1
00:00:00,210 --> 00:00:00,550
Hi.

2
00:00:00,660 --> 00:00:06,990
Welcome back to the course in this section, we'll take a look at a little last object detector, which

3
00:00:06,990 --> 00:00:10,290
is called the plant doctor object detector.

4
00:00:10,950 --> 00:00:12,240
Why plan DR?

5
00:00:12,270 --> 00:00:15,820
Well, what is that what this dataset comprises of which we'll take a look.

6
00:00:15,840 --> 00:00:17,820
So just open notebook 52.

7
00:00:18,660 --> 00:00:21,540
You can see these are these look like sick leaves.

8
00:00:21,570 --> 00:00:23,430
This is a plant with the disease.

9
00:00:24,090 --> 00:00:30,300
So this dataset was compiled basically labeling a lot of different plant lives and whether they're diseased

10
00:00:30,300 --> 00:00:30,870
or not.

11
00:00:31,350 --> 00:00:36,570
So it's a very good tool that DeFi in the wild, if you wanted, if you had a phone up just to check

12
00:00:36,570 --> 00:00:40,980
and see if a leaf was sick or not of a plant or tree was sick or not.

13
00:00:41,580 --> 00:00:46,470
Then you can upload it and use this model to predict where it is, what it is.

14
00:00:47,490 --> 00:00:48,910
So now let's begin.

15
00:00:48,960 --> 00:00:55,920
So this is a yellow vision five model, probably d best ability that to model in 2022 right now.

16
00:00:56,850 --> 00:00:59,640
And this is a notebook from a room of thought yet again.

17
00:01:00,270 --> 00:01:01,590
So what we'll do here?

18
00:01:01,590 --> 00:01:05,220
Firstly, we install the dependencies and set up the model.

19
00:01:05,700 --> 00:01:07,200
So you just clone this there.

20
00:01:07,410 --> 00:01:13,170
This is this is the you will be fired from ultrabooks and then you install the requirements, which

21
00:01:13,170 --> 00:01:17,880
is basically setting up all of the packages that we need in modules.

22
00:01:18,060 --> 00:01:23,100
Next, you can export your own data set using the rebel floor Ramaphosa API.

23
00:01:23,700 --> 00:01:29,730
However, I conveniently have this function here, which is hosted on my Google Drive, so you can automatically

24
00:01:29,730 --> 00:01:34,350
download the dataset here, which this line does next.

25
00:01:34,440 --> 00:01:36,210
We can just navigate to the directory.

26
00:01:36,210 --> 00:01:42,120
Check out the data Y.A. File, which has the training parts, the number of classes and the class names

27
00:01:42,120 --> 00:01:44,370
here, but you can see can all see them here.

28
00:01:45,180 --> 00:01:48,060
Strawberry leaf number of different things here.

29
00:01:48,870 --> 00:01:56,100
So and it tells you tomato more leaf tomato leaf so you can see when the class names that indicate disease

30
00:01:56,100 --> 00:01:57,300
or not sometimes.

31
00:01:58,470 --> 00:02:01,740
So next, we defined a model of architecture and configuration.

32
00:02:02,400 --> 00:02:04,590
So we have done this before in the previous lesson.

33
00:02:04,590 --> 00:02:06,870
So we want go over this again.

34
00:02:07,680 --> 00:02:09,390
Next, let's just go through.

35
00:02:09,390 --> 00:02:12,030
All of this has to do with the model config here.

36
00:02:12,540 --> 00:02:14,550
Now we're ready to train our model.

37
00:02:14,640 --> 00:02:17,730
So let us put the data into the right directory is here.

38
00:02:18,090 --> 00:02:23,760
So you just move to a little five trained to the mainstream folder here so you can see what it looks

39
00:02:23,760 --> 00:02:24,210
like here.

40
00:02:24,220 --> 00:02:32,700
So we have to and valid as a directory for training from here and now we can just call the Python training

41
00:02:33,240 --> 00:02:36,540
function both file so you can see the script here.

42
00:02:36,570 --> 00:02:43,710
You can specify imitates met size number of epochs where the data no file is located, where the CFG

43
00:02:43,710 --> 00:02:48,150
model file is created, and widths where we're not using any weights here.

44
00:02:48,330 --> 00:02:50,940
And then we are results here as well.

45
00:02:50,940 --> 00:02:56,430
And Cashman's you're putting the data into RAM, so which is a lot faster, by the way.

46
00:02:57,630 --> 00:02:59,550
So you can see this.

47
00:02:59,730 --> 00:03:05,280
We've created four hundred epochs right now and you can see after each epoch, you see the training

48
00:03:05,280 --> 00:03:05,970
results here.

49
00:03:06,580 --> 00:03:11,460
You can take a look at the important ones are basically these four right here.

