1 00:00:00,570 --> 00:00:06,840 In order to create our convolutional neural network, we're just going to make use of this sequential 2 00:00:06,840 --> 00:00:12,560 API, which we had built previously, and then we'll define the input layer. 3 00:00:12,570 --> 00:00:18,390 So here we have this input shape to 24 lets we add define in size actually. 4 00:00:18,390 --> 00:00:25,270 So we have M size, that's it by M size by three. 5 00:00:25,290 --> 00:00:29,800 We could also define a number of channels, but let's just let it to be equal to three. 6 00:00:29,880 --> 00:00:36,600 Then from here we have the normal laser and then we have this column to the layer right here. 7 00:00:36,600 --> 00:00:40,020 So we have this arguments which are default. 8 00:00:40,020 --> 00:00:42,890 So we'll then bother to check on this. 9 00:00:42,900 --> 00:00:47,160 We will take this off, take this off, and that's fine. 10 00:00:47,160 --> 00:00:49,230 So we'll have this. 11 00:00:49,680 --> 00:00:50,580 There we go. 12 00:00:50,580 --> 00:00:58,980 We have our come to DX will come to D which we copy or rather we cut that off and then we put this after 13 00:00:58,980 --> 00:01:01,440 the input layer right here. 14 00:01:01,440 --> 00:01:03,900 So there we go, We have our input layer. 15 00:01:03,900 --> 00:01:08,790 We want six filters and then we can size of five. 16 00:01:09,270 --> 00:01:10,320 So that's it. 17 00:01:10,320 --> 00:01:13,530 Our stride, we could take this to be equal one. 18 00:01:13,830 --> 00:01:14,880 There we go. 19 00:01:14,880 --> 00:01:18,840 Pattern is valid, so that's what we expect. 20 00:01:18,840 --> 00:01:22,530 And then just here we have our activation. 21 00:01:22,530 --> 00:01:29,760 So we have activation, we could see sigmoid, we have the sigmoid activation since we're replicating 22 00:01:30,300 --> 00:01:33,600 the rail loop or rather the net architecture. 23 00:01:33,600 --> 00:01:39,960 So there we go, we have the sigmoid and that's our come to D from the continuity. 24 00:01:39,960 --> 00:01:41,550 We have a max pooling layer. 25 00:01:41,550 --> 00:01:46,410 Let's take this off, let's take this dense layers off from this curve to D, We have the max pull and 26 00:01:46,410 --> 00:01:46,890 layer. 27 00:01:46,920 --> 00:01:50,160 Now what we could do is we could just simply import all this. 28 00:01:50,160 --> 00:02:00,540 So let's go ahead and say in part the come to the max pull to and the dense layer coming back to our 29 00:02:00,540 --> 00:02:06,900 documentation we could check out on the max pull two D So there we go we have Max pull to DD. 30 00:02:08,040 --> 00:02:14,910 We'll specify the pull size, which is equal to and then the number of stripes equal to pattern is valid 31 00:02:14,910 --> 00:02:16,260 and data format. 32 00:02:16,260 --> 00:02:17,730 We're not going to specify that. 33 00:02:17,730 --> 00:02:22,620 So we have our MAX pull to DD and then this pull in size with the stripes. 34 00:02:22,620 --> 00:02:23,640 We copy that. 35 00:02:24,060 --> 00:02:27,090 We actually don't know the stride of two, so we replace that. 36 00:02:27,090 --> 00:02:28,440 We have a straight of two. 37 00:02:28,470 --> 00:02:35,160 The next step is this one right here, similar to what we had seen previously with a difference that 38 00:02:35,160 --> 00:02:42,150 now we have 16 number of filters and so we could simply copy this and then paste this out right here 39 00:02:42,150 --> 00:02:46,110 we have come to the max, pull two DD Now come to the max two. 40 00:02:46,110 --> 00:02:49,290 DD And then here we have 16 number of filters. 41 00:02:49,290 --> 00:02:52,140 The pattern is valid activation sigmoid. 42 00:02:52,140 --> 00:02:56,700 Then this next max pull to DD is still the same as what we had previously. 43 00:02:56,700 --> 00:02:59,400 And so we move on to the flattened layer. 