Pytorch Data Augmentation using Torchvision

[Music] in this video we want to learn how to use PI torch inbuilt transforms on images so first if you're unfamiliar with data augmentation or wonder why you should use it essentially more data is always better when we train Internet and if we can get more for free by doing some transformations to our images it's almost always a good thing in this case we're loading our data using a custom data set which I've chose showed how to do in a previous video but that's really not really the focus in this video anyway you managed to load the data is fine in this case the data that we're working with is two pictures of cats and we want to apply some transformations to those and we want to see how they look like after we've transformed the transformations and perhaps most importantly we want to see how do we actually apply the transformations so let's see what we do first is we can actually just for image common label in dataset we can do print image dot shape yeah so we get two images which are colored so three channels RGB and 224 by 224 size the first thing I want to show you how to do is we're gonna use transforms dot compose alright let's go back so we're you can see we do one single transformation which is we convert the in this case numpy array to a tensor but let's say we want to do more transformations than this happen here okay so transforms that compose is what we're going to use to combine several different transformations I'll talk a little bit more about that later first of all we want to do is we're just going to use transforms that compose and we're gonna do transforms that to pill image that's the first thing we're gonna do because all the transformations work on this I guess format when we do it to pill image so that's usually what we do first then there are a bunch of this different transformations you can apply to two images we're gonna go through some of them but really there are many configurations of them you might want to read the documentation for those specifics but let's say we want to do transforms that random horizontal flip and we can input a probability here with which is default 0.5 and then in the end we can do transform start to tensor so for example the name is pretty self-explanatory but what it does is it flips the image horizontally so what we can do is we can do we'd be able to visualize this we can do save we can use the torch vision that utils save image that I've imported here so save image image with the name image plus string of image num which we're going to define here then we're gonna do plus PNG format and then we're just gonna do image num plus equals one and let's say we do that for we do that ten times so in total we will have 20 images and we just run that now if we go back to our folder here we have 20 images we can see that yeah so this one is flipped horizontally I believe that one yeah so some of them are flipped horizontally and and and that's kind of like a simple transformation to do one thing we could do as well is we could do transforms that color jitter with brightness 0.5 so what this does is apply some random brightness like change to the color to the image there are more things you can input here as well but yeah really I just want to go through some of the most common ones and you can choose for your specific case which one you think is best another one we can do is transform that resize so we first resize the image to let's say 256 by 256 and then we do some random crop of that image we do the crop to 24 - 24 then let's say we want to apply we could also apply rotation to the image right a rotated image of a cat still a cat so we can do transforms that random rotation and we input degrees with 45 for example um one more thing we could do transforms that random vertical flip but perhaps you know it's more common to see it's quite uncommon to see vertically flipped images of cats so maybe this is a low probability of 0.05 and yeah so those are some examples what we can do actually we can add some more we can do transforms that random gray scale with a probability of let's say twenty percent so this will convert the image to grayscale with a twenty percent probability another thing that is quite important to do which improves the training and quite a bit is a transform start normalized what you do after you have two tensor and for this you input a mean and a standard deviation and essentially what you want to do here is so for each channel in this case we have three channels you want to find the mean for that specific channel across all training examples and for all of the over all of the pixel values you would find that mean value and you would also find the standard deviation for all training examples across all pixel values and you would define those for each channel so there will be three values in this case then you would input them like this and like this now of course I don't know the mean value in the standard deviation for those two two images that I have in this case but you find those values first and then you do that and what it does it takes each value for that channel and then it subtracts it with the mean that you input it and then it divides by the standard deviation so in this case this would actually not do anything right since this would just subtract zero and divide by one perhaps we can write note this does nothing yeah but in practice you would find those values first and then you would use them I just want to add a comment here about transforms that compose that what it does is it applies all the transformations that we wrote inside to the image that we send in and it also does it in the order that we wrote it so it performs to peel image before resize and random crop after resize etc and since we have a lot of random transformations in the transform compose each time we send in an image we will get another image as output so ya know we've done a lot of transforms on our images let's run this I'll see yes we need a comma here anything and yeah another comma here it should be as small as yeah okay so now we've run it and let's see so here we got a rotation and it's also a grayscale grayscale again yeah so essentially you can see that the images still all look like cats except there are variants of that image and this is a bit weird but yeah still a cat so yeah that's an example of how you would use it now this is just we just do the images but really what you would do when training is you would just do a training loader here using the data loader and then you would train a network yeah so if you have any questions leave them in the comment below hopefully this was useful and thank you so much for watching the video