1
00:00:03,460 --> 00:00:08,580
Now, the next step is to create the structure of a watch CNN model.

2
00:00:11,900 --> 00:00:14,870
For this, we will for this problem.

3
00:00:15,110 --> 00:00:19,940
We will be using four different conv layers with Max pulling.

4
00:00:20,690 --> 00:00:25,520
And after that we will apply our dense layer and then an output.

5
00:00:27,390 --> 00:00:28,920
So for our first Lere.

6
00:00:30,170 --> 00:00:34,940
We will have gone layer with 30 to 40 deaths and three by three with No.

7
00:00:36,600 --> 00:00:41,100
And our input sizes, one four feet by four feet by three.

8
00:00:42,650 --> 00:00:49,690
Remember, we mentioned our target size while data preprocessing as one four feet cross 150 feet.

9
00:00:50,480 --> 00:00:54,980
And since this images are Collodi images, we have RGV, we will lose.

10
00:00:55,020 --> 00:00:59,090
That's why we are providing a third dimension as three.

11
00:01:00,330 --> 00:01:02,250
After this, we want to max pulling.

12
00:01:02,970 --> 00:01:03,280
Here.

13
00:01:03,300 --> 00:01:05,040
We will use a window of two way to.

14
00:01:06,900 --> 00:01:13,740
After this, we'll use another layer which succeeded for four years and three by three window.

15
00:01:14,310 --> 00:01:15,980
Then another pulling layer.

16
00:01:17,810 --> 00:01:25,430
Then we will use another con there with 128 features and add another layer of way to window.

17
00:01:26,920 --> 00:01:34,780
Then we will have another layer of 128 filters and a pulling their weight two by two window.

18
00:01:38,430 --> 00:01:46,530
Then we will use flight then and then a single dense layer with 512 neuron and then a single output

19
00:01:46,530 --> 00:01:48,570
layer with one neuron.

20
00:01:48,720 --> 00:01:51,660
Since we only want to predict two classes.

21
00:01:55,090 --> 00:01:59,590
One important thing here is to note the number of tweeters that we are using.

22
00:02:00,550 --> 00:02:07,880
You can see as we are going on in our network, we are increasing the number of four years in what kindlier?

23
00:02:10,100 --> 00:02:14,820
So this is a general practice with each con layer.

24
00:02:15,320 --> 00:02:19,310
You have to double the number of filters you are going to use.

25
00:02:21,220 --> 00:02:28,600
And another important thing is that if you remember with Max pool of window to way to.

26
00:02:29,700 --> 00:02:33,060
You are reducing the size of your images.

27
00:02:33,930 --> 00:02:43,030
So if you ever image of 150 by 150 after max pulling, you will have a made up of 75 but 75.

28
00:02:44,830 --> 00:02:52,000
So after this, a step after this layer, we will have images in the dimension of so and if by 70 feet.

29
00:02:53,020 --> 00:02:57,400
Then after this, another MECs bullying, we will have images of.

30
00:02:58,710 --> 00:03:00,060
37 by 37.

31
00:03:01,080 --> 00:03:06,630
Then after this, we will have images of it by 18 and then nine by nine.

32
00:03:08,440 --> 00:03:18,340
So if you see the number of features are increasing as we go along on our network and the work and the

33
00:03:18,340 --> 00:03:21,490
lens of forward images are decreasing.

34
00:03:23,410 --> 00:03:25,090
This is a standard practice.

35
00:03:26,470 --> 00:03:35,050
You are a mere size should decrease with each con layer and the feature map height should increase as

36
00:03:35,050 --> 00:03:36,690
you go along the network.

37
00:03:41,310 --> 00:03:45,000
So this is the structure we are following for this problem.

38
00:03:47,450 --> 00:03:49,460
Let's just run this.

39
00:03:55,070 --> 00:03:59,070
Another thing is that earlier we were using Keita's dot Lear.

40
00:03:59,360 --> 00:04:05,030
So now we have imported Lears and we don't have to put guitars, dot layers and so on.

41
00:04:05,570 --> 00:04:07,320
So another little thing.

42
00:04:11,330 --> 00:04:12,620
Let's look at the somebody.

43
00:04:15,570 --> 00:04:22,110
You can see we have for corn layer with four different pulling layers and then.

44
00:04:23,060 --> 00:04:24,990
A dense layer and then output layer.

45
00:04:26,790 --> 00:04:32,220
The total number of parameters that we are screening in our model is at over three million.

46
00:04:35,850 --> 00:04:38,740
Now, the next step is to come by Lamartine.

47
00:04:39,480 --> 00:04:44,460
And this time, instead of using a ZULI, we are going to use them as prop.

48
00:04:47,850 --> 00:04:54,420
Oremus Propp have a little advantage or as Julie while performing image processing.

