1
00:00:05,970 --> 00:00:07,500
Now are is trained.

2
00:00:08,880 --> 00:00:11,730
You can see that it has done for all the 30 epochs.

3
00:00:13,680 --> 00:00:19,180
The loss value we are getting is point to body training data.

4
00:00:20,910 --> 00:00:24,300
Accuracy is nearly 92 percent, which is very impressive.

5
00:00:25,950 --> 00:00:31,090
This is what the training data, the validation set loss is point to fight.

6
00:00:31,860 --> 00:00:37,650
And the accuracy is nearly 90 percent using the callbacks.

7
00:00:38,280 --> 00:00:42,870
We would have stored the best model out of all the 30 epochs.

8
00:00:44,370 --> 00:00:48,780
But you can see that this last epoch is sufficiently good.

9
00:00:49,020 --> 00:00:51,630
It is giving us a ninety point nine percent accuracy.

10
00:00:52,200 --> 00:00:56,820
So we are going to check the best performance of this particular epoch.

11
00:00:59,760 --> 00:01:06,270
Another thing to notice is the validation, loss and validation error.

12
00:01:06,930 --> 00:01:08,340
Have not stopped improving.

13
00:01:09,360 --> 00:01:11,290
Although there are some dips here.

14
00:01:11,790 --> 00:01:18,660
But you can see that overall the pattern is going up for the validation set also for training.

15
00:01:18,690 --> 00:01:19,890
It will definitely go up.

16
00:01:20,220 --> 00:01:23,830
But even for the validation set, it is still going a little bit down.

17
00:01:25,440 --> 00:01:33,170
So it will make more sense to run this for more number of epochs and simultaneously UTI at least topping

18
00:01:33,180 --> 00:01:39,150
criteria so that we can go up to that point where we get the best accuracy.

19
00:01:42,210 --> 00:01:49,050
Now, let us take the performance of this last train model, the final value at 30 epochs.

20
00:01:50,760 --> 00:01:53,120
How does it perform on our test it?

21
00:01:54,840 --> 00:02:00,880
So we will use the evaluate function to store the value in CNN school.

22
00:02:01,830 --> 00:02:05,430
So let us evaluate on the test images and test labels.

23
00:02:11,230 --> 00:02:18,020
And now we can print out the lost value and the accuracy value on our deficit.

24
00:02:18,340 --> 00:02:21,130
And it is coming out to be 90 percent.

25
00:02:22,270 --> 00:02:30,460
So you can see that there is a small improvement just by adding one convolutional and one pulling layer

26
00:02:30,880 --> 00:02:33,920
in our network earlier without these two layers.

27
00:02:35,220 --> 00:02:43,510
When we ran a normal artificial neural network, we had an accuracy ranging from 85 percent to 87 percent

28
00:02:43,510 --> 00:02:44,350
for the desert.

29
00:02:46,150 --> 00:02:50,740
Now we are getting four percent increase in test accuracy.

30
00:02:51,550 --> 00:02:59,680
And now it is at 90 percent, probably adding more convolutional layers, extracting more features,

31
00:02:59,770 --> 00:03:01,920
applying more craters, probably.

32
00:03:02,320 --> 00:03:08,140
There is a chance that we can further increase this accuracy to maybe even 95 percent.

33
00:03:14,070 --> 00:03:19,740
You can also use this model, this train model, to predict on new data.

34
00:03:20,610 --> 00:03:23,790
So here I'm going to predict on the test images only.

35
00:03:26,120 --> 00:03:29,630
So the class of prediction will be stored in class bread.

36
00:03:30,390 --> 00:03:31,980
And we use predict classes.

37
00:03:31,980 --> 00:03:32,820
Functions what that.

38
00:03:37,160 --> 00:03:40,760
I'm printing out the first 20 values in this class print variable.

39
00:03:43,010 --> 00:03:54,860
So the first image is classified as ninet object by our model to give name to these predicted values.

40
00:03:55,790 --> 00:04:00,110
If you remember, we created an array of class names.

41
00:04:00,650 --> 00:04:02,570
This is where we are going to use it.

42
00:04:03,470 --> 00:04:07,520
We are going to use class names to give names to these predicted values.

43
00:04:11,280 --> 00:04:15,540
So these are the first 20 predictions by our system.

44
00:04:16,560 --> 00:04:19,920
What are the actual values for these first 20 images?

45
00:04:22,290 --> 00:04:23,880
These are the actual values.

46
00:04:24,720 --> 00:04:29,550
You can see that most of the observations are exactly matching.

47
00:04:30,000 --> 00:04:31,620
These are predicted values.

48
00:04:31,890 --> 00:04:34,900
These are actual values and most of them are same.

49
00:04:36,600 --> 00:04:38,520
I can identify this one.

50
00:04:38,610 --> 00:04:40,060
This is predicted that sandal.

51
00:04:40,160 --> 00:04:41,580
But it is actually sneaker.

52
00:04:42,600 --> 00:04:45,570
And there is no other observation which just predicted wrongly.

53
00:04:47,010 --> 00:04:51,150
If you want to plot this particular image, we're just getting predicted wrongly.

54
00:04:52,710 --> 00:04:56,080
We can use this blog function here.

55
00:04:56,250 --> 00:04:59,040
I'll have to mention which image I want to plot.

56
00:05:00,170 --> 00:05:02,720
So this is 13 image.

57
00:05:04,350 --> 00:05:07,500
So I put 13 here and run this.

58
00:05:10,090 --> 00:05:11,940
And this is the image.

59
00:05:12,660 --> 00:05:14,270
Our model is predicting sandal.

60
00:05:14,760 --> 00:05:16,530
And it is actually sneaker.

61
00:05:17,460 --> 00:05:25,320
But I think because of the similar characteristics of a standalone sneaker, this could happen.

62
00:05:25,740 --> 00:05:29,850
So for most of the predictions, our model is predicting very, very.

63
00:05:33,160 --> 00:05:40,810
This is how we add convolutional layers and pulling layers to make a convolutional neural network,

64
00:05:41,920 --> 00:05:48,370
convolutional neural network also includes part of multi-level Perceptron that we built earlier.

65
00:05:49,060 --> 00:05:53,200
So initial few layers are convolutional years later on.

66
00:05:53,530 --> 00:05:57,130
We have a normal neural network which gives us the final predictions.

67
00:06:00,470 --> 00:06:03,380
This was a simple example, fashion.

68
00:06:03,440 --> 00:06:11,030
Amnesty is a standard data set, which everyone works on when they start learning, deepening in the

69
00:06:11,030 --> 00:06:19,190
coming lectures will take up one project and we're trying to build a convolutional network on it and

70
00:06:19,310 --> 00:06:23,450
improve its accuracy, as we have to do in real life problems.
