.Net Foundation ML.NET, 6 Aug 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: August 6, 2019 - Trying something new: ML.NET and recognizing Jeff's hats!

Description

Jeff wants to learn about Machine Learning and starts going through the ML.NET tutorial with an intended goal of writing a feature that detects the hat he is wearing on stream and reports it in chat.

He doesn't get very far with his goal feature, but learns a lot about machine learning along the way -- Watch live at https://www.twitch.tv/csharpfritz

A

Hey there, we are good morning, good afternoon, good evening, whatever time it might be, where you are welcome back to the live stream, my name is Jeff fritz. Today, today is August 6th 2019 we're gonna read a little bit of code today, how's it go in their chat room good afternoon good evening great to see folks look who's here. Paulo is here, welcome portal 7 from Argentina thanks so much for joining us today.

A

That's great mm Maria, double-o, hello, hello, stels e good to see you code stencil and musical book worm code stencil connecting from Canada. That's awesome, always great to hear from my neighbors to the north. Thank you for tuning in I wanted to do something a little bit different today, we've been working on doing a little bit of right, we're doing that forms over a data type of thing and we've been going through and we're doing a little bit of yak shaving.

A

At this point, just trying to connect the dots I wanted to get a little bit outside that take it take a stream or two here and let's go just a little bit outside our comfort zone. Let's try something new and I want to try ml dotnet I've heard a lot of folks from Microsoft talking about ml dotnet and how important machine learning is.

A

You hear this from other folks in tech also how important machine learning is the PAL oh good evening from Portugal a good evening to you thanks so much for tuning in and I want. I want to learn a little bit about this and we've done some work around our chat bot. So it does all kinds of neat customized things you can execute the help command there in chat and it will interact with you right. It'll answer some questions based on.

A

What's in our FAQ it'll to the best of its capabilities, do some analytics and try and match up and get you a good answer for things well, I'm, always wearing a different hat and I'm wearing my visual studio, 2019 hat tonight, because I want to get in and look at this feature that comes with dotnet now and learn a little bit about it and I think we might be able to learn how to use ML dotnet, extend it a little bit teach it.

A

What some of my hats are and maybe then we can have a hat command in chat that will recognize what hat I'm wearing on that stream and tell you what's significant about that right. This is the Visual Studio 2019 hat I wore this for the Visual Studio 2019 launch on April, 2nd 2019 done right. We can load that data about which about you know the little blurb the little you know story about why that hat is important. We could put that into a database somewhere.

A

It may be even a CSV just what's sitting out there on disk real easy to load, but identifying which hat it is that we can eat, do a little machine learning and have it just look at the image and maybe identify what it is based on the color and the pattern. If we can recognize it, we should be able to work on that um hello. There Kevin good to see you I'm doing well, thanks so much for joining us.

A

So, let's head over to the code and we'll get our music set up here, hey there, we go thanks so much for tuning it see over here. This is the new camera right. I still have problems with the color on coming off the razor key. Oh it doesn't quite color right and I've gotta deal with that. A little bit more see. We can do it's good to see you Dede thanks so much for joining us. So let me get some music running in the background here. I'm gonna share.

A

This is the getting started tutorial for ML dotnet and we're gonna walk through this. It should take about 10 minutes, it says- and maybe we'll learn enough here- that we can start to extrapolate and go after building building a model, training that model and maybe even getting it into our our stream tools into our chat bot by the end of the session. If.

B

A

We'll get to it tomorrow and I have a guest coming up later this week. Let me get some music playing and we'll come back and talk about the guests and we might have an announcement along the way here as well another announcement along the way- let's see here, let me go over to my music code. Bye Bank, you know what let's play let's play indigo tonight here it comes. This is music, the code by it's written by our friend mr. Carl Franklin.

A

It's designed to get you in the groove to get you in the flow so that whatever task it is that you're.

B

A

On you can focus on, you can get stuff done faster because all your other worries will kind of fade away because the music, it's just groovy- and it gets you into that sense of flow check it out. Mtc b.com.

A

You can execute the music command there in chat. Do me a favor drop our friend Carl Franklin, a message on Twitter he's at Carl Franklin. Let him know how much you enjoy the music. We want to thank him for letting us listen to his music. While we write some code here together, live on stream. Alright, so this is, this is the sample. This is the ML dotnet tutorial I want to walk through this a little bit. So we understand what this is delivering.

A

What this gives us and we'll see how we can use that to apply to our BOTS into this scenario, where we want to train it on some images to identify some hats know what I mean. What do you think chat rooms that sound like a good, a interesting experiment? We go through here tonight, good to see you blazer. Mr. mr. Magoo sequel, mr. Magoo is stuck in relational database land but blazer. Mr. Magoo is here with us so good to see you so two things we are at work. My stream elements work.

A

My with this thing go. We lost our gauge here, no right. Mr. Koetter right, you know what I'm saying the little gauge below me here is gone. Oh no, I gotta go fix that real, quick I want to put that book up there because we are at 71 99 followers.

A

That is tremendous. I am so thrilled to see that. Let me turn back on the gauge here. I, don't know why it's turned off, but we will turn that back on, follow our goal there. It is, and it see the follower goal even has the wrong number on it right now it is 71 on 99 Save, Changes, Save, Changes and boom there. It is 71 99 for each multiple of 100. As we approach the magic 8000 number I'm gonna give away a set of stickers randomly to somebody who's in the Trant room.

A

We are one follower away. So, if you're out there, if you're a friend, you enjoy what you're seeing here on stream, if you like writing code, you want to learn something and you want to be notified next time.

A

I go live click that follow button; click that heart, just above my head there, if you're on the twitch website, if you're in the app it's somewhere up there, you have to click on the video and it will appear, click that heart, it's completely free to sign up and follow, and it'll help us get one more follower closer to that. 8000.

A

Wait! It's sick! He was already following us. Wasn't she I think she's playing a game here.

A

Portal that is music to code by from Carl Franklin, see that there we go now we got our I think that's the seventy-two hundredths followers Alva's Avex, see it in did I pronounce that right. Let me make sure that I'd see this flip over to 7,200, because this doesn't count unfollow Spallone me. It doesn't resolve that I see $71.99 still sitting over here, exit accidentally unfollowed says dee dee sure sure was just an accident.

B

A

Don't get the number our own twitch hasn't, updated, I think I think yeah both of you unfollowed and followed there. Isn't that what you did we're close we're very, very close: hey y'all, Dorian good to see you so yeah, that's the music to code by you can well. You can go to the website and then download copies if you want to listen to it on your own, but I have so I found a bunch of stickers here, not just our stickers that will be thrown in the sticker pack.

A

This time and I'll send these anywhere in the world. I've got, of course the dotnet bought there is I've got super c-sharp there.

A

It is we're gonna make those we're gonna put those back in the emote rotation here so that folks get to use those I've got the rainbow bearded octocat I've got the shaded Fritz pot right he's, so cool and I've also got the rainbow bearded paperclips I, know well, I, don't think Scott's here, I think it's just us good to see you janeski HR book good to see you can I send that to Denmark absolutely I can't that is not a problem. I am more than happy to send stickers to Denmark.

A

We sent some stickers last time to the UK and I sent some stickers to even to Russia, so I'm, not sure if they've made it to Russia yet but I sent them we'll see if they and I've also got some new get stickers. Look at those! Oh there's a throwback new get stickers, so throw them all into a bag and get them out there into a back into an envelope and send them out.

A

So what we're gonna talk about ml dotnet here for a little bit, we'll see if we get some new friends joining us click that follow button and I'm gonna reset that number, because that numbers not right and I, don't want anybody confused! Try to see my PC I can't see your PC eh our book. What should I see on your PC?

A

Let's see here. This should be C stream elements. It isn't a real number that they're looking at here there we go, they more or less draw a line and they say who follows they don't measure when people unfollow and correct it at all. So so we're gonna take a look at ml net here and see what we can learn about this and hopefully we'll learn enough, so that we'll be able to train a model to recognize images.

A

Now I've seen some other folks that have used similar technologies, similar machine learning technologies, to train models, to do all kinds of things, to help train around image, recognition for people's eyes and- and there was even some wandering during an event- they were using it to train to learn about make up that women were wearing and be able to recognize the colors that were on woman's face impressive, very impressive use of the technology. That kind of image, recognition and I think if we can do it, it's something as simple as, what's on my hat.

A

What's that on my noggin? What do you have there.

A

Guy as though with oh, my gosh all kinds of stickers on your laptop there, I'm looking I, don't see any any stickers, I recognize, but we can definitely oh yeah. We would you a rainbow bearded paperclip and some of the other ones that I have you I have here would look kind of kind of good on there right there's what there's, what HR books laptop looks like that's a.

