1
00:00:00,740 --> 00:00:03,270
In the last we knew, we trained our model.

2
00:00:04,920 --> 00:00:08,670
This is what happened at each epoch.

3
00:00:09,650 --> 00:00:11,880
We got accuracy and loss.

4
00:00:12,960 --> 00:00:15,740
And this was for book training.

5
00:00:16,190 --> 00:00:21,630
And the validation to this was also plotted on the right hand side.

6
00:00:22,410 --> 00:00:26,180
You can see in this graph, this top graph is not lost.

7
00:00:26,620 --> 00:00:30,660
It has blue line for lost on the training set.

8
00:00:31,350 --> 00:00:33,710
And this green line for validation loss.

9
00:00:35,130 --> 00:00:37,710
The graph on the bottom is for accuracy.

10
00:00:38,790 --> 00:00:42,870
This blue line is for training accuracy and the green line is for validation accuracy.

11
00:00:45,800 --> 00:00:49,590
You can see that after you've been to people.

12
00:00:50,180 --> 00:00:53,180
There is not much improvement in the accuracy.

13
00:00:55,040 --> 00:00:58,550
However, we have still run this program for 30 epochs.

14
00:01:01,000 --> 00:01:08,290
Now, as I said in the beginning, the real performance of a model is gauged on previously unseen data.

15
00:01:09,430 --> 00:01:14,140
This is the reason we kept a set of 10000 observations separately.

16
00:01:15,760 --> 00:01:23,200
So now we can predict the class for these test images and compare the accuracy against the actual class

17
00:01:23,500 --> 00:01:24,580
of those images.

18
00:01:27,240 --> 00:01:35,070
Just knowing that the training accuracy that we have achieved is nearly 87 percent.

19
00:01:36,000 --> 00:01:41,940
Now we are going to see if there is different accuracy achieved on the test set.

20
00:01:44,570 --> 00:01:51,930
Now, when we are looking at this performance, instead of just giving these test images and getting

21
00:01:51,930 --> 00:02:02,970
the predicted values, we can give boot images and the actual labels this way using this evaluate function.

22
00:02:03,640 --> 00:02:10,830
Our model can straightaway evaluate and show us the error and accuracy of our model on the test.

23
00:02:13,380 --> 00:02:19,140
So I stole the design of this evaluate function in the score variable.

24
00:02:19,940 --> 00:02:24,720
This evaluate takes two parameters test images and labels.

25
00:02:26,910 --> 00:02:34,110
It uses the model to predict on test images and compares the predicted level against these test labels

26
00:02:34,620 --> 00:02:35,700
to Lexus underscored.

27
00:02:42,830 --> 00:02:49,920
Now, if you want, go see the best loss and test air quality, you can run these two lines of code.

28
00:02:54,300 --> 00:03:04,140
You can see that the loss on test set is due to point forward, one that is on banning say it was zero

29
00:03:04,140 --> 00:03:04,950
point three seven.

30
00:03:06,480 --> 00:03:12,730
And the accuracy on this, it is nearly 85 percent, whereas it was 87 percent.

31
00:03:13,170 --> 00:03:21,800
On the training say this means that there is a little bit of overfitting but will not talk about that

32
00:03:22,400 --> 00:03:23,220
in detail.

33
00:03:23,230 --> 00:03:23,610
It hit.

34
00:03:26,330 --> 00:03:34,100
Now, if you are interested in actual predictions of the classes on the test that we can use to predict

35
00:03:34,100 --> 00:03:36,860
function and just in predict test images.

36
00:03:38,940 --> 00:03:40,580
So this is the predict function.

37
00:03:40,910 --> 00:03:43,430
And it just takes the input of test images.

38
00:03:43,880 --> 00:03:48,740
We do not need to give the test labels it when we get past images.

39
00:03:49,550 --> 00:03:53,750
It applies the model on it and gives us deep predictions.

40
00:03:54,320 --> 00:04:01,910
Next to underscore, you can see that we have predictions on our ten thousand best images.

41
00:04:02,900 --> 00:04:06,770
And it has ten values corresponding to each prediction.

42
00:04:08,120 --> 00:04:09,480
Let's look at those 10 values.

43
00:04:09,530 --> 00:04:19,220
Forty first image in the now since we used an outwardly was the dust softmax function.

44
00:04:20,630 --> 00:04:22,640
If you look at the first set of predictions.

45
00:04:22,910 --> 00:04:25,160
This is a set of ten probabilities.

46
00:04:27,140 --> 00:04:35,060
This is the probability of the first test image belonging to the first class.

47
00:04:36,770 --> 00:04:44,990
This is the probability for the second class and so on to find out which of these probabilities is the

48
00:04:44,990 --> 00:04:52,640
largest so that we can assign that class to that image we used, which taught Max function.

49
00:04:55,010 --> 00:04:57,100
You can see here which dark Max.

50
00:04:58,280 --> 00:05:02,040
And we input these predictions variable.

51
00:05:06,300 --> 00:05:14,910
When I've done this, it tells me that the end value, which is this seven point six six in compendious

52
00:05:14,990 --> 00:05:15,870
to one minus one.

