Transcript:
Hi, Aaron. So today we’re gonna be looking at tents aboard and yeah, so lets. Go start with it. Alright, so what? I have over here is a script where I’m going to be looking at an Mis fashion. So not the image data set that you’re probably used to we’re looking at a bit harder than set mania, and so, Tessa Flo has tents aboard to visualize your most functional things that that but know many tutorials talk about the the projector inside, right so good to be looking at the projector to look at embeddings especially alright, so embedding’s pictures and so on. Alright, so so the first thing that I’m gonna be doing is I’m going to be reading in this labels or CSV file, which tells you what each label corresponds so 0 called Constitution 5 on to sandals and so on, okay, the the Donald itself I downloaded using this. She’ll script so like it’s. All doctor eyes for you. So you can play around with the later if you need and this bar over here, don’t worry about it too much because I’m literally copying and pasting from the official in this fashion site on how to read in the data. OK, so now with tents aboard the you need to create a logs directory, right, so we recreate from the current directory. We’re gonna be crane this logs thing, and we’re gonna call slash one to make it just to save yourself a situation. Okay, but once we done that, and once we created our love-story, there’s there’s two things that we need to be doing. The first thing is to create this metadata for so the metadata file, but it’s gonna contain is, so we have 6,000 images, so for all the 6,000 I need to write in a row. That will have the class, which is a number from 0 to 9 and the corresponding name of it now. The name is not compulsory, but it usually helps right so so as a thumb rule. Just this key to this convention, right, so write down the class and a name of the class for each row. So so that turn me over here? Then the second thing is this thing called a sprite image. Now, so we have to sit out an image. Let me just show you what they look like. So while this loads. Yeah, so did this one looks like right, so we have a whole bunch of pictures of shoes, shirts and so on and the Sprite. So this is the sprite dot. PNG thing that we need to create it doesn’t need to be caused by the PNG, but what it has to be is a square image that contains all the all the images that you’re about to show right so now 60,000 like it’s not a square number like 9 or 16 and 25 So what’s gonna happen is what’s gonna happen? Is we’re going to be taking the square root of 60,000 okay, so it’s gonna be – 1044 so it’s gonna be a 245 by 245 image. Okay, so so each row will contain two are 45 images. Yep, so, anyway, this function over here, which, which I’ve used from from someone else is creating exactly that image, right, so it’s creating that square image and for what I was remaining. It’s just gonna put zeros, okay, so if I go back to my picture and I go to my last row. Yeah, you see, you see a blank bit, right, So which means it’s not filled and that’s fine so anyway, coming back to this this, there’s. Two things that you need to do first is write the metadata file and the second is to write this right picture. Okay, so this is just me showing the calculations on how I end up getting the size of image would turn which turn out to be six thousand eight hundred sixty-eight thousand sixty because each image is a 28 by 28 image and so on, you need to factor that in as well anyway, so let’s get back to test bored so with tents aboard now to get the tents aboard working. I need to write those images into a into a tensor flow variable. All right, so that’s what I’m doing over here. Okay, so I’m calling it features, but keep in mind. Test board is designed for embedding’s not images. So I’m reaching it so that I flatten it out, right, So the first the first row is the number of images that I have the first number in shapes, so I will have 60,000 images here and this number. I put a minus 1 because it’s a slump. I go and calculate yourself when you restate things, but in this in our case, we’ll end up with 784 okay, so it will end up being 60,000 by 784 matrix. All right, so these are my embedding scale. So what we gonna do? Is we’re going to sew. Tf not trained at Saber. That’s where you tell 10 suppor to look at and look at thislook in our case because this variable. So that’s what we’re gonna be saving this. You have to run the initializer because keep in mind. All variables in this flow needs to be initialized, even though you’re sending in a constant number. Yeah, and then with I’m telling you to save it into this checkpoint file. Alright, anyway, so the rest of it is fairly standard, so you tell it the metadata part right, so where it is and then you tell this the sprite image part now this, and then you tell