C

A

Of stickers, that's a lot! So alright, let's see if we can go through this yeah I know we's, boss or West boss, no I, don't so! Let's take a look first thing it wants us to do. Is it wants us to install the ML dotnet model builder extension and it's currently in preview? Alright, so I'm gonna click over here, the ml dotnet model builder. You know what I already have a Visual Studio open. So let's go to extensions over here.

A

Let's see if we can find it this side a little bit easier, so I'm gonna go to online and let's see if we can find ml net model builder there, it is so I want to download that downloading downloading.

A

Thank you. Alright need to restart visual studio, so it applies I. Don't he sale sticker.

A

Let's see so model builder is a simple UI tool for developers to train, to build train and ship custom machine model learning models in their applications.

D

A

Dave, thank you so much for that subscription. Look at the look at the emote phone here! Thank you so much for that and we'll we'll make a donation to coder dojo. Thank you very, very much for that. We're making donations all quarter long for all of our cheers, all of our subscriptions to coder dojo, who are helping to build facilities, train trainers and give folks all around the world an opportunity to learn and grow their careers in tech thanks.

A

So much I really appreciate this up and that and brought your twitch prime to if you have an Amazon Prime account, you can link that to your twitch account. You'll get one sub free anywhere on twitch. If you use it here, you I'll make that donation and you'll get all the cool emotes. We have 19 emotes you get 17 by default. If you want to subscribe at a higher level, that's okay, we'll give you some extra remotes for that.

A

Do all these stickers add enough steadiness to the laptop do for it to survive falling from 2 meters.

A

Let's see here so we're gonna use the IMP model builder to build train and ship custom machine learning models. Well, if we're gonna teach some code about my hats, we're gonna need some custom models because I can guarantee. Nobody knows what my hats are know. What I mean there I mean you know all right. So this is yep. Go ahead and modify visual studio. Do the thing load, the stuff yeah go ahead, shut down all those other processes. Do it do it come.

B

A

Where my music go, it just stopped. Are you kidding.

A

A

Are we still here.

B

A

Visual Studio installer literally just took down the entire stream.

A

That was to kill a process. Are you kidding.

A

That's not gonna work know what I mean, let's see, if it I, don't think it installed. That extension.

B

A

Papa bah: no, it's not here.

A

Right, modelbuilder download right: it's gonna go through this thing,.

A

A

Thanks so much for tuning in I know and I appreciate that cheer.

A

Do I need to restart it, it's gonna do the thing there we go.

A

I need a notepad thanks, so much Kerry for that cheer.

A

We'll log that, with the rest of our code as we go through it here and we'll make a donation to coderdojo thanks so much yeah. What the heck is this thing doing.

C

A

Gonna click that in task button and see if it crashes me again, I'm clicking nothing's happening. Oh here we go I'm dying, Visual Studio, closing shop.

A

This is insane.

A

A

This is somebody's.

A

A

What the heck is it doing.

A

A

Like when this settles down, my CPU is usually in great shape. My memories pegged right now, because I've got so much Firefox open, but look let's I'm not using that right now, like you're kidding.

C

A

Let's try and do it this way, I mean what else is it trying to close I? Don't.

A

I'm not sure what not sure what that is right. There's karnak running, there's discord, I mean dotnet, isn't running in the background here anywhere I'm, not quite sure what those are okay, Skype I can kill. Skype cuz we're not on Skype right now.

A

But still I'm like thank you blazer mr. Magoo.

B

A

Okay, I want to put it just in the preview version.

A

Brother here we go again: there are no open instances of Visual Studio. It's not running.

A

It's not there, it's not up here.

A

So what why does it think, there's yeah, there's Cortana's processes working absolutely, but why does it think that there's Visual Studio running you know good evening fixture Jake great to see you yeah yeah reboot, my foot, there we go. Finally, oh.

A

Thank you, you even put it right on Twitter.

A

Let's see, you would say, kill anything everything related to visual studio, download the v6 instead of download on visual studio. It should be the same. We should be going through. The same thing is that key rx time to reboot now now we're fine HR book bought some stickers from West boss with and I've boss from other places from Denmark anywhere else in the US and UK I usually go to sticker mewls, where I get stickers and I'll get them from there.

A

Alright, so that looks like it says it installed, continue, create your Apple visual studio, create a new dotnet core Console app might be happy to. Let's do that. There comes Visual, Studio scoot this over to that side and I'll be running over here now, I'm running at 1920 by 1080.

A

Feedback about the docs I want to be able to put this side-by-side with a with the project. I'm working on be great. If I could have this collapse up right, let's call this first ml net and I: don't I never put stuff there I, don't know why it never remembers that I always put stuff in seed dev.

A

So why preview asks lies, live good to see you I always use the preview version of Visual Studio, because I want to show the latest in coolest things that are available, that we can work with so alright, oh thrown in with the PHP badge. What is that no kidding? We stand? We stand break out. The horses Chris Jones, Chris.

D

Jones just resupply before 12 months, 12.

A

Months welcome to a year. Thank you. So much Chris for a year's worth of subscribing I, really appreciate that we're gonna make a donation to coderdojo and you're gonna. We're gonna put you in a new badge, a new loyalty badge. You're gonna get the hat the rainbow hat with the propellor badge, fantastic thanks. So much for your support. Chris, let's see lies, live asks. Do I work on that team. Yes, I do as a matter of fact.

A

Yes, I do all right, so I have a console app with dotnet core we've opened over here change the project name to my ml lap.

E

A

One works: add machine, learning, right click on the project in solution Explorer. So let me go back to the solution. Explorer I, don't need live, share. Right now, go away. I can get rid of some of these other ones. I'll get rid of server Explorer, because I can sequel server objects, Explorer yep! You can go away too I'm. Not using that right. Now, solution, Explorer, all right, right, click and say add machine learning there. It is that's easy!

A

Okay, this opens the ml dotnet model builder in a new docked tool window inside visual studio model builder will guide you through the process of building a machine learning model in the following steps. Okay, let's see what we got here so I'm working on it, the card: okay, okay, you have to work a bit more with stripe, c-sharp and JavaScript says HR book. You just write me if you think you want to send a sticker to Denmark. Well we're gonna I'll raffle one off and we get one more follower. I will raffle off.

A

I've got ya handful like six different ones. Let's see I should live shared. The lies live, make sure enable console run as admin.

A

What do you really think? I'm gonna fall for that. Do you really think? Oh, my goodness, Michael jolly fifteen Raiders from Michael Jolly Street thanks so much for joining us. It is very cool to see you well c'mon and Raiders. My name is jeff fretts and we're writing a little bit of code. I write code here on Twitch and we're learning all about ML dotnet machine learning with dotnet we're using the new model builder and ml dotnet framework. That's available with dotnet court runs on Windows, Mac and Linux we're walking through the tutorial right now.

A

Let me see if I can click that link. Let's share that link back to the chat room. You.

C

A

Think I have it on the stack there there. It is there's a link to the tutorial if you're interested in in picking it up and following along what we're working on today, Michael great to see you retro CRT, is here electric havoc great to see thanks so much for joining us Carrie. Thanks so much for dropping that shout out there umm settings projects and solutions, locations, see I thought I had that set clear.

B

A

That's data Lake.

A

Thank you for the follow is that Valda nigh I appreciate the follow. Let me make sure that ticks over it. It says 7200 here there we go. That's 7200! Let me final, this one set of instructions here where was it one? Second projects and solutions.

A

B

Oh there, it is.

A

Right and locations projects locations are you kidding, that'll make my life a ton easier. Michael jolly was working on a custom credit roll that displays everyone, who's, subbed, followed, cheered or contributed plate, sounds etc with their name and profile image. That's pretty cool that is really cool. Make sure you check out Michael jolly he's doing some really cool stuff over there, making twitch better thanks. So much for the rate here.

E

A

Ml net is cool. You have you, have some history behind them, alright, we'll take it we'll discuss that in a bit, but we hit our 7200 followers at every 100 followers at every 100. As we approach 8,000 I'm gonna give away a pack of stickers and I have I ran through the list of them. Here. I've got stickers based on all the emotes that we have here and I've even got the new get the new get logo I can give out as a sticker. So I've got the lash tools here.

A

Look at that I'm gonna, clear. The list drop something into tak into chat, I, don't care what you write? Words, something something PG, something friendly drop into chat and it'll put you into into the the collection here the participants list so that you could win a sticker pack, I'll mail, it out to you I'll mail, these anywhere in the world. Yes, even Denmark, no I'm, not mailing out Snickers Snickers Snickers.

