Image classification using CNN (CIFAR10 dataset) | Deep Learning Tutorial 24 (Tensorflow & Python)

in the last video we looked at what is convolutional neural network if you have not seen it i highly recommend watching that before you continue on this video today we are going to do image classification of 60 000 small images and this data set is coming from tensorflow library itself is called cifar10 database it has various objects such as aeroplane ship frog host etc and we will be doing image classification using convolutional neural network so we'll straightaway jump into coding and then in the end i have an exercise for you so make sure you watch till then and you do the exercise on your own the data set that we are going to use in this video has 60 000 32 by 32 colored images with three rgb channels and we'll be doing a classification into one of the 10 categories so these are all the 10 categories that we have i have loaded some essential libraries here and the first thing i'm gonna do is load the cfar10 database into my jupiter notebook so when you call this load data method what you get as a return value is this so you have x strain your y train and then you have x test white is x taste y taste and i'll just quickly check the shape of the extreme you see that the training samples are 50 000 each sample is 32 by 32 image and three is for rgb channels now let's look at the test so in the test we have 10 000 images so this data set is kind of decent size you know it is not too small now i want to check each of the training samples so when you of course do this x train let's say zero you know you get this three dimensional array 32 by 32 into three you know rgb channels uh i want to just quickly plot this to see you know how this thing looks so you can do method you know matplotlib we have already imported here as plt and this has a function called i am show and you get this is actually a frog and if you do one this is a truck but the image is pretty big so i want to just control the size of the image so you know when the image is little smaller you can clearly see that this is a truck and this is a frog just for convenience i'm going to write a function called plot sample here and this function is taking x and y and index and printing that particular uh image sample so here it will be this and i want to also on x label i want to print the label basically whether it's a frog or a ship now i know that the labels are basically zero to ten so airplane will be zero automobile will be one and so on so i'll just define the list variable called classes to be this okay and then uh we can use this classes here classes okay how do we use it so let's think about this so we need to analyze first the [Music] y actually you know so let's check what is there in the y so if you do y train dot shape you get 50 000 samples and one dimensional array so let's just quickly check the first five samples so what i notice is these are the numbers you know six will be for what okay six is zero one two three four five six six is frog then nine we saw it was a ship so nine is a ship but this is a two dimensional array actually we don't need two dimensional array here because we just need direct uh category right so i'm going to reshape this particular uh y train so i what i will do is so see right now it looks like this okay now i will do if you know about numpy arrays you know that to reshape y train you can just call y train dot reshape and the way you want to reshape is the first dimension which is 10 000 you want to keep it as it is so when you don't want to change that dimension you just say minus one and the second dimension is you want to flatten this so instead of 6 being an array you want a simple six so then you will just leave this blank and as a result what will happen is now when you do y train file okay it is saying oh i need to do reshape actually okay so you notice the difference instead of two dimensional array now this is a one dimensional array perfect so now simply if i have classes array like this and in that if i pass lesson nine i get truck okay uh so i can then use that index here and that index will be y of index okay so this will be y of index and i execute this and now when i say plot sample let's say in my training sample and here i will do y train let's say i give zero so for zero i get frog see this x label is just printing this particular label and if you do let's say one you get a truck this is a truck it's not a ship maybe i misspoke but it's a truck so nine number is struck eight number is ship so you can check various samples what the sample this is also a truck see this is a then this is a deer and so on so this is a just quick data exploration part that we did now uh we want to normalize our data we saw in our previous tutorials that whenever you have an image you want to um divide each pixel value by 255 because the pixel value ranges from 0 to 255 for each of the channels rg and b and if you divide it by 255 you will be normalizing it into a 0 to 1 range so let's again quickly check this you see the values here are 103 59 and so on okay now if i do this you see see this is the power of numpy array you can just divide it by 255 it will divide every element in the entire array so now i can simply say x strain is extreme divided by 255 next taste is okay so the values are normalized now we'll build a simple artificial neural network first to train the model i want to see the performance of how artificial neural network works and then we will do a convolutional neural network in one of the previous tutorials we did gpu performance for the same c par 10 data set so i'm going to use that notebook it is on my github here and i'll just copy paste some code from there so here uh we built a simple artificial neural network so this is the same neural network i will build and see how it performs so here you see it's very simple the input layer is a flattened layer it's a first layer which accept the shape of 32 by 32 by 3 then we have two deep layers one having three thousand neurons the other having one thousand neurons and the last layer is having ten uh categories because we have total ten categories right see this is like 10 so when you train this neural network it will use a dense artificial