44 00:02:59,970 --> 00:03:06,420 We add this here we have flatten, there we go and we run that. 45 00:03:06,420 --> 00:03:12,870 So while that's running, we just simply come right here and then add that flattened layer. 46 00:03:12,870 --> 00:03:19,950 So we have this flattened layer which is in charge of flooding this or converting this into this one. 47 00:03:19,950 --> 00:03:26,310 The output from this flatten, we now have a dense layer, which is what we've seen already. 48 00:03:26,310 --> 00:03:32,670 So we just have to put this dense layer here, stick this like this, and there we go. 49 00:03:32,670 --> 00:03:37,560 So we have our dense layer and then the activation is sigmoid. 50 00:03:37,560 --> 00:03:42,450 So we're respecting the activation using the low net paper. 51 00:03:42,450 --> 00:03:50,580 Then we have this here, take 1000 and then we have that, we add another dense layer and this other 52 00:03:50,580 --> 00:03:51,560 dense layer. 53 00:03:51,570 --> 00:04:01,590 Now we'll make sure that as we create this final dense layer right here, it has an output of two neurons 54 00:04:01,590 --> 00:04:06,060 since we are dealing with a binary classification problem. 55 00:04:07,230 --> 00:04:08,310 So that's fine. 56 00:04:08,310 --> 00:04:11,850 Take this here and then run our model. 57 00:04:12,390 --> 00:04:13,020 We have it. 58 00:04:13,020 --> 00:04:14,880 That input layer is not defined. 59 00:04:14,880 --> 00:04:24,270 So let's simply add that here we have the input layer, we run that and then check that out again. 60 00:04:24,960 --> 00:04:25,590 There we go. 61 00:04:25,590 --> 00:04:30,180 We have this input layer which we run, which we add, and then we run this. 62 00:04:30,180 --> 00:04:31,770 Now everything is fine. 63 00:04:31,770 --> 00:04:32,730 So that's it. 64 00:04:32,730 --> 00:04:39,450 We have 45 million parameters and no non trainable parameter. 65 00:04:40,290 --> 00:04:47,820 Notice how the this dense layer here is responsible for a huge percentage of our parameters. 66 00:04:47,820 --> 00:04:49,440 So we could reduce this. 67 00:04:49,440 --> 00:04:52,740 Let's take this to 100 and then this to like ten. 68 00:04:52,770 --> 00:04:54,090 Let's run that again. 69 00:04:54,090 --> 00:04:55,320 And there we go. 70 00:04:55,320 --> 00:04:57,780 We have a smaller model this time around. 71 00:04:58,560 --> 00:04:59,970 A point and notice is not. 72 00:05:00,300 --> 00:05:02,750 Parameters which we pre calculated here. 73 00:05:02,750 --> 00:05:06,480 Yea, we have 456 and we have 2460. 74 00:05:07,110 --> 00:05:12,510 And we'll see how we get exactly this number of values here for the dense layer. 75 00:05:12,540 --> 00:05:20,460 It's quite obvious as we just have 100 times ten plus the ten biases giving us 100 times 10 to 1000 76 00:05:20,460 --> 00:05:23,180 plus ten giving us 1010. 77 00:05:23,190 --> 00:05:28,710 And then here we have ten times two, 20, 20, plus two biases giving us 22. 78 00:05:28,740 --> 00:05:32,970 It should also be noted that there was a slight error here, as we don't have. 79 00:05:32,970 --> 00:05:40,320 Like with this, we have five by five by three, but there is five by five by six since we have the 80 00:05:40,320 --> 00:05:42,660 input number of channels equals six. 81 00:05:42,660 --> 00:05:49,800 So this means this is five by five by six and here is five, by five by three, All this five by five 82 00:05:49,830 --> 00:05:51,900 by three and all this five by five by six. 83 00:05:52,290 --> 00:05:59,490 And if you take five by five by six and multiply it 16 times, you should have 2416. 84 00:05:59,520 --> 00:06:04,470 It should be noted that here we're trying to replicate the learned architecture. 85 00:06:04,470 --> 00:06:10,530 And this is in no way the state of the art kinds of models we use today. 86 00:06:10,530 --> 00:06:16,110 So in the section on the corrective measures, we are going to use even better models. 87 00:06:16,110 --> 00:06:19,140 For now, let's just work with this.