1 00:00:00,090 --> 00:00:03,420 Now, let's move on to building of a model in Paris. 2 00:00:03,930 --> 00:00:09,750 So if you recall from those slides in our previous pay to watch lesson, this is a CNN model that we're 3 00:00:09,750 --> 00:00:10,410 about to build. 4 00:00:10,800 --> 00:00:11,910 We take the input image. 5 00:00:11,910 --> 00:00:13,650 We have tree by tree filters. 6 00:00:13,950 --> 00:00:18,480 We get it to the twenty to twenty two feature map set of 26 by 26. 7 00:00:18,780 --> 00:00:20,370 Then we have another conflict. 8 00:00:20,760 --> 00:00:22,900 So we actually use a tree battery filter there. 9 00:00:23,310 --> 00:00:25,080 And then we get these feature map sort of it. 10 00:00:25,080 --> 00:00:26,760 We use 64 filters in that case. 11 00:00:26,760 --> 00:00:28,590 So now we get 64 feature maps. 12 00:00:28,950 --> 00:00:31,890 We apply réélu activation layers to these. 13 00:00:32,280 --> 00:00:33,440 Then we have a max pool. 14 00:00:33,450 --> 00:00:35,190 We also play re-look to this as well. 15 00:00:35,610 --> 00:00:38,520 We have 12 by 12 because it's half that size. 16 00:00:38,520 --> 00:00:39,480 We have 64. 17 00:00:39,510 --> 00:00:43,530 Because the 64 filters flatten it for the connected layer. 18 00:00:43,560 --> 00:00:48,900 All of these nodes are connected and upwards to the soft max, which is the output, and we get the 19 00:00:48,900 --> 00:00:50,100 probabilities out of it. 20 00:00:50,700 --> 00:00:56,580 So in Keros, it's actually quite easy to construct complicated CNN's. 21 00:00:57,120 --> 00:01:00,750 So these are just some details of what we're doing here, what I just went over. 22 00:01:01,200 --> 00:01:03,990 Now let's take a look at the code where we actually build our CNN. 23 00:01:04,530 --> 00:01:06,780 So it's quite a lot of text to have. 24 00:01:06,780 --> 00:01:09,240 There's a lot of comments that explain what we do. 25 00:01:10,170 --> 00:01:15,030 So it would be helpful if you paused sometimes or just take a look at your own notebook and go through 26 00:01:15,030 --> 00:01:15,870 this on your own. 27 00:01:16,470 --> 00:01:17,920 So these are the inputs here. 28 00:01:17,940 --> 00:01:23,070 So in Keros, when TensorFlow, we have to import all of the objects. 29 00:01:23,430 --> 00:01:26,640 These are the building blocks of the layers that we would be using and see it in the CNN. 30 00:01:27,180 --> 00:01:33,030 So we have the sequential sequential basically says that we're building a sequential model, which means 31 00:01:33,030 --> 00:01:35,100 that it's each layer is connected to the next. 32 00:01:35,700 --> 00:01:38,700 Then these are two of the model is building as well. 33 00:01:39,120 --> 00:01:40,530 From the layers library. 34 00:01:41,040 --> 00:01:44,940 We have then struck out flatten builds have convex pool. 35 00:01:44,940 --> 00:01:47,950 This can all be in one layer, but is just on two lines. 36 00:01:47,970 --> 00:01:53,880 So to make it shorter, Keros we import the back end is key because we occasionally need to access that. 37 00:01:54,390 --> 00:01:58,920 And then we have this stochastic gradient descent from the as part of the library. 38 00:01:59,490 --> 00:02:05,100 So we have all of the building blocks we need if we imported them and look how easy this is in the US 39 00:02:05,100 --> 00:02:10,620 to construct a model with this new model dot ad after we create the model up here. 40 00:02:11,430 --> 00:02:13,020 We added a clumsily are here. 41 00:02:13,020 --> 00:02:19,410 We specify how many filters we specify the kernel size here, tree by tree specify the activation we 42 00:02:19,410 --> 00:02:26,580 want to use redo and input shape, which is input shape we took from up here or there. 43 00:02:26,580 --> 00:02:26,880 Was it? 44 00:02:27,600 --> 00:02:32,400 Yeah, up here 28 by 28 comma one, which is the depth of the image. 45 00:02:34,080 --> 00:02:34,920 And that's it. 46 00:02:35,220 --> 00:02:37,290 So we just cleared of this conflict. 47 00:02:37,830 --> 00:02:39,960 Let's make our second conflict now. 48 00:02:40,380 --> 00:02:41,070 So now we can. 49 00:02:41,070 --> 00:02:44,610 We want to make a 64 filter tree by tree kernel size. 50 00:02:45,090 --> 00:02:46,920 We don't actually have to write content, says Eq.. 51 00:02:46,920 --> 00:02:47,220 You can. 52 00:02:47,550 --> 00:02:52,920 The function knows that a second parameter is going to be kernel size, and then we specify the activation 53 00:02:53,720 --> 00:02:56,080 activation layer again using real. 54 00:02:56,520 --> 00:02:58,500 Now you may have wondered why we didn't. 55 00:02:58,710 --> 00:03:01,060 We didn't need to set padding or straight. 