neural network with all these parameters optimizer is sgd and there is pass categorical cross entropy value as an input in that tutorial by the way i use a categorical cross entropy and if you want to know the difference between categorical and sparse categorical then see i have this nice image that can explain you so here uh whenever uh you have let me just go in our presentation mode so whenever you have um your y as one hot encoded vector so let's say you have ship here which is number nine and if your y is something like this which is one heart encoded you will use categorical cross entropy but if y is directly a value which is number eight you use sparse categorical cross entropy in that video of gpu performance we converted into categorical that's why we use categorical but here we are directly using the value 8 and so on and that's why we are using sparse categorical cross entropy here after the training uh you see that accuracy i am running just five box but accuracy is pretty low 48 percent you see 48.58 on training samples when you evaluate it on test samples it is 47 percent so artificial neural network is performing really bad on this data set with five epochs i have also printed here a classification report and this classification report gives precision recall an f1 score on each of the classes so for example this nine nine is what truck or a ship maybe ship no it's a truck actually okay so for truck class the precision is 59 percent recall is 48 percent and you can see this matrixes if you don't know about precision recoil and f1 score again in my this tutorial series there is a video on what these terms are so it is better if you watch these videos in sequence now we are going to use cnn uh to improve the performance of this model so how do you use cnn well in our previous video we saw that cnn so let me pull that presentation here so cnn will have um couple of layers so those layers would be usually you'll have convolutional layer then you have value which is the activation and then pulling then again convolution value and pulling and then you have a dense network so you will keep the dense network in front of your current dense network all we are going to do is add some convolution and a pulling layers so let's uh do that so let me just copy this code here okay so here this will be cnn and so you will have some cnn layers here okay and this will be your dense network so when uh you are in the middle layer you don't need to specify the shape because the network can figure it out automatically and just to keep things simple i'm just going to keep only one dense network because my cnn would have done most of the work so i don't need so many neurons and so many deep layers okay and i'm going to use a soft mix function here this is one of the popular activation function the difference between soft mix and sigmoid is that so soft max mix will normalize your probability so it will um if you have probabilities like let's say your two classes class one is let's say it gives 0.45 value and class 2 gives whatever 0.67 value if you use soft mix you get this as an output uh you if you use sigmoid sorry then you get this as an output but if you soft max then you get this divided by this as an output so what happens is now you are normalizing let me just show you quickly so you get an idea so you're normalizing it to a range between zero to one so your total sum basically will be z one so if you sum this point this one and this one you get total sum as one so that is soft max but if you are using sigmoid uh and if you sum this up you don't necessarily get one all the time now in this code here i'm going to add first convolutional layer and then the max pooling layer uh this goes back to our presentation earlier so it's pretty straightforward convolution is are detecting the features in your image now when we saw this presentation for the image of nine in the previous video we saw that uh we can have this kind of three filters so the first one is detecting the loopy pattern and the second one is detecting the vertical edge which is the middle part third one is detecting the tail the best thing about convolutional neural network is you don't need to tell it what what the filters are it will figure out the filters for you you only need to tell the filter size and how many filters you want see in this case we had three filters one two three so if you look at this image we get these three filter stack filters feature maps here we will use you know just random like maybe 32 filters so this can detect 32 different features or different edges in your image and then the actual filter size is specified here so let's say we are using three by three filter see here also we had three by three filter this green box so that's what this is and then you need to specify activation which is value okay so here i have activation as value and then my input shape is 32 by 32 by 3 okay because that is one image so going back to the architecture once again all right what did we have we had convolutional value and then we had pulling so we have convolution and value and you have pulling so pulling i'm going to specify a max pooling of 2x2 max pooling is very popular pulling a method people use average pooling as well and for activation also you know value is quite popular it is less expensive to calculate now you can have only one layer or you can have multiple doesn't matter you know like you kind of figure this out by trial and error so just for fun i'm going to have another set of convolution and max pulling layer okay so see here we had like convolution pooling convolution pooling so that's what we have now now we do our usual model compile thing this is something we have been doing in all our videos you need to specify optimizer loss and matrix i already explained why you use past categorical cross entropy here there are various optimizers misers sgd adam and adam adam is the kind of popular one that gives the good accuracy so that's the reason i'm using it here and i need to say cnn actually here okay and then i will do uh cnn dot bit and i will run it for tan epochs there are 50 000 images to train so it's gonna take some time so please have some patience here you saw that after tan epochs it gave me 83 percent