A

Is a candy bar that Fritz can't mail out? Sorry, the more you know: yes, stickers code stencils, something works. Absolutely SMED, word good to see you Denmark! Yes, the lies life. Denmark is an amazing place. It's a magical place, it's as magical as Minecraft, maybe even a little bit more magical than that. Okay,.

F

Really the questions that I was.

A

Called here to answer yes,.

B

A

Shown I, you are feeling lucky, you may win, you could win.

A

So how much code is behind? That array add shake. Oh no, this is this was written with with flash so stickers.

B

A

Loves stickers right all kinds of these ones that we'll be giving out here. The super C sharp one there. It is I'm gonna put that back in the emote rotation here, so everybody gets a chance to use that I actually want to I've been wanting to make that into a cheer mode so that you can cheer with C sharp and throws C short bits at my face.

A

Know that the old dotnet bot right and, of course, the current emotes that we have there's the github octave cat, with a rainbow beard clipping with a rainbow beard and, of course, the frits pot, with the cool shades, alright ready. Let's do this, I wouldn't finished some stickers, yeah yeah, tell you more about Snickers, says Michael jolly. It's a chocolate bar, that's made with nougat caramel peanuts, and it really satisfies you. It's amazing, Simon Says flash run away, I kind of agree with you on that one, but so it may all right.

A

Let's, uh let's do this! You love Snickers I, can't can't say that that I would turn down a Snickers bar here. Alright, let's do this.

A

I got to find the right slider for that holy crow was that loud.

A

Alright, let's give out some stickers here: new get chocolate bar yeah.

A

And we're gonna mail, some snicker, some Snickers darn it we're gonna mail, some stickers to fix, stir Jake, congratulations, fixture jig.

A

You still there Jake.

A

Let me know you're in the chat room.

A

Yeah there we go: congratulations, Jake, I'm, gonna drop! You a whisper.

A

Fantastic, every 100 followers is when we get to 70. 300 will give out another sticker pack here and thanks.

E

A

Really appreciate all the support from everybody in chat and, of course, if we get to 8,000 before September 15th I will dye my beard rainbow for.net, conf and twitchcon they're the same week. I will have that beard as a salute to you, our twitch community, for both events. How cool is that I'll be interviewing people like Scott hunter Scott, Hanselman, other people named Scott, and even the DD Walsh and I'll have a rainbow beard on this here face as a salute to you, and everybody will know that I support our twitch community out there thanks.

B

A

Much bucharest again says nothing else, matters yeah, there's one of the Scott's, none of the Scott's in the chat room, terrific Jake thanks so much for sending that information along so promptly, you're hype, time to add on to the C sharp one you got from work. Oh yeah, there you go page, another Scott, there's plenty of Scots. Did you know there are thirteen mentions of Great Scott throughout the Back to the Future movies.

A

Alright, let's do this. I am let's get back into this, no worries there Jake, not at all. Okay. How do I know? Oh, we went looking for it and someone's edited together. Eh someone's edited together a supercut, there's 12 different, and then six of them were actually usable and I've got all six of them loaded up for you, both on my stream deck and on the chat bot. So fire away, the Scott commands the end end commands and oh yeah it'll run its mouth for you.

A

Nope didn't list the project, so, let's put in a project here, real quick, Jeff, I, know learning about ml, dotnet and attempting to add a hat recognition feature to our chat box. I think that'll be cool. That one was funny. Oh yeah, oh they're, all good, all right, good, everything's, fine! So let's continue here so we're continuing through we've added in and we're walking through the ml net tutorial.

A

So we added the model builder extension into visual studio here, and this is this is something that that you want as an IDE to enhance and and make your development experience a little bit easier right. These types of things you're going to expect from an IDE from a text editor you want to just add a text. You don't want full-featured extensions that walk you through processes like this. You know what I mean to generate your model. You need to select your machine learning scenario. Okay model builder offers several templates.

A

Price prediction sentiment, analysis, custom scenario: okay, so binary classification predict positive or negative sentiment of comments. That sounds like something that our chat bot already does and write. With that sentiment. Analysis number over there is there's a sentiment command actually now in the chat bot right that should be working or not.

A

I thought I'd deployed that didn't I, deploy that.

B

Maybe I didn't yet.

A

Is it in there there we go. The sentiment over the last minute was 77. One percent active positive trending downward. No, it's trending up compared to the last five minutes. The most recent messages were positive.

B

Yes, there we go.

A

There's all of our commands, so I don't want that I'm, not predicting a so we're not going into two categories: we're not predicting a numeric value, custom models with Diddy's and classification regression in other tasks. That feels like we're I'm going to want to be when I cook when I classified that- or maybe it's this one issue classification putting data into one of several different categories. Well, maybe it's the custom when we get there for the hats, I, don't know, let's figure this out when we get there I'm going to finish walking through the tutorial.

A

We're gonna predict sentiment based on the content of customer reviews. Okay, so we're gonna choose the sentiment. Analysis scenario, all right, which is a binary classification, machine learning tasks.

B

A

Right continue, download and add data download the Wikipedia detox data set and save it as this in the directory we created all right, so it looks like I can right click and Save Link As, alright, see dev and.

C

E

C

A

What was it like? First, first ml in that there it is and I'll put it down, and let's put it yeah put it down in here- I guess: right: I, guess that makes sense all right, so I saved it in there each row represents a a different review left by user on Wikipedia.

A

The first column represents the sentiment of the text. Zero is non-toxic. One is toxic, yeah they're, saying things like that's pretty toxic right, but then again.

A

That's positive in how he said that, but it's definitely negative I don't know so this is the just. Do the clever thing in my app shrink-wrapped ml, any box, flavor of ml I, think so yeah right. Let's, let's make machine learning easy for folks that want to do the simple scenarios right if we can make that kind of thing easy for folks.

B

To use yeah now.

A

It becomes it becomes impressive and easy for folks to do. This is kind of like where database access was 25 years ago. Right people need to use database. How do we make it easier for folks to use database from their applications and things like oh D, BC evolved and the other API is that we now have that are common for databases? That's how I end all my commit messages says: Michael jolly I wish you would know. I end all of my commit messages. Now, with its positive. It's welcoming it's better than.

A

That's different right: okay, so each row, it's a it's a comment and then the sentiments of things like this one: zero one zero, whether it's toxic or non toxic, okay columns, are separated by tabs. Dateable looks like the following: blah blah blah blah blah Wow.

A

Okay, add data in model builder: you can add data from a local file or connect to a sequel, server database. In this case, you will add the TSV from a file. Alright, so file select the file and that's not where I want to be. Why did you go there? I mean literally that's.

A

Whatever you know, okay, there, it is come to predict all right under come to predict, select sentiment, Hey.

E

Look at that sentiment.

A

Label is what you're predicting okay, so we're predicting this value, which is in this case the sentiment found in the first column, the rest of the columns are features which attributes which are attributes that help predict the label. Okay. So it's gonna, look at this feature the sentiment text and use that to help predict this value. Okay, got it after adding your data, go to the Train step. I think that's! This here continue.

A

You alternate between between ending your messages, I wish you would die and as the prophecy foretold. Oh that's good. Okay,.

B

A

Like that, you could also answer answer your your work. Ticket assignments with.

A

See wrong guy assign it to someone else and when they finally do assign you the right thing. You can say.

E

A

Hans Gruber Hans Gruber's, pretty good. The model builder evaluates many models with varying settings. To give you the best performing model the default time to Train the amount of time you would like model builder to explore. Various models is 10 seconds for larger data sets. You should set a longer training time, select, start training to start the training process, I'm feeling lucky right. I'm gonna go out on a limb here: I'm gonna.

A

Try something different, give you 20 seconds.

A

A

Right and we're.

D

A

We're waiting I mean. Does it really add that much more if we doubled the amount of time here, I mean right, it doesn't say not to double the amount of time. I mean. Let's see what happens here, isn't the traditional ending? Well, it works on my machine.

A

That's a good point, I think I'm. You know what I may have to reach out to the boss.

A

And see if I can get him to record that as a sound Oh sound as a voiceover for us, what is the maximum to train the model? I, don't know, let's see what this little eye thing is. Logger training time will give better accuracy.

A

It's still letting me type these nines.

A

That's a lot of nines I just put in the chat room. That's like that's a lot of nines, that's more than 10! That's like 25 nines! That's a lot of nines.

F

Shéi shéi shéi.

F

A

May want to put a cap on that value, I'm, not sure but I'm, just saying okay, it it might. It might be the right idea, because.

C

A

I'm just using the textbox what about in 64 max oh we're bigger than in 64 max there.