56 00:03:01,060 --> 00:03:08,670 If we can, we can actually add in a padding or straight parameter, like putting padding equal to one 57 00:03:08,670 --> 00:03:09,300 or two. 58 00:03:09,810 --> 00:03:15,450 However, we using a default values and default values, we didn't use any padding and we used stride 59 00:03:15,450 --> 00:03:17,280 of one for both conflicts. 60 00:03:17,970 --> 00:03:20,100 Next, we use a max boot again. 61 00:03:20,550 --> 00:03:27,600 And similarly, with Max boot, we can specify inside here we could put Stride Eq. four or whatever 62 00:03:27,600 --> 00:03:28,410 we wanted to use. 63 00:03:28,950 --> 00:03:32,190 However, we're not going to just yet because we're going to use default one. 64 00:03:33,210 --> 00:03:37,490 So we add the Max Boulia here and then we add the flattened layer here. 65 00:03:37,500 --> 00:03:42,270 We just basically flattened latex the previous input and flattened it. 66 00:03:42,570 --> 00:03:48,960 I notice we don't have to link like have X equal to this and have an input, but would carry us when 67 00:03:48,960 --> 00:03:49,680 we do model it. 68 00:03:49,680 --> 00:03:53,100 Add it adds sequentially each layer. 69 00:03:53,100 --> 00:03:55,350 So we know each layer feeds into the other layer. 70 00:03:56,040 --> 00:04:01,380 And when we do flatten here, it just flattens the output of the Max Boulia right here. 71 00:04:01,590 --> 00:04:05,730 That gives us nine thousand two hundred sixteen by one dimension of that. 72 00:04:06,210 --> 00:04:07,230 Come and take in this. 73 00:04:07,590 --> 00:04:09,750 Think of this next. 74 00:04:10,260 --> 00:04:15,090 We have this dense layer, which is once in twenty eight layers using an activation of rlU again. 75 00:04:15,090 --> 00:04:20,650 So we just take that and feed it into there, take that output and then fit it into the soft max layer 76 00:04:20,650 --> 00:04:22,770 here now in the loss of max layer. 77 00:04:23,010 --> 00:04:28,680 We need to specify the input size and the input size as a number of classes, which is 10, which is 78 00:04:28,680 --> 00:04:30,980 what we set that variable up here. 79 00:04:31,260 --> 00:04:31,540 10. 80 00:04:32,460 --> 00:04:33,960 Consider printed it out just to make sure. 81 00:04:34,710 --> 00:04:38,490 So that's what we need for the soft max layer right now. 82 00:04:39,180 --> 00:04:42,870 And then all we have to do after that is just to model that compile. 83 00:04:43,380 --> 00:04:49,920 This creates an object that stores a model that we just created that's not above the code right there. 84 00:04:51,570 --> 00:04:58,140 So what we do here in the model to compile, we can specify, just like in Potawatomi, specified all 85 00:04:58,170 --> 00:04:59,430 across an arbitrary and. 86 00:05:00,080 --> 00:05:06,440 Entropy, lost function and optimizer, we specify those things here in carrots, we have to attitude 87 00:05:06,440 --> 00:05:11,680 metric here, metric being accuracy, which is what we want to monitor during the training process. 88 00:05:11,940 --> 00:05:17,000 It's quite simple and you can put because it's a list, you can put multiple things here, like if you 89 00:05:17,000 --> 00:05:19,700 wanted to do loss as well, you could have done it like that. 90 00:05:20,570 --> 00:05:21,980 So let's take it out for now. 91 00:05:22,940 --> 00:05:26,570 And then at the end, this is a very good practice that I tend to do a lot. 92 00:05:26,930 --> 00:05:32,630 Print the model of summary and it gives you a breakdown of the little the output shape number of parameters 93 00:05:32,630 --> 00:05:32,900 in it. 94 00:05:33,260 --> 00:05:39,470 And if you remember from all the slides, these were the parameters we got got when we created this 95 00:05:39,470 --> 00:05:40,850 neural net to CNN. 96 00:05:42,140 --> 00:05:47,150 And it gives you the total trainable parameters, total parameters, non-tradable parameters which are 97 00:05:47,160 --> 00:05:48,230 zero in this case. 98 00:05:48,620 --> 00:05:49,520 So it's pretty cool. 99 00:05:49,670 --> 00:05:54,850 So we'll stop there in the next section, though we'll start training on model. 100 00:05:54,860 --> 00:06:00,980 But this here just wanted to show you this is the same could above here without all the comments, and 101 00:06:00,980 --> 00:06:05,150 you can see how short and easy it is to construct that fairly. 102 00:06:05,450 --> 00:06:10,580 I mean, it's not a complicated CNN, it's a basic CNN, but you can see how easy it was to construct 103 00:06:10,580 --> 00:06:14,600 at CNN using cameras right now and defined a loss and optimizer. 104 00:06:14,960 --> 00:06:19,630 So it's quite good, quite nice, and hope you took a lot from this lesson. 105 00:06:20,000 --> 00:06:23,570 In the next lesson, we'll take a look at how we start treating this model. 106 00:06:23,990 --> 00:06:25,280 So stay tuned for that.