1 00:00:01,260 --> 00:00:10,480 ‫Now in this lecture we are going to be on model using Cato's before that we will do one small thing. 2 00:00:10,510 --> 00:00:18,660 ‫We will take out end of this small set of training examples as validation set of validation set is used 3 00:00:18,660 --> 00:00:22,000 ‫to tune the model hyper barometers. 4 00:00:22,170 --> 00:00:30,720 ‫So we'll just take our first 5000 examples as validation set and the remaining we will see them in another 5 00:00:30,720 --> 00:00:39,630 ‫variable called partial training say so here and setting the noses of postpaid him in a variable called 6 00:00:39,630 --> 00:00:44,090 ‫value indices. 7 00:00:44,210 --> 00:00:50,600 ‫Now I knew this variable could take all the information off plus 5000 and stored it into the validation 8 00:00:50,600 --> 00:00:51,070 ‫images 9 00:00:54,050 --> 00:01:00,660 ‫for the first 5000 brain images are now stored in value images. 10 00:01:00,660 --> 00:01:05,030 ‫The remaining part of the images is stored in part underscored brain images 11 00:01:08,780 --> 00:01:11,280 ‫we do the same thing with labels. 12 00:01:11,600 --> 00:01:19,670 ‫The First 5000 labels are stored in VAT labels and the rest of the labels are stored in part print labels 13 00:01:23,000 --> 00:01:25,050 ‫so we have created two parts. 14 00:01:25,080 --> 00:01:32,060 ‫Although these trading examples one is validation say and one is the partial dating reality partial 15 00:01:32,060 --> 00:01:33,340 ‫dating site to trendy model. 16 00:01:33,980 --> 00:01:39,920 ‫And the validation set will be used to do only hyper parameters and you will see the use of validation 17 00:01:39,930 --> 00:01:44,700 ‫set in becoming victims. 18 00:01:44,740 --> 00:01:50,210 ‫Now there are two ways in which we can define our model using get us. 19 00:01:50,560 --> 00:01:55,240 ‫One is using sequential EPA and the other is using functional EPA 20 00:01:57,890 --> 00:01:58,460 ‫sequencing. 21 00:01:58,460 --> 00:02:05,090 ‫EPA is used when we want to make a normal neural network with linear stack of layers. 22 00:02:05,090 --> 00:02:12,290 ‫The functional EPA is used for a complex network structure where you have multiple usages of several 23 00:02:12,290 --> 00:02:21,020 ‫small one it looks will see an example of functional EPA when we build a regression model for this classification 24 00:02:21,020 --> 00:02:25,980 ‫example we will use sequential EPA. 25 00:02:26,570 --> 00:02:30,750 ‫Now we'll be taking three steps in defining our model. 26 00:02:30,770 --> 00:02:34,250 ‫The first step is defining the network structure. 27 00:02:34,250 --> 00:02:42,710 ‫This includes setting up the number of layers number of neurons in each layer and the activation function 28 00:02:42,800 --> 00:02:45,150 ‫to be used in each layer. 29 00:02:45,170 --> 00:02:52,100 ‫This is captured and this part of decode. 30 00:02:52,300 --> 00:03:00,510 ‫The second part is configuring the learning process this includes selecting the lost function and optimizer 31 00:03:01,200 --> 00:03:05,050 ‫and some metrics to be monitored. 32 00:03:05,070 --> 00:03:12,050 ‫This is this part of the code. 33 00:03:12,130 --> 00:03:17,340 ‫The third is specifying the operations and blending the model. 34 00:03:17,350 --> 00:03:21,010 ‫This is done in this last part of this code. 35 00:03:23,710 --> 00:03:28,330 ‫So let's start discussing each line of code one by one. 36 00:03:28,670 --> 00:03:37,490 ‫Here we start by creating a new variable called model and this variable will contain the information 37 00:03:37,610 --> 00:03:41,020 ‫of the structure of our network. 38 00:03:41,030 --> 00:03:50,030 ‫This is the function we use for using sequence and EPA gave us model underscored sequences and then 39 00:03:50,030 --> 00:03:51,960 ‫to start defining the structure. 