A

No wait. That's over it's! Oh look! Now I got a little red box on it. Oh yeah, there's an error there. What do you think of my Pig anyways?

A

That's better. It put a red box on it, but how long should I train for? What's that? Do it opened a browser over here there we go longer. Training periods, allow Auto ml to explore more models with multiple trainers and settings, so Auto ml I dunno is a command-line tool that you can also use to build some of these models. So if you have more than a gig three plus hours, yikes.

A

Types of columns text versus numeric, the type of machine learning test regression versus classification.

A

I'm learning this: how do I understand my model performance model builder by default, splits, the data you provide into train and test data, the training data, eighty percent split is used to train your model and the test data. The other 20 percent is used to evaluate your model, ah so four out of five records are used to teach the model and it uses the fifth record as validation to see if it worked properly.

A

That's cool, okay, that makes sense to me: hey brave, Cobra, good, to see you, okay, so I clicked that I did the training thing and I mean 70 percent, accurate, pretty good right, that's a see right in in grade school I would get a see if I got a 70, see right, that's something you can keep track of the progress by going in the progress section. No, it's it's done shows the accuracy of the best model that model builder has found.

A

So far higher accuracy means the model predicted more correctly on the test data you think, shows you which algorithm performed the best so far during model builders exploration averaged perception, binary. Okay, this shows the last algorithm that was explored all right, so go to the evaluate step, continue, evaluate your model after model builder selects the best model it'll. Take you to the evaluate step which shows you various output, including how many models were explored in the ML tasks. In the case, binary classification.

A

Let's see here, make sure make sure I did everything right here.

B

A

Okay, so there's the models. Explored model builder also displays top 5 models there. They are including the AUC au, PRC and f1 score, which you can learn more about here. I'm, not gonna click through that right now, because I don't know is that Oh X Valley I see what you're trying to do there. That's event: that's hex!

A

This is ml. Dotnet, we're learning a little bit about using the machine learning libraries that come with dotnet Cornell. These will run on Windows, Mac and Linux, and you can use them to build and teach your applications how to make predictions based on a model that you train and we're working through the intro to ml dotnet. That's right there on the ml net. Tutorial I should have put that in the project command shouldn't I.

A

Let's put that on the project command: where is it the tutorial? Is there we go now it's on their neural networks, now, not quite not quite so. 15 models explore training time. Ten point: five: seven seconds no I spent 20 seconds over here I went a little further.

E

A

It only went through five models. I feel ripped off, move on to the code step. Alright, move on to the code, step that machine learning model and projects for model, consumption and training to your solution.

A

Select add project, so it's not really a select it's more like quick, but okay. Let's see model builder adds both machine learning model and the projects for training and consuming the model to your solution in the solution. Explorer. That's this dealio over here. You should see code files that were generated by model builder, including well I named my project slightly differently. So I see, model builder and I see my model down here. Data models in a zip file that I guess has that right that trained model living inside of it key oryx.

A

Let me keep me honest here, make sure I'm learning that I'm pronouncing your name properly there. What I don't get is, if you're, using the same data preset of the tutorial, why the results aren't the same.

A

Different processors runs different right, different speeds, faster I, don't know, or at least get closer results. Well, they went a little bit longer. This seems like building IKEA furniture, says Simon jeering easy enough to open all the boxes and fit some bits together, but sooner or later you need to go back and read. The manual I agree right for the basic things like this, not bad.

A

We can get something that fits our needs, but when we go and we try and expand this a little bit, we and we want to get into this training it to understand what my hats are. Might.

C

Be a little bit.

A

Different, let's see we can learn here. If you increase the time the train doesn't increase, the number of models explored and hence the accuracy I would think that Paulo I think that would make that make sense to me.

A

So the console app is the dotnet console app, which contains the model builder CS used to build train the model and program CS, which is used to run the model. The ML dot model is a dotnet standard class library and it's standard. Of course. This runs on every environment and it when it gets compiled when it's referenced in compiled it'll be adapted adapted adopted so that it runs on the appropriate runtime that you're targeting with your application, so in this case I'm targeting Windows Windows console.

A

So this will be compiled with the windows console runtime information, so it runs on windows. Console I could also make this run on notches neck or butt on the neck or on Windows I can make it run, dotnet core on Linux and have it run with the Linux runtime when I compile Valley, says c-sharp seems so confusing, Oh or dotnet as a whole, rather Li. So many different things: what are your? What are your concerns there? So C sharp is designed to be able to do a whole bunch of different things.

A

As a language you can apply. Different frameworks apply and use it with different frameworks, whether it's building websites, building mobile applications with xamarin building, IOT apps, working with machine learning models working with database with entity framework so as a language and how you use it with the framework and it's the same thing with Java and C++, depending on those different framework components that you choose you're gonna be able to compile and build for different things.

A

You need different runtimes that are appropriate for those other for those other operating systems and dotnet has the same thing: different runtimes that you can run your app on top of whether it's Windows Mac Linux, with the.net core runtime or the mobile devices with xamarin. So things are gonna. Get a lot simpler here in the next year when dotnet five lands and the promise of dotnet 5 is instead of having xamarin and mono and and dotnet core runtimes you'll just have one runtime, the.net runtime.

A

It runs everywhere and it will adapt appropriately for whatever operating system. You want to target and run on top of very much like C++ and Java. At some point, you need to declare. Oh, this is running on this thing, this piece of hardware and because processors and operating systems are different. Of course, we need to make sure that our our code knows how to operate in those places.

A

C++ is more like taking compiler, doesn't matter which include the headers and link to live, and now you can use a framework, yes you're, including the headers and things that are appropriate for that operating system, that you're targeting your binding to the runtime at that time. That you're writing your code with dotnet you're, swapping in that runtime just before it runs.

A

So this is a just-in-time compiler, that's in in effect versus an ahead of time. Compiler like C++, you compile all the way down so that it runs this native code dotnet.

A

Unless you're running in web assembly, you're running on a mobile device doesn't compile to native code, it compiles to some to some intermediary language code that is then interpreted at runtime and delivered exactly.

B

A

Yep pick your poison, so hey pac-man, jr. good to see you so I've got my I've got these two new projects here the console app now and the model try the model. You can run the console app to predict the sentiment of a single statement within the model. Okay consume your model: okay, mom. What hang on wait a sec to try the model run, the console that run the console, number very fun, debug start new instance, and it will start the console lab here for us.

A

But BA here we go.

A

Is it doing something here we go single prediction, actual value, true predicted value. True.

B

A

A

All right, I, don't know what that did. Model builder has generated trend model for.

B

You well thank you.

A

City can use the model another dotnet languages by using the following steps from your ml app yeah from the app reference the generated model, okay,.

B

A

Do that this is what that looks like.

B

A

In your app add reference to DES generated class library, install the ml net NuGet package.

B

A

Which is Microsoft and mantle, it opened it in another browser earlier. So let's do this. I'm gonna come back over to first ml dotnet manage NuGet packages right. These are all of those packages right, libraries, things that folks have written that we can reference in use. There's Microsoft MLM when I install this the latest stable version, which is one three one they've been iterating on this every month, adding new features so.

A

Problem with a lot of modern languages is that their way to run time dependent both on a low and high level.

A

Okay, at some point you got to depend on a run time. At some point, you got to depend on something: that's interpreting and running your code. So what are you gonna depend on so there's my reference? Okay, so right now this is just doing hello world, but that's stupid. It wants me to just call consume model and to paste in a consume model method. Here, alright, don't mind if I do bar control dot on this, so I get my using statement. I want those to be bars.

A

New model input, maybe I'm missing. It add reference to the model class. Yes, please make that a bar, also because I okay, so it's gonna load the model in ml model, dot, zip, okay,.

B

A

So it wants me to just call consume model from here. Okay, there we go and it'll run that in your MLM modify the sample text, input, dot, sentiment, text in program, CS.

A

That's silly, why don't I do this right, string, input text and write.

A

Input text and up here I, can just say.

A

Console dot readline come on now.

A

Input text equals that's and I'll put that in here and I'll I, don't need to console.readline. It's already gonna. Do that for me so now, it'll take whatever I key in and read it and force it through. Here, let's see what that, though, right.

A

You can always write your own language if you're not satisfied the current available ones, yeah.

A

C-Sharp is compiled, so it should not have to rely so heavily on the runtime. Well, so you.

A

Write the the just-in-time compiled languages, all the dotnet languages are like that and because that's that's the heritage of dotnet and Java is like that. Also you compile you get just in time code in dotnet, we call it intermediary language and at runtime it will complete. It will compile that for the appropriate runtime that you're on now xamarin and mono.

