>> In today's Visual Studio Toolbox, we're going to continue our look at machine learning. Veronika's going to show us how easy it is to create and use a model inside Visual Studio. [MUSIC] >> Hi welcome to Visual Studio Toolbox. I'm your host Robert Green and today we are continuing our discussion of ML.net this is part 2 of a three-part series featuring Veronika Kolesnikova. There she is. Hey Veronika. >> Hey Robert. Good to see you. >> Good to see you again. In the previous episode, we did a high-level overview. We didn't really get into Visual Studio yet but I think it was very good to set the stage what is machine learning? Where does it fit into artificial intelligence? Today we're ready to dive in, fire up Visual Studio and start building a model and then seeing how we can use it. >> Yeah that's right. Let me share my screen and show my Visual Studio. >> Which Visual Studio is this? Is this the old Visual Studio 2019 or the brand new just released Visual Studio 2022? >> That's the brand new Visual Studio 2022. >> Good answer. >> I've been using the old one and new one in parallel for some time, but now I fully switch to 2022. >> Excellent. >> Here I want to create just a simple empty console application. Creating new project window and choosing just a console app. I have a bunch of them so it's number two. I'm using C-sharp here and then creating a new solution. You can always name it something more meaningful. I think I'll just keep it as ConsoleApp2. >> Next one. Then here you can choose framework. I'm sure .Net developers have seen that window and they can choose the right .Net version that they prefer. >> It's in all about .Net 6, they can go back and watch .Net conferences a couple of weeks ago where it was talked about in great detail. >> Exactly. I will choose .Net 6 and create a new app. It is empty. Doesn't have anything, just Hello world. >> We actually recorded this before the launch so that's why it says Preview and it says watch the .Net launched on November eighth which I said we published this will have already happened. Not that anybody is surprised that when we recorded an episode at a time we actually did it ahead of time. >> So maybe we can cut out that yellow notification but then it still assess preview here. >> Leave it in. >> You can see in the Solution Explorer at the ConsoleApp, doesn't have anything other than ProgramCS file and now I'll be adding machine learning magic here. Click on "Add" and then add machine learning model. >> Now did you have to install? Did you have to select a particular workload or add a specific component to get that or does it come by default? >> ML.Net ModelBuilder is available inside Visual Studio you just need to make sure that when you are installing it this component is checked. >> It's not a workload, it's one of the components? >> Yes. Or you can install it later if for some reason it's not there you can always go back and add it to your installation. Here when we are adding that ML.Net ModelBuilder we are able to change the name. I'm not going to change name just leave it by default since it's not the main goal of this demo but when you're working with production code working on your glam project, definitely don't forget to update the names. It just takes a couple of seconds and then you can see that beautiful UI that is ML.Net ModelBuilder. Now you can see the scenarios available data classification, value prediction and limited scenarios that I mentioned in our previous episode like anomaly detection, forecasting, and clustering. I want you to pay attention to those little labels here. It says locals for some of them and then Azure and local. For example, for image classification, you might not have enough compute power on your local machine in order to perform that model training and model creation so if that's the case then you can always connect to Azure and use that Azure magic. But today I want to start with the most basic, in my opinion, scenario. It is widely used in all applications, so I'll start with the data classification. >> This would be a good place for people new to this to start out and learn how this stuff works? >> Yes. You can use your local processor and the memory you have, you even need to connect to Azure you'll be using just text data so nothing too complex there or no heavy data is involved usually. Before moving forward you can see on that screen that it shows my local CPU environment just to make sure that that is working correctly, that I have enough computer power to actually move forward with that scenario. Then for next step, I have two options here connect to a file uploaded here or connect to SQL server. I have all the data inside ATSV file, so let me actually find it. I have a couple of options available I just want to use an option with less data so it is not taking too long. >> Did you create that data? Is it from a sample? >> It's a sample data. You can find it on GitHub there are also lots of third party tools and datasets that are available. I chose that Wikipedia detox file with only 250 lines. That is not enough for a production model but for demo purposes that's good enough, it will run through it faster. I will save a little bit of time and I got that data from GitHub but they are also online for those with datasets so you have lots of options for demo purposes. Here you can see data preview or what's happening in the dataset. It's not going to show you the whole dataset obviously because it might be too long. But just 10 rows out of 250. Then here we need to choose column to predict and we'll be predicting sentiment based on the text we are trying to figure out if it's a rude comment or non-rude comment, non-toxic comment. Also, you have that advanced data options. I have only two columns herein the dataset but if you have more then you can pick and choose what columns you're going to use and other data options that might be needed before that actual training. Now we need to move to next step and next step is training. Here is that higher tricky part where you need to choose how many seconds you want to train it. Out of the box and by default it tries to estimate the amount of time you might need but also they have that link to documentation so you can read more about training time and figure out what is better for your data set. Here I can start the training. In the output window I see what's happening there. It actually checks different model types and tries to figure out the best and the best model that was selected is that model here. I'm not going to try to pronounce it but again you can read a lot about types of models for all data that documentation. You can see that actually went only through two types because we had only 10 seconds but usually if you have more time to train it, then it will go through more types of models and pick the best one for your data. Also a good thing that you can see here on the screen is the accuracy. You can see that the accuracy's pretty low. I won't to recommend using the model with that accuracy unless you're building something not important at all. >> But then what would happen if you trained it for longer? >> That's a good question. >> Does it depend on the data? >> It depends on a data, we can definitely try to train it longer and just check if that accuracy