

Oh absolutely., like a lot of people, are using ai..

B

They mostly put it into production in some other way.

B

like they put a rest endpoint,, but people are using it for fraud. Detection, people are using it for recommendation, systems., there's a lot of stuff like,. Think of any algorithm that you've had to write where there's a lot of, if statements and for loops, and you feel like you need to go home and take a shower because you just kind of guessed at what it would be.

B

yeah.. Those are the kind of algorithms that really fit well with machine learning, because, instead of you coming up with a series of steps to solve the problem,, you give it examples of the right thing.

B

like in this case, numbers and the actual digit pixels and the digit..

B

Then it comes up with an algorithm. awesome.

A

All right. bud. okay. well,, thanks for coming by. yeah.

A

All right. bud we'll see you next time.

A

okay.. This has been another episode of on.net.

A

About machine learning with seth juarez. thanks.