1
00:00:00,740 --> 00:00:05,410
But what the next step is to create architecture for our model.

2
00:00:05,520 --> 00:00:08,970
We will be using almost the same architecture.

3
00:00:09,120 --> 00:00:12,070
We will be using sequential guitars US A.

4
00:00:12,360 --> 00:00:15,960
And we will be adding four con layer.

5
00:00:15,960 --> 00:00:18,050
Will Assad would touch it for years.

6
00:00:18,240 --> 00:00:20,070
Then 60 votes were lost.

7
00:00:20,070 --> 00:00:21,800
Then for filters.

8
00:00:21,820 --> 00:00:29,030
Then again 128 filters and after each of this on layer we will.

9
00:00:29,070 --> 00:00:37,330
We are also applying willing layer and as always the activation function is that a loop for all these

10
00:00:37,330 --> 00:00:38,780
layers.

11
00:00:39,430 --> 00:00:45,770
And after that this time we are also playing global player.

12
00:00:46,080 --> 00:00:56,330
So what this layer will do is it we deactivate 50 percent of neurons during each epoch.

13
00:00:56,520 --> 00:01:03,030
It will randomly pick for people sent off our neurons and it will deactivate them and we will be creating

14
00:01:03,030 --> 00:01:08,770
model with the remaining 50 percent off neuron during each poll.

15
00:01:08,940 --> 00:01:17,620
So for each book we are deactivating randomly 50 percent of our total neuron.

16
00:01:17,760 --> 00:01:24,800
We are using robot here because doubled is a very effective layer to avoid or operating in our more

17
00:01:24,800 --> 00:01:27,970
than.

18
00:01:28,300 --> 00:01:32,900
So this is of a model architecture now.

19
00:01:33,080 --> 00:01:39,800
The next the surplus to come by for lost function we will be using binary cross entropy since we have

20
00:01:39,800 --> 00:01:41,020
two different classes.

21
00:01:42,640 --> 00:01:50,890
If you remember earlier in Amis we had 10 different classes and there we were using is spot's categorical

22
00:01:51,250 --> 00:01:52,590
across entropy.

23
00:01:52,750 --> 00:01:59,350
But here since we have under two classes we are using binary cross and property for optimize it.

24
00:01:59,380 --> 00:02:06,880
We are using Artemus prop with learning rate of zero point zero zero one and since this is a classification

25
00:02:06,880 --> 00:02:10,390
problem we are calculating accuracy metrics as well.

26
00:02:13,590 --> 00:02:23,200
The next step is to create our model since we are taking our data from Green generator.

27
00:02:23,230 --> 00:02:26,420
We have to use for the generator to fit our model.

28
00:02:26,530 --> 00:02:29,460
So we'll be using model dot for generator.

29
00:02:30,040 --> 00:02:38,480
Then green gender to this doesn't matter which will continuously generate data in the batches of 32

30
00:02:38,780 --> 00:02:49,460
images and then here we are using these steps but epoch as hundred earlier in our last model we were

31
00:02:49,460 --> 00:02:51,670
using a bed size of 21 steps.

32
00:02:51,680 --> 00:03:01,160
But epoch as hundred because we only had 2000 images for training purposes but this time since we are

33
00:03:01,160 --> 00:03:10,370
randomly generating images from this transformation we can use more than 2000 images as well.

34
00:03:10,370 --> 00:03:19,190
This time we are using a bed size of potato and steps but epoch as hundred So overall in e.g. pork we

35
00:03:19,190 --> 00:03:31,630
are feeding around three thousand two hundred images the number of epochs this time is hundred and similarly

36
00:03:31,660 --> 00:03:38,470
we will use validation generator to get the validation data.

37
00:03:39,100 --> 00:03:42,430
Now since we are running this four hundred epochs.

38
00:03:42,670 --> 00:03:49,840
If you are using the system with less than 16 GDP of RAM and without any graphic card it may take up

39
00:03:49,840 --> 00:03:54,950
to one and half to two overs to create this model.

40
00:03:55,020 --> 00:04:02,070
That's what I have already green this model and I have the data of what a hundred epochs yet.

41
00:04:04,250 --> 00:04:13,780
You can see that our validation accuracy is increasing with each book and at our own ninety two hundred

42
00:04:13,830 --> 00:04:25,180
but we are getting the validation accuracy between 80 to 84 percent and a training accuracy of 84 85.

43
00:04:25,240 --> 00:04:35,190
But say so if you compare in our last model we were getting a training accuracy of around 95 to 98 percent

44
00:04:36,000 --> 00:04:42,040
and a significantly lower validation accuracy of our own 79 percent.

45
00:04:42,690 --> 00:04:49,020
In this model we are getting almost the same validation and lower training accuracy overall it before

46
00:04:49,020 --> 00:05:00,070
passing so you can see that with our image processing and creating dummy images we have created or fitting

47
00:05:00,160 --> 00:05:00,910
in our more than

48
00:05:03,660 --> 00:05:04,770
after running this.

49
00:05:04,800 --> 00:05:13,710
You can save your model by model lots save method and lets us create this graph to see how our valuation

50
00:05:13,710 --> 00:05:19,990
accuracy and training accuracy are changing with each epoch.

51
00:05:20,160 --> 00:05:24,240
So this orange and red lines are for accuracy.

52
00:05:24,250 --> 00:05:30,510
This orange line is for training accuracy and red line is for validation accuracy.

53
00:05:30,510 --> 00:05:37,110
You can see here that the validation accuracy is more than 80 percent as well as the training accuracy

54
00:05:37,110 --> 00:05:42,630
is also more than 80 percent and both are moving together.

55
00:05:42,720 --> 00:05:51,090
So there are no evidence of what fitting in our model and this is still increasing.

56
00:05:51,090 --> 00:06:01,350
So if you run it for say 40 or 50 more epochs the validation accuracy may reach around 85 86 percent

57
00:06:01,380 --> 00:06:01,830
as well.

58
00:06:03,760 --> 00:06:12,460
So that's all for this will do we see that by augmenting our initial dataset by applying shear rotation

59
00:06:13,060 --> 00:06:16,210
we shift height shift and flips.

60
00:06:16,480 --> 00:06:24,610
We can treat or what fitting in our data and we can get a higher validation accuracy from our model.

61
00:06:25,180 --> 00:06:25,630
Thank you.