1
00:00:00,870 --> 00:00:03,570
Now, let's take a look at Colonel size and depth.

2
00:00:03,960 --> 00:00:10,470
These are some parameters that actually control the size of our convolutional filter and controls the

3
00:00:10,470 --> 00:00:13,230
size of our feature map as well.

4
00:00:13,770 --> 00:00:14,730
So let's get started.

5
00:00:14,850 --> 00:00:21,030
So some parameters that control the cone filter, as I mentioned, are colonel size and depth depth

6
00:00:21,100 --> 00:00:24,960
meaning meaning one for greyscale skill tree fire RGV, which you've seen previously.

7
00:00:25,470 --> 00:00:29,010
There's also a stride and padding, which we'll do in the following chapters.

8
00:00:29,850 --> 00:00:31,470
So let's take a look at this.

9
00:00:32,130 --> 00:00:37,740
Previously, when we were using our filters, we only used a tree by tree filter.

10
00:00:38,190 --> 00:00:39,670
Can there be order sizes?

11
00:00:39,930 --> 00:00:42,840
Can it be a four by four, five by five or seven by seven?

12
00:00:43,140 --> 00:00:44,430
Does it have to be square?

13
00:00:44,980 --> 00:00:46,020
What's going on?

14
00:00:46,680 --> 00:00:49,800
Actually, yes, it has to be square, and it has to be odd.

15
00:00:49,800 --> 00:00:50,880
No filters.

16
00:00:51,330 --> 00:00:52,470
And I'll explain to you why.

17
00:00:53,400 --> 00:00:59,880
So firstly, before I go on to that explanation, what I want to show you is if you use a five by five

18
00:00:59,880 --> 00:01:05,850
filter here on a six by six image here you're going to get a two by two output and hope.

19
00:01:06,420 --> 00:01:07,740
So imagine this here.

20
00:01:07,830 --> 00:01:12,360
Imagine we overlay this five by five matrix on our six by six grid here.

21
00:01:12,630 --> 00:01:16,350
It's going to occupy this range here to this range here.

22
00:01:17,040 --> 00:01:20,900
So you can you can see that point is going to be one available position.

23
00:01:20,920 --> 00:01:21,900
We can move it to the right.

24
00:01:22,350 --> 00:01:28,860
So we shifted one to the right suit now covers this area here and then we shifted down dome and it covers

25
00:01:28,860 --> 00:01:29,580
this area here.

26
00:01:29,580 --> 00:01:31,140
And then we shifted to the right again.

27
00:01:31,620 --> 00:01:33,540
So you can see there's only four positions.

28
00:01:33,540 --> 00:01:39,450
This matrix can be overlaid onto this input image, and that's how you get two by two here.

29
00:01:40,410 --> 00:01:45,510
So here's a little simple formula for calculating the future map size.

30
00:01:46,050 --> 00:01:48,450
So the end is the dimensions of the image.

31
00:01:48,450 --> 00:01:52,560
And remember, the images have to be square in convolutional neural networks.

32
00:01:53,040 --> 00:01:58,050
So it's six by six here and then minus five by five, which is a filter here.

33
00:01:58,050 --> 00:02:02,940
Plus one, which is two two, is the dimension of our feature map output.

34
00:02:03,480 --> 00:02:05,820
So that's an easy, simple way to calculate it.

35
00:02:06,810 --> 00:02:11,190
So I said we couldn't have on even numbered filters.

36
00:02:11,670 --> 00:02:12,420
Why is that?

37
00:02:12,480 --> 00:02:18,600
Well, if you go back to the slide here, you can see when you put the five by five matrix here, there's

38
00:02:18,600 --> 00:02:20,730
a center point in the middle right here.

39
00:02:20,970 --> 00:02:23,010
This a little point, right right here.

40
00:02:23,670 --> 00:02:30,630
However, if it was a four by four matrix, there's no center point that does cause a problem now because

41
00:02:30,630 --> 00:02:36,990
what it does, it basically messes up the symmetry of our or filters, and that can lead to distortions

42
00:02:36,990 --> 00:02:37,680
across layers.

43
00:02:37,680 --> 00:02:44,130
So while it can work and there's no mathematical reason why you can't have an even numbered filter in

44
00:02:44,130 --> 00:02:49,710
theory, in practice, convolutional networks don't perform well with even sized filters to work much

45
00:02:49,710 --> 00:02:55,500
better with odd size filters because of that symmetry around the Centrepoint or anchor point.

46
00:02:56,160 --> 00:02:58,560
So let's move on to debt.

47
00:02:58,770 --> 00:03:03,450
Remember, I referred to depth as being the depth, the number of channels in an image.

48
00:03:03,930 --> 00:03:07,770
Well, that doesn't always refer to that in some nomenclature.

49
00:03:08,220 --> 00:03:13,890
Depth can refer to how many, like how many kernels of filters you're using, which technically it's

50
00:03:13,890 --> 00:03:17,130
wrong because it's actually four dimensions instead of a two dimension.

51
00:03:17,520 --> 00:03:20,790
But nevertheless, some researchers tend to use it often.

52
00:03:20,790 --> 00:03:25,710
Maybe when you're talking about grayscale images than depth would refer to, the two dimension of it

53
00:03:25,710 --> 00:03:27,510
can also refer to the four dimension.

54
00:03:27,960 --> 00:03:30,540
Generally, depth does refer to the two dimension.

55
00:03:30,870 --> 00:03:39,150
It kind of means that there's kind of always indicates the color of the image what it's color or grayscale.

56
00:03:39,690 --> 00:03:42,510
So that concludes this chapter on depth.

57
00:03:42,900 --> 00:03:45,690
And let's see what what we did initially.

58
00:03:45,690 --> 00:03:47,760
Just rehash it kernel size.

59
00:03:48,210 --> 00:03:52,770
So in the next chapter, we're going to take a look at putting.

60
00:03:52,890 --> 00:03:55,380
And then afterwards, we're going to take a look at straight.

61
00:03:55,920 --> 00:03:59,040
So stay tuned for those lessons and we'll see you in the next section.

62
00:03:59,250 --> 00:03:59,730
Thank you.