40 00:03:52,250 --> 00:03:55,330 ‫We use this type symbol. 41 00:03:55,730 --> 00:04:01,890 ‫This pipe operator comes with the magnetite package which is auto installed when we installed the cameras 42 00:04:01,940 --> 00:04:03,920 ‫package. 43 00:04:03,950 --> 00:04:08,790 ‫This is used for passing values as arguments to function. 44 00:04:09,050 --> 00:04:17,090 ‫We can do away with a decimal also but the use of this symbol makes the code more readable and compact. 45 00:04:17,090 --> 00:04:20,950 ‫So as a good practice we will use this operator too. 46 00:04:20,990 --> 00:04:23,490 ‫This is a pipe operator. 47 00:04:24,860 --> 00:04:34,340 ‫And if you remember even this operator that we used by assigning new plane images and green labels this 48 00:04:34,340 --> 00:04:40,370 ‫multiple assignment operator this has come into the alert package which was also part of the clearance 49 00:04:40,400 --> 00:04:42,540 ‫package. 50 00:04:42,740 --> 00:04:53,090 ‫We're using this also because it makes the code compact so first we flatten the input layer my flattening 51 00:04:53,450 --> 00:05:00,860 ‫I mean that we have 28 micro indeed image we can turn it into one dimension by putting all these pixels 52 00:05:01,070 --> 00:05:02,650 ‫in one line. 53 00:05:02,660 --> 00:05:12,800 ‫For example if you have this three by 3 two dimensional Eddy you can flatten it by putting these multiple 54 00:05:12,800 --> 00:05:14,440 ‫roles in front of each other. 55 00:05:14,780 --> 00:05:17,060 ‫So this will become a one dimensional edit 56 00:05:20,270 --> 00:05:26,630 ‫this step is important as we have to give a straight shape of input values in place offered to the input 57 00:05:28,460 --> 00:05:31,790 ‫you can converted in one dimensional using edit easier. 58 00:05:31,790 --> 00:05:36,830 ‫Function also when we have hit us why should we bother. 59 00:05:36,920 --> 00:05:47,190 ‫Just specify here layer flatten like this and specify the input shape. 60 00:05:47,750 --> 00:05:52,380 ‫That is what is the kind of input that this layer is having. 61 00:05:52,460 --> 00:06:00,380 ‫It will automatically convert this to endured by two indeed input in 2 7 a report pixel values for the 62 00:06:00,380 --> 00:06:03,660 ‫next layer. 63 00:06:04,050 --> 00:06:09,660 ‫Next we specify any details of our dense hidden layer. 64 00:06:09,660 --> 00:06:17,730 ‫That is we are telling that this layer is dense meaning each neuron is connected to all neurons of the 65 00:06:17,730 --> 00:06:18,290 ‫next layer. 66 00:06:19,290 --> 00:06:26,760 ‫And then this layer we want one twenty eight neurons and the activation function for all these neurons 67 00:06:27,090 --> 00:06:29,010 ‫will meet the loop. 68 00:06:29,010 --> 00:06:33,270 ‫That is rectified linear Unit 2 in this way. 69 00:06:33,330 --> 00:06:35,730 ‫We have defined one hidden layer. 70 00:06:35,850 --> 00:06:37,320 ‫It's a densely. 71 00:06:37,320 --> 00:06:43,020 ‫It has 120 neutrons and they look at and function. 72 00:06:43,190 --> 00:06:45,560 ‫Next we specify the output layer. 73 00:06:45,740 --> 00:06:55,790 ‫You can add more layers also but here I am using only one had a layer and one output in this last net. 74 00:06:55,890 --> 00:07:03,750 ‫We have 10 neutrons whiten because we have 10 classes to be predicted. 75 00:07:03,750 --> 00:07:09,840 ‫Each of these 10 neutrons will be predicting the probability of one class such as whether it is a shirt 76 00:07:10,020 --> 00:07:11,410 ‫or a boot. 77 00:07:11,700 --> 00:07:19,410 ‫And if you remember from your theory lecture this soft Max activation just make sure that the sum of 78 00:07:19,440 --> 00:07:22,770 ‫all the probabilities come out to 1. 79 00:07:22,860 --> 00:07:28,330 ‫So this last layer has 10 neurons with soft Max activation. 