Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,670 --> 00:00:06,790 Hello and welcome to this new tutorial today we're going to start with the big first step of the implementation 2 00:00:06,790 --> 00:00:11,760 of our deep convolutional Ganns that consists of defining the generator. 3 00:00:11,950 --> 00:00:13,210 That's the big first step. 4 00:00:13,330 --> 00:00:18,700 And the second big step will be to define the discriminator then we'll have our brains and then we will 5 00:00:18,700 --> 00:00:21,650 train these brains to generate fake images. 6 00:00:21,790 --> 00:00:23,890 But today we're going to start with this big first step. 7 00:00:23,900 --> 00:00:30,790 Defining the generator and we'll do it in two steps steps first we will make the class that will define 8 00:00:30,790 --> 00:00:33,830 the architecture of the neural network of the generator. 9 00:00:33,850 --> 00:00:39,790 So we will put in a sequence to different modules that is the different layers of the neural network 10 00:00:40,150 --> 00:00:45,130 then we'll make the forward function to propagate the signal inside and will that work. 11 00:00:45,130 --> 00:00:51,280 And then once this class is done we'll be able to create the generator itself that is the neural network 12 00:00:51,400 --> 00:00:57,690 of the generator by simply creating an object of the class an instance of the class. 13 00:00:57,910 --> 00:01:00,980 So let's do this let's make this class. 14 00:01:01,120 --> 00:01:08,290 So to define a class in Python We start with class and then we give the name of the class which is going 15 00:01:08,290 --> 00:01:14,550 to be G for regenerator and then we'll call the class for the discriminator D. 16 00:01:15,010 --> 00:01:21,840 And now we're going to use a trick an object oriented programming language which is inheritance. 17 00:01:21,840 --> 00:01:28,660 We are going to inherit from in that module which contains all the tools that allow us to build a neural 18 00:01:28,660 --> 00:01:35,260 network and it's called an end or modules because the modules are the different applications and connections 19 00:01:35,260 --> 00:01:37,390 you can make inside a neural network. 20 00:01:37,450 --> 00:01:40,730 For example one module can be a convolution. 21 00:01:40,840 --> 00:01:43,420 Another module can be a full connection. 22 00:01:43,600 --> 00:01:49,450 Therefore the name and to apply the inheritance there is nothing more simple we just need to put as 23 00:01:49,510 --> 00:01:51,060 argument of this class G. 24 00:01:51,190 --> 00:01:56,220 The module we want to inherit from which is and that module. 25 00:01:56,680 --> 00:01:58,570 There we go this one. 26 00:01:59,410 --> 00:02:02,510 And in that module so that introduces a new class. 27 00:02:02,530 --> 00:02:07,300 And now inside this new class we have to define the whole architecture. 28 00:02:07,480 --> 00:02:12,230 And besides we'll include the forward function to be able to afford propagate a signal. 29 00:02:12,250 --> 00:02:18,640 That's the advantage of making a class A class is an advance structure inside of which you can include 30 00:02:18,640 --> 00:02:22,540 some properties of the future instances of the class that will be created. 31 00:02:22,540 --> 00:02:24,570 That is the future object. 32 00:02:24,580 --> 00:02:30,250 And also some different functions that will be like some tools of your object. 33 00:02:30,280 --> 00:02:35,490 Quick and the best example for those of you who are seeing your class for the first time. 34 00:02:35,530 --> 00:02:42,460 Well we could define a self-driving car with a class in one class we could include all the properties 35 00:02:42,550 --> 00:02:48,610 of the Sandrine car and then different functions for example to move forward function to turn the function 36 00:02:48,730 --> 00:02:51,700 to turn right function the stop function. 