In the last video we loaded our data and then we did some pre-processing to make it ready for model training. In this video we are going to build a convolutional neural network and train that network on our trained data set and then we'll measure the accuracy using our test data set. If the accuracy looks okay we will export that model to a file on a disk so that it can be used later on in our fast api based server for making prediction. Let's get started! Our data set is pre-processed and we are ready to build the model. We are going to use convolutional neural network which is a famous type of neural network architecture. If you are solving you know image classification problem and for that I made a video actually so if you go to Youtube search for this you'll find my video. I highly recommend you watch this because there are a lot of concepts that I'm clearing in this video and we are going to use this architecture you know in our code today. So when you have convolutional neural network there are two parts. There are convolutional layers so there are convolutional and pulling layers so convolutional pulling, convolutional pulling there could be a number of layers and then there is a dense layer where you just flatten it so we are going to follow this architecture and that therefore I am saying you need to watch this video if you haven't to understand what we are doing in today's session. So here I'm now building basically my model. So I'll say model is equal to this and we will have set of layers here. So the first layer would be resize and rescale whenever you feed an image first thing will be you will resize it then you scale it by dividing by 255, then you do data augmentation so that you do you know horizontal vertical flip random rotation to generate new samples? The third layers is your convolutional layer so here I will say Conv 2d. Okay. What are the arguments here? you can say tensorflow conv2d layer you find a help the first argument is number of filters and kernel size. So this kernel size is basically the size of the filter so if you looked at this video you know I'm talking about filters here in this video and the filter.. Let's see this is that filter in that video so this is three by three filter. So the second argument is what filter this is. This is kernel size which means this is the three by three what is filters then? Okay. Let's see. So when you build a convolutional neural network, I'm just quickly skipping this ahead you are applying different filters. So let's say one filter. Let's say you are doing koalas image detection one filter could be to detect eye, one filter could be to detect nose this way you detect eye, nose, ear and if you have eye, nose ear of koala you can say the image has koala's hair too! Similarly if you have hands and legs of Koala you can say Koala the image has Koala's body too! And if you have koala's head and body it means it can be a koala's image. So you filled flatten that and you kind of figure that out. So here you have one, two three, four, five.. five layers in our case we are going to use 32 layers. So here how did I come with 32? Well this was trial and error. You need to have a lot of layers where you can detect edges. you know the small features? The second one is the actual filter size and the beauty of convolutional neural network is that in this example I had to give specific ionosphere filter there. Here the neural network just figures it out for you. Then the activation layer the popular activation layer for hidden layers is always value because it's very fast to compute and the input shape here is 256 by 256, right? So the 256 by what is it's image size. Okay. I'm going to maybe create a variable using this because that could be useful little later and basically I created this variable assign it here pretty straightforward. After convolutional layer you generally have pulling layer. So if you look at again this video this video here you see convolutional plus value then pulling conversion value you just keep on repeating it and pulling is basically there are different types of pulling. Here, this image is showing you the max pooling. Max pooling means you have 4 by 4 pixel area you take the maximum which is 8 eight another four by four area what is maximum nine. So nine here. What is maximum-three? three three two two when you do this you are preserving the features and you are reducing the size of the image which can be very helpful you know computation wise. So that's exactly what we did here. So here I'm saying I want a max pooling layer of size 2d and then I before preparing this video by the way I did some trial and error and i figured maybe I need to stack few max pooling and convert 2d layers. Okay. This was trial and error you can remove couple of layers and figure it out. Again 30 to 60 to 64 all these numbers are a little bit based on that trial and error and then once you have stacked the max pooling you know convert to 2d layer max pooling layer. If you look at this image once again you need to now flatten it so whatever data you get here you need to flatten it so that it's an array of neurons and you can have then a hidden layer here. You know a dense layer and the final layer is your final output classification layer. So I'm going to flatten it here and then let me add a dense layer of you know 64 neurons and then I'm going to add okay let me say that my number of classes is 3 and my last layer will have three neurons with soft max activation function. So soft max activation function if you don't know about it it will normalize the probability of your classes okay? After that you need to do model dot build. I mean this is just the API. API requires that okay we're getting images must have either three or four dimensions. Also looks like the input shape that we gave here is not just image it's see what we're supplying is actually there is channels. And there is a batch size too. That's the format of our input. Okay? Now. Our model architecture is ready I will just do a quick summary we are not done training yet by the way so this summary just prints the quick summary, you know. The parameters these are the weights that you need to train that's why it says trainable parameter. You know you are doing back propagation on all this weight basically number of weights. In deep learning we always first define the neural network architecture so that is this. Then we do compile using optimizers and all that like adam is a famous optimizer. Then you define your loss function your matrix in each epoch. what type of metric will you use to track the gradient descent basically? to you know.. So the accuracy is the metric