increases. It might increase a little bit. I'm just saying that based on my experience but it's not going to increase a lot because of a pretty small dataset and you can see the same output but now because we have more time it went to three types. >> Interesting. >> It increased a little bit but still not production grade accuracy but that's okay we are using it just for demo purposes so I'm going to go to next step. Next step is evaluation. Here by default it's just picking out the first row and it gets that sentiment text from there. Ideally in the real world case, I wouldn't recommend just using the first row. Usually what data scientists do, they split the dataset into two parts, one they're using for building the model and the second one they using for evaluation or you need to have at least a couple of options for evaluations that were not part of that training dataset. But here I'm going to use just the default option click on "Predict". It is toxic. Its pretty sure 99 percent so that's pretty good. We can move to next step. Next step is consume. There are really good options here if you already have your application built up and you are just waiting for that machine learning parts then you can just copy that snippet that they are providing and start using the model. But if you want to play a little more with the model, see how it works, see a good example of usage and you don't mind creating a separate app that you will delete later most likely, then you can just add a ConsoleApp or a web API. I'm going to use the ConsoleApp here, and again you can rename it here. I'm not going to do that. Here you are getting basically the whole console app which is connected to ML. Net model. It is already using it so we can just run that console application and see how it actually worked. Let's go. Here, I'm going to close that and this part 2. You can see in the program file it is automatically passing that first line so whatever we did during that evaluation step it is trying to do the same thing here. It is using the text to predict and passing it to the model than some writelines in the console so they are helping to understand what's happening there. It provides them sentiment. Also repeats the sentiment texts and then provides their prediction. It's a good way to understand how you can use the model but you definitely don't have to include a console application in your solution. If you know what you're doing you just grabbed in that snippet that was there and then connect the ML.Net model. The model itself is here. It's that MLmodel1.zip. Can see it in both applications. I think I can just run it and make sure it is working. >> We basically created a model and then we're using that model in a couple lines of code and we're just passing in some text which could be read from a file or picked off a entry in a form or on the screen and then we're applying it against the model and coming back and saying that's rude. That's it? It's that easy? >> Yeah. That's it. >> That's amazing. Then of course all the difficulty comes in obviously the dataset and then somebody ahead of time needed to decide for all of this whether, what is a rude statement, what's not a root statement? What the values are, so the initial dataset initially this might be a manual process. Somebody just has a quick screen, brings all the comments and they say negative, positive, neutral? Then just go through that and eventually you'd have enough of that data in there to create an accurate model. You can create a model from any amount of data but the more data you have, the more good data you have, the better your model so then you replace that manual process with your ML model, review it to keep training the data and training the model and eventually whether it's a week, a month, a year who knows, hopefully sooner rather than later, you'd have a model that's good enough that you don't need to do that manual process anymore. You probably just on a regular basis check it to keep training of the model but eventually, you've automated this process. >> You need to keep logs as for any software for machine learning, you need to have logs to make sure that the predictions are accurate or maybe user feedback if you're especially building bots then usually people need to have that opportunity to complain or maybe say that something didn't work correctly and then you are keeping those logs. They can be automatic. They don't have to manually type anything but based on those logs, you can actually provide more data and retrain your model with that new data and make it more accurate and just overall better. >> That's basically all it takes to get started doing this with a very simple model very simple scenario but the key takeaway is about ModelBuilder which is part of Visual Studio you just have to make sure it gets installed. There is an amazing amount of power in there and of course, it gets better and better and better as the team behind it who knows an awful lot about this subject continues to improve it. >> I started using the ModelBuilder may be just as they released it a couple of years ago and I can see big changes definitely gets more convenient. Better faster, more convenient to use and I'm really impressed with the work amount that that team is putting on. >> What are some of the other basic scenarios that people could look into or that or that you've seen? Here we just did sentiment analysis. I know we've seen image classification. Is this a dog? Is this a cat? What type of cars is? But what are some of the other things that you've seen people be pretty successful with? >> I think all scenarios are definitely a popular that's why they decided to include them all here. Value prediction is popular. You can predict the value of your house if you're trying to sell your house or maybe buy a house. Data classification is simple and powerful and it is easy to set up, maintained with image classification and there are lots of tools that are available so you can mix and match maybe you are creating some parts using the model builder and then you can add cognitive services on top of it, doing speech recognition or something like that. >> If you wanted to do something with image classification you just need to know what the categories are ahead of time and then have enough images and identify that this type of images, this category and the more images you have ahead of time to go into each category, the better job the model would be able to do to figure things out? Then it's just a question of constantly monitoring how good your data is and how good your model is and your model is only going to be as good as the data you put into it, so it becomes this feedback loop of my models no good that's because my data is no good so you got to get better data which leads to a better model which leads to better data leads to better model and you get this virtuous cycle I would imagine. >> Yeah, data is constantly changing so we need to make sure that you're staying up to date. >> This would be a good place to wrap up this episode and what are we going to see next week? >> Next week we'll see a more advanced scenario. I want to show image classification. >> Hope you join us for that. We will see you next time on Visual Studio Toolbox. [MUSIC]
No comments:
Post a Comment