WEBVTT 00:00.240 --> 00:02.670 Instructor: Hello and welcome to this Python tutorial. 00:02.670 --> 00:04.890 Now we have to define the five variables 00:04.890 --> 00:06.240 of this init function. 00:06.240 --> 00:08.040 That is the three convolutions, 00:08.040 --> 00:09.513 and the two full connections. 00:10.636 --> 00:11.469 So let's start with the first one. 00:11.469 --> 00:16.080 Convolution one applies convolution to the input images. 00:16.080 --> 00:18.210 So that's the original images. 00:18.210 --> 00:20.905 And now you're gonna see how everything will become 00:20.905 --> 00:22.740 so simple to create this convolution. 00:22.740 --> 00:25.590 Well, what we have to do is actually create an object 00:25.590 --> 00:27.030 of some specific class. 00:27.030 --> 00:30.240 And this class is taken from NN, 00:30.240 --> 00:32.523 and then the class is COM-2D. 00:34.560 --> 00:37.800 2D because we're working with 2D images. 00:37.800 --> 00:40.890 And now as you can see, we need to input several arguments. 00:40.890 --> 00:45.890 First one is in channels, so let's input it in channels. 00:45.960 --> 00:49.320 The second one is out channels. 00:49.320 --> 00:52.650 The third one is kernel size. 00:52.650 --> 00:55.140 And the rest of them are stride, the betting, 00:55.140 --> 00:57.060 the dilatation groups, and bias. 00:57.060 --> 00:59.220 And we have default values for all these ones. 00:59.220 --> 01:00.510 So we're not going to input them. 01:00.510 --> 01:02.460 We're gonna keep the default values. 01:02.460 --> 01:05.018 But what's important is these three arguments in channels, 01:05.018 --> 01:06.777 out channels, and kernel size. 01:06.777 --> 01:09.870 And so do you guess what they correspond to? 01:09.870 --> 01:12.030 Well, very simply, in channels correspond 01:12.030 --> 01:14.400 to the input of the convolution, 01:14.400 --> 01:16.650 and out channels correspond to the output 01:16.650 --> 01:17.910 of the convolution. 01:17.910 --> 01:20.070 So in channels, what is it going to be? 01:20.070 --> 01:22.500 Well, very simply, that's going to be the number 01:22.500 --> 01:24.570 of channels in our images. 01:24.570 --> 01:26.310 And actually we are gonna work 01:26.310 --> 01:28.860 with black and white images because basically 01:28.860 --> 01:31.410 we don't need colors to recognize the monsters. 01:31.410 --> 01:32.880 The AI is totally capable 01:32.880 --> 01:35.430 of recognizing the monsters in black and white. 01:35.430 --> 01:36.930 So we don't need the colors, 01:36.930 --> 01:39.180 it'll just recognize them by their shape. 01:39.180 --> 01:41.700 And therefore we're gonna use one channel. 01:41.700 --> 01:44.220 So one channel is when you have black and white images 01:44.220 --> 01:46.890 and three channels is when you have colored images. 01:46.890 --> 01:48.360 And therefore, since we're working with 01:48.360 --> 01:50.340 black and white images in channels, 01:50.340 --> 01:51.783 is going to be equal to one. 01:52.770 --> 01:54.210 Then out channels. 01:54.210 --> 01:57.208 So out channels is going to be equal to the images 01:57.208 --> 02:00.330 you want to have in the convolutional layer, 02:00.330 --> 02:02.940 which is the output of this convolution one. 02:02.940 --> 02:05.820 And so basically this is equal to the number 02:05.820 --> 02:09.210 of features you want to detect in your original images, 02:09.210 --> 02:12.150 because we will create one image per feature 02:12.150 --> 02:15.090 we want to detect because basically you know how it works. 02:15.090 --> 02:18.690 We apply one feature detector to the input image to 02:18.690 --> 02:21.240 detect a specific feature in the input image, 02:21.240 --> 02:23.070 and therefore the number of outputs images 02:23.070 --> 02:25.988 here is the number of features we want to detect. 02:25.988 --> 02:28.830 So now the question is how many features 02:28.830 --> 02:30.210 do we want to detect? 02:30.210 --> 02:32.460 Well, a common practice is to start 02:32.460 --> 02:35.040 with 32 feature detectors. 