Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,450 --> 00:00:03,220 Hello and welcome back to the course on computer vision. 2 00:00:03,420 --> 00:00:06,260 Today we're going to find out how Gannes work. 3 00:00:06,270 --> 00:00:08,830 So this is gonna be an interesting tutorial. 4 00:00:08,880 --> 00:00:11,620 Some exciting slides prepared ahead. 5 00:00:12,120 --> 00:00:14,760 It's going to be quite long so prepare yourself for that. 6 00:00:14,820 --> 00:00:17,590 And let's dive straight into it. 7 00:00:17,610 --> 00:00:17,970 All right. 8 00:00:17,970 --> 00:00:24,360 So as we discussed Gannes stands for generative adversarial networks and we're going to go through these 9 00:00:24,360 --> 00:00:30,420 letters one by one and then we'll get to the training part of the gang which is the exciting part. 10 00:00:30,420 --> 00:00:39,140 All right so G stands for generative and this is a model which takes as input a random noise a signal 11 00:00:39,420 --> 00:00:43,400 and then it can output an image. 12 00:00:44,160 --> 00:00:47,560 The adversarial part stands for the discriminator. 13 00:00:47,640 --> 00:00:53,950 It's another model which is going to be rivalling the generators going to be the opponent of the generator. 14 00:00:54,180 --> 00:01:02,350 You can think of the generator as like a thief or somebody who's trying to forge dollar bills and discriminator 15 00:01:02,390 --> 00:01:04,960 as as the police officer as the detective. 16 00:01:05,220 --> 00:01:11,820 Oh there you go Dee for Detective so is going to be the revival of the generator and diesel model which 17 00:01:11,820 --> 00:01:21,480 is capable of learning about objects animals or people or basically certain features or a cable or blurring 18 00:01:21,480 --> 00:01:21,810 features. 19 00:01:21,810 --> 00:01:27,700 For instance if you show with lots of dogs and then you show it lots of non dogs it will be able to 20 00:01:27,700 --> 00:01:31,410 offer that discriminate between dogs and not dogs. 21 00:01:31,410 --> 00:01:38,080 So in an ideal world if you show the discriminator of this image it'll say zero. 22 00:01:38,150 --> 00:01:41,040 I will give it a zero probability that it's a dog. 23 00:01:41,190 --> 00:01:45,540 And if you show it this image it will give it a warning and give it a 100 percent probability that it's 24 00:01:45,540 --> 00:01:46,210 a dog. 25 00:01:46,500 --> 00:01:49,530 So that is the ideal world. 26 00:01:49,530 --> 00:01:52,030 Now we're going to get oh now we have it. 27 00:01:52,050 --> 00:01:57,000 And so that's why this is where we have this little animation you see. 28 00:01:57,000 --> 00:02:00,030 There we go the network letters jumping around. 29 00:02:00,090 --> 00:02:05,310 And that stands for the fact that these two models are actually neural networks so we're going to replace 30 00:02:05,310 --> 00:02:09,120 them with neural networks because that's what they are. 31 00:02:09,120 --> 00:02:16,350 So we've got a neural network on the left a neural network on the right and why that's important is 32 00:02:16,350 --> 00:02:24,480 because as we discussed in annex to the course about neural networks hopefully you checked that out. 33 00:02:24,690 --> 00:02:30,480 They are capable of learning that's where you know the weights of those synopses that we see in blue. 34 00:02:30,480 --> 00:02:36,690 They are that's where information that's where the training is going to go that's where the those weights 35 00:02:36,690 --> 00:02:39,270 are going to be updated so that the neural networks. 36 00:02:39,630 --> 00:02:39,840 All right. 37 00:02:39,840 --> 00:02:41,220 So we discussed the letters. 38 00:02:41,220 --> 00:02:42,360 Let's go through the training. 39 00:02:42,360 --> 00:02:44,160 Exciting times. 40 00:02:44,160 --> 00:02:44,430 All right. 41 00:02:44,430 --> 00:02:45,490 Step one. 42 00:02:45,750 --> 00:02:47,160 Basically all of the steps. 43 00:02:47,160 --> 00:02:53,000 It's not like step one step two step three that you need to Conseco really not that different. 