that we use to kind of track the training your training process. And then finally this is the third step which is model dot fit. Here you actually train your network so in the epochs you know we have 50 epochs. We have batch size or batch size. We have I will do verbose one so that it just prints lot of output and we can see what's going on and we have validation data as well. By the way so this validation will be used in you know during the each epoch it can help you track the accuracy. And by the way I will record all the history of every epochs in the history parameter so that we can do some plot we can chart some plots later on so if this is gonna run it's gonna take some time based on what computer you have. Sometimes people have you know gpus. So gpus makes this process faster if you don't it might take time you can try with less epochs in like 50s it's taking time. 50 might be high so you can try with less epoch but let me explain you all these parameters that you're seeing here. So in the first go the loss the accuracy that you got on a training data set was 51. But the accuracy on validation data set was 71 percent. You can understand so it will train your model first on training data set it will measure the accuracy and then it will kind of take this is like running little test so it will run little test using validation data and it gets 71 accuracy. Okay. Then it just keeps on running and you'll see as we have more epochs the accuracy keeps on improving see validation accuracy 84 now we started with what 71. Even the training accuracy just keeps on improving. So right now I'm here at 45th epoch 46 47 and my validation accuracy is already close to one. See one it yeah it jumps up and down but it's pretty good actually so now I will run a test on my test data set. See, before you deploy any model you want to run a test you want to figure out how well your model is performing by trying it out on a test data set. So that it's not biased. Your model has not seen this data set. This is the first time we are trying it and it can give us good understanding. The accuracy comes to be 98 percent which is actually pretty good. And if you look at this course parameter it's a python list which has first parameters as loss. The second parameter is accuracy. Now I want to play with this history parameter, okay? You see history so we store history. Let's see what this history is. Actually, it tells me it's a tera's callbacks history and immediately I'll go to my friend do you know who is my friend? Well, Google tensorflow keras callback history. And it will give me some documentation history has some parameters. So let me try this parameter see what's going on here. Okay it's just telling me you know that I had 50 epochs 54 steps verbose was one and so on it also has history.history So history object has another element called history and if you look at all the keys see you get four parameter loss, accuracy validation loss and validation accuracy. Those are this parameter loss accuracy, validation loss, validation accuracy and we ran 50 epochs. So for each parameter we have 50 value. See one, two, three for loss we have 50 accuracy view 50 validation 50 and so on and that's exactly what this is. So if you look at let's say accuracy let me look at accuracy. It's a python list 50 values so see if you do lan 50 value correct and what are these values? Let's check it 0.513.612 Okay let's go here 0.513.612. So all these numbers during epochs it will just record them in the history and once we have those we can see what we can do is this. We can plot them in a nice matplotlib chart you know so I will first get all these errors stored in different variables and then I will plot the accuracy. So let me plot training and validation accuracy. So this is how you plot it. I'm not going to go into details. This is not matplotlib tutorial but I'm just plotting training versus validation accuracy and you can see that these are epochs on the x axis. As number of epochs so we started with very less accuracy, one accuracy is like perfect. So we started with point five you know and it kept on jumping jumping jumping and it went here. This graph shows that even if you train for let's say 30 or 40 epochs is fine you don't need to go all the way till 50 because you already achieved very high accuracy. And if you want to plot another chart next to it, let's say the loss chart you know. Then see the loss initially is high but it just loss is basically an error. So an error in back propagation keeps on reducing as you proceed forward in your epochs. So far it looks good. Now I want to make a prediction so again we will go through test data setup. We'll run our for loop. So we are taking just one batch. So this will be 32 images and I will take the first image. Okay, what is my first image? My first image is this. okay? Let me just print it this is my first image. You see it's a tensor and if you want to convert tensor to numpy it's just a different format. Okay. I want to now display this. So if you do as type and 8. and if you do you know this function it will show you the actual image. So this is an actual image and if you print it you see it's a three dimensional array rgb. Okay. I will call it first image. This is my first image, okay? If you do this you get second image, third image and so on. This is my first image. What is my first label? My first label is this. Okay? Now I will say this print first image to predict and then just that image and the first image actual label is you know what? Let me do this here. So this is my first label so actual label is one but I want to know the class type. So maybe you can do class names here and you get the class type. So it will say okay the actual label is this. This is the actual label I want to find the predicted label, okay? What is my predicted label? My model is ready I can call predict function The predict function expects the image batch. Fine. I will give the image batch and whatever I get is the batch prediction so this is this will be the prediction for 32 images. If I want the prediction for the first image I need to do this badge prediction, zero correct. Oay so my image is potato early blight and my prediction is this okay why is it a three dimensional array? one, two, three.. Well because of this you see our model. Or model architecture is we have three neurons and activation is soft max. So soft max is just a probability so there are like three probabilities and whatever is the highest probability that is the class. So you see e raised to minus 0 1 e raised to minus 7 e raised to 18. So this is the actual class. Okay and if you want to get like which is