1
00:00:01,660 --> 00:00:09,430
Now one of the problems that originate with large strayed is that we may not be able to cover the entire

2
00:00:09,430 --> 00:00:11,070
image.

3
00:00:11,260 --> 00:00:19,690
For example let us try to cover this image with a five by five window moving with a straight of three

4
00:00:22,280 --> 00:00:27,920
so the first neuron covers these first 25 pixels.

5
00:00:28,070 --> 00:00:34,400
Now in the next trade we will cover these next 25 pixels.

6
00:00:34,400 --> 00:00:36,160
So we have a trade of three.

7
00:00:36,770 --> 00:00:44,570
After the fourth pixel we called an electron divide pixels and so on till this point where our neuron

8
00:00:44,630 --> 00:00:54,190
is seeing these 25 pixels now in the next trade we do not have quantified pixels.

9
00:00:54,190 --> 00:00:59,510
We will be out of pixel columns which can be covered.

10
00:00:59,710 --> 00:01:07,220
So if I take next trade there are only 20 pixels available and not 25.

11
00:01:07,240 --> 00:01:16,180
This is a problem because we want uniformity that is each neuron should have same receptive field of

12
00:01:16,180 --> 00:01:20,820
25 pixels in this situation.

13
00:01:20,850 --> 00:01:23,260
We have two options.

14
00:01:23,310 --> 00:01:27,420
First option is to ignore these extra pixels at the border.

15
00:01:29,820 --> 00:01:38,250
So since we cannot cover the last two pixel columns we leave out one pixel column from left and one

16
00:01:38,250 --> 00:01:41,350
pixel column on the right.

17
00:01:41,520 --> 00:01:51,300
So instead of a 16 by 16 image we will consider only 14 by 14 image one line of pixel removed from all

18
00:01:51,300 --> 00:01:51,940
the sides.

19
00:01:53,570 --> 00:02:05,380
This 14 by 14 image can be covered with a fiber five window and a of three this particular option is

20
00:02:05,380 --> 00:02:15,790
called valid birding which actually means that we are not using any birding and that we only use valid

21
00:02:15,790 --> 00:02:21,110
window locations and ignore the extra pixels at the border.

22
00:02:23,030 --> 00:02:34,830
Second option is adding extra rows and columns of dummy pixels or blank pixels for example here we add

23
00:02:34,830 --> 00:02:37,650
one more layer of pixels.

24
00:02:38,010 --> 00:02:46,440
We can then move forward with a straight of three and the last neuron will also have a view of 25 pixels

25
00:02:47,150 --> 00:02:56,280
guarantee from the original image and five offer artificially generated blank pixels.

26
00:02:56,280 --> 00:03:06,600
This type of padding is called Same padding which means bad in such a way so as to have an output which

27
00:03:06,600 --> 00:03:16,780
can be covered by a window of same Burton height by default padding arguments in our software are set

28
00:03:16,790 --> 00:03:19,240
to valid.

29
00:03:19,400 --> 00:03:26,420
If you think the border of your image stores important information in those scenarios only we will change

30
00:03:26,450 --> 00:03:30,070
this parameter to see it's valid works well.

31
00:03:33,700 --> 00:03:37,360
Here is the definition of two arguments that we have covered.

32
00:03:39,330 --> 00:03:50,950
First this trade straight denotes how many steps we take in each step of convolution so in the first

33
00:03:50,950 --> 00:04:00,340
step that is our first neuron is looking at this red rectangle and a second neuron is looking at this

34
00:04:00,340 --> 00:04:01,780
blue rectangle.

35
00:04:01,870 --> 00:04:10,710
Then we have a straight off to we can specify both horizontal and vertical straight.

36
00:04:10,800 --> 00:04:14,230
Separately by four straight.

37
00:04:14,230 --> 00:04:18,630
Value is set at 1.

38
00:04:18,690 --> 00:04:27,010
The second concept that we discussed is off fighting padding is the process of adding zeros to the input

39
00:04:27,130 --> 00:04:37,850
image to maintain the dimension of output as an input if we decide to ignore the border values which

40
00:04:37,850 --> 00:04:41,320
could not be covered due to our large trade.

41
00:04:41,330 --> 00:04:48,810
In that case we are not using any padding for which the argument is valid padding.

42
00:04:49,100 --> 00:04:55,480
If we are adding additional black pixels so that our window can cover the bordered pixels.

43
00:04:55,490 --> 00:05:01,310
Also in that scenario we are using same padding.

44
00:05:01,310 --> 00:05:06,740
So these are the two arguments that we will need to specify when we train our convolution earlier.