1
00:00:00,630 --> 00:00:08,250
And welcome back to Lesson 45, we will build a pothole detector, which could be a very cool thing

2
00:00:08,250 --> 00:00:13,680
if you're thinking about if you're thinking about deploying self-driving cars and that type of robotic

3
00:00:13,680 --> 00:00:14,670
vehicle type stuff.

4
00:00:15,360 --> 00:00:18,300
And we're going to use the tiny yellow V model.

5
00:00:18,750 --> 00:00:25,950
Tiny UAV for is basically a smaller, lightweight version of the of the 44 model.

6
00:00:26,340 --> 00:00:33,600
It's meant for deployment in embedded devices and mobile phones and that type of low computational requirement

7
00:00:33,600 --> 00:00:35,520
or low power device.

8
00:00:35,670 --> 00:00:36,810
So let's take a look.

9
00:00:36,810 --> 00:00:41,280
Let's open up this notebook notebook 45, and we'll get started.

10
00:00:41,370 --> 00:00:45,750
So again, this is a rebel floor notebook, and this is what the dataset looks like.

11
00:00:46,440 --> 00:00:54,000
This is actually an output of the object attached to inference because you can see this level this twice.

12
00:00:54,300 --> 00:00:56,340
So maybe we need to adjust the animus.

13
00:00:56,820 --> 00:01:02,730
That's not a maximal suppression settings, but nevertheless you can see how it actually works and you

14
00:01:02,730 --> 00:01:03,720
can see it's quite cool.

15
00:01:03,870 --> 00:01:06,720
It's still going to possible, even though it counted, it has to.

16
00:01:07,590 --> 00:01:08,960
So let's take a look.

17
00:01:08,970 --> 00:01:13,380
So I've run all this before, and this model is actually being trained right now.

18
00:01:14,010 --> 00:01:21,090
You can see, firstly, we need to configure oh, cool DNA, which is a CUDA library for deep learning

19
00:01:21,090 --> 00:01:21,960
neural networks.

20
00:01:23,100 --> 00:01:28,110
It also depends on which video card you get in call up here, which would be the card you're using.

21
00:01:28,680 --> 00:01:34,800
So you can see those have some environments, things that changes depending on the video card or CPU

22
00:01:34,860 --> 00:01:35,430
you have.

23
00:01:36,180 --> 00:01:37,950
So let's run off that.

24
00:01:38,430 --> 00:01:44,100
Then we install darknet because darknet is a framework where you're going to be using to train in attaining

25
00:01:44,100 --> 00:01:45,210
your lawful model.

26
00:01:46,140 --> 00:01:50,970
So we clone there, get triple the darknet one honorable, flawless AI.

27
00:01:51,510 --> 00:01:55,920
Then we set up all the environment variables here for to make files.

28
00:01:56,550 --> 00:02:03,220
Then we also load the cat pre-trained widths, as well as the configuration file for the tiny.

29
00:02:05,220 --> 00:02:07,740
And then what do we do next?

30
00:02:07,890 --> 00:02:08,830
We take a look.

31
00:02:08,850 --> 00:02:09,930
We get the data here.

32
00:02:10,440 --> 00:02:11,340
This is the dataset.

33
00:02:11,340 --> 00:02:15,660
We'll be using the portal dataset and let's scroll down now.

34
00:02:16,170 --> 00:02:21,930
So now we need to set up our directories and point to folders in the correct places and load different

35
00:02:21,930 --> 00:02:22,500
files.

36
00:02:22,620 --> 00:02:23,970
So we do that, dear.

37
00:02:24,750 --> 00:02:28,950
Next, we're going to write a custom training configuration for your V4.

38
00:02:29,610 --> 00:02:33,930
So this is going to be this isn't actually about how we just run this here.

39
00:02:34,380 --> 00:02:39,030
Make sure you get a number of classes as correct here, because that's that's important.

40
00:02:39,030 --> 00:02:44,370
We only have one class in this portable training file, which you can see it printed out here just in

41
00:02:44,370 --> 00:02:44,820
case.

