1
00:00:00,300 --> 00:00:07,860
In this lesson we're gonna be learning how to make predictions on individual images using Charisse after

2
00:00:07,860 --> 00:00:10,720
our model has been trained.

3
00:00:10,800 --> 00:00:15,920
I'll add a markdown cell here to create a subsection for us.

4
00:00:16,150 --> 00:00:22,830
And now what I'll do is I'll quickly pull up the documentation from Caris to see how it is that we can

5
00:00:22,830 --> 00:00:24,500
make a prediction.

6
00:00:24,600 --> 00:00:31,230
So scrolling down on this page here the model functional API we had our compile method we had our fit

7
00:00:31,230 --> 00:00:38,610
method and then down here we have evaluate which will we'll use later on and here predict this seems

8
00:00:38,610 --> 00:00:39,620
what we're looking for right.

9
00:00:40,710 --> 00:00:45,260
This method will generate output predictions for the input samples.

10
00:00:45,570 --> 00:00:52,990
And if you give it a lot of input samples then it will do the computation in batches as an output.

11
00:00:53,010 --> 00:00:56,970
This method will give us a name pi array of predictions.

12
00:00:57,630 --> 00:00:59,190
So let's try this out.

13
00:00:59,190 --> 00:01:02,310
Let's try and predict a single image.

14
00:01:02,310 --> 00:01:09,450
I want to use my validation data set him X Val and I'll quickly remind ourselves of the shape of the

15
00:01:09,450 --> 00:01:11,040
validation data set.

16
00:01:11,220 --> 00:01:18,810
It had 10000 images and each image had three thousand seventy two values in it that corresponded to

17
00:01:18,810 --> 00:01:20,560
the individual pixels.

18
00:01:20,610 --> 00:01:27,630
So that means if I wanted to pull out the very first image in the validation data set then I could do

19
00:01:27,630 --> 00:01:31,920
so with square brackets and then zero.

20
00:01:31,920 --> 00:01:36,140
This here will give us the very first image all right.

21
00:01:36,150 --> 00:01:41,150
So now that we've got a single image let's take a look at what the shape of this image is.

22
00:01:41,310 --> 00:01:44,310
And here you can see that this is actually flattened right.

23
00:01:44,310 --> 00:01:46,560
It's a one dimensional array.

24
00:01:46,770 --> 00:01:54,960
If we want to make this a two dimensional right then we could use num PI's capabilities and use N.P.

25
00:01:55,050 --> 00:01:57,810
dot expand after school.

26
00:01:57,910 --> 00:02:00,050
DRM is four dimensions.

27
00:02:00,420 --> 00:02:08,910
So here we could supply our X Vale square brackets zero and then we can say it should add another dimension

28
00:02:08,910 --> 00:02:16,470
here with access equals zero shift tab on my keyboard and bring up the quick documentation then I can

29
00:02:16,470 --> 00:02:25,130
see that axis is an integer and it's the position in the expanded axis where the new axis is placed.

30
00:02:25,420 --> 00:02:32,850
I want to store this whole thing in a variable called test and I'm very quickly going to take a look

31
00:02:32,850 --> 00:02:36,300
at what the shape of test now is.

32
00:02:36,300 --> 00:02:42,120
And here you can see that all we've done is add an extra dimension here so that we can now feed this

33
00:02:42,120 --> 00:02:43,920
into our predict method.

34
00:02:43,920 --> 00:02:44,640
Why is that.

35
00:02:45,030 --> 00:02:49,610
Well it's because it actually usually expects more than one prediction.

36
00:02:49,610 --> 00:02:56,100
It will want to make predictions on a number of samples here's how we would use the predict method.

37
00:02:56,100 --> 00:03:00,050
I want to pick model number to model on a score to predict.

38
00:03:00,520 --> 00:03:04,100
And if I wanted to make a predict on my test image.

39
00:03:04,320 --> 00:03:06,270
So that's our argument.

40
00:03:06,270 --> 00:03:08,610
Let's take a look at what the output is here.

41
00:03:09,060 --> 00:03:12,480
What we get here are actually our probabilities.

42
00:03:12,480 --> 00:03:14,310
All of these should sum to 1.

43
00:03:15,660 --> 00:03:20,990
So if I put that sum afterwards then I can prove this to you.

44
00:03:21,000 --> 00:03:25,660
So these here are the probabilities for each of the different categories.

45
00:03:25,740 --> 00:03:32,190
Each of the different classes for this image the reason we've got these nice probabilities is because

46
00:03:32,250 --> 00:03:38,360
we've used soft Max in our output layer for our activation function right now.

47
00:03:38,550 --> 00:03:44,880
I think this is some really ugly formatting and I think we could do a little better than this and the

48
00:03:44,880 --> 00:03:50,970
way we're going to change the way these numbers are formatted in our Jupiter notebook is to change the

49
00:03:51,270 --> 00:03:53,970
print options for num pi.

50
00:03:54,030 --> 00:04:01,860
So with MP dot set on a scope print options we can actually say how precise how many decimal points

51
00:04:02,010 --> 00:04:07,580
we want to put a notebook to print out so we can do that by setting the precision.

52
00:04:07,710 --> 00:04:15,450
And if I said that 2 3 and hit shift enter on this and hit shift enter again on my prediction then you

53
00:04:15,450 --> 00:04:20,780
can see that I'm only printing the probability up to three decimal points now.

54
00:04:20,790 --> 00:04:24,360
What if we wanted to predict more than one image.