the shape of the images and then you go projected up Visualize embedding’s all right, so this is really key key thing. Now you don’t have to really memorize any of these functions or you, or you really have to do is pretty much copy and paste right because even. I got this from some other side, and I will link that in in my my gay people. Yeah, anyway, so once that once it’s all done, all we have to is still 10 support, – lock directory and the part of the directory. Once you do that, we end up in this nice visualization of the 60,000 images that we had now, it said. So you can play around. So this is the principle component analysis frigate on three three dimensions right, so remember, it was seven high eighty four dimensional space, and then we projected onto three dimensions. Okay, so one thing one cool thing the thing you can do is you can color it, right so you can color it by class. So when you look at it, you can see like the shoes are starting to ankle boot, so it’s not to come in one place so and the other thing of the cool thing is once. If you click on one, then you can see what the nearest points are in that space. So ankle boots are really post ankle boots. So let’s look at T-shirt So T-shirt teacher. / Top is close to shirts as well so. I’m assuming she. When this mean shirts, they mean long-sleeve shirt’s, so it’s. The point is, this is an easy data set to deal with all right so now. The cool thing that we really want to do is use. Teasley, now this might take a while. So let me let me talk. What tees need us so? T’s knee is a bit more complicated at PC, so PC simply pushes them, which is a high dimensional vector onto the small dimensional vector, but tasty actually looks at the manifold. So what is the manifold? So if I if I had a shape like a set of data points? That’s that’s like this, right, Lets let’s curve it a bit. So a data point somewhere there somewhere here. What is it does is it takes into account that that they take there in this surface rather than just saying this point is. Euclidean distance is closed, right, so it takes into account the fact that this curve, which is and stuff like that Anyway, so right, so let’s let’s see, so it’s yeah, so you can set the learning rate so it is keep on, keeps on learning what the manifold looks like, right, we need when you’re projecting on a three dimensional space. Yeah, so that we can see better. Yes, so you should always get color, right. Yeah, it starts showing. Yeah, there. You go right so yeah, so the pants is on -. Yeah, so you can see. The colors are separating much better than it did in PCA, which was one Which was one big blob. So, yeah, you can twist this around and, yeah, so you can see this nice shape panning out and how it’s separating out these classes, and by the way so once you once you start to see some separation should really turn down your learning rate, So that mean, yeah, so let me just pause it there, all right, so, yeah, we tease me. You get this nice, nice thing. And for unfortunately, for some reason, if I go back to PCA. This is gonna start all over again. I’m not really sure how to freeze that and just come back to this thing anyway. So you have, so you can see pull over. Yeah, so pull over. It is close to show rather than pull or but yeah, so you can see you can see that similar. Similar kind of clothes are in similar areas. Right and yes. I can ankle boots in one place, which is which seem to be the easy thing to classify, but it’s doing a lot better than PCA. So you can see all these classes now. Keep in mind, we’re not doing any classification, right, So the the chastening algorithm doesn’t actually take into account class class of the of the image that are sent right so yes, it happens to color it, but that’s because I provided this metadata file, right, so it’s not actually doing any classification so yeah, so that’s pretty much it when it comes to test board, you can help play around with the other stuff that’s in there, so if you had a second second embedding to look at it, it will be youll. Be up here and you can. We can make your top mode and you can select things and look at what’s the closest point. So, yeah, so this is a bit to play around with. But those are the main things about 10 support, So hopefully that all made sense to you. Oh, and last thing before. I go now too. Once you run once you run this thing either in in a batch in your terminal or on the Javascript. And so you have to go to localhost six double. O six. Okay, so I pretty much created a link over here going to look, Of course, it’s level six, but yeah, otherwise, if you want to manually, that’s what you do, but again. I’m going to be putting the git repo in the comments below, Please to like share it. And if you have any questions, please do comment, but otherwise thanks for watching you.