50
00:03:12,360 --> 00:03:18,510
These four bit difficult to see because of the format, but precision recall not point at point five

51
00:03:18,510 --> 00:03:20,820
at and map at point five two point ninety five.

52
00:03:21,510 --> 00:03:24,770
So you can see as we train, those scores get better.

53
00:03:24,780 --> 00:03:26,130
They don't get that good.

54
00:03:26,280 --> 00:03:33,420
To be honest, this model is having some trouble with this dataset, although overall it's moderately

55
00:03:33,540 --> 00:03:36,510
OK, which is a fair thing to say.

56
00:03:36,510 --> 00:03:38,010
But nevertheless, that's what it is.

57
00:03:38,520 --> 00:03:44,760
You can probably get more data or trained us with more the party to improve this and saying you can

58
00:03:44,760 --> 00:03:52,380
train this with more data or treated for more e-books or on a deeper will network can see the training

59
00:03:52,380 --> 00:03:53,850
results visualized here.

60
00:03:54,690 --> 00:04:01,200
Then you can see some of the other charts basically different illustrations as a hardcoded PPG file

61
00:04:01,200 --> 00:04:04,170
that is produced at the end of training from the Google V5.

62
00:04:05,310 --> 00:04:08,070
And next, we can see some of the differences here.

63
00:04:09,690 --> 00:04:14,700
Well, this is actually our training level, so this is all in the correct labels here says not inferences

64
00:04:14,700 --> 00:04:15,060
yet.

65
00:04:15,300 --> 00:04:22,470
And some more of them are right here and these are the basically you'll be five has data augmentation

66
00:04:22,470 --> 00:04:23,010
built in.

67
00:04:23,520 --> 00:04:27,020
So this is how the augmented data looks when it goes into the model.

68
00:04:27,730 --> 00:04:30,660
Now we have to see how a model actually performs.

69
00:04:30,670 --> 00:04:32,310
So let's run some inferences here.

70
00:04:32,310 --> 00:04:38,940
So run these cells here, and this generates and saves all the results to do Star Trek overruns.

71
00:04:39,420 --> 00:04:40,740
Let me just point us to you.

72
00:04:43,420 --> 00:04:48,370
Runs and detect and you can see experiment on the experiment, too.

73
00:04:48,400 --> 00:04:54,220
That's because I have run two experiments here, but nevertheless, let's take a look at the results

74
00:04:54,220 --> 00:05:00,310
of the experiment one you can see it's a belief you can see generally, there's a lot of, I guess you

75
00:05:00,320 --> 00:05:07,000
could say, missed detections, which means that I will recall it's probably quite bad.

76
00:05:07,510 --> 00:05:11,230
But also the precision would be quite bad, too, because it's just not even predicting much at all.

77
00:05:11,920 --> 00:05:16,510
So maybe we can lower the confidence threshold, perhaps or the stream of more data, but you can see

78
00:05:16,510 --> 00:05:17,080
it overall.

79
00:05:17,080 --> 00:05:23,200
It's performing OK, it's not performing too great, but at least we have a model that we can work with

80
00:05:23,290 --> 00:05:30,880
and iteratively iteratively improve either with more data, better data augmentation of configuration,

81
00:05:31,240 --> 00:05:34,360
configuring the hyper parameters of the model, just messing with them.

82
00:05:35,110 --> 00:05:36,700
But that's about it.

83
00:05:36,760 --> 00:05:43,020
Now for this lesson, they hope you enjoyed it and this was our last object detector tutorial.

84
00:05:43,030 --> 00:05:48,100
I know a lot of them were quite quick, and that's because it didn't go into the detail because all

85
00:05:48,100 --> 00:05:49,330
of them are so different.

86
00:05:49,900 --> 00:05:55,720
And to be honest, honestly, I would just recommend you don't have to understand all of them.

87
00:05:56,140 --> 00:05:57,770
Just pick one or two.

88
00:05:57,790 --> 00:06:03,550
I would definitely suggest look at your load if you want to get some of the best results out, as well

89
00:06:03,550 --> 00:06:09,400
as if you want our CNN's actually quite widely used still, so you can take a look at RC lenses well

90
00:06:09,820 --> 00:06:12,100
and maybe efficient detect as well.

91
00:06:12,250 --> 00:06:18,580
Those are all quite good, quite popular and well-established models that have good of real world performance.

92
00:06:19,300 --> 00:06:20,740
So that's it for this lesson.

93
00:06:20,740 --> 00:06:27,310
In the next lesson, we'll move on to deep segmentation or segmentation by deep learning.

94
00:06:27,520 --> 00:06:30,580
So thank you for watching, and I'll see you in the next lesson.

95
00:06:30,940 --> 00:06:31,240
But.