1 00:00:03,460 --> 00:00:08,560 Now the next step is to cleared the structure of what CNN more 2 00:00:11,910 --> 00:00:15,070 for this we will for this problem. 3 00:00:15,090 --> 00:00:19,940 We will be using four different kinds of layers with Max pooling. 4 00:00:20,670 --> 00:00:30,530 And after that we will apply our dense layer and then an output Nero so for our first layer we will 5 00:00:30,530 --> 00:00:35,180 have one layer with 32 feeders and three by three window. 6 00:00:36,590 --> 00:00:42,580 And our input sizes 150 by 150 by 3. 7 00:00:42,620 --> 00:00:45,750 Remember we mentioned our target size. 8 00:00:45,750 --> 00:00:49,740 While data processing as one for pick was one of these. 9 00:00:50,510 --> 00:00:55,010 And since this images are colored images we have added we will lose. 10 00:00:55,010 --> 00:01:00,990 That's why we are providing a third dimension as three after this. 11 00:01:00,990 --> 00:01:03,270 We want the max pooling layer here. 12 00:01:03,300 --> 00:01:06,600 We will use a window of two way too. 13 00:01:06,900 --> 00:01:14,150 After this we will use another layer with six to 44 years and three by three window. 14 00:01:14,340 --> 00:01:23,600 Then another willingly then we will use another con layer with 128 features and add another pooling 15 00:01:23,600 --> 00:01:32,980 layer of the White tool window then we will have another layer of 128 filters and a pooling there with 16 00:01:33,100 --> 00:01:34,750 two by two window 17 00:01:38,410 --> 00:01:46,090 then we will use like ten and then a single dense layer with five and value at all and then a single 18 00:01:46,090 --> 00:01:48,710 output layer with one neuron. 19 00:01:48,730 --> 00:01:51,640 Since we would only want to predict two glasses 20 00:01:55,090 --> 00:02:00,420 one important thing here is to know the number of features that we are using. 21 00:02:00,550 --> 00:02:07,240 You can see as we are going on in our network we are increasing the number of four years and now what 22 00:02:07,240 --> 00:02:09,930 can layer. 23 00:02:10,100 --> 00:02:15,290 So this is a general practice with each cone layer. 24 00:02:15,320 --> 00:02:18,070 You have to double the number of filters. 25 00:02:18,080 --> 00:02:28,270 You are going to use and another important thing is that if you remember the max pool of window 2 way 26 00:02:28,270 --> 00:02:33,860 to you at reducing the size of fewer images. 27 00:02:33,940 --> 00:02:46,310 So if you have a image of 150 by 150 after Max bullying you will have a major of 75 but 75 so after 28 00:02:46,380 --> 00:02:52,970 this is step off that this layer we will have images in the dimension of zone divide by 70 feet. 29 00:02:53,010 --> 00:03:00,910 Then after this another Max pulling we will have images of 37 by 37. 30 00:03:01,080 --> 00:03:11,770 Then after this we will have images of 18 by 18 and then 9 by 9 so if you see the number of features 31 00:03:11,890 --> 00:03:23,320 are increasing as we go along on our network and the work and the lens of forward images are decreasing. 32 00:03:23,390 --> 00:03:26,450 This is a standard practice. 33 00:03:26,450 --> 00:03:35,060 You are image size should decrease with each one layer and the feature map height should increase as 34 00:03:35,060 --> 00:03:40,980 you go along the network. 35 00:03:41,310 --> 00:03:49,430 So this is the structure we are following for this problem let's just run this 36 00:03:55,060 --> 00:03:59,100 another thing is that earlier we were using guitars got laid. 37 00:03:59,350 --> 00:04:05,520 So now we have important layers and we don't have to put guitars or layers and so on. 38 00:04:05,570 --> 00:04:07,330 So another little thing 39 00:04:11,330 --> 00:04:15,100 let's look at somebody. 40 00:04:15,560 --> 00:04:24,280 You can see we have four can layer with four different pooling layers and then a dense layer and then 41 00:04:24,290 --> 00:04:32,220 output layer the total number of parameters that we are creating in our model is that on three million 42 00:04:35,850 --> 00:04:36,530 now. 43 00:04:36,660 --> 00:04:39,070 The next step is to compile the model. 