37 00:02:51,700 --> 00:02:55,670 All these functions that would allow to control the self-driving car. 38 00:02:55,810 --> 00:03:01,300 Well here that's exactly the same we're making in class to not only define the architecture of the neural 39 00:03:01,300 --> 00:03:07,120 network of the generator but also we'll add the forward function which is like a tool that we can apply 40 00:03:07,210 --> 00:03:08,410 on the neural network. 41 00:03:08,530 --> 00:03:13,340 And this tool will be used to propagate the signal inside the neural network. 42 00:03:13,690 --> 00:03:19,360 So let's do this let's first start by defining the architecture of the neural network. 43 00:03:19,360 --> 00:03:28,990 And as for every neural network that we make with a class we do this inside that in its function the 44 00:03:28,990 --> 00:03:32,320 innate function is the function you always start with a class. 45 00:03:32,350 --> 00:03:37,660 It is the function that defines the properties of the future objects that will be created from your 46 00:03:37,660 --> 00:03:38,660 class. 47 00:03:39,190 --> 00:03:44,810 And speaking of this object that's exactly the argument that we have to employ right now. 48 00:03:44,860 --> 00:03:51,610 And this is called self self has been a mystery for a lot of people but there is absolutely no mystery 49 00:03:51,610 --> 00:03:52,270 about it. 50 00:03:52,270 --> 00:03:56,280 It just refers to the future object that will be created. 51 00:03:56,350 --> 00:04:02,500 It's just a common way to specify that when we create a variable we use the self to specify that this 52 00:04:02,500 --> 00:04:04,920 variable belongs to the object. 53 00:04:04,930 --> 00:04:06,870 So each time I put self. 54 00:04:06,910 --> 00:04:08,850 That means I am referring to the object. 55 00:04:08,890 --> 00:04:13,070 That means I'm taking the object and setting some properties to it. 56 00:04:13,680 --> 00:04:14,910 So self. 57 00:04:15,010 --> 00:04:22,780 And now we are arguing to activate the inheritance because we import the n and module to inherit from 58 00:04:22,780 --> 00:04:23,130 it. 59 00:04:23,260 --> 00:04:31,210 But now we have to activate this inheritance and to do this we use the super function inside of which 60 00:04:31,420 --> 00:04:37,410 we input our Class G which is inheriting from the end in that module right here. 61 00:04:37,660 --> 00:04:45,160 And then the second argument is self because we will use the tools of that module on our object because 62 00:04:45,160 --> 00:04:49,360 our object is nothing else than the new one that work at the generator and the neural network and the 63 00:04:49,360 --> 00:04:52,630 generator will be composed of the different modules. 64 00:04:52,780 --> 00:04:57,760 And in that module the conclusions the full connections and more. 65 00:04:57,760 --> 00:05:03,990 All right and then that's not over we just need to add a that and underscore underscore in it double 66 00:05:04,010 --> 00:05:06,820 underscore again and parenthesis. 67 00:05:06,840 --> 00:05:07,590 All right. 68 00:05:07,670 --> 00:05:10,570 Nothing more important to know about this. 69 00:05:10,610 --> 00:05:17,250 That's just what we have to type here to activate the inheritance of the end and that much our next 70 00:05:17,250 --> 00:05:19,060 step. 71 00:05:19,070 --> 00:05:23,870 Now the next step is to make a meta module what is a metal module. 72 00:05:23,870 --> 00:05:28,540 That's a huge module that will be composed of several modules. 73 00:05:28,670 --> 00:05:34,580 I remind that by module I mean the different layers the different connections inside the neural network. 74 00:05:34,580 --> 00:05:41,150 So now we'll create a big metal module which will be a sequence of several modules some convolutions 75 00:05:41,360 --> 00:05:46,390 full connections inversed convolutions and more integrate this metal module. 76 00:05:46,430 --> 00:05:53,300 Well we're going to take our object because this metal module is going to be a property of the object 77 00:05:53,750 --> 00:06:00,080 and therefore I'm taking my object and this metal module of the object will be called Maine. 78 00:06:00,750 --> 00:06:07,050 Main will be the metal module that will contain all the different modules in a sequence of layers. 