02:35.040 --> 02:39.660 And so that will lead us to 32 processed images 02:39.660 --> 02:41.460 in this first convolutional layer. 02:41.460 --> 02:45.537 So the input is one black and white image, a real image 02:45.537 --> 02:46.920 and the output. 02:46.920 --> 02:51.450 And the first convolution layer is 32 processed images. 02:51.450 --> 02:53.370 And by processed, I mean of course, 02:53.370 --> 02:55.080 that the convolution was applied 02:55.080 --> 02:58.230 to the input image to get 32 new images 02:58.230 --> 03:00.120 with detected features. 03:00.120 --> 03:03.360 And then we need to specify a kernel size, 03:03.360 --> 03:04.680 which is nothing else than 03:04.680 --> 03:06.660 the dimensions of this square 03:06.660 --> 03:09.600 that will go through the original image. 03:09.600 --> 03:12.180 And in common practice we use either 03:12.180 --> 03:15.660 two by two, or three by three, or five by five. 03:15.660 --> 03:17.137 And for the first one 03:17.137 --> 03:21.090 we're gonna use a five by five feature detector. 03:21.090 --> 03:22.800 That is a feature detector that will have 03:22.800 --> 03:24.780 five by five dimensions. 03:24.780 --> 03:26.460 And then we will reduce the size 03:26.460 --> 03:29.310 of this kernel for the next convolution layers. 03:29.310 --> 03:30.750 And speaking of it, this is 03:30.750 --> 03:32.370 exactly what we're gonna do now. 03:32.370 --> 03:34.556 We are going to copy this to 03:34.556 --> 03:37.253 define the second convolution, 03:37.253 --> 03:40.140 and therefore I'm facing that here, 03:40.140 --> 03:42.180 and now it's very funny and very easy. 03:42.180 --> 03:43.470 It's like a domino. 03:43.470 --> 03:44.370 The input channel 03:44.370 --> 03:47.910 of the second convolution layer is the output channel 03:47.910 --> 03:49.920 of the first convolution layer. 03:49.920 --> 03:52.080 So this number of outputs 32 here, 03:52.080 --> 03:55.440 is the same number of inputs, 32 here. 03:55.440 --> 03:57.660 And that's because we have 32 images 03:57.660 --> 04:01.410 in the input convolution layer of the second convolution. 04:01.410 --> 04:04.170 And so the second convolution is applied 04:04.170 --> 04:07.620 to this second convolutional layer, 04:07.620 --> 04:10.500 to return a third convolutional layer. 04:10.500 --> 04:13.410 And so now the question is how many new images do we want? 04:13.410 --> 04:15.990 Well same, let's create 32 new images. 04:15.990 --> 04:18.360 32 is actually a very common number 04:18.360 --> 04:19.830 in convolutional neural networks. 04:19.830 --> 04:21.540 If you look at the architectures 04:21.540 --> 04:23.970 you will find 32 in many of them. 04:23.970 --> 04:26.160 And then for the kernel size, well 04:26.160 --> 04:28.050 we need to reduce the kernel size. 04:28.050 --> 04:30.750 That is the dimensions of our feature detector. 04:30.750 --> 04:32.817 And so now we're gonna go from five to 04:32.817 --> 04:35.910 either four or even three, 04:35.910 --> 04:37.890 and then we'll go even smaller. 04:37.890 --> 04:40.800 All right, so our second convolution is ready. 04:40.800 --> 04:43.860 It takes as inputs, 32 processed images, 04:43.860 --> 04:46.770 each one detecting a first feature 04:46.770 --> 04:48.600 of the original input image 04:48.600 --> 04:51.240 and it creates 32 new images, 04:51.240 --> 04:55.110 thanks to this reduced dimensions of the feature detector. 04:55.110 --> 04:57.300 And so now let's push this even more. 04:57.300 --> 04:59.007 So I'm copying this, 04:59.007 --> 05:01.110 and pasting that here to 05:01.110 --> 05:05.460 create a third convolution to detect some features. 05:05.460 --> 05:08.010 And so now that's the same, the input channels 05:08.010 --> 05:11.340 here is the number of input images at the left 05:11.340 --> 05:13.710 of the convolution connection, and that is the number 05:13.710 --> 05:15.719 of processed images that was at the right 05:15.719 --> 05:18.314 of the previous convolution connection. 05:18.314 --> 05:20.040 So that's 32, therefore we keep 32 here. 05:20.040 --> 05:21.180 That's perfect. 