A

Allow you to further äôt ahead of time, compile all the way down to a target runtime, so that you're not specifically well because you're gonna target a specific machine, a mobile device or web assembly at this point. So let me key in something here. This is wonderful, and now it's gonna. Do the analysis on that and that's not right. That's really not right.

A

I feel like it's wrong.

A

Dotnet core 3 adds the ability to do package applications and its runtime inside a single exe. Yes, it does I need to update my hat to account for the new visual studio logo. No, that's the Visual Studio 2019 logo.

B

A

Like cheese, let's see what it says toxic, are you kidding that doesn't make sense at all? You know what I mean.

B

A

B

A

Love butterflies and unicorns are my favorite things.

A

Ok write the input text, that's passing along or that okay, well, are those things: yeah, yeah, yeah right, so I'll unclick that f5 not toxic. Finally,.

B

A

If you look at the trend of what's going on we're heading towards a further abstraction of language and platform with webassembly, meaning the browser is fast becoming the new operating system. Yes, oh yeah interpreted languages further interpreted languages like Python JavaScript make it a little bit further out that that we're pushing away from the virtual machine and how we're compiling and running playing with machine learning. Yes, we are Frank. So all right so.

A

It's working, it's not necessarily always right, but it's working right, I mean I could break this up. This consume model so that it right so that the engine is loaded and we run it through the prediction here right I mean.

A

Right so I ran that next steps. There's other data sets to work on all right. So I guess we don't work. So, let's run something you're called. Let's call it setup model, let's see if we can make this a little bit easier to work with right, I'm, gonna move this part.

A

A

Actually, I should leave that there and move the other stuff right.

A

Public static void, setup model right, I forgot the L and setup model. How could I all right so there's my context. The model loading that up the prediction engine is gonna, come off of that. Okay and I need that down here. So let's make this.

A

Right: that's going to create a prediction: engine of model, input, model, output, okay,.

C

B

A

I control dot on this yep generate that field there. It is good so now that works and I can should be able to just do it now right. Let's move this up here, so now we're gonna console.readline and consume text. So let's do this say, while not.

A

Let's save our read text while not read text equals is easy.

A

A

Yeah that should work no, not silly. Okay, so now it should read the text and then push into the consume model. Let's see, if that, how that works, alternative corn, hello, hello, you heard about ML dotnet. Is it any good we're trying it out here? Let's, let's see what happens so, it's running in sentiment, analysis off of some toxic non toxic indicators from Wikipedia, so I wrote earlier I like cheese, and it was telling me that's.

C

A

Hate cheese, toxic cheese is my best friend, okay, not toxic. Okay cheese is not my friend.

A

Now I'm worried, I hate cheese, oh good.

A

Cheese is toxic, cheese is toxic.

A

A

We love unicorns, okay, good, so whatever is generating that muddled assembly. Was it doing that locally on your box or in the cloud, or only your app is running locally. My app is only running locally. This is all local there's nothing reaching out to the cloud it's running with a small data set here.

A

When you look at right, the data set that we downloaded here. It only downloaded 250 comments.

A

So it's it's not a huge data set that we're training with right.

A

We would want to train with a heck of a lot more data to get much more accurate, much more accurate analysis of all. The statements that we could be keying in here depends on the kind of cheeses, brief, Cobra yeah, there's, definitely some more toxic cheese than others. Okay, now you see our machine learning. Ken stinks. Sometimes, yes, I kind of grew with you, brief Cobra says that's a cheesy model to me.

A

Okay, I do think it's cool right to get to have a machine kind of make a prediction based off of what it's been trained with. That is pretty cool now.

A

The next thing that we wanted to figure out is: is there a way that we can train the engine? What my different hats are each each one of my hats is a different set of colors has a different logo on it. Can we isolate images of those and have it predict? B I mean I'm, always in the same location here on one stream, but can we have it predict what hat identify which hat it is that I'm wearing based on a train set of images of all the different hats that I've worn here on street.

D

A

That's what we want to figure out. That's what we want to try and learn here together. So I'm going to go back over here to this gamer 23:13 asks any good resources for getting into c-sharp. For someone with Java knowledge, there's out here on dotnet microsoft.com, you can click through to learn and there's a whole bunch of getting started bits here, depending on how you want to get involved here.

A

You can click through and learn the basics with getting started with dotnet getting started with c-sharp there's a great group of folks here on Twitch that are walking through and answering questions and doing samples here if you're familiar with Java, you already know a lot of the structure for this Java and c-sharp. Look very very similar. It's just a question of learning the base class library, the BCL, all those things that come with dotnet that you can reference so C sharp corner is a great location. Absolutely alternative!

A

Korn, the C sharp corner is a great location to get started. Learning more about it. They have a tremendous repository of articles over there features that you can learn for that. You can learn from write, C, sharp corner.

A

There you go, so you can absolutely check out that website great articles from all kinds of folks that are that are MVPs that are specialists. Some Microsoft folks show up here from time to time, writing articles telling you how to make the most of the technology. So it's not exactly a training website, but it's more of a more of a blog and a recipes website. Right I need a little bit of code that does this there's great articles. There that'll show you how to do those things so I want to I.

A

Think I want to do model builder guide. I. Think I want to figure out how to make a more robust model here. That'll learn about images that we're gonna train it with right model builder is a simple UI tool that runs locally for developers. There you go to the question earlier. This is running on my machine once locally to build train and ship custom machine learning models in your applications. Try it now do it try it now.

A

F

A

Take a big step back, and you can learn more about this. That was that was my producer. Len Grossman installation scenario. Data train evaluate improve, improve that other one needs them improving things. You can try to improve your model performance train for a longer period of time. This will provide automated machine learning to try out more trainers and find a better model for your machine learning scenario.

A

Sometimes the amount of data or quality of data is not sufficient. I think that's what we saw with that sample. We only had 250 records if we had 2500 records now we're talking now, we've got a little bit to work with try adding more dated for model builder to operate on in case of classification tasks. You can make sure that there is a good amount of data available per classification category. Thank you for the follows. Let me see hack the planet, TV I think I know who that is thanks.

A

So much for the follow LJ Palmer. Thank you for the follow. Oh, my gosh MCD India welcome. Thank you for the follow and I, not sure how to pronounce this last one. Thank you for the follow as well appreciate all of our new followers guys here today.

A

So this is the first preview has some limitations we're working on improving the experience we currently support, TSV, CSV and sequel as the data types alternative, corn thanks so much for the cheer, we'll log that how to make it learn from images. That is the next step. Thank you.

A

Alternative corn will make a donation to coderdojo who were supporting all quarter long for all of our cheers, and subs really appreciate that, so the file should have a header row, so this is only supporting text is what this is looking like here, in which case we're gonna need to build a model somewhere else. We're not gonna be able to use this to build a model rat's, because we're gonna we're using a binary data type writing image. It's not a TSV, csv or sequel.

A

The plot thickens, so how do we train ml net if we don't have a text? Data set right, we're doing we're doing image. Analysis right? Is that something we is that something we can do right.

A

Sample sentiment, analysis, spam detection. This was something that I thought they demoed a while ago: flowers, classification, iris flowers, classification that might be right.

A

Movie, recommender price predictions demand, clustering, flower, I, don't know what clustering means: anomaly, detection, that's cool anomalies, okay, image classification that feels like where I need to be right: computer vision right: this is the kind of thing yeah it's like it is very much a cognitive service, but I mean I could do that with Azure cognitive services, I think, but I'd rather have it all run local to the box. Think I think that makes sense. Let's see what this is image. Classification.

A

You're kidding it's a link to the image.

A

Very funny: okay, I, don't know what the difference between these two things are samples and to end apps, deep learning, image classification. This is sample c-sharp, deep learning, image, classification they're the same.

A

Okay, let's see what this is image classification scoring sample image classification is a common case in many business scenarios. For those cases well, I hope it's a case in my scenario. For these cases you can either use pre train models or train your own model to classify images specific to your custom domain. That's what I want to do and I'll train my own model, alternative corn, says I've always been impressed with the great examples and documentation Microsoft offers yeah I think that's definitely something to be said there.

A

When you have a great set of documentation and examples, it really enables folks to get started faster and better.

A

Your internship ends on Thursday getting ready to graduate now that you've gotten a taste of the real world, says fixture Jake more power to you. Good luck to you, Jake all of the m/l net tutorials I could find for image. Classification start with a pre, trained, tensorflow or onyx model, not sure. If there's a way to train a model on image data using just ML dotnet, we'll find out, one is supposed to be a console. The other is of is a forum application. It is.