79 00:06:07,070 --> 00:06:08,870 And speaking of sequence of layers. 80 00:06:08,960 --> 00:06:17,000 That's exactly what we'll introduce right now with the N-N that sequence show class inside of which 81 00:06:17,150 --> 00:06:20,360 will create the different modules. 82 00:06:20,360 --> 00:06:20,990 All right. 83 00:06:21,020 --> 00:06:26,870 So you probably recognize that sequential is actually a class so self that main is actually a new object 84 00:06:26,960 --> 00:06:33,080 of the sequential class but of course this object represents this method module composed of the different 85 00:06:33,080 --> 00:06:34,100 modules. 86 00:06:34,100 --> 00:06:40,460 All right so now we're starting to define the architecture of the neural network and let's start with 87 00:06:40,460 --> 00:06:41,450 the first module. 88 00:06:41,750 --> 00:06:47,690 According to you according to the intuition lectures you saw with curial what is going to be this first 89 00:06:47,690 --> 00:06:48,520 module. 90 00:06:48,890 --> 00:06:54,380 Well of course that's going to be a non-first convolution not a convolution. 91 00:06:54,440 --> 00:06:58,030 And in verse 1 that is exactly the inverse of a convolution. 92 00:06:58,250 --> 00:07:01,620 And why do we have to start with an inverse convolution. 93 00:07:01,820 --> 00:07:06,480 Well that's because the role of the generator is to generate some fake images. 94 00:07:06,680 --> 00:07:13,730 And therefore since CNN takes as input some images and returns as output a vector. 95 00:07:13,820 --> 00:07:16,940 Well and then CNN will do exactly the opposite. 96 00:07:17,030 --> 00:07:23,240 It will take as inputs a vector we will create that vector we'll call it noise later on in the training 97 00:07:23,570 --> 00:07:29,410 that will be the input of this inverse CNN and it will return a fake image. 98 00:07:29,570 --> 00:07:36,740 So you're going to see that the input of this inverse CNN that we're about to make will be a random 99 00:07:36,740 --> 00:07:39,380 vector of size 100. 100 00:07:39,830 --> 00:07:46,760 And to get the inverse of a convolution in PI torch we're going to use the can transpose to the class 101 00:07:47,030 --> 00:07:52,610 that is that this inverse convolution module that we're going to create will be an object of the current 102 00:07:52,820 --> 00:07:54,600 transpose to the class. 103 00:07:54,800 --> 00:07:55,850 So let's do it. 104 00:07:55,970 --> 00:08:02,210 We take our end in that module because this can transpose to the class is taken from the end module 105 00:08:02,610 --> 00:08:08,670 so end then dot com capitals C transpose to the. 106 00:08:08,670 --> 00:08:10,350 Here is this one. 107 00:08:10,350 --> 00:08:12,080 That's the inverse convolution. 108 00:08:12,150 --> 00:08:18,510 The first module of our big Maira module that is the first module of our neural network can be transposed 109 00:08:18,510 --> 00:08:19,030 to D. 110 00:08:19,050 --> 00:08:21,870 And now we need to input several arguments. 111 00:08:21,870 --> 00:08:23,670 Get ready for them as you can see. 112 00:08:23,670 --> 00:08:25,560 There are many of them. 113 00:08:25,590 --> 00:08:25,890 All right. 114 00:08:25,890 --> 00:08:28,110 So the first argument is 100. 115 00:08:28,260 --> 00:08:36,850 That's the size of the input 100 meaning that the inputs are inverse CNN will be a vector of size 100. 116 00:08:37,020 --> 00:08:39,840 Then the second argument is 512. 117 00:08:39,900 --> 00:08:43,170 And that's the number of feature maps of the output. 118 00:08:43,230 --> 00:08:45,400 Then the third argument is four. 119 00:08:45,540 --> 00:08:50,790 And that's the size of the kernel which means that the kernels will be squares of size four by four 120 00:08:51,330 --> 00:08:53,100 then the next argument is one. 121 00:08:53,190 --> 00:08:59,190 And that's the stride and the last argument is zero and that's the padding. 