05:21.180 --> 05:23.910 And now the question is again, how many new images 05:23.910 --> 05:25.290 do we want to detect? 05:25.290 --> 05:27.567 We are gonna take now 64, 05:27.567 --> 05:31.230 and therefore 64 output processed images. 05:31.230 --> 05:34.830 And of course now we take a smaller kernel size, 05:34.830 --> 05:36.630 and we're gonna take two. 05:36.630 --> 05:39.510 And so that's a very classic architecture 05:39.510 --> 05:42.480 of a convolutional layer, and it's very efficient to 05:42.480 --> 05:45.856 have a high level of feature detection inside images. 05:45.856 --> 05:48.930 All right, and so, now that we have our 05:48.930 --> 05:51.960 three convolution layers, thanks to our three 05:51.960 --> 05:53.610 convolution connections here, 05:53.610 --> 05:56.790 well now it's time to get our two full connections that 05:56.790 --> 06:00.300 I remind will take this huge vector that we obtain 06:00.300 --> 06:05.010 after flattening all the 64 times 32, times 32 again 06:05.010 --> 06:08.130 images that we got from all these convolutions. 06:08.130 --> 06:10.890 So we flatten all the pixels of these images 06:10.890 --> 06:14.190 and we gain one huge vector that will become the input 06:14.190 --> 06:16.800 of a new, fully connected neural network. 06:16.800 --> 06:19.320 And so that's when we have to make these full connections 06:19.320 --> 06:22.440 between first this huge vector and a hidden layer, 06:22.440 --> 06:25.620 and then a second full connection between the hidden layer 06:25.620 --> 06:28.320 and the output layer composed of the output neurons 06:28.320 --> 06:31.950 each one corresponding to a Q value of the possible actions. 06:31.950 --> 06:33.930 So let's make these two full connections. 06:33.930 --> 06:35.190 You know how to do that. 06:35.190 --> 06:37.530 That's exactly what we did for the self-driving car. 06:37.530 --> 06:38.970 So let's do that again. 06:38.970 --> 06:41.472 Well, first we take our NN module, 06:41.472 --> 06:45.660 then we take the linear class because again 06:45.660 --> 06:47.730 the full connection we create is an object 06:47.730 --> 06:49.260 of the linear class. 06:49.260 --> 06:51.960 And then in parenthesis, well, that's the same. 06:51.960 --> 06:54.180 First, we input the input features, 06:54.180 --> 06:58.830 that is the number of them, then the output features. 06:58.830 --> 07:02.100 And so the input features for the first full connection 07:02.100 --> 07:03.330 where is it going to be? 07:03.330 --> 07:05.520 Well, that's going to be equal to the number 07:05.520 --> 07:09.390 of pixels there are in this huge vector obtained 07:09.390 --> 07:11.940 after flattening all the processed images 07:11.940 --> 07:13.830 after the three convolutions. 07:13.830 --> 07:15.180 And so what is this number? 07:15.180 --> 07:17.340 Well, actually, there is a trick here. 07:17.340 --> 07:19.650 This number is actually hard to get. 07:19.650 --> 07:22.920 We actually need to make a function to compute that number. 07:22.920 --> 07:25.560 We don't have a variable that will get us this number. 07:25.560 --> 07:26.940 We have to compute it. 07:26.940 --> 07:29.100 And therefore, what we're gonna do now, 07:29.100 --> 07:31.230 and now it's very important to understand 07:31.230 --> 07:34.020 the mindset of programming that we must have 07:34.020 --> 07:36.960 and trying to bring to you the mindset that is 07:36.960 --> 07:38.850 what you must be thinking right now 07:38.850 --> 07:41.130 to do to overcome this obstacle. 07:41.130 --> 07:43.170 Because the first time you might say, hey 07:43.170 --> 07:45.720 I don't have this number of neurons in the flatten vector. 07:45.720 --> 07:46.620 What should I do? 07:46.620 --> 07:47.760 I'm stuck here. 07:47.760 --> 07:49.080 Well, no, actually 07:49.080 --> 07:52.950 because what you can do now is simply input any 07:52.950 --> 07:56.250 name here that will represent this number of neurons. 07:56.250 --> 07:59.610 So I'm calling it number neurons, number of neurons. 