E

A

Right, this is machine learning samples, blob master sample, c-sharp, getting started deep learning, image, classification, tensorflow, end-to-end apps, not just getting started; well we're getting started. Okay, there are two data sources, the TSV file and the image files. The TSV files contains two comps. The first one is the image path and the second one is the label corresponding to the image.

A

So if we have images of that right that are trimmed to the same size of my hats, then we can generate right. Here's a list of the images right, then maybe it's maybe it's the visual studio, 20 19, hat dot, jpg and it says visual studio next to it visual studio, 20 19, as the label, as you can observe, the file does not have a header row and looks like this. Okay, the training and testing images are located in the assets folders. These images belong to Wikimedia Commons pre-trained model.

A

There are multiple models which are pre-trained for classifying images in this case, we'll use a model based on an inception, topology and trained with images from imagenet. This model can be downloaded from.

A

Okay, sure: how do we train our own model? That was the question.

A

How many mattes, how many hats have you to photograph and have a usable sample? Dataset? Well, I have right: I've got 40, 50, different hats, but I think the question is going to be getting different pictures of those hats so that they're in different lighting different image sizes. So we can say these are all that hat right, not just one image that says: oh, that's, the visual studio hat. We want to have like 10 different images that are different sizes, different Lighting's. That would say that's this hat! That's what this one is!

A

So there's an explanation about the two types at the beginning of the readme. Just go to the beginning of the readme.

A

Getting started ml, netcode, focus samples and end to end apps, end user sample web and desktop apps infused.

D

A

Machine learnings, well, that explains a little bit better. Okay. Thank you. Hey thanks for looking at up Kyra Kyra, Kyra 'kz,.

A

Big thanks to you, there we go points. Stick yours see that give out coins here. All right here, I've got to find all gotta find all my buttons here. Has anyone used ML dotnet for audio analysis? Yet that's a great question. Two deaf packs.

A

Really it's another binary file here and I know. During build last year, Bill 2018 there were. There were some samples where they analyzed music and MIDI files.

A

We had we had some of the folks from the ml net team on the show and they were working through them. So they're, using this model, okay trained with images from imagenet well I want to train it with my own images but hey whatever.

A

Okay note that this sample only uses consumes a pre-trained tensorflow model with ml DPI. Therefore it does not train any ml dotnet model. Tensorflow is only supported. An ml dotnet for scoring predicting with existing tensor flow trained models. Okay, you need to follow the next steps in order to execute the classification test set, the Visual Studio default startup project to this project as starting in Visual Studio hit f5 and Visual Studio. You think the output will be similar to this screenshot reading the model.

A

Okay, classifying the images there is a single project in the solution named classic image, classification, dot score, which is responsible for loading, the model in tensorflow format and classifying images model scoring. So here's how its right image path and label that's yeah! That's the format of that TSV right. This thing that it's loading from okay I got it. So we've got a column, that's the images, a column, that's the labels that we're going to apply, and it's got a static method here.

A

To read from that CSV file and put it it's not a CSV, it's a TSV, it's a wrong name. There load the data using text loader to read from the image file used to load images has two columns. First, one is the image pad. Second, one is the label it's important to highlight the label in the image net data class is not really used when scoring with the tensor flow model. No, no! It's a label that we're gonna use to reference. It looks like this thing: give it this label it's used when testing a prediction.

A

So you can compare the actual label of each sample data with a predicted label.

E

A

You can observe the file does not have a header right. You mentioned that earlier, okay, I get it. The inception. Model has several different parameters. You need to pass in image, height image width channels. Last okay, can you detect on the fly from video, so I, don't brave, Cobra I, don't think we're going to be able to detect on the fly from video. However,.

A

Twitch takes a screenshot and.

A

Uses that screenshot for that the card for the channel, while we're broadcasting so either that screen shot or maybe we can open open a headless browser, go to twitch and take a screen shot. Maybe we can do something like that to get this, there's got to be a way to get that screen shot automatically.

A

Drexell dave says: I feel like ml net is sort of dangerous I've seen people training models that really don't make logical sense, interesting mathematically. You can create models that predict data accurately, but don't really mean anything.

A

Interesting yes, headless browser. If you look at phantom right, the phantom I believe it's called phantom Jas browser. Try.

E

A

There too scripted headless browser. You can actually run this inside of a docker container. I said docker container robert tables usually appears, but I know he's busy moderating another very busy channel right now suspended.

B

A

Notice, oh! No! What.

B

A

But you can create point to a website until it render this page out to an image exit. The file.

A

Archiving the project suspend development, oh no, due to lack of contribution. Oh no.

A

Well, that's a shame! It's a really great project, phantom j/s.

A

It looks like it's still available out there, but we're power to the who look at that guy.

A

Puppeteers the way to go, you think I'm, not.

C

A

Puppeteer puppeteer no headless crumb, node API. Really. What do we got here and it's at PPR T done PPT our dev okay.

A

Usage require example, go to take a screenshot, close, fantastic, that's easy! Let me get a easy screenshot and we'll be able to zoom in on just the video right. That's not bad you're excited something here. To get the screenshot alternative corn always ask questions. They have never have to announce you're asking a question. Please always welcome to ask questions here in the channel that goes for everybody.

A

You'll need a trained model of my various hats from tensorflow, yes, musical bookworm I completely agree, and we have a link or other tensorflow tutorials as a starting point for image classification. Let's see we have here so this is musical. Bookworm is pointing us to here train your first neural network la-dee-da all right. What do we got?

A

This guy trains, a neural network to neural network model, to classify images of clothing? Hey great example, musical bookworm.

E

A

We're on track here, I think we're okay on this one. Let's see here. This is the fast-paced overview of a complete tensorflow program. With the details explained as we go good. This guide uses TF Kerris, a high level API to build and train models and tensorflow. Okay.

B

A

This very much looks like Python we're gonna be using here to build this. Am I right import the fashion minused data set. This guide uses the fashion minused data set, which contains 70,000 grayscale images in ten categories. The images show individual articles of clothing at low resolution 28 by 28. As seen here. Oh man, look at all that right can I. Can I zoom in no I can't zoom view image right holy crow? Look at all that stuff, alright, makes sense, cool.

A

Do I know of any dotnet core compatible drawing libraries that will work with Linux.

A

Dotnet, what do you mean by drawing libraries you're trying to render an image? There are a couple of them out there that folks are using right.

A

That are not right that are not just on Windows. If you do an image processing right, core compat system, drawing yeah I, don't think you want that, but it's the mono implementation of system, drawing.

A

Yeah there's appropriate runtime, so use image. Sharp is one that a lot of folks use there go cross-platform library for processing of image files, so.

A

Oh brave Cobra I'm, not gonna, read that.

B

A

That's not right. No, no! No! No.

A

Just a total coincidence, you've just been used in puppeteer, sometimes funded, really straightforward, yeah.

B

A

Okay, so fashion MN ist is intended or as a drop-in replacement for the classic minused is that I pronounce that dataset often used as the hello world of machine learning programs for computer vision contains images of her handwritten digits. Yes, I've seen that in an identical format to the articles of clothing, we'll use here, okay, we'll use 60,000 images to train the network in 10,000 images to evaluate how accurately the network learned to classify images. Okay, you access the fashion MN ist directly from tensorflow just import and load the data.

A

B

A

E

A

Available for you and you just say, load data, okay, loading, the data set returns for numpy, arrays, train images and train labels or rays, or the training set the date of the model uses to learn the models tested against the test set test images in the test labels arrays. The images are 28 by 28 numpy arrays, with the pixel values, ranging between 0 and 255, because it's black and void the labels are an array of integers raising from 0 to ranging from 0 to 9.

A

So we may have to cast these images to black and white gareth asks. How am I planning to get enough images to train a model for my hats, I only need 70,000 images of hats to train it. How's it going Gareth good to see. That's Gareth, it's great to great to see him because everything's awesome make sure you check out Gareth stream.

A

You can find him streaming as a member of the live coders team click that live coders link just below me over there and you can learn about all 65 members of the live coders that are writing. Code live here and teaching working through software projects like this and other interesting things that they're working on as well. Would you need network access this way? No I would not need network access, brave, go brave, I'm training, all this stuff locally and I have my model. All trained and saved I wouldn't need to hit the network.

A

I could run directly against a local instance of this image. Sharpe is exactly what alternative corn was looking for. Congratulations, that's great! To hear alternative corn I'm glad that's gonna work for you! Oh I have all that sound bites. Gareth I have all of them. Let's say: I have sound bites. Let's say no! A lot of sound bites.

A

Duke of twit Duke twitch, or is that Duke of twitch, but welcome Duke twitch great to see you thanks for joining us here so thinking out loud here if we're gonna, if we were to build in train using a very similar data set here, we'd have to use black and white images of hats.

