8. Build CNNs in Python using Keras/12. Building a Simple Image Classifier using CIFAR10.srt

1
00:00:02,190 --> 00:00:07,970
OK so you've just finished looking at your honest dataset you're a first model.

2
00:00:08,070 --> 00:00:13,050
So basically you've imported the data you've created your model you've trained it and if you've analyzed

3
00:00:13,050 --> 00:00:17,310
the results you've seen that you've loaded it and you've tested it on some really low.

4
00:00:17,640 --> 00:00:19,500
So now let's do the same for Safar 10.

5
00:00:19,530 --> 00:00:23,290
We're not going to break it down into all the steps together and we're actually going to show you know

6
00:00:23,310 --> 00:00:26,220
how we actually just do it all in one sequence of code.

7
00:00:26,370 --> 00:00:29,180
Similarly to what we did at the end of the chapter.

8
00:00:29,500 --> 00:00:31,230
But now a different data set.

9
00:00:31,230 --> 00:00:32,840
So let's get to it.

10
00:00:32,970 --> 00:00:35,290
But first let's talk a bit about c14.

11
00:00:35,340 --> 00:00:41,430
So foughten basically it's a huge image data set to get support 60000 images as well or maybe more.

12
00:00:41,790 --> 00:00:48,210
And basically it has 10 classes here Epley an automobile blade cat or dog frog horse chip truck.

13
00:00:48,210 --> 00:00:53,760
These are some sample images in it and we're going to try and classify that actually the techs what

14
00:00:53,760 --> 00:00:55,050
is seen in the image.

15
00:00:55,050 --> 00:00:59,350
So if you showed a picture of a car should notice a car a dog should know it's a dog.

16
00:00:59,580 --> 00:01:01,170
And for the other categories here.

17
00:01:01,260 --> 00:01:02,430
So let's get to it.

18
00:01:03,650 --> 00:01:04,020
OK.

19
00:01:04,050 --> 00:01:11,110
So opening our 8.1 one building a CNN image justification Safad 10 that's a violation on the books.

20
00:01:11,140 --> 00:01:12,350
The book full the here.

21
00:01:12,990 --> 00:01:15,070
Basically I just defined the categories again.

22
00:01:15,150 --> 00:01:17,360
And now we get straight to it.

23
00:01:17,490 --> 00:01:19,110
Let's just go through this quickly.

24
00:01:19,140 --> 00:01:22,650
We have all our imports here probably some that aren't even necessary.

25
00:01:22,650 --> 00:01:25,270
I tend to do that sometimes and I'm testing stuff.

26
00:01:25,350 --> 00:01:28,430
My boss gets irritated with me quite a bit for that.

27
00:01:28,450 --> 00:01:29,840
But fair enough.

28
00:01:29,850 --> 00:01:31,870
I still do it from time to time.

29
00:01:31,920 --> 00:01:33,360
We have a size.

30
00:01:33,420 --> 00:01:35,650
We have a number of classes we have.

31
00:01:35,770 --> 00:01:40,930
He parks it in right here which we imported from here as well.

32
00:01:40,980 --> 00:01:44,920
We just display our ships just to get a handle on what we're doing with our data.

33
00:01:44,970 --> 00:01:51,840
It's always a good idea to see just quantify how many data sets are called because your dataset.

34
00:01:51,900 --> 00:01:55,110
What's the shape what's the size of your test data set as well.

35
00:01:55,530 --> 00:02:01,620
So then we just do a quick formatting here because if turn is true dimensions we do actually have to

36
00:02:01,620 --> 00:02:03,040
add the mentioned on to it.

37
00:02:03,050 --> 00:02:07,830
It just comes up automatically here and then we just change it to float.

38
00:02:07,830 --> 00:02:14,580
We divide by 255 and we change the training to to test labels on the training labels too categorical

39
00:02:14,580 --> 00:02:16,280
or hot one in coding it.