44 00:02:53,010 --> 00:02:55,530 There is always going to be repeating the same step. 45 00:02:55,530 --> 00:03:01,080 We're going to be doing it many times in fact here in this intuition tutorial we're going to do it three 46 00:03:01,080 --> 00:03:06,940 times in the practical tutorials are going to do with hundreds and thousands of times. 47 00:03:07,520 --> 00:03:09,930 But intuition is going to be enough. 48 00:03:10,050 --> 00:03:14,220 But basically you will see how the networks evolve over the iteration. 49 00:03:14,220 --> 00:03:17,180 So every step is an iteration. 50 00:03:17,460 --> 00:03:23,140 Don't confuse step with espagnol and pork is after you've done your steps. 51 00:03:23,190 --> 00:03:27,760 That's one epal and then you do another book that's you do all the steps again and again. 52 00:03:28,050 --> 00:03:32,570 So you'll see that in the practice but back here what we have is step one. 53 00:03:32,580 --> 00:03:41,320 We input the noise signal into a noise signal a random noise signal into our generator. 54 00:03:41,540 --> 00:03:45,210 It just basically kind of like gives it some random this to generate from it. 55 00:03:45,240 --> 00:03:50,240 It needs to start somewhere and then it generates some images. 56 00:03:50,250 --> 00:03:55,530 These images are as you can see completely useless completely random images. 57 00:03:56,610 --> 00:04:02,040 And at this stage what we want to train is we want to train the discriminator. 58 00:04:02,040 --> 00:04:07,290 So this is what we're going to be doing and you will see exactly the same steps in the practical tutorials 59 00:04:07,290 --> 00:04:12,300 and this is how the generative adversarial networks concept or 60 00:04:14,970 --> 00:04:18,330 algorithm was really like was designed in the first place. 61 00:04:18,330 --> 00:04:20,120 So first we train the discriminator. 62 00:04:20,340 --> 00:04:25,950 And as we discussed we want the descriptor to be able to distinguish between for instance we're going 63 00:04:25,950 --> 00:04:29,060 to stick with the example of dogs between dogs and non-dog. 64 00:04:29,070 --> 00:04:36,420 So we need to give it in addition to these random images we generated we need to give it some dog images. 65 00:04:36,420 --> 00:04:37,130 So there you go. 66 00:04:37,140 --> 00:04:43,740 We've got some dog images and so we're going to input these a batch of batch of non-dog images and a 67 00:04:43,740 --> 00:04:45,850 batch of dog images. 68 00:04:45,850 --> 00:04:51,720 We're going to put them into the discriminator and we'll see what probabilities this community will 69 00:04:51,720 --> 00:04:52,160 spit out. 70 00:04:52,170 --> 00:04:56,240 And remember the discriminator knows nothing at this stage hasn't been trained at all. 71 00:04:56,250 --> 00:04:57,490 This is the very first step. 72 00:04:57,540 --> 00:05:03,460 None of these two models have been trained so the discrimination might spit out numbers like this for 73 00:05:03,460 --> 00:05:08,710 instance zero point or point trees or open fire like quite high probabilities for the images at the 74 00:05:08,710 --> 00:05:09,920 top that they're dogs. 75 00:05:09,940 --> 00:05:14,500 Even though we can see they're not dogs discriminate has never seen a dog before so it might give them 76 00:05:14,830 --> 00:05:20,170 some you know just based on its initial configuration which is you know some probably doesn't really 77 00:05:20,170 --> 00:05:21,070 worry us at this stage. 78 00:05:21,070 --> 00:05:23,770 It's got to start somewhere again. 79 00:05:24,070 --> 00:05:32,950 But you're going to start somewhere and we will fix this as a goal as a train that will fix this and 80 00:05:32,950 --> 00:05:35,920 then it give some probabilities to the images below. 81 00:05:36,310 --> 00:05:39,290 Even though we know their dogs discriminate has never seen a dog before. 82 00:05:39,340 --> 00:05:44,860 It just gives them some believe that happened to be the outputs of this neural network at this point 83 00:05:44,860 --> 00:05:48,920 in time or at this stage. 84 00:05:49,210 --> 00:05:51,760 At the very beginning at the very first stage. 