Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,480 --> 00:00:07,560 Hello and welcome to this new tutorial in the previous oil we define the architecture of our discriminator. 2 00:00:07,560 --> 00:00:11,320 So now we are one step left to get our second brain. 3 00:00:11,400 --> 00:00:17,070 Before that we need to make the forward function we're going to make a new form function this time for 4 00:00:17,070 --> 00:00:19,140 the brain of the discriminator. 5 00:00:19,230 --> 00:00:25,070 This is going to be almost the same as the forward function we made for the generator. 6 00:00:25,200 --> 00:00:27,630 But there is going to be a slight difference. 7 00:00:27,810 --> 00:00:30,400 A slight trick not to forget to apply. 8 00:00:30,480 --> 00:00:32,300 We'll see that in this tutorial. 9 00:00:32,310 --> 00:00:38,220 All right so let's define a new function that we're going to call again forward. 10 00:00:38,280 --> 00:00:42,360 There is no danger to call it again forward. 11 00:00:42,660 --> 00:00:49,040 And this word function is going to take two arguments the same as before self to refer to the object. 12 00:00:49,050 --> 00:00:54,720 And because we're going to use the metal module main to form propagate the signal inside the neural 13 00:00:54,720 --> 00:01:02,020 network and the second argument is going to be the input of the discriminator neural network. 14 00:01:02,190 --> 00:01:08,790 So keep in mind and keep it well understood the input of the neural network of the discriminator is 15 00:01:08,790 --> 00:01:13,720 an image that is going to be one of the images created by the generator. 16 00:01:14,040 --> 00:01:19,320 All right so an input of three dimensions corresponding to the three channels. 17 00:01:19,560 --> 00:01:25,470 And let's also keep in mind that the output of the discriminator and therefore of this forward function 18 00:01:25,800 --> 00:01:32,910 is a discriminating number is going to be of value between 0 and 1 and that will do the discrimination 19 00:01:33,120 --> 00:01:39,030 according to which a number close to zero will reject the image and the number close to one will accept 20 00:01:39,180 --> 00:01:40,080 the image. 21 00:01:40,080 --> 00:01:41,480 So let's recap. 22 00:01:41,490 --> 00:01:47,300 The process is actually very easy to understand for discriminator the discriminator takes as inputs 23 00:01:47,610 --> 00:01:50,060 an image created by the generator. 24 00:01:50,220 --> 00:01:56,910 And for this image it will decide if it wants to accepted or rejected and to make that decision it will 25 00:01:56,910 --> 00:01:58,560 return the output. 26 00:01:58,560 --> 00:02:01,250 That is the discriminating number between 0 and 1. 27 00:02:01,320 --> 00:02:07,110 And if this output is close to zero it will reject it and if it is close to 1 it will accept it. 28 00:02:07,110 --> 00:02:13,490 So that's why in some way the discriminator is discriminating the creations of the generator. 29 00:02:13,560 --> 00:02:20,000 And so now we perfectly understand the name of what we're implementing the generative adversarial networks. 30 00:02:20,010 --> 00:02:26,750 Well that name is perfectly chosen because the discriminator is an adversary of the generator. 31 00:02:26,820 --> 00:02:33,030 It is like some kind of an adversary judge judging the creations of the generator judging whether they 32 00:02:33,030 --> 00:02:35,490 should be accepted or not. 33 00:02:35,490 --> 00:02:40,110 All right so now we understand clearly what's the input and what's the output. 34 00:02:40,110 --> 00:02:43,870 Let's go inside the function and let's define what we wanted to do. 35 00:02:44,040 --> 00:02:51,270 So the first thing that we need to do is get the output right the output that is returned by the main 36 00:02:51,330 --> 00:02:56,130 METAR module of our object which is referred by self. 37 00:02:56,130 --> 00:03:02,820 So I'm taking my objects self and then I'm taking the main metal module inside which of course I have 38 00:03:02,820 --> 00:03:05,340 to input Well the. 39 00:03:05,660 --> 00:03:09,990 All right the input image and image created by the generator. 40 00:03:10,400 --> 00:03:17,580 So that returns the output and in the end of course we must not forget to return the output because 41 00:03:17,580 --> 00:03:20,230 that's exactly the role of the forward function. 42 00:03:20,310 --> 00:03:26,760 Not only it propagates the signal inside the neural network but also and mostly it returns the output 43 00:03:26,930 --> 00:03:29,990 that is a discriminating value between 0 and 1. 44 00:03:30,060 --> 00:03:34,160 But here comes the little trick that we must not forget here. 45 00:03:34,290 --> 00:03:36,210 If you have guessed about it. 46 00:03:36,210 --> 00:03:37,500 Congratulations. 47 00:03:37,500 --> 00:03:39,880 It's actually slightly technical. 48 00:03:39,990 --> 00:03:46,350 This Trig has to do with the result of the convolutions if we have a better look at the architecture 49 00:03:46,350 --> 00:03:52,940 of the neural network of the discriminator we see that it's actually a sequence of convolutions. 50 00:03:53,130 --> 00:04:00,660 But if you remember how a CNN works that is a convolutional neural network composed of several convolutions 51 00:04:00,960 --> 00:04:02,780 that is exactly what we have right here. 52 00:04:02,940 --> 00:04:08,820 Well at the end of the CNN we need to flatten the result of all the convolutions that is the result 53 00:04:08,880 --> 00:04:12,710 of what we get after applying the last convolution. 54 00:04:12,720 --> 00:04:21,330 So the trick well actually the thing that we have to do now is exactly this flattening we have to flatten 55 00:04:21,510 --> 00:04:28,560 the result of the convolutions and we need to do this so that all the elements of the output are along 56 00:04:28,680 --> 00:04:30,300 one same dimension. 57 00:04:30,510 --> 00:04:35,320 And by the way this time engine corresponds to the dimension of the batch size. 58 00:04:35,380 --> 00:04:39,300 So now the question is how do we flatten using pine torch. 59 00:04:39,330 --> 00:04:42,490 Now that is the result of several convolutions. 60 00:04:42,660 --> 00:04:47,350 Well it's actually very easy once you know the trick we have to add here that. 61 00:04:47,520 --> 00:04:54,490 And then we need to use the View function to which we input minus one. 62 00:04:54,660 --> 00:05:01,560 This means nothing else than we want to flatten the result of the convolutions that at the end or in 63 00:05:01,580 --> 00:05:07,460 2D them mentions into one same dimension along one same flattened vector. 64 00:05:07,780 --> 00:05:08,480 All right. 65 00:05:08,650 --> 00:05:09,520 And that's done. 66 00:05:09,520 --> 00:05:11,360 We are done with the forward function. 67 00:05:11,500 --> 00:05:18,570 So congratulations you have made the architecture of the second brain of the deep convolutional Ganns. 68 00:05:18,820 --> 00:05:20,270 So that's quite a big deal. 69 00:05:20,410 --> 00:05:26,350 And now since the architecture is made we can create as many brains of discriminator as we want and 70 00:05:26,350 --> 00:05:28,780 we'll create one in the next to Tauriel. 71 00:05:28,780 --> 00:05:30,550 Until then enjoy computer vision. 7860