>> On today's Visual Studio Toolbox, we wrap up our mini-series of ML.NET. Veronika is going to show us how to do image classification. [MUSIC] >> Hi. Welcome to Visual Studio Toolbox. I'm your host Robert Green. Today we're going to wrap up our three-part series on ML.NET, and my guest is Veronika Kolesnikova, there she is. Hey, how are you? >> Hey, Robert. Good to see you again. >> The first episode, you provided us with a really nice overview of machine learning and where it fits in the grand scheme of things. Last week, we saw how easy it was. I've seen this many many times, but it still just shocks me how it is easy it is to do. We built the model and we did some sentiment analysis and it was a handful of lines of code and it's just so cool. Then today, we're going to do a bit more advanced scenario, we're going to do image classification, which I know people have seen a bunch of times, but it's the next one that is an obvious thing that you might want to be able to do in your app. We're going to do that today and that'll make our nice little three-part series, and then, of course, we'll have you back on the show later on to do more stuff. >> Thank you, Robert. I can share my screen again. Here we have the same project we were working on last time, that empty console application with that ML.NET model connected to it and also the sample console application where we actually tested the ML.NET model that we created using the model builder, and now I want to move to a more advanced scenario and show them image classification. You can start by again, hitting machine learning model. Here again, you can update your name. It is definitely recommended to do that, but I am not going to do it just for demo purposes. Here, we see that beautiful ML.NET model builder's screen with all available scenarios. Last time we used data classification, but today we'll be using image classification and going back to those little labels here. You can see Azure and local, it is definitely beneficial for that type of scenarios to have that option. Sometimes you are using just more high-quality high-resolution images or just lots of images for training and you don't have enough that machine power on your local machine so Azure is there to help you. I'm going to pick that scenario here, and actually we're not only two options, we're getting three options here, that local GPU, that is something that was introduced recently. You need to actually do a separate set up in order to use that local GPU option. It is all documented, so if you're interested in trying that option, you can definitely set it up, connect to your GPU, and use it. I know some people they are hesitant to use features in preview, and that's totally normal. If you don't feel like you want to use that in production or with production-level models, then you don't have to use it. Then Azure is also in preview right now, but it's just amazing how easy it is to connect to Azure and use all the power of the Cloud right from your Visual Studio. Today I will just move forward with the local CPU option. It's not in preview, it is widely used everywhere so you don't need to worry about changes there or something not working. You can see the environment settings here just so you can make sure that you have enough compute power to manage your data set. Here, I really like that screen because they have really good example on the right, how you can organize your images. You need to have the top-level folder with images and then sub folders with categories >> What you're doing is you're taking the images that you're going to use to build the model and you're organizing them. You put all the pictures of dogs in the dog's folder, all the pictures of cats in the cat's folder, all the pictures of lamas in the llama's folder, and then that's how it knows what a dog looks like, what a cat looks like, what llama looks like, and that's how you build your model, right? You take all the images you're going to base the model on and organize it like that, and then with future images, then it's trained on how to detect things, right? You pick the categorization upfront. >> Yes. >> Okay >> Really, good thing here is that you don't need to go through each and every image and tag them separately. I know that for object detection scenario, you need to actually do relatively a lot of work ahead of time. You need to tag your images, you need to select objects using special tool, and then you can use that document with tags for your machine learning and training. It's a little tricky there that's why I'm really happy with this scenario, how easy it is to organize your data. Here, I can select that top-level folder, make sure that you have images that are compliant with those supported file formats. I have two categories, cats and dogs, and I am choosing that top-level folder here. Here the data preview is great. You can see all the images. I have only five, so you can see all of them, but obviously, if you have more than probably seven or eight, then you won't see all of them, but at least you see the preview, you see what images you have. You can see I have five images of dogs and five images of cats. They are high-resolution images. I just got them from Internet. Here, I can go to next step. You can see the difference with the previous scenario that you actually don't have that time set up, it's not up to you anymore to set up the amount of time because with images it gets more complex and it is really hard to figure out how long you need to train it and it might be expensive to actually train it and retrain it several times, especially when you are connected to Azure then it is consuming resources. Even in this scenario where you use in your local CPU, it is still consuming your local resources and just might stop some ongoing processes. That's why you don't have options to set up the time, it will do it automatically. It is easier, you don't need to worry about one more thing. >> Cool. In general, how long does it take to train images? >> Usually it takes quite some time. It depends on how many images you have. With high-resolution images, it takes longer, with lower-resolution images, it takes a little faster. I tried it with 10 images in each category, it took around maybe 20 minutes. >> We won't watch this for 20 minutes, so we'll come back when it's done. >> We can see that training is complete. We see the output with that magic completed word, some specifics of what happened there. >> It is just edification that