85 00:05:51,820 --> 00:05:59,240 So what happens next is we take these values and we look at what they should have been. 86 00:05:59,350 --> 00:06:08,230 So we as humans because we are training this neural network we know what each of outputs in the ideal 87 00:06:08,230 --> 00:06:12,850 scenario and ideal scenario the top values should have been zeros the bottom values should have been 88 00:06:12,850 --> 00:06:13,330 ones. 89 00:06:13,360 --> 00:06:20,560 So the top owlish top our pictures are not dogs so they should have caught on a 0 percent likelihood 90 00:06:20,560 --> 00:06:21,410 of being dogs. 91 00:06:21,470 --> 00:06:24,970 Bottom one should have grown 100 percent like a big dog so that's ideal world. 92 00:06:25,000 --> 00:06:30,120 Obviously our model isn't ideal but it can learn from this. 93 00:06:30,130 --> 00:06:37,210 So what's going to happen is the era is going to be calculated so basically in every single case the 94 00:06:37,240 --> 00:06:43,960 values will be subtracted more from each other so zero zero point AIDS or about 0.3 or or 0.5 and 1 95 00:06:44,140 --> 00:06:46,660 0.1 minus 1 and so on. 96 00:06:46,660 --> 00:06:52,780 So all those will be subtracted then the cumulative error will be calculated and will be back propagate 97 00:06:52,810 --> 00:06:55,280 through the network of the discriminator. 98 00:06:55,450 --> 00:07:02,800 So if you have and you're not familiar with the term back propagation then I highly recommend checking 99 00:07:02,800 --> 00:07:07,630 an annex on this course where you talk about artificial neural networks and can do it on your own networks 100 00:07:08,380 --> 00:07:15,130 and you will be up to after that you'll be up to speed with everything we discussed including back propagation. 101 00:07:15,310 --> 00:07:21,070 It's quite a lengthy consequence we're going to discuss it here but if you are we're back propagation 102 00:07:21,090 --> 00:07:22,950 and this is exactly what's happening here. 103 00:07:23,080 --> 00:07:28,840 These errors are back propagated through a network and the weights on the network are a bit dated. 104 00:07:28,870 --> 00:07:29,620 So there we go. 105 00:07:29,620 --> 00:07:34,240 Now the discriminator has learned that that was basically the learning process of this has learned from 106 00:07:34,240 --> 00:07:36,920 his mistakes and next time is going to do a better job. 107 00:07:37,770 --> 00:07:41,480 And now we want to train the generator. 108 00:07:41,500 --> 00:07:47,560 So this part of our Gamme and how we're going to trend generation is we're going to take the same image 109 00:07:47,650 --> 00:07:53,860 images that we've created this batch of images which are supposed to be dogs and we're going to use 110 00:07:53,860 --> 00:07:54,430 them again. 111 00:07:54,430 --> 00:07:59,520 So a quick note here is that you'll notice that this is exactly what we did in the practical Tauriel. 112 00:07:59,530 --> 00:08:05,680 We took the images that were indeed there we had generated for the training of the discriminator and 113 00:08:05,680 --> 00:08:09,270 we use the same images in his paper. 114 00:08:09,610 --> 00:08:15,360 Ian Goodfellow recommends taking image generating new images. 115 00:08:15,430 --> 00:08:21,160 So running the noise part and the generation part again and using those images for the train or generator 116 00:08:21,520 --> 00:08:24,170 we just use the same images and it worked fine. 117 00:08:24,430 --> 00:08:25,910 It's up to you how you want to go. 118 00:08:25,930 --> 00:08:30,520 You might want to experiment with this part in the practical if you like or just follow along with our 119 00:08:30,520 --> 00:08:35,230 tutorials and the networks worked really well. 120 00:08:35,230 --> 00:08:38,760 But just so that you're aware of what it is. 121 00:08:38,770 --> 00:08:43,160 Originally in the paper by in good form. 122 00:08:43,620 --> 00:08:48,700 OK so we're going to take those images and we're going to run them through the discriminator again but 123 00:08:48,700 --> 00:08:55,960 this time we don't need any dog images because the way that the generator learns is by trying to trick 124 00:08:55,990 --> 00:09:02,110 the discriminator and based on whether it succeeds or not it will update its way so it doesn't need 125 00:09:02,110 --> 00:09:08,680 to compare to any dogs or dogs it just needs