44 00:04:39,480 --> 00:04:44,430 And this time instead of using as Zilly we are going to use Adam's prop 45 00:04:47,850 --> 00:04:49,470 Adam's prop. 46 00:04:49,470 --> 00:04:55,140 Have a little advantage or as Julie while performing image processing. 47 00:04:55,140 --> 00:04:57,810 That's why we are going to use Adam's prop. 48 00:04:57,810 --> 00:05:04,880 And here we have also mentioned the learning rate of zero point zero 0 1 by default. 49 00:05:04,880 --> 00:05:13,240 The learning rate is point 0 1 and we have mentioned the learning date of zero point 0 0 1 7 this 50 00:05:17,500 --> 00:05:18,000 now. 51 00:05:18,050 --> 00:05:29,220 The next step is to fit training data in our model and earlier we were using model load fit but now 52 00:05:29,310 --> 00:05:37,560 since we have data in the form of green generator we are getting our data in the batches of 20 directly 53 00:05:37,560 --> 00:05:39,660 from our directory. 54 00:05:39,740 --> 00:05:49,890 That's why instead of using dot fake we are going to use dot generated for using image data generated 55 00:05:49,980 --> 00:05:51,040 as our input. 56 00:05:51,060 --> 00:05:53,360 We have to use for dinner. 57 00:05:54,690 --> 00:05:58,170 The process is almost same instead of DOT fed. 58 00:05:58,170 --> 00:06:07,080 You have to use dot underscore the data then then you have to provide the object that is generating 59 00:06:07,080 --> 00:06:07,780 the data. 60 00:06:07,800 --> 00:06:16,010 In our case that is green generator that we have created earlier. 61 00:06:16,640 --> 00:06:26,110 Now now another difference is that this train generator is generating data continuously. 62 00:06:26,730 --> 00:06:33,450 So we have to mention this stopping point and we are mentioning the stopping point in the form of number 63 00:06:33,450 --> 00:06:35,270 of steps. 64 00:06:36,000 --> 00:06:42,720 As you remember here while creating our train generator we mentioned that the bed size should be of 65 00:06:42,720 --> 00:06:44,220 20. 66 00:06:44,220 --> 00:06:51,140 So this train generator is generating data in the packets of 20 images. 67 00:06:51,210 --> 00:06:59,660 Now our crane dataset is of 2000 images so how many steps are required to cover all that to hold on 68 00:06:59,660 --> 00:07:01,050 the steps. 69 00:07:01,140 --> 00:07:07,490 That is 2000 divided by the number of images in each bet which is 20. 70 00:07:07,590 --> 00:07:15,610 So that's where we want under step in the hundreds of steps we are going to cover all the 2000 images 71 00:07:15,640 --> 00:07:18,450 that we have in our train directory. 72 00:07:19,870 --> 00:07:24,050 So just remember to use the steps but epoch barometer. 73 00:07:25,470 --> 00:07:34,140 And this you should go away by dividing the total number of images that you have by the number of images 74 00:07:34,260 --> 00:07:37,170 you have in your each batch. 75 00:07:37,440 --> 00:07:40,590 So 2000 divided by 20 equally 200. 76 00:07:42,300 --> 00:07:45,150 We want to create this model for 20 bucks. 77 00:07:45,150 --> 00:07:50,290 That's why we are providing box equal to 20 and similar to fate. 78 00:07:50,310 --> 00:07:56,910 We can provide our validation data as well here the validation data is in the form of validation and 79 00:07:57,000 --> 00:07:59,940 data. 80 00:08:00,180 --> 00:08:08,550 We have created validation generated Aliyah and here also we are using a bed size of 20 and in our validation 81 00:08:08,550 --> 00:08:14,120 formula we have it on poles and may images so tall then divided by 20. 82 00:08:14,320 --> 00:08:17,260 We need to run this for steps. 83 00:08:17,260 --> 00:08:20,260 So we have to provide validation additional steps equally 250 84 00:08:24,170 --> 00:08:26,280 let's run this. 85 00:08:26,300 --> 00:08:31,910 So if you see this is very similar to fake function and fake function will directly provide the data 86 00:08:32,540 --> 00:08:39,200 here and set off providing data directly we are providing a generator that is generating the data and 87 00:08:39,200 --> 00:08:46,310 we are mentioning the steps or the limit on the number of times that generator is going to generate 88 00:08:46,310 --> 00:08:46,760 the data. 89 00:08:47,930 --> 00:08:50,600 So the training may take several minutes. 90 00:08:51,290 --> 00:08:54,440 So I am skipping the training part.