80 00:07:28,980 --> 00:07:31,210 ‫That's all for the structure. 81 00:07:31,230 --> 00:07:39,590 ‫In short it is a 780 for hyphen 128 different in neural network. 82 00:07:39,810 --> 00:07:46,020 ‫Once we have run the entire code I would suggest that you come back to this point and experiment a little 83 00:07:46,020 --> 00:07:52,560 ‫bit here try to see what is the effect of having more lives and what is the effect of increasing or 84 00:07:52,560 --> 00:07:56,810 ‫decreasing the number of neurons in different layers. 85 00:07:56,940 --> 00:07:59,650 ‫Next underscored. 86 00:08:01,350 --> 00:08:08,740 ‫You can see that a new variable called model is created and it has the structure stored in it. 87 00:08:08,760 --> 00:08:13,400 ‫Now let's look at the second step at this step. 88 00:08:13,460 --> 00:08:18,880 ‫We configured the learning process in this. 89 00:08:18,970 --> 00:08:23,650 ‫The first thing is specifying optimizer. 90 00:08:23,760 --> 00:08:28,080 ‫We have discussed the concept behind stochastic gradient descent. 91 00:08:28,140 --> 00:08:29,280 ‫This is deadly. 92 00:08:29,340 --> 00:08:38,850 ‫Is that only data other optimizer also with small differences other optimizes include Adam Artemus prof 93 00:08:39,510 --> 00:08:41,720 ‫and few others. 94 00:08:42,030 --> 00:08:45,710 ‫In fact in the coming team we may see few more added to this list. 95 00:08:46,980 --> 00:08:54,420 ‫But to answer the question which should be used then ideally the optimizer depends on the shape of the 96 00:08:54,420 --> 00:08:56,300 ‫edit function go. 97 00:08:56,880 --> 00:08:58,780 ‫But we do not know that shape. 98 00:08:59,100 --> 00:09:08,360 ‫So we do not know the ideal optimize it but practically in most of these scenarios all of these work 99 00:09:08,400 --> 00:09:09,040 ‫very well. 100 00:09:09,930 --> 00:09:17,160 ‫It's just that for some scenarios as did converges it and for some situations outermost prop converges 101 00:09:17,160 --> 00:09:19,160 ‫faster. 102 00:09:19,170 --> 00:09:24,540 ‫So my suggestion would be labeled readonly model once with duty. 103 00:09:24,840 --> 00:09:30,540 ‫If you think it is taking too long to convert and you are also not seeing much improvement in training 104 00:09:30,930 --> 00:09:36,720 ‫it is always worth a shot to Brio Artemus prop optimizer also. 105 00:09:37,400 --> 00:09:41,870 ‫Let's move on to this again parameter which is lost function. 106 00:09:42,090 --> 00:09:49,500 ‫We have this because this is the theory part for classification models views cross entropy and for regression 107 00:09:49,500 --> 00:09:57,990 ‫models we use the means squared added values but within cross entropy you will find three options. 108 00:09:58,230 --> 00:10:03,930 ‫Which of these three should you use depends on the type of problem you have. 109 00:10:03,930 --> 00:10:10,830 ‫These are the three names sparse categorical cross entropy binary cross entropy and categorical cross 110 00:10:10,880 --> 00:10:21,430 ‫entropy if your problem has two classes to be predicted like whether a is spam of North spam we use 111 00:10:21,430 --> 00:10:24,190 ‫the binary cross entropy. 112 00:10:25,400 --> 00:10:32,210 ‫If you have multiple classes such as this problem where we have ten fashion objects and each example 113 00:10:32,390 --> 00:10:42,260 ‫is exclusive meaning each image contains only one object to be predicted then we use sparse categorical 114 00:10:42,260 --> 00:10:43,630 ‫cross and Ruby. 115 00:10:44,700 --> 00:10:52,160 ‫So that is why I am right and logical loose pass categorical course and repeated if we have multiple 116 00:10:52,160 --> 00:11:00,380 ‫classes and one observation can belong to many classes for example if we are labeling whether an email 117 00:11:00,380 --> 00:11:06,410 ‫is from someone you know or not and we are also labeling whether the email is important or not. 118 00:11:07,700 --> 00:11:10,730 ‫Here one email can be bought. 