1 00:00:00,330 --> 00:00:07,130 Hi and welcome back in this lesson, we'll be using Harris to implement a simple again something called 2 00:00:07,140 --> 00:00:13,290 a D.C. gun, which is short for a deep convolution gun and will be using it on the amnesty to set. 3 00:00:13,300 --> 00:00:17,310 So we're going to generate synthetic samples of that data set. 4 00:00:17,790 --> 00:00:20,430 So it's a very good instructional lesson on guns. 5 00:00:20,430 --> 00:00:21,420 So let's get started. 6 00:00:21,450 --> 00:00:25,890 So open notebook, too, the one here and we can begin to listen. 7 00:00:26,520 --> 00:00:29,110 So firstly, let's take a look at what we're doing. 8 00:00:29,130 --> 00:00:31,920 So let's do a quick recap on guns. 9 00:00:32,400 --> 00:00:37,980 Remember, what we have here is a generative network, and it discriminates in that we have two main 10 00:00:37,980 --> 00:00:38,700 components. 11 00:00:39,090 --> 00:00:44,400 The generator generates fake images, and it is committed to try and tell the difference between real 12 00:00:44,400 --> 00:00:45,900 images and fake images. 13 00:00:46,350 --> 00:00:48,780 And then it uses that to update its wits. 14 00:00:49,260 --> 00:00:56,880 And then the generator sends more images again, more a more mixed images dataset here to that discriminator 15 00:00:56,880 --> 00:00:59,940 with the frozen widths, and then he uses that knowledge. 16 00:01:00,160 --> 00:01:06,300 The results of that two updated tweets so that it can then in turn, keep improving. 17 00:01:06,600 --> 00:01:08,520 So this is how we train again. 18 00:01:09,060 --> 00:01:10,310 So let's get started. 19 00:01:10,320 --> 00:01:14,160 So first, let's download and install some of the packages. 20 00:01:14,160 --> 00:01:14,670 We need it. 21 00:01:15,300 --> 00:01:19,710 So we'll do that now shouldn't take too long so we can move on while we do that. 22 00:01:20,310 --> 00:01:30,060 Next, import of libraries that we will be using, and then we can go on and loading and preparing our 23 00:01:30,060 --> 00:01:32,420 data sets, including the amnesty to set. 24 00:01:32,430 --> 00:01:35,430 Here we get the tree in images, tree and labels. 25 00:01:35,850 --> 00:01:41,760 Then we reshape and normalize the data here, so we normalize it between minus one and one and converting 26 00:01:41,760 --> 00:01:43,170 it to a float to two. 27 00:01:44,010 --> 00:01:46,680 We set up a batch and buffer sized buffer. 28 00:01:46,680 --> 00:01:49,950 Size is the number of images in the training dataset. 29 00:01:49,950 --> 00:01:53,250 60000 to that size is 256. 30 00:01:53,790 --> 00:01:58,860 And then we just create this dataset to see dataset load effectively. 31 00:01:59,280 --> 00:02:05,220 That gives us a batch of data from that that that input that training data we have. 32 00:02:05,790 --> 00:02:07,650 So let's run this. 33 00:02:10,010 --> 00:02:11,540 And it should finish shortly. 34 00:02:11,630 --> 00:02:12,140 There we go. 35 00:02:12,770 --> 00:02:15,710 So now we can define what generates a model. 36 00:02:16,220 --> 00:02:23,660 Now remember, the generator model takes random noise as an input and generates a synthetic output from 37 00:02:23,660 --> 00:02:23,870 it. 38 00:02:24,320 --> 00:02:30,620 So to do that, basically, we just have the input here, input chip being some random noise vector 39 00:02:31,340 --> 00:02:33,110 and then which is the densely here. 40 00:02:33,110 --> 00:02:34,790 So we have a fully connected dense layer. 41 00:02:35,210 --> 00:02:42,200 However, this then goes to convert to the transpose, which is an upsampling layer, so you can see 42 00:02:42,200 --> 00:02:43,280 that it up samples. 43 00:02:43,280 --> 00:02:49,520 It gradually took the network so that the outputs final size is a 28 by 28 image. 44 00:02:49,610 --> 00:02:51,140 That's the generator network to remember. 45 00:02:51,140 --> 00:02:57,230 It's generating synthetic images so we can take a look at how generated network looks. 