A

28 by 28 is really small 28 by 28 pixel where'd. It go right. There.

A

So brave Cobra what I understand from just a little bit of reading that I've done over here. The model is data, that's been configured, that's been loaded and analyzed. That knows how to take some example. That's been given in the in in write in this case the first example the the first column of the model- analyze it in apply a label. That's on the second part of the model and we're gonna apply different algorithms to figure out how to correlate those things.

A

You then use the results of that model. That piece of code that's sitting on disk, to run your predictions through to run your right, so it's done all of its predictions and it's figured out how to make a good comparison between the two we're gonna run through some new test data and say well using that model that you've analyzed and built some correlation between run this new piece of data through it and tell me what you think it is based on those other factors in the model.

A

That's what I understand a machine learning model to be Simon Says. Surely you only need a bunch of images, free stream from your twitch streamers, of each of the hats from different angles as the input Simon.

C

A

Guess who has 200 in 77 videos sitting out there on YouTube begging to be analyzed? huh What do you think of that?

A

So my thinking is when it comes time to build the model we can add, we can write, have it navigate to those different videos right. We know where the hat is in. The model weakens right because I'm always sitting in the same spot, we can have it grab that area of the video take a screenshot and.

A

Right turn that into black and white, and then we need to go through and assign what the name is of that hat to 70,000 records.

A

May be okay for the same videos, we're gonna, be able to copy and paste and at least reap heat across those.

A

Kind of like the old sounds like method, but more complex is brave, Cobra, I think so.

A

ha Simon jeering yeah we to quote to misquote the short circuit, more input its and hopefully we get it trained enough that it'll go really fast. It'll go so fast, they'll go to plan, janeski was seeing onyx AI, perhaps I'm not familiar with onyx I I would need to learn a little bit more about that. Onyx is supported by ml, Duffman duked, which says I'm watching a workshop.

A

That I did of asp net core, hey thanks, so much for tuning in there's two workshops out there on youtube, there's one that goes to 2.1 and there's a new one that we just recorded about a month ago that targets asp net 2.2 they're, both great they're, both about the same length 6 to 8 hours check about. Let me know what you think of them: yeah we're we're really excited about those we're looking at doing some more workshops in the future.

A

I've got tons of content that that we're gonna be building around and recording these longer form. Oh, my.

D

Godness kittens just recent finds: we've had a twitch baby. Oh my.

B

God javascript.

D

A

Teller Ganza thanks so much for the follow I appreciate you joining us fierce kittens is here. Welcome! That's my good friend fierce kittens she's a she's, a creative streamer. She does once or twice a month, oh yeah.

A

She does. She writes code, live on her channel and I. Think I think we might set up to raid her a little bit later and she loves JavaScript right. Yet you see she doesn't. She doesn't she'll freak out with JavaScript um and she's a good friend of the channel. We've done a pair programming stream with heard the past she's tremendous besides doing code here on Twitch. She also does some other some other creative things, some other art, some other crafts she's done.

A

Painting um making handbags makeup amazed if she's working on cosplay, sometimes really great stuff, you gotta check out fierce kitten stream. Can somebody throw a shout-out to fierce kittens out there for me, but nine months? Yes, the right the twitch baby, meme and because just because we know first kittens loves, JavaScript, so much to name the baby, Java JavaScript ins, that's a thing you might be late, though. Oh well, we'll see out we'll see if we can catch it.

A

If we can raid right into you new glasses, also duke twitch says it's great well, I'm glad you enjoyed the the workshop there, Duke Duke twitch, we're around here four or five times a week on this stream.

A

Males who sing? Okay, thanks for the great content just rush home, do not want to miss this well gosh. Thank.

B

A

You never have to have to rush there's always video and domain here on Twitch and we archive things over to YouTube lots of shout outs for Fears kittens. Thank you so much, no, not script net. Oh no! You missed the stream just go over there. Did it did I miss it. Let's say: do you miss the stream it it's? It's.

C

A

Might be, but I mean just for missing the stream I've got to say.

F

Shame. Shame shame fear.

A

Kitten says every time people mention JavaScript I know a demon gets its horns. What are you thinking? No.

C

A

Wrong, it's so wrong.

C

A

Course, on the other hand, chimes in because horses love, JavaScript love, it I, don't know who likes sam'l, we'll figure out who legs Amal. So we are gonna, be wrapping up here in just a few minutes, just checking in with the family over here cool.

A

Okay, let's see here all right, let's see what we can do here so going back to the tensorflow example, so I think we're gonna have to get a whole bunch of images and we can probably get those from from my from my youtube videos explore the data, explore the format of the data set before training the model following shows. There are 60,000 images in the training set, see it's not 70,000. It's 60,000.

E

A

Let's see here, let's see divulge er says.

A

Is asking for looking for some help getting started with Azure.

A

Let me there there are some startup programs for Azure that you can look into take a look at the website. There's there's ways to get to get startup credits for Azure for small companies that are starting up and you can. You can get a lot of credits as a startup. There, I'm not gonna, dig into that right now, but.

A

Drop me a message on Twitter and I'll, see what I can do to hook you up with some some links. Every language has its uses as fierce kittens. Yes, the requirements of you. Your users drive the language choices. You agreed you've been developing at discord botnets as alternative corn for a while.

A

Absolutely no experience with VPS hosting.

A

Will VPS hosting allow running dotnet core applications? The VPS Virtual Private servers? Well, you can run dotnet core inside of a docker container, so put it inside of a container. You can run the container and on those we.

C

A

You'll be good to go but dotnet core applications. You can compile with the runtime and deploy the whole enchilada and you should be good, so yeah.

E

A

A look at that alright, so it looks like.

A

It looks like we're: gonna need to yeah. Each date must be pre processed before training network. If you inspect the first image, you will see that the pixel values fall in the range of 0 to 255.

A

B

A

Why isn't that black and white? Why they make it a we scale these values to range of 0 to 1 before feeding the neural network model? For this we divide the values by 255. It's important that the training set in the training that the training set and the testing set or pre processed in the same way, yeah okay, display the first 25 images of the training set and display the class name below each image, verify that the data is in the correct format and we're ready to build and train the network so ankle boot.

A

Okay, t-shirt, t-shirt, top dress, t-shirt top sandal, sandal, okay, so the different articles of clothing that make sense building the neural network requires configuring. The layers of the model then compiling them all. Okay, the basic building block of the neural network is the layer layers extract representations from the data fed into them, and hopefully these representations are more meaningful for the problem at hand. Most of deep learning consists of chaining together, simple layers. Most layers, like TF Karis layers, dense, have parameters that are learned during training.

A

Okay, the first layer in this network flatten transforms the format of images from a 2d array of 28 by 28 pixels to a 1d array of 28 by 28 or 784 pixels sure that makes sense. Okay, think of this layer is unstacking rows of pixels and the image in lining them up right. So you get just a string of them. Instead of a square. Okay, I can follow that startups. Microsoft comm there you go that's another good location for folks to go.

A

You can do color and tensor flow as well. Just a bit more complicated to process the images musical bookworm says: they're a tensorflow beginner cool, so working through tutorials, but I've already done this one now fantastic before the models ready for training. It needs a few more settings. These are added during the models compile step, loss, function, optimizer and metrics. This measures how accurate the model is during training cool. We want to minimalize the function to steer the model in the right direction, and that makes sense.

A

This is how the model is updated, based on the data it sees and its lost function. Okay, metrics used to monitor the training and testing steps. Okay, I understand that.

A

The following example uses accuracy: the fraction of the images that are correctly classified sure, train the model training. The neural network model requires the following steps: feed the trading model to the data, the train, images and train labels, arrays sure the model learns to associate images and labels. Ok, we asked the model to make predictions about a test set. In this example, the test images array we verified that the predictions match the labels from the test, labels array, cool to start training called the model plot fit method.

A

The model is fit to the training data alright. So let me think this through here. First kitten's is now hosting us. Well, thank you so much there friend I really appreciate you hosting us.

F

A

More than happy to be here on the fierce kittens Channel live. Look at me, I'm talking about JavaScript, actually I'm talking about Python, that's! Ok! We like Python.

F

What'd you do I'm.

A

Learning about Python and training models here so thinking out loud.

A

Tell me if I'm not oversimplifying, this I feel like you'd lose relationships between the vertical pixels when you flatten them out like that I kind of agree with you, Sam Obama, wonder how training two models against the transpose and comparing the outcomes would go. Maybe what if, if.

A

It's taking the data and it's flattening them out right each one of these images right and it's got numbers 0 to 255 for how bright each one of the images each one of the pixels is for all 28 by 28 of this, and it turns it into just the 7. Well visit 700, byte long array, 784 byte long array.