55
00:04:24,480 --> 00:04:31,500
Well in this case we use our model model number two and then as an argument to the predict method we

56
00:04:31,500 --> 00:04:35,580
can actually supply the entire evaluation dataset.

57
00:04:35,580 --> 00:04:41,640
In this case what we'd get is an enormous array with all the predictions.

58
00:04:41,670 --> 00:04:43,050
Let me show you what I mean.

59
00:04:43,170 --> 00:04:47,970
If I had shift enter on the cell then we get something like this.

60
00:04:48,190 --> 00:04:53,980
These probabilities are in scientific notation which is why you see this e afterwards.

61
00:04:53,980 --> 00:05:01,540
But if I take a look at the shape of this array then what you'll see is I get 10 predictions or 10 probabilities

62
00:05:01,840 --> 00:05:09,040
for each and every single of the 10000 images in our valuation dataset.

63
00:05:09,040 --> 00:05:12,720
Now what if we're just interested in the predicted class.

64
00:05:12,730 --> 00:05:16,120
What if we're just interested in the largest number of these 10.

65
00:05:17,260 --> 00:05:21,050
Well of course there's a handy method from Caris once again.

66
00:05:21,220 --> 00:05:30,830
So model underscored to dot predict underscore classes and then test will give us the predicted class.

67
00:05:30,850 --> 00:05:34,800
So in this case it will predict class number six.

68
00:05:34,840 --> 00:05:38,090
Now of course the question is did it get it right.

69
00:05:38,350 --> 00:05:42,900
The place we would look for that of course is why on the score Val.

70
00:05:43,000 --> 00:05:50,230
So the labels for evaluation valuation dataset y underscore Val square bracket 0 will give us the actual

71
00:05:50,230 --> 00:05:54,170
class the actual category that this belongs to.

72
00:05:54,250 --> 00:06:02,160
In this case our model predicted the correct class for the very first image in the validation data set.

73
00:06:02,250 --> 00:06:10,210
Now as a challenge can you write a for loop where you print out the actual value and the predicted value

74
00:06:10,510 --> 00:06:14,760
for the first 10 images in the validation dataset.

75
00:06:14,880 --> 00:06:19,620
You will get bonus points if you format this nicely and use an F string.

76
00:06:19,630 --> 00:06:26,940
I'll give you a few seconds to pause the video and give this a go.

77
00:06:26,990 --> 00:06:27,800
Ready.

78
00:06:27,800 --> 00:06:29,360
Here's the solution.

79
00:06:29,480 --> 00:06:36,040
It's gonna be four and then let's say number in range 10.

80
00:06:36,350 --> 00:06:40,700
And here what I'll do is I'll say test on a score.

81
00:06:40,700 --> 00:06:44,600
Image is equal to number pi.

82
00:06:44,780 --> 00:06:53,350
Expand dimensions parentheses x on a score Val square brackets No.

83
00:06:53,630 --> 00:06:59,960
So as I'm iterating through my loop I'm going to pull out the corresponding image from my validation

84
00:06:59,960 --> 00:07:01,410
data set.

85
00:07:01,430 --> 00:07:05,990
Put a comma C axis is equal to zero.

86
00:07:06,470 --> 00:07:18,820
Then I'll say predicted underscore Val is equal to model on a score to don't predict classes.

87
00:07:19,040 --> 00:07:20,870
Test image.

88
00:07:21,650 --> 00:07:28,490
And since the output from predict classes is an array I'm going to pick out the first element here with

89
00:07:28,580 --> 00:07:38,540
square brackets zero and now I'll write my print statement using an F string or I'll see actual value.

90
00:07:38,540 --> 00:07:45,650
Curly braces y on the score bough square brackets No.

91
00:07:46,280 --> 00:07:57,930
And since this is also an array I'll add another pair of square brackets here put a 0 c versus predicted.

92
00:07:58,040 --> 00:08:03,510
Curly braces predicted underscore Val.

93
00:08:03,590 --> 00:08:04,890
There we go.

94
00:08:04,910 --> 00:08:07,130
Let's see if I got this right.

95
00:08:07,130 --> 00:08:14,060
Here we are and what we should see is that our model got approximately 50 percent of the predictions

96
00:08:14,090 --> 00:08:15,680
correct.

97
00:08:15,680 --> 00:08:20,060
We've already seen that the first image in the valuation data set was predicted correctly.

98
00:08:20,060 --> 00:08:21,470
But the second one I got wrong.

99
00:08:21,470 --> 00:08:22,990
The third one to get right.

100
00:08:23,120 --> 00:08:24,530
Fourth one I got wrong.

101
00:08:24,530 --> 00:08:25,580
Fifth one.

102
00:08:25,580 --> 00:08:27,980
Wrong again and so on.

103
00:08:27,980 --> 00:08:34,730
So that's how you would use the predict method and the predict class's method from Caris.

104
00:08:34,940 --> 00:08:43,880
You can use these methods on individual items or you can use these methods on entire arrays in the next

105
00:08:43,880 --> 00:08:44,430
lesson.

106
00:08:44,450 --> 00:08:48,710
We're going to be evaluating our favorite model and a bit more detail.

107
00:08:48,830 --> 00:08:56,390
We're gonna be judging it based on its accuracy its precision its recall and its F school for all of

108
00:08:56,390 --> 00:08:57,180
that and more.

109
00:08:57,230 --> 00:08:58,880
I'll see you in the next lesson.

110
00:08:58,880 --> 00:08:59,450
Take care.