46 00:02:57,770 --> 00:03:03,470 You can just use this function that we defined to make code generator model and returns the model here. 47 00:03:03,950 --> 00:03:05,450 So we have generator here. 48 00:03:05,960 --> 00:03:07,770 We have over random noise vector. 49 00:03:07,790 --> 00:03:15,680 This is a one dimensional 100 item feature vector, and we just feed that noise into the generator here 50 00:03:16,190 --> 00:03:20,520 and we get the image out of it, and we did not run this function above. 51 00:03:20,870 --> 00:03:23,470 Unfortunately, that's why we got the error. 52 00:03:23,530 --> 00:03:26,870 So now we can run this and get that nice image we saw before. 53 00:03:27,710 --> 00:03:31,310 So how close is this image to the amnestied? 54 00:03:31,310 --> 00:03:33,240 Is that not very close isn't. 55 00:03:33,800 --> 00:03:35,870 So we need to improve that generator. 56 00:03:36,260 --> 00:03:41,720 Remember when we created the model, it's initialized with random weights, so effectively it's a dumb 57 00:03:41,720 --> 00:03:43,760 model that hasn't learned anything yet. 58 00:03:44,450 --> 00:03:49,850 Next, we can create a function to define or create a discriminative model. 59 00:03:50,450 --> 00:03:54,020 So remember, the discriminator model takes an input image. 60 00:03:54,020 --> 00:04:01,070 That's why it takes an input image of this size and then uses a regular CNN network to classify whether 61 00:04:01,070 --> 00:04:04,490 that network, whether that image is real or fake. 62 00:04:04,700 --> 00:04:06,590 That's what it does, and that's what returns of. 63 00:04:06,590 --> 00:04:12,650 Basically, that's one which is a single binary node, so zero would be real, one would be fake or 64 00:04:12,650 --> 00:04:13,430 vice versa. 65 00:04:14,120 --> 00:04:15,470 So let's run that. 66 00:04:15,860 --> 00:04:17,420 So we've created that function. 67 00:04:17,960 --> 00:04:20,960 Now let's test our discriminator on this image. 68 00:04:21,350 --> 00:04:22,760 So remember this variable? 69 00:04:22,760 --> 00:04:24,980 This image was called generated image here. 70 00:04:26,120 --> 00:04:32,720 So let's feed that in now to our discriminative model, which we create here and get its decision. 71 00:04:32,900 --> 00:04:35,010 And you can see what it tells us. 72 00:04:35,540 --> 00:04:42,260 It gives us a value of point zero zero one, which means that it's a real image because we've trained 73 00:04:42,260 --> 00:04:47,430 it to basically output positive values for real images and negative values for fake images. 74 00:04:47,430 --> 00:04:52,970 So you can see it just I mean, this is a good untrained discriminator, so it's not really of any relevance, 75 00:04:52,970 --> 00:04:59,120 but at least you understand the outputs and inputs of these building blocks for our overall good. 76 00:04:59,810 --> 00:05:02,750 So now we can define the lost function and optimizer. 77 00:05:03,080 --> 00:05:06,830 So firstly, that both functions will be using its binary cross entropy. 78 00:05:07,430 --> 00:05:10,280 So let's create that next. 79 00:05:10,280 --> 00:05:12,980 We have to create a function for the discriminative loss. 80 00:05:13,430 --> 00:05:17,810 So remember, Discriminator is trying to determine whether an image is a fake. 81 00:05:18,260 --> 00:05:24,500 So that loss function also uses cross entropy here, and the inputs are the real output and the fake 82 00:05:24,500 --> 00:05:25,310 output here. 83 00:05:25,760 --> 00:05:28,910 And then it gets distorted instead of total loss right here. 84 00:05:30,350 --> 00:05:37,070 Next, we have a generator lost function, which quantifies how well it was able to trick the discriminator. 