to try and trick the discriminator as the other discriminator 126 00:09:08,890 --> 00:09:13,900 will provide an output as you can see the output is different to what it had before because it has already 127 00:09:14,470 --> 00:09:20,260 being trained itself and therefore knows what to what a dog can kind of more or less it knows what a 128 00:09:20,260 --> 00:09:24,080 dog looks like and it knows what a dog doesn't look like. 129 00:09:24,580 --> 00:09:31,950 And as you can see here the numbers are not all like 0 0 0 and they're kind of close to zero they're 130 00:09:31,960 --> 00:09:38,730 lower but they're not just older or 0 0 because it takes time for the discriminator to learn is one 131 00:09:38,740 --> 00:09:44,470 this illustrates the concept that why these two should be training at the same time because if you for 132 00:09:44,470 --> 00:09:50,920 instance the discriminator was very well trained like it was it had been trained for a very long time 133 00:09:50,940 --> 00:09:57,310 where I was trained separately and up until two to a very good level then and then and the generator 134 00:09:57,310 --> 00:10:04,150 was only being trained now than the discriminator would always like completely completely get leave 135 00:10:04,180 --> 00:10:06,490 no chances for the from the generator. 136 00:10:06,730 --> 00:10:11,920 And that's why the will be like You're too smart for the generator the it would have no chances of tricking 137 00:10:11,920 --> 00:10:13,050 those discriminator. 138 00:10:13,270 --> 00:10:14,920 But that would be not good for anybody. 139 00:10:14,920 --> 00:10:16,590 They have to they have to train together. 140 00:10:16,600 --> 00:10:22,400 They have to kind of get better and better and better with time together and so. 141 00:10:22,770 --> 00:10:27,130 So what we're going to do now is we're going to take these values and we're going to compare them to 142 00:10:27,130 --> 00:10:28,740 what it should be. 143 00:10:28,990 --> 00:10:34,920 And you know what these values should have been like what would you what would you compare them to. 144 00:10:34,990 --> 00:10:36,380 You compare them to zero. 145 00:10:36,470 --> 00:10:40,710 But this time we're going to compare them to one because now our objective is different prove you are 146 00:10:40,720 --> 00:10:46,620 training the discriminator now entering the generator and for the generator it wants to trick the discriminate. 147 00:10:46,620 --> 00:10:51,730 Once those values to be ones as it can see those values are much less than ones who calculate the error 148 00:10:51,870 --> 00:11:00,640 and calculate the difference between between each value on one and you'll take the aggregate era and 149 00:11:00,730 --> 00:11:09,780 it will buy and that era will be back propagated through the network of the discretion of the generator. 150 00:11:09,790 --> 00:11:11,030 This time around. 151 00:11:11,480 --> 00:11:24,280 And so that's going to allow the generator to now update its way its They go so as you can see it in 152 00:11:24,290 --> 00:11:25,710 a day that it's weights there. 153 00:11:25,920 --> 00:11:27,280 And now this time. 154 00:11:27,390 --> 00:11:32,760 Now next time round they will know it will do a better job will be to do a better job of generating 155 00:11:32,750 --> 00:11:37,210 the images because now it knows that this wasn't wasn't good enough. 156 00:11:37,350 --> 00:11:43,080 So it knows how to hold that awaits him up there to improve that result. 157 00:11:43,560 --> 00:11:52,200 And again this is this all boils down to neural networks and the process of back propagation and gradient 158 00:11:52,200 --> 00:11:55,850 descent and the casting gradient descent which happens in the neural network. 159 00:11:55,850 --> 00:12:00,960 So those are concepts that are required in order for if you like to understand in in detail how this 160 00:12:00,960 --> 00:12:03,890 is happening and those concepts we discussed in the annex. 161 00:12:04,250 --> 00:12:04,770 OK. 162 00:12:04,860 --> 00:12:07,320 So that was the first step. 163 00:12:07,320 --> 00:12:07,580 All right. 164 00:12:07,590 --> 00:12:12,140 So let's move on to Step 2. 18202