07:59.610 --> 08:02.551 And then we will simply make a function that will return 08:02.551 --> 08:05.160 in this number of neurons variable 08:05.160 --> 08:07.320 this number of pixels we're looking for. 08:07.320 --> 08:08.880 So we can totally do that. 08:08.880 --> 08:10.620 We can totally put this variable. 08:10.620 --> 08:13.020 Well, of course we'll get a warning 08:13.020 --> 08:14.160 because it doesn't exist yet, 08:14.160 --> 08:17.310 but we will create it afterwards with a function 08:17.310 --> 08:19.470 and we are totally allowed to do that even 08:19.470 --> 08:21.120 if the function comes afterwards. 08:21.120 --> 08:23.370 So that's a typical programming thinking 08:23.370 --> 08:25.980 you must have when you get that kind of obstacle. 08:25.980 --> 08:28.863 Well you can make a function to get what you're missing. 08:29.730 --> 08:31.890 All right, and then out features, 08:31.890 --> 08:34.170 and out features that the number of neurons 08:34.170 --> 08:37.830 in the hidden layer and that this time is up to you. 08:37.830 --> 08:39.270 That depends on the architecture 08:39.270 --> 08:41.220 of the neural network you want to create. 08:41.220 --> 08:44.220 And so a good number would be not a too small number. 08:44.220 --> 08:46.980 So for example, 40 neurons might be fine. 08:46.980 --> 08:48.750 We can try to increase it. 08:48.750 --> 08:51.300 If the training is not too slow, you can try to increase it. 08:51.300 --> 08:53.760 Maybe that will improve the predictions. 08:53.760 --> 08:54.900 But let's start with 40, 08:54.900 --> 08:57.120 maybe we'll increase that afterwards. 08:57.120 --> 09:00.060 All right, so that's it for the first full connection. 09:00.060 --> 09:02.283 Then we'll copy this, 09:03.300 --> 09:05.370 paste that here for the second full connection. 09:05.370 --> 09:06.203 That is the connection 09:06.203 --> 09:09.330 between the hidden layer and the output layer. 09:09.330 --> 09:13.050 And so the in features here becomes the out features 09:13.050 --> 09:15.750 of the previous layer, and that is 40. 09:15.750 --> 09:18.150 So here we input 40. 09:18.150 --> 09:20.850 That's of course the number of neurons in the hidden layer. 09:20.850 --> 09:24.060 And our features here is going to be equal to the number 09:24.060 --> 09:27.300 of output neurons there should be in our neural network. 09:27.300 --> 09:29.490 And since each output neuron corresponds 09:29.490 --> 09:30.780 to one Q value, 09:30.780 --> 09:33.180 and one Q value corresponds to one action, 09:33.180 --> 09:35.640 well the number of alpha neurons here is of course 09:35.640 --> 09:36.930 the number of actions. 09:36.930 --> 09:39.990 And we have one variable for this, which is number actions. 09:39.990 --> 09:44.813 And therefore here we input number actions. 09:44.813 --> 09:46.080 And there we go. 09:46.080 --> 09:47.370 Congratulations. 09:47.370 --> 09:49.410 We defined the architecture 09:49.410 --> 09:50.820 of our neural network. 09:50.820 --> 09:52.830 Our neural network is composed 09:52.830 --> 09:56.100 of three convolutional layers and one hidden layer. 09:56.100 --> 09:58.680 All this in one big CNN. 09:58.680 --> 10:01.380 And this CNN will detect the features in the game 10:01.380 --> 10:03.960 so that the AI will know what it has to do 10:03.960 --> 10:06.870 where it has to go, and where it needs to shoot. 10:06.870 --> 10:08.310 So there we go for this step, 10:08.310 --> 10:10.650 that's a first very important step done. 10:10.650 --> 10:13.290 Now we're gonna move on the next step, which is of course 10:13.290 --> 10:16.980 to get this number of neurons that is still missing. 10:16.980 --> 10:18.900 That's actually why we have the warning here. 10:18.900 --> 10:21.930 And define name number neurons, but no worries. 10:21.930 --> 10:24.600 Now we will make a function that will return the number 10:24.600 --> 10:27.450 of neurons in this huge vector, and we will put that number 10:27.450 --> 10:30.240 in a variable that we'll call number neurons. 10:30.240 --> 10:32.070 So let's do this in the next tutorial. 10:32.070 --> 10:33.180 That's our next step. 10:33.180 --> 10:34.893 And until then, enjoy AI.