1
00:00:00,890 --> 00:00:08,520
But what the next step is to create architecture for our model, we will be using almost the same architecture.

2
00:00:09,120 --> 00:00:11,890
We will be using sequential as they appear.

3
00:00:12,360 --> 00:00:15,630
And we will be adding four on layer.

4
00:00:15,930 --> 00:00:17,980
We will start with 32 filters.

5
00:00:18,240 --> 00:00:19,740
Then 64 filters.

6
00:00:19,920 --> 00:00:21,480
Then 128 filters.

7
00:00:21,810 --> 00:00:23,880
Then again, 128 filters.

8
00:00:25,780 --> 00:00:28,660
And after each of this con led, we will.

9
00:00:29,080 --> 00:00:31,080
We are also playing with willingly.

10
00:00:33,370 --> 00:00:37,840
And as always, the activation function is the RELU for all these layers.

11
00:00:39,430 --> 00:00:43,630
And after that, this thing, we are also playing Grop Otilia.

12
00:00:46,080 --> 00:00:55,650
So what this lawyer will do is it will deactivate 50 percent off neurons during each epoch.

13
00:00:56,490 --> 00:01:01,170
It will randomly pick for people sent off our neurons and it pretty directly with them.

14
00:01:01,770 --> 00:01:08,190
And we will be training model with the remaining 50 percent off neurons during each people.

15
00:01:08,910 --> 00:01:15,860
So for each epoch, we are activating randomly 50 percent of our total neurons.

16
00:01:17,720 --> 00:01:24,780
We are using Grabbled here because Rubbled is a very effective leader to avoid overfitting in our more

17
00:01:24,780 --> 00:01:24,990
than.

18
00:01:28,290 --> 00:01:30,710
So this is of a modern architecture.

19
00:01:32,390 --> 00:01:34,670
Now, the next, the surplus to come by.

20
00:01:35,660 --> 00:01:38,650
For lost function, we will be using binary cross entropy.

21
00:01:39,050 --> 00:01:41,030
Since we have two different classes.

22
00:01:42,640 --> 00:01:50,860
If you remember earlier in Emmis, we had 10 different classes and there we were using spots categorical

23
00:01:51,250 --> 00:01:52,100
across entropy.

24
00:01:52,750 --> 00:01:59,040
But here, since we have one the two classes, we are using Leonetti, Cross and Groppi for optimize

25
00:01:59,050 --> 00:01:59,260
it.

26
00:01:59,410 --> 00:02:03,820
We are using Oremus prop with learning rate of zero point zero zero one.

27
00:02:05,120 --> 00:02:10,390
And since this is a classification problem, we are calculating accuracy metrics as well.

28
00:02:13,590 --> 00:02:16,140
The next step is to create our modern.

29
00:02:17,830 --> 00:02:26,390
Since we are taking our data from Katrina and Rita, we have to use for generator to fit our model.

30
00:02:26,530 --> 00:02:31,240
So we'll be using model dot for gender to Dan Crane, gender to.

31
00:02:32,270 --> 00:02:39,370
This dirty doesn't matter, which will continuously generate butat in the batches of 32 images.

32
00:02:41,120 --> 00:02:44,950
And then here we are using Insteps but Epoch as hundred.

33
00:02:46,970 --> 00:02:51,510
Earlier in our last model, we were using a bad size of 20 steps.

34
00:02:51,680 --> 00:02:57,120
But Epoch as hungry because we only had to pose an immediate.

35
00:02:57,140 --> 00:02:58,550
For training purposes.

36
00:02:59,420 --> 00:03:07,880
But this time, since we are randomly generating images from this transformation, we can use more than

37
00:03:08,160 --> 00:03:09,680
2000 images as well.

38
00:03:10,370 --> 00:03:16,060
This time we are using a bed size of potato and steps, but epoch as hundred.

39
00:03:16,770 --> 00:03:22,850
So overall, in each epoch we are feeding at all three thousand two hundred images.

40
00:03:25,270 --> 00:03:28,780
The number of epochs this time is hundred.

41
00:03:30,240 --> 00:03:36,450
And similarly, we will use validation generator to get the validation data.

42
00:03:39,090 --> 00:03:46,050
Now, since we are running this four hundred a box, if you are using this system with less than 16

43
00:03:46,050 --> 00:03:52,050
DGP of RAM and without any graphics card, it may take up to one and half to two or worse to create

44
00:03:52,050 --> 00:03:52,580
this model.

45
00:03:55,020 --> 00:04:01,980
That's what I have already train this model and I have the data off or one hundred epochs here.

46
00:04:04,230 --> 00:04:13,740
You can see that our validation accuracy is increasing with each book and at their own ninety two hundred

47
00:04:13,820 --> 00:04:14,110
book.

48
00:04:14,160 --> 00:04:20,340
We are getting the validation accuracy between 80 to 84 percent and.

49
00:04:21,440 --> 00:04:25,650
A training accuracy of it before a debate, but say.

50
00:04:26,940 --> 00:04:35,190
So if you compare in our last model, we were getting training accuracy of our own 95 to 98 percent

51
00:04:36,000 --> 00:04:41,140
and a significantly lower valuation accuracy of our own 79 percent.

52
00:04:42,690 --> 00:04:49,460
In this model, we are getting almost same regulation and lower training accuracy from it before person.

53
00:04:51,370 --> 00:05:00,070
So you can see that with what image processing and creating dummy images we have created or what fitting

54
00:05:00,160 --> 00:05:00,910
in our modern.

55
00:05:03,660 --> 00:05:11,220
After running this, you can save your model by model lot save method and lets us create this graph

56
00:05:12,150 --> 00:05:18,150
to see how our valuation, accuracy and training accuracy are changing with each epoch.

57
00:05:20,160 --> 00:05:23,280
So this orange and red lines are for accuracy.

58
00:05:24,240 --> 00:05:28,950
This orange line is for training accuracy and red line is for validation accuracy.

59
00:05:30,510 --> 00:05:36,600
You can see here that the validation accuracy is more than 80 percent as well as the training.

60
00:05:36,600 --> 00:05:38,760
Accuracy is also more than 80 percent.

61
00:05:40,320 --> 00:05:42,560
And both are moving together.

62
00:05:42,720 --> 00:05:45,420
So there are no evidence of overfitting.

63
00:05:45,660 --> 00:05:46,890
And then what model?

64
00:05:48,690 --> 00:05:50,670
And this is still increasing.

65
00:05:51,090 --> 00:06:00,930
So if you run it for, say, 40 or 50 more Reeboks, the validation accuracy may reach around 85, 86

66
00:06:00,930 --> 00:06:01,820
percent as well.

67
00:06:03,760 --> 00:06:05,230
So that's all for this review.

68
00:06:06,310 --> 00:06:14,620
v.C that by augmenting our initial dataset, by applying sheer rotation, will shift, hide, shift

69
00:06:15,120 --> 00:06:15,910
and flips.

70
00:06:16,480 --> 00:06:24,590
We can treat overfitting fitting in our data and we can get a higher valuation accuracy from our model.

71
00:06:25,180 --> 00:06:25,630
Thank you.
