1 00:00:01,860 --> 00:00:05,340 As you might remember, model has three parts. 2 00:00:05,830 --> 00:00:16,620 The model architecture, the model configuration and fitting the model first will define the model architecture. 3 00:00:17,460 --> 00:00:21,210 I've shown you the architecture that we are creating for this problem. 4 00:00:22,740 --> 00:00:28,500 After the input layer, the first layer that we will have is a convolutional to deliver 5 00:00:31,280 --> 00:00:32,310 in this layer. 6 00:00:32,490 --> 00:00:34,920 We are going to have 32 feature maps. 7 00:00:35,250 --> 00:00:45,450 So we are going to use 32 filters the city first parameter each week that has a window size or a receptive 8 00:00:45,570 --> 00:00:48,420 field of three by three pixels. 9 00:00:50,070 --> 00:00:57,300 So the first neuron in this convolutional layer will look at only three and two three. 10 00:00:57,330 --> 00:01:01,230 That is a square of nine pixels in the input image. 11 00:01:04,130 --> 00:01:12,390 So that is the kernel size or the window dimensions of our feet that the activation function for our 12 00:01:12,390 --> 00:01:13,370 convolutional layer. 13 00:01:13,920 --> 00:01:15,720 We are using here is Relu 14 00:01:18,570 --> 00:01:24,900 and the input shape of the images is 28 by 28 and one. 15 00:01:25,230 --> 00:01:29,250 So these two are the height and weight of images. 16 00:01:29,730 --> 00:01:31,500 And this is the channel. 17 00:01:35,190 --> 00:01:37,290 Next it is a max pooling layer. 18 00:01:38,400 --> 00:01:40,920 Max pulling leered has a pool size of two by two. 19 00:01:41,310 --> 00:01:49,050 That is the window of neurons in the previous layer will be of two and two two pixels out of these four 20 00:01:49,050 --> 00:01:55,970 pixels, whichever values maximum will be taken as the output value of the cell in max willingly. 21 00:01:58,380 --> 00:02:02,700 By default, this layer will have a straight off to this convolutional layer. 22 00:02:02,790 --> 00:02:07,160 We'll have a by default straight value one if you want to change these trade. 23 00:02:07,170 --> 00:02:13,530 Also, you can add additional barometer of straight is equal to two or three, whatever you want to 24 00:02:13,560 --> 00:02:20,200 set it to in the pulling layer, the default value status two. 25 00:02:21,150 --> 00:02:29,100 So this pulling layer will reduce the dimensions by two in both directions horizontally and vertically, 26 00:02:30,210 --> 00:02:37,050 which means that the output layer will have one fourth of the neurons as compared to the convolutional. 27 00:02:41,200 --> 00:02:43,150 So I separated this section. 28 00:02:43,390 --> 00:02:47,710 This is the convolutional section from our normal neural network. 29 00:02:48,670 --> 00:02:54,160 The output of this convolutional section is going into our normal neural network. 30 00:02:55,170 --> 00:02:56,460 So these are the next layers. 31 00:03:00,160 --> 00:03:06,430 You could have listed all the layers in one go only, but just to make it clear that this part is the 32 00:03:06,490 --> 00:03:07,510 convolutional part. 33 00:03:07,930 --> 00:03:11,920 And this is where we do normal neural network processing. 34 00:03:12,310 --> 00:03:13,900 I have separated these two. 35 00:03:15,790 --> 00:03:19,030 So the output of this layer will be three dimensional. 36 00:03:19,480 --> 00:03:22,450 We need to flatten it as a layer. 37 00:03:22,930 --> 00:03:24,670 So we use a flatten layer. 38 00:03:25,960 --> 00:03:33,700 The output of this flattened layer goes into a dense layer of 300 neurons with activation function relu. 39 00:03:35,350 --> 00:03:40,510 The output of these three under neutrons goes in to next to the layer of under neurons. 40 00:03:41,590 --> 00:03:43,510 These also have activation function below. 41 00:03:45,420 --> 00:03:53,650 Then the last layer, that is the output neuron layer has only 10 output neurons with activation function 42 00:03:53,680 --> 00:03:54,420 softmax. 43 00:03:57,340 --> 00:03:59,530 This part is exactly same. 44 00:03:59,980 --> 00:04:04,000 So this structure of the model is exactly same as we discussed last time. 45 00:04:07,690 --> 00:04:08,890 So let's run these two. 46 00:04:13,240 --> 00:04:14,890 Now, if you look at your model, 47 00:04:20,830 --> 00:04:22,900 we have created a sequential model. 48 00:04:23,950 --> 00:04:27,040 First layer is a convolutional to deliver. 49 00:04:31,530 --> 00:04:38,110 Its output ship is 26 by 26 input, the image was 28 by 28. 50 00:04:38,640 --> 00:04:42,290 But because by default, padding is set to valid. 51 00:04:42,720 --> 00:04:45,900 That is no additional padding is done. 52 00:04:46,500 --> 00:04:50,650 The additional boundry pixels were removed in the convolutional. 53 00:04:51,840 --> 00:04:56,160 So Vendig receptive field of convolutional layer was moving to the input layer. 54 00:04:57,360 --> 00:05:00,060 There were some pixels which could not be covered. 55 00:05:00,570 --> 00:05:02,010 So those were ignored. 56 00:05:02,250 --> 00:05:05,730 So one pixel from all of the boundaries were removed. 57 00:05:06,000 --> 00:05:11,160 That is why this ship is now 26 by 26 instead of 28 by 28. 58 00:05:13,380 --> 00:05:17,010 These 32 are defeater maps because we had 32 printers. 59 00:05:19,590 --> 00:05:26,370 Max, pulling Pullinger reduces these dimensions by two and we still have 32 feature maps. 60 00:05:27,660 --> 00:05:34,650 Then we have a flattened layer, then dense layer with 300 neurons, then another dense layer with hundred 61 00:05:34,650 --> 00:05:35,190 neurons. 62 00:05:35,700 --> 00:05:37,700 And lastly, the output layer retender. 63 00:05:37,740 --> 00:05:46,140 Don't the total number of parameters to be predicted is one point six million in the next lecture. 64 00:05:46,320 --> 00:05:54,660 When we remove the pooling layer or we try to add additional convolutional layer, we will see how this 65 00:05:54,660 --> 00:06:03,090 number of parameters changes and it will give us a hint of how pulling layer reduces the computational 66 00:06:03,210 --> 00:06:04,470 load on our system.