122 00:08:59,190 --> 00:09:03,440 By the way the architecture that I'm about to define right now I didn't invent it. 123 00:09:03,510 --> 00:09:09,450 It is coming from a lot of experimenting that was done by researchers and the machine learning and AI 124 00:09:09,480 --> 00:09:10,510 contributors. 125 00:09:10,530 --> 00:09:12,500 So I did not invent it. 126 00:09:12,540 --> 00:09:18,840 I just took an architecture that is open source and that turns out to work very well for adversarial 127 00:09:18,840 --> 00:09:19,560 networks. 128 00:09:19,680 --> 00:09:21,820 So don't worry about the choice of the numbers. 129 00:09:21,870 --> 00:09:24,560 They just come from experimentation. 130 00:09:24,610 --> 00:09:30,090 All right and then we have one last argument and I'm specifying it because it is actually about the 131 00:09:30,090 --> 00:09:32,740 bias which by default is equal to true. 132 00:09:32,880 --> 00:09:36,330 But I'm going to set it equal to false because we don't want any bias. 133 00:09:36,330 --> 00:09:37,970 It works better this way. 134 00:09:38,250 --> 00:09:43,060 And therefore I'm saying the bias here to false. 135 00:09:43,110 --> 00:09:49,770 So one hundred for the size of the input 512 feature maps and the output meaning the output of the inverse 136 00:09:49,770 --> 00:09:52,170 convolution not the final output. 137 00:09:52,170 --> 00:09:59,560 Then a kernel of size for four by four is stride of 1 a batting of zero and no buyers. 138 00:09:59,580 --> 00:10:02,810 That's the first Mudgal of our neural network. 139 00:10:02,820 --> 00:10:03,920 Congratulations. 140 00:10:03,960 --> 00:10:13,030 You just applied an inversed convolution all right and now we are going to normalize all the features 141 00:10:13,150 --> 00:10:17,900 along the dimension of the Bachche the diamonds and of the batch is 512. 142 00:10:18,010 --> 00:10:26,410 We have 512 feature maps and we're going to Bachan on each of these 512 feature maps and to do this 143 00:10:26,410 --> 00:10:33,210 we take again our and then Mudgal and then we apply the batch we should find in here. 144 00:10:33,210 --> 00:10:40,090 Here we go Bachan on 2D and as arguments we need to input the number of feature maps we want to bet 145 00:10:40,210 --> 00:10:40,560 on. 146 00:10:40,570 --> 00:10:47,490 And since the 512 feature maps Well we are going to bet on 512 feature maps. 147 00:10:47,610 --> 00:10:48,240 All right. 148 00:10:48,270 --> 00:10:52,100 So playing batched normalization is pretty simple. 149 00:10:52,120 --> 00:10:59,200 Then we are going to apply Ereli to rectification for the non-linearity of the neural network that is 150 00:10:59,200 --> 00:11:02,720 to break the linearity and to do this. 151 00:11:02,770 --> 00:11:10,500 Nothing more simple things to our end and module we take are in a module that and we apply the review 152 00:11:11,320 --> 00:11:18,360 rectifier activation function and as arguments we just need to input just like that. 153 00:11:18,390 --> 00:11:26,750 That applies really to rectification then we are going to apply another inverse convolution. 154 00:11:27,180 --> 00:11:33,280 So I'm simply going to take my end and can transpose to the module. 155 00:11:33,360 --> 00:11:34,390 Pasting it here. 156 00:11:34,440 --> 00:11:38,150 And then here we go with the series of arguments. 157 00:11:38,190 --> 00:11:45,270 So this time we don't have 100 for the size of input because the input is now the output of the previous 158 00:11:45,450 --> 00:11:53,400 module which was an inverse convolution but in the output we got 512 feature maps and these feature 159 00:11:53,400 --> 00:11:58,050 maps become the new input of our new inversed conclusion. 160 00:11:58,050 --> 00:12:02,430 So this time we don't input 100 year when put 512. 161 00:12:03,000 --> 00:12:06,680 That's the number of inputs of this new inverse convolution. 162 00:12:06,900 --> 00:12:12,860 Then we need to choose a new number feature maps in the output of this new inverse convolution. 163 00:12:13,020 --> 00:12:15,630 And this time we're not going to pick five phonons well. 164 00:12:15,630 --> 00:12:19,020 We're going to pick two hundred and fifty six. 165 00:12:19,020 --> 00:12:23,140 Again that's a choice motivated by experimentation. 