85 00:05:37,550 --> 00:05:44,210 So intuitively, if your generators performing well, the discriminator were classified real and fake 86 00:05:44,210 --> 00:05:47,060 images as real, meaning that the output won. 87 00:05:47,660 --> 00:05:49,390 So that's our aim for this. 88 00:05:49,430 --> 00:05:50,480 So let's create that. 89 00:05:50,480 --> 00:05:57,170 What generates a loss and then we declare optimizes this for both of them will be using Adam with very 90 00:05:57,170 --> 00:05:58,370 small lending rates here. 91 00:05:58,370 --> 00:06:01,220 This is one zero zero zero one for both. 92 00:06:01,490 --> 00:06:08,060 So let's create that and then we create our checkpoints here, so we'll be saving on model after every 93 00:06:08,060 --> 00:06:08,510 run. 94 00:06:09,110 --> 00:06:10,430 And let's do that. 95 00:06:10,820 --> 00:06:16,930 So we created our checkpoint using TensorFlow and then now we could create a shooting loop. 96 00:06:16,940 --> 00:06:21,020 So let's just set the parameters for our training experiment here. 97 00:06:22,010 --> 00:06:27,710 This is our noise dimensional vector, which is 100, which we saw above number of examples to generate 98 00:06:27,710 --> 00:06:30,980 an on a batch and the number of epochs here. 99 00:06:31,160 --> 00:06:37,060 So and we just set a random seed so we can recreate this after and use this to visualize that. 100 00:06:37,070 --> 00:06:40,100 Yet the animated GIF afterwards as well. 101 00:06:40,970 --> 00:06:45,920 Next, we're going to use gradient tip to basically execute our trading step. 102 00:06:46,310 --> 00:06:49,670 So you would have seen this similar in PI touch to PI touch. 103 00:06:49,670 --> 00:06:57,170 So reading where we have where we get a batch of data and then we just pass it to the generator here. 104 00:06:57,710 --> 00:06:59,630 So let's take a look at what's happening here. 105 00:06:59,720 --> 00:07:07,010 So we get a noise vector here just generated from one noise and to that which is 100 and the size. 106 00:07:07,460 --> 00:07:09,470 So we get enough noise for the batch at. 107 00:07:09,510 --> 00:07:17,760 You need then it gradient tube, which basically allows us to calculate degree and so on a step, so 108 00:07:17,760 --> 00:07:24,090 we have a generator here, so we give it a noise, we get the generated images outfit, then we give 109 00:07:24,090 --> 00:07:29,460 those images here, we'll give you the generated discriminator, some real images and then we give off 110 00:07:29,460 --> 00:07:32,400 a generator or discriminate to some generated images here. 111 00:07:32,850 --> 00:07:35,220 That's the real output and fake output here. 112 00:07:35,220 --> 00:07:38,160 And then we just get the loss from the generator here. 113 00:07:38,700 --> 00:07:40,170 So we get to fake output. 114 00:07:40,620 --> 00:07:46,530 So we get our generator loss and then we get our discriminator loss by comparing the real output to 115 00:07:46,530 --> 00:07:47,070 the fake. 116 00:07:47,070 --> 00:07:52,470 Up until now, remember, the generator needs to output the discriminator. 117 00:07:52,890 --> 00:07:55,980 That's why it takes to fake output here to train itself after. 118 00:07:55,980 --> 00:07:57,900 That's where it uses to this loss here. 119 00:07:58,710 --> 00:08:03,360 Then we use the gradient tip here to update the gradient for them, for the boat. 120 00:08:03,870 --> 00:08:07,440 And then we apply to gradients, to the new model. 121 00:08:07,950 --> 00:08:11,660 So that's our gradient, sort of that's our trailing step function there. 122 00:08:12,570 --> 00:08:15,120 Next, we can actually create a training loop now. 123 00:08:15,120 --> 00:08:22,890 So for the number of epochs in the box we have set, we get images out of it here and then we just actually 124 00:08:22,890 --> 00:08:24,810 we're producing a GIF as we treat it. 125 00:08:25,260 --> 00:08:32,340 So we use this generated to save images here, and we received a model every 15 epochs and print some 126 00:08:32,340 --> 00:08:38,940 stats where we train and then we'll clear the output so we can actually see the images as the train, 127 00:08:38,940 --> 00:08:41,700 as they progress to the training experiment. 