1
00:00:00,090 --> 00:00:00,510
Hi, guys.

2
00:00:00,540 --> 00:00:06,480
Welcome back in this lesson, we'll be treating a object detector to detect different chess pieces,

3
00:00:06,480 --> 00:00:08,250
which is actually a really cool project.

4
00:00:08,710 --> 00:00:12,700
I will be using PyTorch vision of your vision tree.

5
00:00:12,720 --> 00:00:13,890
So let's get started.

6
00:00:13,890 --> 00:00:16,710
So I've already set up this notebook here.

7
00:00:17,040 --> 00:00:17,580
I already run.

8
00:00:17,580 --> 00:00:22,200
Some of these files can see how long they each take to run six seconds one second.

9
00:00:22,740 --> 00:00:26,480
Here we just clue into robots the way gilovich and tree ripple.

10
00:00:27,280 --> 00:00:29,730
But then we're done with our dataset here.

11
00:00:29,880 --> 00:00:33,150
Let me just check our directory as make sure all images up there.

12
00:00:34,320 --> 00:00:37,920
Then we just move around images to the correct directories.

13
00:00:37,920 --> 00:00:41,310
Then we just check to make sure terrible, flawed text is all right.

14
00:00:42,270 --> 00:00:49,410
Then again, we just do the same thing with the trees and the other validation directories, and now

15
00:00:49,410 --> 00:00:51,960
we can set up our model config here.

16
00:00:52,200 --> 00:00:56,070
So you just navigate into this directory display labels.

17
00:00:56,070 --> 00:00:57,750
These are the labels we'll be looking at here.

18
00:00:58,210 --> 00:01:00,120
This is a Bishop Black Bishop.

19
00:01:00,390 --> 00:01:02,000
Well, you can see a bulletin here.

20
00:01:02,290 --> 00:01:04,020
Standard chess pieces that we all know.

21
00:01:05,280 --> 00:01:10,470
And now we can convert all labels to the automatic specification Ultralight X, by the way.

22
00:01:10,710 --> 00:01:14,670
Maybe I mentioned that in previous lessons, but they're a group.

23
00:01:14,880 --> 00:01:21,210
I guess they're a company that has open sourced some of their your pie torch implementations, and it's

24
00:01:21,210 --> 00:01:22,500
quite well done.

25
00:01:22,590 --> 00:01:26,130
It has a lot of features and a lot of bells and whistles and some fine tuning.

26
00:01:26,820 --> 00:01:28,800
It's a really, really well done package.

27
00:01:28,800 --> 00:01:32,910
I encourage you to check out the Get Up report, and you can explore it there.

28
00:01:32,970 --> 00:01:36,400
You can take a look at the YOLO Vision five models I did.

29
00:01:36,420 --> 00:01:41,850
I did mention to you in the previous section that there's a reason sometimes you would want to train

30
00:01:42,000 --> 00:01:45,750
your fish tree model, and it's not because it's faster.

31
00:01:46,230 --> 00:01:51,840
It's because in some OpenSea the implementations, it's actually quite easy to get a view of which entry

32
00:01:51,840 --> 00:01:54,870
model up and running using the DNN module in OpenCV.

33
00:01:55,710 --> 00:02:00,060
So if you wanted to do that, I mean open, see if we probably will support your vision five.

34
00:02:00,660 --> 00:02:04,470
Maybe, maybe if it does right now and haven't checked, but it didn't.

35
00:02:04,830 --> 00:02:09,540
Maybe a couple of months ago when I last checked, but we can see that there's a lot of reasons why

36
00:02:09,540 --> 00:02:10,410
you would want to use it.

37
00:02:10,410 --> 00:02:16,430
You'll love which industry still has good, very good performance, and it's quite easy to create prototype

38
00:02:16,440 --> 00:02:19,020
prototypes using the elevation tree.

39
00:02:19,050 --> 00:02:21,420
So anyway, let's get back to the notebook.

40
00:02:21,870 --> 00:02:26,160
So once you've run all of these blocks of code, you can start beginning the training model and that's

41
00:02:26,160 --> 00:02:27,060
what's going on here.

42
00:02:27,510 --> 00:02:29,600
So you can see this is from beginning output.

43
00:02:30,090 --> 00:02:33,480
It's what it looks like, and you can see the trading summary here.

44
00:02:33,480 --> 00:02:38,310
You can see we've only completed 60 books, and that was pretty quick, actually, because I started

45
00:02:38,310 --> 00:02:46,500
this maybe about 10 minutes ago, and you can see some of the losses here, as well as some other details.

46
00:02:47,340 --> 00:02:48,610
So this is quite good.

47
00:02:48,630 --> 00:02:53,370
I can actually see what it is your memory usage here, which is also a good thing to note.

48
00:02:54,210 --> 00:02:56,580
So this is quite good here.

49
00:02:56,760 --> 00:03:03,810
Now let's take a look and see some of the performance metrics we get from an ultra that's x open source

50
00:03:03,830 --> 00:03:04,860
to you report.

51
00:03:05,700 --> 00:03:10,590
Did they provide similar metrics to for the all of which in five models actually

52
00:03:13,860 --> 00:03:16,350
can also see some of the inferences here?

53
00:03:16,990 --> 00:03:21,570
No, I actually did run some inferences here which were on the previous model I trained.

54
00:03:22,170 --> 00:03:24,580
And you can see this is how it looks right here.

55
00:03:24,600 --> 00:03:28,800
You can see all of the pieces I've been doing to fight correctly, which is quite good.

56
00:03:28,890 --> 00:03:29,370
Amazing.

57
00:03:30,270 --> 00:03:32,430
And you can see it actually works quite well.

58
00:03:32,460 --> 00:03:40,020
Now, one thing you will notice here is that this model works quite well on this chessboard, and that's

59
00:03:40,020 --> 00:03:41,910
because it's actually just what it's trained on.

60
00:03:42,360 --> 00:03:48,060
If you were to try this model on a different chessboard, a different looking chess pieces, it may

61
00:03:48,060 --> 00:03:48,810
have some issues.

62
00:03:48,810 --> 00:03:49,410
Definitely.

63
00:03:50,040 --> 00:03:55,860
However, if you wanted to combine several different data sets, they could get a group of like maybe

64
00:03:55,860 --> 00:04:00,300
20 of the most popular styles of chess board, just boards and different pieces.

65
00:04:00,300 --> 00:04:04,560
And to get them different lighting conditions, different angles, then I would say you would you would

66
00:04:04,560 --> 00:04:10,500
have a fairly good object detector that can recognize all these chess pieces quite accurately.

67
00:04:11,400 --> 00:04:14,080
So that's it for this chess piece model.

68
00:04:14,100 --> 00:04:15,220
I hope you enjoyed it.

69
00:04:15,240 --> 00:04:20,460
And in the next lesson, we'll take a look at a hard hat detector.

70
00:04:20,760 --> 00:04:22,020
So stay tuned for that.

71
00:04:22,410 --> 00:04:25,620
This is useful for construction sites, by the way, which I mentioned here.

72
00:04:26,430 --> 00:04:27,260
All right, thank you.

73
00:04:27,270 --> 00:04:29,070
And I'll see you in the next lesson.

74
00:04:29,310 --> 00:04:29,610
But.