accuracy actually if you compare it with a n see after phi epoch i get 73 accuracy in a n after 5 i got only 48 so you can see that using cnn helps you tremendously and now i will test this out on my test set and here you know i got 70 percent accuracy which is pretty good okay uh if you training for more epochs you can probably get more accuracy and you can do little fine tuning but for the images which are like this which are kind of you see these images are like kind of random actually this is actually a tough data set uh and in this kind of complex stock data set also we get 70 percent accuracy which is pretty nice you know i'm like kind of happy with that so good job you just built your first convolutional neural network well done i hope you are also typing this code as you watch this video uh so coding in parallel with me is going to be definitely useful now i will do a plotting of some samples uh so let's see okay we got some error and i think this is happening because we did not reshape you know we had to reshape our y test because our white is okay before executing this let me show you it is a two dimensional array i want to convert it into one dimension so if you do this it is now one dimensional array and in my y test i want to see what is my first image okay my first image is ship now let's use our model to predict this and see how it performs so what i am going to do now is i will just say i will predict all x test samples and i will get y predicted values so if you do y predictor by the way the values in each of these elements are actually probability distribution between 0 and 1. it is using softmax and you want to find out which element is the max so for that you can use arg max function okay so the way arc max works is if you do let's say np dot arg max see let me just quickly show you you know why it gave one because 12 is the maximum element and the index of 12 is one see if i do this one let's say i make this one then say it is getting two so that's what it is doing so here if i supply let's say y print 0 then the index is 5 which means number 5 is the maximum element i'm going to convert this into y classes so what i will do is i will say okay why classes is equal to i'll use list comprehension in uh python so the way you do that is it's like running a for loop so for element in y predicted or you're computing arg max for each of these so you know like what you get as a result is something that you can compare with y test so now if i have y test here you see this is how so the first sample it got wrong it was five got three but the second third port it got right and then see it will make some errors okay because our accuracy test accuracy is 70 so it's fine all right so now we already saw that test sample one is a ship which is this eight okay and if you look at now classes you know we have this classes what is classes class is just a simple list of all the classes and if you supply y class is one see so this is the actual value and this is the predicted value so for zeroth element first element it got it right if you look at zeroth element it would have made a mistake so it was less a cat and it said it's a dog but even for you it is hard to recognize that this is a cat right so cnn making mistake is probably okay here because this image is very very difficult if you look at let's say zero one two three the third image third image is aeroplane see it's flying like airplane and if you look at three see arrow blind so here it did not make a mistake i also printed a classification report using y test and y classes so now using cnn you're getting better numbers see your f1 score is overall better here 80 81 percent here is less 50 percent 70 percent and so on but when you looked at simple a n the score was a quite low i know i did only five epochs but even if you try tiny box you still get lower score in a n and better score in cnn in cnn also computation is less because we are using max pooling layer and that reduces the dimension the jupyter notebook that we covered in this video is uploaded on my github i am going to provide this link in the video description so make sure you check the video description and then in the end uh what we are going to do is uh i think it's not this it's this one yeah so it is this particular notebook and the link is in the video description and if you go towards the end there is an xss for you so what you have to do is and by the way this is the most interesting part of this tutorial i i always say that exercises are the most important thing just by watching videos you're not going to learn anything you have to practice it practice makes the man or woman perfect so in one of the previous video we did handwritten digits classification for mnist data set and i have this notebook here so what i want to do what i want you to do is take this notebook and do the same digit classification using cnn this one is using a n okay artificial neural network see here and you have to just use a cnn and just see how the accuracy and the classification report improves when you use cnn after you have worked on your solution you can match your solution with my solution by clicking on this solution link so i hope you're liking this deep learning tutorial series so far if you do please give it a thumbs up share it with your friends i am trying to make this this tutorial series very very simple so that even a total new beginner can understand so please check out my previous videos if you have not already going in sequence is important because i am covering the topics one by one so please watch everything practice it and i wish you all the best and thank you for watching

Channel: codebasics

Views: 135,112

Keywords: yt:cc=on, cnn image classification python, tensorflow cnn image classification tutorial, cnn deep learning, image classification, convolutional neural network, image classification using deep learning, cnn image classification, image classification using tensorflow, cnn model for image classification, image classification using cnn, image classification in python, image classification using neural network, image classification using cnn in python, tensorflow image classification

Id: 7HPwo4wnJeA

Length: 28min 11sec (1691 seconds)

Published: Sat Oct 24 2020