the maximum number, there is a function called np dot arg max. This gives you the index. So arc max here the answer will be see if you have this array let me show you if you do np essentially we are doing np dot argmax on this. [Music] Oh! import numpy as correct the answer is zero. Meaning this the item at the zeroth location is maximum. Okay that will be your class? Now again you want to convert it to class name so you will say class names like this. Okay you do one by one this is not confusing by the way it's pretty straightforward actually. So this one is a predicted labor and you will be like okay why image is changing well we are doing some shuffling here. That's why image keeps on changing. see? late blight late blight only bled out so my actual label is early blight my predicted label is also early blight so my model is performing awesome. see? We are trying all these cases and it is exactly telling me what image is this. You see this is the power of convolutional neural network. Okay. Nw what I'm going to do is this. I will write a function and I'm not going to go into detail. I already wrote this here so this is just a simple function that is taking model and image as an input and telling you what is the predicted class and what is the confidence. Hundred percent confidence means say I am the accuracy of prediction is hundred percent, you know. So that's your confidence score and if you look at this we converted image into an image array and then we created a batch out of it after that we called predict function and then we figured out predictor class. This code is same as this code, okay? So we we figured it out and we found the confidence so we run this and now I'm going to run my prediction on the entire batch and I'm going to print those so we'll run a for loop on the first batch and I don't want to do prediction on all 32 images. Let me just pick only nine images you know so I will say for i in range nine okay I will first show that particular image so you see. But wait so it's not showing all the images and that's because you need to Ithink have a subplot then you know so see it shows all the images. Okay, I don't want all these axes so I will look plt dot access off. So now I don't see it. I also want to increase the dimension so that each image looks a little bigger. Now each image is bigger because of this line and now I want to put a title what should be my title? Well actual predicted label and confidence okay. So let's first so I'm calling predicted function using the model that we have and using the image, okay. It's giving me predicted class and confidence and my actual class is this. So I have predicted class actual class and confidence and all these three things one two three. I will print as a plot title. So I will say plot title okay what is my plot title? I will use python's format string so if you want to display just the actual label this is how you do it. If you use python format string then you know predicted class will be this. So see actual class and predictor class Okay that came now I want one more line and I want to introduce confidence as well. Perfect! You see my actual image is early blight my prediction is only right confidence is 100. Confidence is pretty high in most of see most of them is 99 this is amazing folks. Amazing amazing performance indeed! Now I would like to save the model. To save the model c we were here in this training file right but I will create a new directory called models and I will save my model here okay? So here my I will show my model by model version so let's say this model working is one I can just say model dot save again I'm using python form string dot dot slash models y dot dot well I'm in this directory if I want to go to modules Ineed to go one step up. That's why dot dot and then model so here then here it will save the model okay? Here again python format string model version is one. When I do that here you will notice it will see it created one. See here it's saved model to here and if you look at this file this is how tensorflow models are saved. All there are all these directories you don't need to worry about what these are you will be able to import model from this one directory and let's say one if you have one model next time what you might want to do is as a data scientist you want to conduct some experiments. Maybe you remove data augmentation, maybe you change model architecture a little bit and you want to export let's say two or three different versions of your model that you can try out. So you can you know you can run it like you can say okay model 2 and that will be my second version and see it will save it here right now they are same but maybe you can make some changes here. You know in your model architecture or maybe change optimizer and save different versions. If you want to keep a running count and kind of figure out. Okay, if there is a two directory then automatically increment the model version then you can do that using OS library. See if you have import OS and if I do OS dot list directories if I do this what's gonna happen. See? It is telling me what directives we have now. This is returning me string how do you convert string to an integer? Well it's very simple. You use list comprehension you will say okay end i. For i in this and you get integer if you want to listen model zero as version you can get that also and eventually we find a max of all this and max of this is two and when you do one you get your next model version. So this is how you get your next model version. So I'm just going to say this is my model version okay? And I want to save it like this so now when I save this and when I run it up it created version 3. if I run it again it automatically creates version four version four and so on. So this way you can auto increment your models. So folks, that's all I had for this video. In the next video we are going to build a fast API server and also we will be using tensorflow serving actually to load all these models. So we'll cover all of that but as an exercise I want you to change this model for tomato classification so we tried potatoes, okay? You need to do this exercise where you are building this model for this tomato classification and there are many classes so it will require some changes into your model architecture. But that's an exercise a sincere student is the one who works on the exercise. If you work on exercise you will learn a lot of useful things. All right, if you like this video please give it a thumbs up and share it with your friends so that they can also benefit and I will see you in the next video in the same tutorial series. Thank you very much.