42
00:02:45,870 --> 00:02:50,730
And no, we have a custom CFG, which is a configuration file.

43
00:02:50,820 --> 00:02:54,120
You can see it's quite long, actually, it has some moral architecture here.

44
00:02:54,660 --> 00:02:59,250
A lot of the model training settings, as well as some of the data.

45
00:03:00,000 --> 00:03:02,310
Well, basically, it's a small architecture mainly.

46
00:03:02,820 --> 00:03:08,090
And some of the actually these are some of the data augmentation settings and other criterion.

47
00:03:08,640 --> 00:03:11,240
They're in school and back to the top.

48
00:03:11,250 --> 00:03:17,940
You can see we have that size, image size channels, momentum for letting read the key for that twinning

49
00:03:17,940 --> 00:03:18,840
rates as well.

50
00:03:19,380 --> 00:03:21,030
So you can see there's a lot of different things here.

51
00:03:21,060 --> 00:03:25,290
This is actually some of the data augmentation settings as well.

52
00:03:25,710 --> 00:03:26,730
So it's pretty good.

53
00:03:27,510 --> 00:03:31,340
So we're ready to train this network with this well.

54
00:03:32,340 --> 00:03:36,450
So let's talk to training so you can see we just run the darknet.

55
00:03:37,200 --> 00:03:43,020
We just pointed this file here, and the document detector starts the training process right here.

56
00:03:43,770 --> 00:03:45,270
And you can see it's training.

57
00:03:45,270 --> 00:03:46,260
It's training pretty quick.

58
00:03:46,260 --> 00:03:48,800
Actually, I believe this model doesn't take that long to train.

59
00:03:49,290 --> 00:03:54,180
We already have a fairly good map score at point zero, five, six or seven percent.

60
00:03:54,690 --> 00:03:55,680
That's pretty good.

61
00:03:56,010 --> 00:03:58,290
Time left started this.

62
00:03:58,650 --> 00:04:00,090
Let's see how long this is taking.

63
00:04:00,120 --> 00:04:01,410
I didn't start this that long ago.

64
00:04:01,470 --> 00:04:06,930
To be fair, the school up to the top because it's printing out a lot of different lines.

65
00:04:07,470 --> 00:04:07,930
Whoops.

66
00:04:09,480 --> 00:04:15,810
So you can see it's probably going to train this entire network in maybe about a half full, roughly.

67
00:04:17,580 --> 00:04:20,880
So, yeah, so the 20 minutes initially had left.

68
00:04:21,510 --> 00:04:22,350
So that's pretty good.

69
00:04:23,100 --> 00:04:27,720
So I don't actually have the inference screenshots for the moment that you'll be training here.

70
00:04:28,380 --> 00:04:32,640
So you just have to wait for this to be finished training and then you can see it.

71
00:04:34,020 --> 00:04:34,590
But that's it.

72
00:04:34,590 --> 00:04:36,600
Actually, it for this lesson is quite short.

73
00:04:37,530 --> 00:04:42,870
It's a lot going on, obviously, but we aren't going to get into the nitty-gritty of training this

74
00:04:43,290 --> 00:04:45,300
one when these notebooks are meant for.

75
00:04:45,840 --> 00:04:51,060
Is that you can put your own datasets in them once you can host them on rebel floor and quite easily

76
00:04:51,060 --> 00:04:57,750
and put them here and then start training all these different models on your data and create a bunch

77
00:04:57,750 --> 00:04:59,900
of different models that you can use and deploy.

78
00:05:00,000 --> 00:05:00,840
Different spaces.

79
00:05:01,410 --> 00:05:03,320
So that's it for this lesson.

80
00:05:03,390 --> 00:05:09,270
We'll stop there, and the next lesson, we'll take a look at a mushroom type detector, which is a

81
00:05:09,270 --> 00:05:14,670
very cool one, and we're going to use detection too from Facebook to create that model.

82
00:05:15,150 --> 00:05:16,440
So stay tuned for that.

83
00:05:16,650 --> 00:05:17,070
Thank you.