166 00:12:23,190 --> 00:12:27,760 That's really not his job to figure out such an architecture. 167 00:12:27,780 --> 00:12:35,880 All right then we're going to choose a kernel size of four four by four then as I tried to this time 168 00:12:36,150 --> 00:12:46,520 and padding of one and then same we don't want any buyers and therefore adding Byars equals false perfect. 169 00:12:46,590 --> 00:12:48,500 That's our next module. 170 00:12:48,510 --> 00:12:51,000 We now have two inversed convolution. 171 00:12:51,060 --> 00:12:53,990 Our generator is getting into shape. 172 00:12:54,150 --> 00:13:00,390 So now we're going to do the same we're going to batched norm each of the new 256 feature maps. 173 00:13:00,390 --> 00:13:12,390 So I'm just going to take this and that batch Enorme 2D it here but then replace 512 by 256 because 174 00:13:12,390 --> 00:13:21,990 we're batch norming each of the new 256 feature maps and then we apply another rectification which is 175 00:13:21,990 --> 00:13:23,610 going to be the loop again. 176 00:13:23,640 --> 00:13:29,230 So here let's just copy this and paste it here. 177 00:13:30,490 --> 00:13:34,100 All right then guess what we're going to do again. 178 00:13:34,240 --> 00:13:37,800 We're going to apply another inverse convolution. 179 00:13:37,900 --> 00:13:40,650 So I am taking this. 180 00:13:40,840 --> 00:13:44,260 And then that can transpose to D. 181 00:13:44,290 --> 00:13:45,740 I am pasting it here. 182 00:13:46,000 --> 00:13:53,410 And then the input is actually the output of the previous operation that is the inverse convolution. 183 00:13:53,590 --> 00:14:02,760 So that's 256 the 256 outputs feature maps become the input of our new inversed convolution. 184 00:14:02,810 --> 00:14:04,550 So 256. 185 00:14:04,720 --> 00:14:07,850 And now guess what we want in the outputs. 186 00:14:08,170 --> 00:14:11,760 Well you probably understood the logic we wanted divided by two. 187 00:14:11,800 --> 00:14:16,540 So we're going to take one hundred and twenty eight output future maps. 188 00:14:16,540 --> 00:14:23,650 In this new inverse convolution and then the good news is that we're keeping a size four by four for 189 00:14:23,650 --> 00:14:24,500 the kernels. 190 00:14:24,640 --> 00:14:28,580 Let's try to add a padding of one and then no bias. 191 00:14:28,690 --> 00:14:29,740 Congratulations. 192 00:14:29,800 --> 00:14:32,870 You just applied your third inverse convolution. 193 00:14:33,220 --> 00:14:38,380 And so then that's the same we're going to apply a batch known to the two batched norm. 194 00:14:38,380 --> 00:14:43,510 Each of the new 128 feature maps. 195 00:14:43,510 --> 00:14:53,980 So here I am replacing tournament 56 by 128 and we are going to apply reglue rectification to make sure 196 00:14:53,980 --> 00:15:00,150 we break the linearity for the non-linearity of the neural network. 197 00:15:00,160 --> 00:15:00,580 All right. 198 00:15:00,580 --> 00:15:02,670 And then guess what we're going to do again. 199 00:15:02,770 --> 00:15:07,520 We are going to apply an inversed convolution again. 200 00:15:07,720 --> 00:15:16,450 So I'm taking this again I'm taking the cones transpose to 2D and I'm pasting in here and this time 201 00:15:16,480 --> 00:15:20,010 we have as input of this new inverse convolution. 202 00:15:20,070 --> 00:15:26,930 One hundred and twenty eight feature maps because that was the output of the previous inverse convolution. 203 00:15:27,130 --> 00:15:32,880 And this time guess what we want for the number of the output feature maps of this new inverse convolution. 204 00:15:33,130 --> 00:15:40,960 Well of course we want 64 new feature maps and then good news again we're keeping a size of four by 205 00:15:40,960 --> 00:15:44,280 four for the Colonel's a stride of two and a betting of one. 206 00:15:44,470 --> 00:15:47,140 And then nobody has a right. 207 00:15:47,140 --> 00:15:49,010 We're getting close to the end. 208 00:15:49,270 --> 00:15:52,140 We need again to apply a batch norm. 209 00:15:52,150 --> 00:15:53,900 So let's copy this again. 