took about five minutes. We just chatted for five minutes while waiting for that. >> Yeah, and it gives see training time right here in seconds. If you want to convert it, feel free to do it. >> Five minutes, that was my estimate. Neural models pretty done with. >> Good job. One model was explored. That is definitely not enough, but again that is because we don't have enough data. As I mentioned before, I tried a similar process with 10 images in every category that definitely went way longer. Properly around 20 minutes but it provided better accuracy because you can see the accuracy is zero percent and also it checked more models and found the best model for the data. As we can see and learn from this example that five images are just not enough both for the demo purposes. We're just going to go ahead and use the model that was created. Here, we can see a similar screen as we saw with the previous scenario where we can evaluate the model. You can see the model type, it's image classification, but less convenient but a better thing is that we actually need to browse for an image, it is not going to provide any images here by default, which is good. We're just forcing that model that I mentioned before where you're supposed to separate your data. One data set you use for model training and the second data set use for evaluation. That is a good thing here. I have a separate data set with different images, I call that folder test. I can pick an image, that image wasn't used for training. Again, it's really convenient here. Really convenient that you see the image that you chose. >> It's not clear what that is apparently. >> That's not good. >> I guess, when it said zero percent accuracy it wasn't kidding. >> It wasn't, yeah. >> With more images, it would do a better job? >> Definitely. As I mentioned, I had tried it before with 10 images, and with 10 images, it provides you the result. I even tried to trick it and try an image with both a cat and a dog on the same image, and then the percentage was 60 percent cat and 40 percent dog, which was probably accurate. Sixty percent of the image was the cat and then 40 was the dog. But here because the accuracy is really low, that is unusually low, it can't detect whether it's a cat or a dog. Also, a good example of what you're suppose to do when you see something like that, so maybe you can ignore the accuracy in the previous step, but when you get to evaluation, you'll get those weird numbers here and then you understand that, okay, that didn't go well and I need to add more images or the images were really low quality and the model wasn't trained enough. >> That's good learning. But the code to consume it is still exactly the same. Let's just pretend this is a good model. >> I'm sure everyone can go and try it themselves, use more high-quality images and figure out how it works, but the process is the same better images there. Here, you can copy the code snippet if you are planning to use it in your existing application or you want to see how it's going to work, see all the code and examples. You can add a console application or web API here. >> Cool. >> I'm going to add another console application, and now I'm going to have two test console applications and one empty, but you can see the model that we created is already here, and it's the second model here in a ZIP file. You don't have to use it here, you can move it around, use it with other applications, so that ZIP file is the model. Here we have the same model MLmodel2.zip. Then when we go to the program file, we get that image source for testing, then we're passing it to the model, and then predicting label so it's the same as the previous scenario. >> It is the same code, that is really cool. It is just so easy to use. >> You can create one of those scenarios see how it works, and then reuse that code with different types of models. >> Then you can decide what to do with the answer, is it validating that it's a picture of what you think or is it more elaborate shooting off an email or even taking the picture and moving it to a different folder that says code you can write once you have the result? >> Yeah. Another really interesting part is that you have that code-behind here, you can see how it was all built. Next time if you feel adventurous and you want to use ModelBuilder and you wanted to try to build them from scratch, you can use it as an example how it actually was built and then reuse the code from the training file and then consumption file and just recreate the training process without ModelBuilder. >> Doing image classification just been thinking about a lot of scenarios. I take a lot of pictures, and let's say I go to the zoo and I take pictures of lions and tigers and giraffes and monkeys, then I take a lot of pictures. Let's say I go to the zoo and I take 200 pictures and I want to have all the pictures of lions in one folder and giraffes and another, I would just need to build a model and then I could just write a little simple app that just goes into that folder, looks at every picture, classifies it, then moves into the right folder. So rather than having to do that by hand, I could create a model and create a little program that moves on the pictures for me, right? >> Yes, exactly. Or you can also filter out maybe some inappropriate pictures, depends on what kind of application you are building maybe it's a public forum where people post pictures and you need to have that option to filter some things out. >> Very cool, and again, very easy to use. That's awesome. In just these three episodes, we've gotten a really good overview of ML.NET. Last time we saw how to do sentiment analysis, today we looked at image classification and what we're learning is the code is pretty straight forward, you just need to build the model, and the better your data, the better the model. We saw last time with sentiment analysis, with a modest amount of data, we can do pretty well. Here today we learned that if you just have five pictures of dogs, five pictures of cats, your model's not going to be that good so you need a lot more data which just means it takes longer to train the model, but ML.NET does all of that hard heavy lifting and work for you. Then to use the model in your code, handful lines of code, so cool. Thanks so much for coming on the show and doing this. >> Thanks, Robert. It's great to be here. >> Hope you guys enjoyed that and we will see you next time on Visual Studio Toolbox. [MUSIC]
No comments:
Post a Comment