128 00:08:41,790 --> 00:08:48,810 So it's good to babysit this training process here, so it's you can monitor and see whether your garden 129 00:08:48,810 --> 00:08:51,060 is doing well next. 130 00:08:51,090 --> 00:08:57,120 We can create that function that we use here generated to see if actually this should be above this. 131 00:08:57,780 --> 00:08:58,530 It's my bad. 132 00:08:59,460 --> 00:09:01,620 Let me just move this up again. 133 00:09:03,390 --> 00:09:06,030 So when you understand or the work? 134 00:09:08,250 --> 00:09:14,400 And no, we are ready to start training the model, so let's do this. 135 00:09:16,500 --> 00:09:22,140 Let's start training, so are we going to train for 50 epochs here and after each epoch, we should 136 00:09:22,140 --> 00:09:23,070 see the images of? 137 00:09:23,610 --> 00:09:25,470 So let's stay tuned for that. 138 00:09:26,580 --> 00:09:27,060 There we go. 139 00:09:27,060 --> 00:09:34,240 So we can see this is the results of our first epoch after it fell roughly nine seconds and it was a 140 00:09:34,260 --> 00:09:40,320 beating again every six seconds faster, though, so you can see, oh, it's getting better. 141 00:09:40,590 --> 00:09:44,880 You can definitely see that we have a five and a one. 142 00:09:44,910 --> 00:09:46,380 Let's see what happens in the next epoch. 143 00:09:46,980 --> 00:09:49,080 I definitely see a big improvement there. 144 00:09:49,290 --> 00:09:50,850 So let's keep what watching. 145 00:10:12,490 --> 00:10:13,950 OK, so we are done. 146 00:10:14,050 --> 00:10:18,960 And you can see after 50 books, the results are actually pretty good. 147 00:10:18,970 --> 00:10:26,020 I mean, some of these look don't look great like this one looks like an e or where it this one looks 148 00:10:26,020 --> 00:10:27,510 like a zero with a dot in it. 149 00:10:27,520 --> 00:10:33,310 But you can see generally it's actually learned to scan has learned to generate some pretty decent images. 150 00:10:33,520 --> 00:10:34,700 It's look OK here. 151 00:10:35,230 --> 00:10:37,930 This looks like a nine, but it's a bit misshapen. 152 00:10:38,530 --> 00:10:40,100 But I mean, this is pretty cool. 153 00:10:40,150 --> 00:10:46,750 You're using it to generate synthetic images and look almost quite real. 154 00:10:47,410 --> 00:10:49,420 This should get you quite excited, in my opinion. 155 00:10:50,260 --> 00:10:54,220 So now we can just restore the world's two largest and latest checkpoint. 156 00:10:54,220 --> 00:10:54,550 Sorry. 157 00:10:55,300 --> 00:10:59,320 And can you use this now to create a GIF? 158 00:10:59,500 --> 00:11:03,400 So let's create a GIF to show our progress. 159 00:11:05,410 --> 00:11:06,940 And then what we do? 160 00:11:07,000 --> 00:11:15,820 We use the image you function well library to create that gif that we're about to create from our training 161 00:11:15,820 --> 00:11:16,180 data. 162 00:11:16,180 --> 00:11:20,440 Because you can see here we have the image, see that every epoch. 163 00:11:20,980 --> 00:11:28,420 So let's close this and then we can generate this here, which I mean, you just so me the below. 164 00:11:28,420 --> 00:11:35,960 But let's do it again, because that was using a previously trained can that I use before we go. 165 00:11:35,980 --> 00:11:39,760 You can see it as we get closer than 50 books there. 166 00:11:40,780 --> 00:11:41,950 So this is pretty cool. 167 00:11:41,960 --> 00:11:48,280 I mean, what you can do, maybe experiment, maybe deepen the networks, add some more parameters. 168 00:11:49,330 --> 00:11:51,610 Next, you can probably trade for longer epochs. 169 00:11:52,150 --> 00:11:56,470 So this is a quite good lesson, and I hope you enjoyed this tutorial. 170 00:11:57,070 --> 00:12:03,370 You can take a look at the source for this tutorial here as well and hopefully enjoy the lesson. 171 00:12:03,370 --> 00:12:08,260 And what we'll do next will do a similar experiment in PyTorch. 172 00:12:08,380 --> 00:12:09,520 So stay tuned for that.