1
00:00:01,620 --> 00:00:05,070
Now, let us see the importance of a pulling lever in our network.

2
00:00:06,600 --> 00:00:13,590
In the last video, we created a network in which there was one convolutional layer and one pulling

3
00:00:13,590 --> 00:00:16,200
layer in the model, somebody.

4
00:00:16,470 --> 00:00:22,500
We saw that the number of total parameters that are to be trained is nearly one point six million.

5
00:00:25,130 --> 00:00:28,770
Now, if I remove this pulling layer.

6
00:00:32,130 --> 00:00:35,170
I'll comment on this part that there is a pulling layer.

7
00:00:43,630 --> 00:00:44,760
Run this part also.

8
00:00:46,550 --> 00:00:57,690
And now if I look at the somebody, you can see that by just removing one pulling leered, the number

9
00:00:57,690 --> 00:01:01,350
of parameters have now reached six point five million.

10
00:01:02,800 --> 00:01:04,570
Earlier, it was one point six million.

11
00:01:05,760 --> 00:01:11,730
But after that, moving this max, pulling lere, the number of parameters is six point five million.

12
00:01:12,660 --> 00:01:17,100
You can imagine the computational load this will have on our system.

13
00:01:19,020 --> 00:01:25,150
If you want to go through with this, let us update model to model to every bit.

14
00:01:35,100 --> 00:01:36,590
It's configured model to.

15
00:01:41,360 --> 00:01:46,010
And reduce the number of epochs to really run it.

16
00:01:53,100 --> 00:02:00,150
So here we will see the no time taken by a network to train for one epoch.

17
00:02:00,900 --> 00:02:06,960
So when we had that pulling layer, the time taken was nearly 32 seconds but epoch.

18
00:02:07,320 --> 00:02:09,160
So it took nearly 15 minutes.

19
00:02:09,330 --> 00:02:12,540
Fifteen minutes to train for 30 bucks.

20
00:02:13,800 --> 00:02:21,080
Now, I have run it for only three blocks because I knew the time taken for E.T. book will be very lunch.

21
00:02:22,850 --> 00:02:24,900
Let us see when it completes one epoch.

22
00:02:30,960 --> 00:02:32,010
There you go.

23
00:02:32,610 --> 00:02:35,070
For each epoch.

24
00:02:35,280 --> 00:02:43,140
Now it is going to take one minute so that computational time has doubled because we have to move this

25
00:02:43,390 --> 00:02:44,020
pooling layer.

26
00:02:46,500 --> 00:02:55,080
Now, imagine having a network with multiple convolutional layers because from basic features, we wanted

27
00:02:55,080 --> 00:02:57,030
to extract higher level features also.

28
00:02:57,540 --> 00:03:05,040
So that is why I told you that we have a set of convolutional layers which bring out higher level features

29
00:03:05,230 --> 00:03:06,960
out of these input images.

30
00:03:08,910 --> 00:03:17,200
Now, if you have to use multiple convolutional layers, you need to have pooling layers in-between.

31
00:03:17,550 --> 00:03:21,270
So that deep computational load stays under control.

32
00:03:22,770 --> 00:03:31,620
So if you add another convolutional layer without adding the max pooling layer, it will increase the

33
00:03:31,920 --> 00:03:36,780
processing time to it even more than this one minute body block.

34
00:03:39,860 --> 00:03:42,590
So this is just the time taken by our system.

35
00:03:43,010 --> 00:03:49,040
It is also taking up resources of our system, like Haraam or Deepu.

36
00:03:51,140 --> 00:03:57,680
So if you have a lot of congressional layers in your network and you do not add pooling layers, you

37
00:03:57,680 --> 00:04:04,700
may reach a point where your network cannot train because of limitation on your processing power of

38
00:04:04,700 --> 00:04:05,780
your CPA or GP.

39
00:04:07,250 --> 00:04:13,630
So it is very important that we keep decompositional law under control by using Max Pully.

40
00:04:15,500 --> 00:04:16,920
So that is all in this we do.

41
00:04:17,090 --> 00:04:25,370
I just wanted to highlight that when we have a large convolutional neural network model, pulling layers

42
00:04:25,550 --> 00:04:31,310
are used along with congressional leaders to keep decompositional load under check.