A

Couldn't we just put that in a sequel, server database.

B

B

A

Where am I here right and just do this same thing: training the thing training, the doodad, because it's just got a byte array. Oh don't get me started about the snakes things, I kind of agree with you fierce kitten, so we're gonna need higher res pics to do training right, 28 pixels by 28 pixels. These are really small here. Right, I feel like we're. Gonna need at least 100 by 100, if not 200, by 200, to get a good picture and I feel like we're.

A

Gonna want not and we're not gonna want a single byte but I feel like we're. Gonna want 3 bytes, so we can set the RGB of whatever that pixel is, which is gonna, make the the image significantly bigger for 200 by 200 right. Each one of those images, then, is 200 by 200 is what 4 and 4 zeros 40,000, and if we put three digits with each one of those, it's 120 thousand.

A

120,000 whites.

D

A

And use that to train- maybe wouldn't I- need images of hats. Only while brave Cobra I want not necessarily of hats, only but hats as they're on my head, because that's what we're gonna be. That's what we will eventually test against, because we'll take a screenshot of where I am where my noggin is here. Isolate that part of the video, because I never move and grab the image, grab the hat image and test against that store them in Azure blob storage.

A

That's not a bad place to store them, but we need to turn them into raw data. So we can analyze.

E

A

Yeah Simon's right more input, surely it's more accurate in color, given that it's differentiating factor in logo, design, I kind of agree with you, Simon we're gonna need to look at that numbers escalating in spent exponentially on the time. It'll take the training, I kind of agree with you, and if we set up right, if we set that up to run training over an hour or so offline, let it do its thing. We might get a nice model.

A

Out of that that we can use I mean this is a very simple making and then I said. Make predictions model predict test images. A prediction is an array of ten numbers. These describe the confidence of the model that the image is corresponds to each of the ten different articles of clothing. Okay. So this is each one of the ten and how close it is well, this one it's pretty close to and that wouldn't it's really close to right.

A

Yeah, the max is the ninth one right: zero one, two three four or five: six, seven, eight nine, this one that's in at 80% yeah.

A

We can check the test label to see. This is correct: test label zero! Nine! We can graph this to look at the full set of ten class predictions.

A

Look at that alright and that it's 80%, accurate, cool, okay,.

A

Plot several images with their predictions.

A

Okay, cut correct color, correct color, correct predictions in blue incorrect in red okay. So, in that one, two three by one two, three four five in these 15 it got. Fourteen of those predictions correct. That's not bad!

A

Doesn't the azure custom vision service? Do all this work for us a bit. It doesn't know necessarily all the things that we do end.

D

Boy just subscribe two months two months, thanks.

A

So much for the resub I appreciate that and we'll make a donation to coderdojo like we are with all of our subs. All of our cheers here all quarter long thanks! So much for that. But yes, you're right as your custom vision does do that, but I need to be connected and using Azure and using it's actually quite expensive to do those tests to to run it. So only if you have I mean this is a toy that we're building here right.

A

This is something that's helping us learn a little bit about machine learning, where that's machine learning as a service go load. All these images up there identify the things until the spit out an answer so yeah right, that's a thing, but I want to get a little bit more hands-on with what's involved in building a model, what's involved in training the model and and then applying it. So, let's see here is there a way to debug sequel code executed by the web server inside Visual? Studio? Yes!

A

Well, you want to get all the way to the sequel code to the SQL code. You want to just see the SQL code that you want.

A

Let me know duked, which either way you can actually run when you hit the debugger in Visual Studio it'll step through and take you to those objects, we're using an object, relational mapper in those samples that you're seeing and we're hiding that SQL you can get the SQL out of entity framework. You can extract that you can log that, if you'd like, so you can see what it's actually executing.

A

Yes, those emojis did go across the overlay. Yes, they did pac-man jr., says custom vision. You could run on prep in a docker container once you train them on Azure. There's there's a thing that it calls back to Azure net logs. How much you use it I believe, but you can export models from custom vision. Nice. You found cognitive services to be too expensive.

A

If you have a lot of data, this would solve it, but you need to make it yourself exactly I want to try and learn something new here get outside the box, something custom. Yes, the emotes show up they. Of course, they show up I.

A

Didn't see those events, that's too many of us try to get sequel trace could help the other good yep, yep, yep, yep I'm, waiting to see those and most go flying by. They should go flying by right.

A

It's it's amazing, all the emotes. I.

A

Thinka, grilled cheeses, it only sends I think the first eight, so I didn't want to flood my screen there. We go there's some more right and only subscribers get that opportunity to do that. A plural stels eat, tacos yep. You can play that game too. There's the tacos, everybody loves, tacos, see.

B

A

And that helps with the sentiment analysis over there so and then we need to build a model. It sounds like in order to work with this. We need to build a really big model, so we can load this up.

A

What's its images transfer learning with wood I have no idea what this is tensorflow hub is a way to share pre-trained model components, yeah an image net class yeah. This is the one that it was mentioning over in the ml net sample. I can't imagine this is going to No. Well, there's Grace Hopper, okay,.

A

What's pix to pix normal style transfer image segmentation? What's this.

B

Well, that doesn't help.

A

Learn how to segment images? Segmentation is the process of generating pixel wise segmentations, giving the class of the object visible at each pixel. We could be identifying the location and boundaries of people with an image or identifying cell nuclei from an image underneath.

A

Large number of car images each car taken from a different angle, in addition to each car image, we have an Associated manually cut out mask our task is to automatically create these cutout masks for unseen data.

A

Fierce kitten says fritz. Please stop interrupting taco talk.

D

A

By channel missing and you're in my chat room.

A

It's Tuesday, please observe its glorious, TOC Onis right. So what I'm? What I'm seeing here is we're gonna need to well. I. Don't want to do this bit here. We're gonna need to get a little bit out there and figure out figure out how to set up and build our model. We're gonna need to capture a lot of images of fritz hats, lots of images.

A

Fortunately, right we've got lots of images. We've got all that video that we can mind. So it sounds like we've got a couple steps here that we need to do. We need to do a little bit of research to figure out how we're gonna generate our model.

A

What's the process here to actually generate that tensorflow model that we're gonna be able to feed into ml net, and have it start to do the identification with clearly once we have the model right here with ml net? This feels like it's as simple as.

A

Or is it new model input send in whatever that model input is, and in our new case, it'll be in the image of my hat predicting result and then output that right, this part seems simple. It's getting that thing. That's gonna be a little bit of a challenge, so I think we're gonna need to spend a little bit of time figuring that out.

A

Yeah I need to wrap up, though I've got I've got something going on here with the family that that came up. That I'm gonna need to take a look at taco model input. Yeah, probably do that, but I'm not analyzing tacos, here, I'm doing something simple: I'm, analyzing hats, to figure out the logo and the color and be able to guess.

A

Oh that's the visual studio 2019, but we're gonna come back to this another time after we do a little bit more research and we figure out a strategy and we get the and we get some images and we'll we'll do something to capture images from our from our videos. We'll have to set up something some sort of a script: that'll analyze those videos and pull out that content.

A

You know what I'm gonna do here. You know, let me see I think we've got somebody that we can raid here. Let me see if we can do this real, quick, oh yeah, oh yeah, he does he's not even gonna know it's coming.

A

Let's set up a raid, let's put out the raid Cole and I need to fix the structure of this grab that first line of text, if your subscriber in the chat room, if you're not that's, okay, grab the second line of text out of there and we're gonna announce our presence, we're gonna visit our friend blue lava.

A

This is yeah. This is another one of our friends from twitch. This is John Bulava, he's working on calls it restful doom over here he's working on and we are going to head over. There he's writing some code and quite sure what it is that he's doing here. My my twitch UI is frozen. There. We go thanks so much everybody for joining us.

A

We're gonna do a little bit more research on this see if we can collect some of those images of my hats so that we can start modeling and loading up a model for this, and maybe I'll come back to this in a in in a week or so when we've got all that data collected and we can start processing thanks for much for joining me for from walking through and learning a little bit of a metal net with me getting ready to say hi to to lulav Oh with me, I'm gonna copy out that raid statement.

A

So we can head over and say hello all right, friends thanks! So much for joining us, I will be back tomorrow. 10:00 a.m. Eastern is when I'll be back and I hope. You join me on Thursday Thursday at 10:00 a.m. I'm gonna have a guest. My guests will be Jeffrey Palermo, we're gonna, be talking about answer: DevOps, building your application, automating all kinds of DevOps things around your applications, so you can get the best bang for your bite that you write and say that I just made that up already friends.

A

I will see you next time take care.