40
00:02:16,530 --> 00:02:21,540
We define our model here which is basically this should be the same model as we did before.

41
00:02:21,540 --> 00:02:28,740
But I think I've added in yet I did it in two more convolution layers here also what's different about

42
00:02:28,740 --> 00:02:34,770
this model is that we have the two filters and the two filters here for the fisty convolutional is that

43
00:02:34,780 --> 00:02:37,350
we have the activation defined outside of it.

44
00:02:37,590 --> 00:02:43,620
I actually have done that on purpose because I remember when I was creating the stable model I actually

45
00:02:43,620 --> 00:02:47,300
wanted to shoot is a variety of ways to actually create these models.

46
00:02:47,310 --> 00:02:48,690
This isn't meant to confuse you.

47
00:02:48,690 --> 00:02:53,940
This is meant to actually show you that Karatz is very flexible and you will find a lot of example Cara's

48
00:02:53,970 --> 00:03:00,000
good where sometimes people find the activation as outside of the conflict here and some names is added

49
00:03:00,010 --> 00:03:04,650
in here you can easily have just gone put activation equal real you here as well.

50
00:03:04,770 --> 00:03:07,020
So going back to this we have to convert.

51
00:03:07,050 --> 00:03:13,200
So a second call here really Max puling dropout and we have two more convolutional is here each with

52
00:03:13,200 --> 00:03:19,330
64 filters and activations relo Max spooling and drop out here again and we flatten everything.

53
00:03:19,380 --> 00:03:22,260
And now we have a much larger dense layer here.

54
00:03:22,500 --> 00:03:27,750
This puts the specific next to 512 notes and then it goes to reload again.

55
00:03:27,810 --> 00:03:33,780
And then we have it connected here to this number of glasses which is 10 which we defined above and

56
00:03:33,780 --> 00:03:36,260
we use a different optimizer.

57
00:03:36,660 --> 00:03:41,370
Armis Propp and it's defined outside here we compile a model and print it.

58
00:03:41,400 --> 00:03:44,120
So let's print this and see how it looks.

59
00:03:44,520 --> 00:03:45,210
Very nice.

60
00:03:45,210 --> 00:03:48,740
And as you can see even though this model is more complicated.

61
00:03:48,750 --> 00:03:51,890
Number of promises doesn't increase that much which is good to know.

62
00:03:52,140 --> 00:03:56,630
It's always good to know to check a number of parameters in the model because the more premises they

63
00:03:56,640 --> 00:03:58,700
are usually the longer it takes to train.

64
00:03:59,740 --> 00:04:04,440
And you can train your model here and we just renamed it so I don't delete the previous models trend

65
00:04:05,260 --> 00:04:06,800
and basically run this.

66
00:04:06,850 --> 00:04:08,160
I'm going to run this now.

67
00:04:10,340 --> 00:04:11,290
There we go.

68
00:04:11,340 --> 00:04:16,920
Sick as you can see because the images are bigger even though this data set as you saw has only 50000

69
00:04:16,940 --> 00:04:21,240
I said 6:6 it doesn't before but does because I added the test and training.

70
00:04:22,010 --> 00:04:30,130
So we are training on 50000 images here and it's honestly not going that slow for CPE use a super training.

71
00:04:30,230 --> 00:04:36,380
We're going to do maybe about 10 epoxied I believe I said it at Ahwahnee book actually just for experimental

72
00:04:36,380 --> 00:04:40,100
purposes and you probably you're already seeing that.

73
00:04:40,490 --> 00:04:46,560
Basically an untrained classifier basically guessing would give you 10 percent accuracy one intentions

74
00:04:47,120 --> 00:04:53,210
and we're already close to 1 in 2 chance of 20 percent accuracy one in five chance of getting it right.

75
00:04:53,250 --> 00:04:56,570
When the training data sets you can see our model models slowly improving and we have a long way to

76
00:04:56,570 --> 00:04:57,160
go.