53
00:05:17,820 --> 00:05:20,530
This value is the largest probability.

54
00:05:21,420 --> 00:05:33,570
So we want to assign the class to this image to find out what is the class associated with this tent

55
00:05:33,600 --> 00:05:34,200
position.

56
00:05:34,920 --> 00:05:43,560
If you remember, we created a class name I.D. which stored dawdy names of the classes here, is that

57
00:05:43,620 --> 00:05:43,790
right?

58
00:05:45,660 --> 00:05:50,100
So we want to find out the end element of the city.

59
00:05:50,940 --> 00:05:55,440
We just put that entire which match function to this class, Demitry.

60
00:05:57,030 --> 00:06:02,640
So basically with this which Max will return the position and corresponding to that position, we find

61
00:06:02,640 --> 00:06:04,620
out the last name from this day.

62
00:06:05,480 --> 00:06:11,220
So when we done this, it tells us that the class is ankle boot.

63
00:06:12,630 --> 00:06:19,860
So basically our prediction is that the first image in the test set is of ankle boot.

64
00:06:21,930 --> 00:06:29,130
If you want to check takes look at that image will again use the block function and we block this first

65
00:06:29,190 --> 00:06:30,450
image of this thing.

66
00:06:32,670 --> 00:06:36,570
So when we block this image, you can have a look at this image.

67
00:06:36,960 --> 00:06:42,840
This looks like an ankle boot and which is also our prediction that it is an ankle boot.

68
00:06:43,500 --> 00:06:45,190
So this looks correct.

69
00:06:46,770 --> 00:06:53,190
Lastly, instead of predicting probabilities, you can straightaway predict the flosses also using the

70
00:06:53,190 --> 00:06:54,780
predict classes function.

71
00:06:57,680 --> 00:07:05,720
Here's how we do that, we create a new video we'll call glass spread, and this gets the information

72
00:07:05,720 --> 00:07:08,840
off all the predicted glasses on the best images.

73
00:07:09,390 --> 00:07:11,610
This function also gets only one parameter.

74
00:07:11,660 --> 00:07:12,920
That is the input.

75
00:07:13,140 --> 00:07:14,000
Best images.

76
00:07:16,020 --> 00:07:20,770
And we've done this and we look at the first grindy predictions.

77
00:07:22,000 --> 00:07:25,860
You can see that for the first image we have predicting nine plus.

78
00:07:27,010 --> 00:07:29,530
So you may think that a lead we predicted then.

79
00:07:29,880 --> 00:07:31,690
Why is it number looking nine?

80
00:07:32,200 --> 00:07:42,880
So this tent is the position of the glass label and glass is actually labeled nine because glass labels

81
00:07:42,880 --> 00:07:43,900
are starting from zero.

82
00:07:45,250 --> 00:07:48,010
So both of these represent the same thing.

83
00:07:48,670 --> 00:07:52,850
Ninet label label glass is ankle boot only.

84
00:07:54,940 --> 00:07:56,770
So these are all declasse predictions.

85
00:07:58,210 --> 00:07:58,890
So we have both.

86
00:07:58,930 --> 00:08:02,050
The method one is to predict the probabilities.

87
00:08:03,190 --> 00:08:04,990
Use those predicted properties.

88
00:08:05,290 --> 00:08:07,080
Find out the maximum out of those.

89
00:08:07,620 --> 00:08:12,690
And that glass having the maximum probability will be the predicted glass.

90
00:08:13,180 --> 00:08:19,500
Or you can straightaway you do predict losses function to get the predicted glass on the new best set.

91
00:08:21,790 --> 00:08:22,370
That's all in.

92
00:08:22,490 --> 00:08:22,840
We do.

93
00:08:23,560 --> 00:08:28,120
So we have created a complete neural network classification model.

94
00:08:29,890 --> 00:08:34,030
So here's a summary of whatever we did to build this classification model.

95
00:08:35,110 --> 00:08:42,910
We installed a cross package and we activated it using install, underscore, get.

96
00:08:43,210 --> 00:08:44,080
We installed it.

97
00:08:44,380 --> 00:08:46,970
US go get us Labidi and Antiflu.

98
00:08:49,630 --> 00:08:58,810
We then imported the data we pre-process did and normalized it using that data.

99
00:08:59,410 --> 00:09:00,440
It was in two parts.

100
00:09:00,880 --> 00:09:03,050
It had a train pass and a test flight.

101
00:09:03,850 --> 00:09:10,510
We used to train part to bring our model when we were defining our model.

102
00:09:10,840 --> 00:09:11,860
We had three parts.

103
00:09:12,310 --> 00:09:14,650
One was to give the structure of the model.

104
00:09:15,640 --> 00:09:18,670
The second was configuring the learning process.

105
00:09:19,240 --> 00:09:22,060
And the third was learning the quick function.

106
00:09:22,230 --> 00:09:28,900
Kuprin de model one time model was train reaching its performance on the basic.

107
00:09:29,650 --> 00:09:35,650
And we found out that our model is giving all the performance of 85 percent, which we were satisfied.

108
00:09:37,930 --> 00:09:45,520
So following this entire process, we have created a classification model using neural networks.