49
00:04:55,140 --> 00:04:57,400
That's why we are going to use Artemus Prop.

50
00:04:57,810 --> 00:05:02,820
And here we have also mentioned the learning rate of zero point zero zero one.

51
00:05:04,240 --> 00:05:06,450
By default, the learning rate is point zero.

52
00:05:07,150 --> 00:05:10,570
And we have mentioned learning data, pseudo point zero zero one.

53
00:05:12,050 --> 00:05:13,240
That's it on this.

54
00:05:17,750 --> 00:05:21,650
Now, the next step is to fit training data in our model.

55
00:05:23,750 --> 00:05:26,550
And earlier we were using model law fit.

56
00:05:28,580 --> 00:05:36,980
But now, since we have data in the form of green data, we are getting our data in the batches of 20

57
00:05:37,070 --> 00:05:38,720
directly from over directory.

58
00:05:39,760 --> 00:05:45,080
That's where instead of using dot fake, we are going to use dot fake gender.

59
00:05:47,250 --> 00:05:52,490
For using homemade detergent, narrator as our input we have to use for gender.

60
00:05:54,690 --> 00:05:58,110
The process is almost same instead of the Fed.

61
00:05:58,170 --> 00:06:00,060
You have to use DOT for it.

62
00:06:00,160 --> 00:06:01,130
Underscore the gender.

63
00:06:02,070 --> 00:06:02,400
Then.

64
00:06:03,600 --> 00:06:07,590
Then you have to provide the object that is generating the data.

65
00:06:07,770 --> 00:06:13,910
In our case, that is green generator that we have created earlier.

66
00:06:16,620 --> 00:06:17,070
Now.

67
00:06:18,960 --> 00:06:25,590
Now, another difference is that this train generator is generating data continuously.

68
00:06:26,700 --> 00:06:29,610
So we have to mention that stopping point.

69
00:06:30,330 --> 00:06:34,320
And we are mentioning the stopping point in the form of number of steps.

70
00:06:36,000 --> 00:06:42,540
As you remember here, while creating our train generator, we mentioned that the bed size should be

71
00:06:42,540 --> 00:06:43,290
of 20.

72
00:06:44,220 --> 00:06:50,310
So this train generator is generating data in the packets of 20 images.

73
00:06:51,240 --> 00:06:55,170
Now, our screen dataset is of 2000 images.

74
00:06:55,890 --> 00:07:00,270
So how many steps are required to cover all the two Holderness steps?

75
00:07:01,140 --> 00:07:06,830
That is two tosing divided by the number of images in each bet, which is 20.

76
00:07:07,560 --> 00:07:09,870
So that's where we want to step.

77
00:07:11,170 --> 00:07:18,250
In the hundreds of steps we are going to cover all the 2000 images that we have in our train directory.

78
00:07:19,870 --> 00:07:23,980
So just remember to use the steps, but epoch parameter.

79
00:07:25,470 --> 00:07:34,140
And this you should provide by dividing the total number of images that you have by the number of images

80
00:07:34,290 --> 00:07:35,320
you have in your head.

81
00:07:35,460 --> 00:07:36,390
Each batch.

82
00:07:37,440 --> 00:07:40,550
So 2000, they were led by 20, equal 200.

83
00:07:42,300 --> 00:07:44,910
We want to create this model for 20 bucks.

84
00:07:45,150 --> 00:07:50,240
That's why we are providing epochs equal to 20 and similar to fact.

85
00:07:50,310 --> 00:07:52,740
We can provide our validation data as well.

86
00:07:53,430 --> 00:07:57,420
Here, the validation data is in the form of validation generator.

87
00:08:00,200 --> 00:08:02,570
We have clear trade regulation, gender to clear.

88
00:08:03,380 --> 00:08:05,970
And here also we are using a bed size of 20.

89
00:08:07,100 --> 00:08:10,740
And in our ventilation formula, we have our on polls, on images.

90
00:08:11,980 --> 00:08:18,430
So then do either quite frankly, we need to run this for peace steps, so we have to provide whether

91
00:08:18,430 --> 00:08:20,260
additional steps equally 250.

92
00:08:24,170 --> 00:08:25,700
Let's run this.

93
00:08:26,300 --> 00:08:31,910
So if you see this is very similar to fact function and food function will directly provide the data

94
00:08:32,480 --> 00:08:35,130
here and sort of providing data directly.

95
00:08:35,270 --> 00:08:41,870
We are providing the generator that is generating the data and we are mentioning the steps or the limit

96
00:08:42,770 --> 00:08:46,540
on the number of times that generator is going to gender generator.

97
00:08:47,920 --> 00:08:50,600
So the training may take several minutes.

98
00:08:51,260 --> 00:08:54,440
So I am skipping the training part.