77
00:04:57,170 --> 00:04:58,510
We just want it back.

78
00:04:58,790 --> 00:05:03,040
So what leaves us I'll leave this view as an exercise.

79
00:05:03,100 --> 00:05:08,450
Once you two actually trained us for different monkeypox training maybe even change up some parameters

80
00:05:08,450 --> 00:05:09,940
here to start playing with us.

81
00:05:09,950 --> 00:05:15,650
It's quite fun playing with learning learning rates as well and can change it to the kids.

82
00:05:15,650 --> 00:05:17,480
Basically how much the leading grade decreases.

83
00:05:17,480 --> 00:05:20,510
I'll explain these concepts later on in others lives.

84
00:05:20,600 --> 00:05:26,030
But for now I just know that we have a variety of things we can tweak and deep learning is basically

85
00:05:26,540 --> 00:05:32,750
people say it's an art form and it kind of is because it's so many variations variations that are dependent

86
00:05:32,750 --> 00:05:33,550
on each other.

87
00:05:33,980 --> 00:05:40,520
You can simply change some layers here a number of layers change sizes change this and there's no hard

88
00:05:40,520 --> 00:05:43,930
science that defines what to do to get the best results.

89
00:05:44,060 --> 00:05:49,490
Mainly because it all depends on your dataset and your dataset is basically naturally occurring images

90
00:05:49,940 --> 00:05:53,190
which have so much naturally a naturally occurring variety in them.

91
00:05:53,570 --> 00:05:58,970
So it's once you can understand your data you can start figuring out how you should treat these parameters

92
00:05:59,470 --> 00:06:04,700
and I'll discuss these things later on in the slides as we build more and more complicated pacifies

93
00:06:07,430 --> 00:06:11,730
so left it could be to plot your charts here and you can run some tests as well.

94
00:06:11,730 --> 00:06:14,440
So for now I'm going to just stop this quickly

95
00:06:17,950 --> 00:06:24,970
what I'm going to do is I'm going to just load Hopefully all the imports I needed a little model train

96
00:06:24,990 --> 00:06:27,330
before and let's see if it works.

97
00:06:28,590 --> 00:06:31,110
Now it did not work because X isn't defined.

98
00:06:31,110 --> 00:06:36,780
So what we can do is just quickly run this block here because X to us and X train and all those things

99
00:06:36,780 --> 00:06:37,310
are here.

100
00:06:39,000 --> 00:06:42,400
And we can run this and this brings up this window.

101
00:06:42,570 --> 00:06:44,430
Here it is.

102
00:06:44,430 --> 00:06:45,640
So clearly this isn't a frog.

103
00:06:45,690 --> 00:06:49,550
This is probably a bit and then to it.

104
00:06:49,560 --> 00:06:50,510
This is not a dog.

105
00:06:50,760 --> 00:06:51,550
This is a dog.

106
00:06:51,580 --> 00:06:53,010
This is a horse not a cat.

107
00:06:53,010 --> 00:06:53,510
This is a frog.

108
00:06:53,520 --> 00:06:55,130
Yes automobile.

109
00:06:55,140 --> 00:06:55,780
Yes.

110
00:06:55,980 --> 00:06:57,420
Yes automobile.

111
00:06:57,570 --> 00:06:59,290
Not a frog.

112
00:07:00,180 --> 00:07:05,580
So we can see businesses 10 10 samples this model they trained here before which I probably trained

113
00:07:05,610 --> 00:07:09,560
maybe just 10 ebox has about 50 percent accuracy.

114
00:07:09,660 --> 00:07:16,140
So let's see if he can be debt accuracy on Safar has reached very very high like 99 percent.

115
00:07:16,140 --> 00:07:17,250
Let's see if you can get that.

116
00:07:17,250 --> 00:07:18,510
Get it out on your own.

117
00:07:18,750 --> 00:07:19,570
Good luck.