119 00:11:10,730 --> 00:11:17,780 ‫It can be from someone you know and it can be important so it may belong to two classes at the same 120 00:11:17,780 --> 00:11:18,970 ‫time. 121 00:11:18,980 --> 00:11:23,670 ‫In this scenario we use categorical growth and Ruby. 122 00:11:24,230 --> 00:11:26,420 ‫I hope you understood this. 123 00:11:26,420 --> 00:11:29,210 ‫Here's the gist of what I just said. 124 00:11:29,600 --> 00:11:36,980 ‫You can look at this commented part here to understand the three cross and rupees. 125 00:11:37,400 --> 00:11:39,920 ‫The third parameter is metrics. 126 00:11:39,920 --> 00:11:45,650 ‫This is not mandatory but we specify this to monitor the performance of model on the training set. 127 00:11:48,350 --> 00:11:54,960 ‫Basically we would like to see the improvement in accuracy of a classification model or the mean squared. 128 00:11:55,000 --> 00:11:59,760 ‫Edit offer the regression model over each epoch. 129 00:12:00,440 --> 00:12:09,350 ‫As I told you we go over the entire training dataset several times each time we will calculate the accuracy 130 00:12:09,590 --> 00:12:10,700 ‫of our model. 131 00:12:10,790 --> 00:12:17,510 ‫At that instant and store it so that we can see if the learning process is having any improvement in 132 00:12:17,510 --> 00:12:18,490 ‫accuracy or not. 133 00:12:20,990 --> 00:12:27,980 ‫So with these three parameters say we can done this part of the code. 134 00:12:28,090 --> 00:12:30,350 ‫Now we have configured the learning process also. 135 00:12:30,700 --> 00:12:38,170 ‫This brings us to the third part where we actually train not more than training is done using different 136 00:12:38,290 --> 00:12:41,660 ‫function within function. 137 00:12:41,680 --> 00:12:45,460 ‫We have to specify the input variable first. 138 00:12:46,120 --> 00:12:54,980 ‫That is the posture training dataset that we will input then comes the actual output corresponding to 139 00:12:54,980 --> 00:12:59,980 ‫those inputs to the actual output is stored in partial train labeled. 140 00:13:00,500 --> 00:13:02,200 ‫So that is the second parameter. 141 00:13:04,690 --> 00:13:07,600 ‫Next we specify the epoch number. 142 00:13:08,140 --> 00:13:13,970 ‫This is the number of times our entire training data will be put into the model. 143 00:13:14,140 --> 00:13:21,030 ‫We set this to 30 for this example then we have that size. 144 00:13:21,040 --> 00:13:27,460 ‫This is the number of observations which will be used during each forward and backward propagation step 145 00:13:27,980 --> 00:13:30,380 ‫so we take a battery of hundred 146 00:13:33,730 --> 00:13:42,030 ‫lastly we tell that we have a separate validation data also which is a list of valid images and value 147 00:13:42,040 --> 00:13:48,140 ‫labels and we would like to see the accuracy score on this validation data. 148 00:13:48,200 --> 00:13:57,280 ‫As keep in mind that only this part brain images and bartering labels will be used to bring any model 149 00:13:58,460 --> 00:14:01,680 ‫this validation data is like this data. 150 00:14:01,700 --> 00:14:09,260 ‫In this scenario that is our model will not have seen the validation data when it calculates the accuracy 151 00:14:09,260 --> 00:14:11,420 ‫on this validation data. 152 00:14:11,680 --> 00:14:15,960 ‫Now let's run this line of code as well and this will be number one 153 00:14:24,960 --> 00:14:31,880 ‫well you can see that neural network model is getting green and the accuracy in lost value is being 154 00:14:32,180 --> 00:14:40,090 ‫recorded for each epoch in the next video we will see the performance of this train more model.