210 00:15:54,220 --> 00:16:04,210 We are going to batched norm each of the new 64 feature maps that are the output of our new inverse 211 00:16:04,210 --> 00:16:05,410 convolution. 212 00:16:05,410 --> 00:16:12,850 And then we are going to apply a really rectification to again break the linearity. 213 00:16:13,360 --> 00:16:16,180 Right there we go almost over. 214 00:16:16,180 --> 00:16:17,580 Don't get crazy too fast. 215 00:16:17,590 --> 00:16:21,410 We have one final inverse convolution to make. 216 00:16:21,430 --> 00:16:23,440 So that's exactly what we're going to do. 217 00:16:23,650 --> 00:16:32,140 We are going to apply our last module the inverse convolution and I'm just copying the name of the module 218 00:16:32,140 --> 00:16:36,200 because this time we are going to input different arguments. 219 00:16:36,220 --> 00:16:41,420 So the first argument is the number of inputs which is the output of the previous inverse convolution. 220 00:16:41,470 --> 00:16:43,160 So that's 64 again. 221 00:16:43,780 --> 00:16:49,630 And then for the output That's important since we're making the generator that is supposed to generate 222 00:16:49,630 --> 00:16:51,160 some fake images. 223 00:16:51,340 --> 00:16:56,530 And since these fake images are going to be with three channels while the output of the generator is 224 00:16:56,530 --> 00:17:02,500 exactly going to be the three channels of the fake images and therefore the number of dimensions for 225 00:17:02,500 --> 00:17:07,740 the output is going to be three corresponding to the three channels. 226 00:17:07,750 --> 00:17:08,150 All right. 227 00:17:08,170 --> 00:17:10,180 And then actually that's the same. 228 00:17:10,180 --> 00:17:14,700 We're going to use kernels of size four by four a stride of two the padding of one. 229 00:17:14,800 --> 00:17:16,250 And again no bias. 230 00:17:16,300 --> 00:17:23,130 So I can just copy this and paste it here and there we go. 231 00:17:23,140 --> 00:17:28,130 We have applied all our inversed convolution that's almost over. 232 00:17:28,150 --> 00:17:34,990 We just need to end this up with a hyperbolic tangent rectification to not only break the law in the 233 00:17:34,990 --> 00:17:40,120 area again but also to make sure we are between minus 1 and plus 1. 234 00:17:40,120 --> 00:17:44,720 Have a look at the age function you'll see that the values are between minus 1 and plus 1. 235 00:17:44,890 --> 00:17:47,620 And also it is centered around 0. 236 00:17:47,650 --> 00:17:53,420 And why do we want to output values to fall between minus 1 and plus 1 and being centered around 0. 237 00:17:53,560 --> 00:17:57,660 That's because we want the same standards as the images of the data set. 238 00:17:57,850 --> 00:18:03,550 And because the created images of the generator will become the input of the discriminator that will 239 00:18:03,550 --> 00:18:04,780 come after that. 240 00:18:04,780 --> 00:18:12,640 All right so now we just need to and this architecture with our n module again and then that we apply 241 00:18:13,180 --> 00:18:20,320 the term age to hyperbolic tangent rectification for the non-linearity of the neural network and to 242 00:18:20,320 --> 00:18:23,930 be between minus 1 and plus 1 centered around zero. 243 00:18:24,010 --> 00:18:24,420 Whew. 244 00:18:24,430 --> 00:18:25,800 Congratulations. 245 00:18:25,810 --> 00:18:28,000 You made a neural network. 246 00:18:28,000 --> 00:18:30,460 That was definitely not a piece of cake. 247 00:18:30,880 --> 00:18:37,150 So I guess now it would be nice to have a break in the next story will make the food function that will 248 00:18:37,150 --> 00:18:45,070 propagate the signal that is the input of the generator through all the different layers of this new 249 00:18:45,070 --> 00:18:46,270 will that work. 250 00:18:46,270 --> 00:18:49,290 So have a good break and I'll see you in the next tutorial. 251 00:18:49,360 --> 00:18:51,370 Until then enjoy the revision. 26621