1 00:00:01,110 --> 00:00:07,050 And welcome back in this lesson, we'll take a look at executing the Google deep dream algorithm using 2 00:00:07,050 --> 00:00:07,640 Keros. 3 00:00:07,710 --> 00:00:08,760 So let's get started. 4 00:00:08,970 --> 00:00:14,910 Open Notebook 25, which I've already done here, actually changed us to 25. 5 00:00:14,940 --> 00:00:15,780 Sorry about that. 6 00:00:16,560 --> 00:00:18,540 So let's take a look at this lesson. 7 00:00:18,660 --> 00:00:20,010 So what we'll do? 8 00:00:20,160 --> 00:00:26,370 Firstly, I'll just say one thing, though, because implementing these algorithms like Google Live 9 00:00:26,370 --> 00:00:33,600 Stream and neural style transfer in a few of these fluid projects, in this course, they can get quite 10 00:00:33,630 --> 00:00:37,410 technical because you're now implementing real research papers. 11 00:00:37,590 --> 00:00:40,920 These researchers have spent years or months working on. 12 00:00:41,430 --> 00:00:49,860 So to do justice to explain these, these theycould and their methods in a short space of time is often 13 00:00:49,860 --> 00:00:50,730 quite difficult. 14 00:00:50,760 --> 00:00:56,340 So what I will do, though, is I'll explain it at a very high level and try to put a lot of comments 15 00:00:56,340 --> 00:00:59,280 in the code to help you understand what's going on. 16 00:00:59,700 --> 00:01:04,140 I'm not going to do a line by line explanations because that can get very tedious. 17 00:01:04,140 --> 00:01:09,600 And to be quite fair, I don't actually fully understand every single thing that they do. 18 00:01:10,200 --> 00:01:12,080 So I apologize for that. 19 00:01:12,090 --> 00:01:15,480 And now we can get started after that disclaimer warning. 20 00:01:16,110 --> 00:01:17,850 So what we're doing here? 21 00:01:17,880 --> 00:01:19,980 We're loading our modules and this image here. 22 00:01:20,460 --> 00:01:23,220 Also, what we're doing, we're setting two layer settings here. 23 00:01:23,700 --> 00:01:26,790 Remember, these are represent different layers of the network. 24 00:01:27,150 --> 00:01:35,130 And this this, which we attribute here, basically see how much we want to this way to contribute to 25 00:01:35,130 --> 00:01:38,070 the visual effect so you can mess with these parameters. 26 00:01:39,020 --> 00:01:44,460 Um, I tend to actually like the scaling it up like this, but you can mess with those parameters to 27 00:01:44,460 --> 00:01:46,320 get different types of effects. 28 00:01:46,980 --> 00:01:49,330 Next, we can do a number of octaves. 29 00:01:49,350 --> 00:01:52,830 This is the skills that you run gradient descent up to more octaves you do. 30 00:01:53,190 --> 00:01:58,230 The more time it takes, however, that you at some point you're not going to see much difference because 31 00:01:58,590 --> 00:02:01,180 when it's too small, it's hard to see that on the image. 32 00:02:01,200 --> 00:02:03,090 So Tree is a good number to use. 33 00:02:03,660 --> 00:02:08,330 The ratio between scales is 1.4 and iterations will do. 34 00:02:08,350 --> 00:02:12,360 That's the Ascend steps, but scale would be 20. 35 00:02:12,480 --> 00:02:15,480 And then we use a max loss of 15 to stop that. 36 00:02:16,050 --> 00:02:19,890 So let's run this and we'll get started. 37 00:02:20,790 --> 00:02:21,240 There we go. 38 00:02:21,330 --> 00:02:24,720 So it's we've loaded an image and we've floated all libraries. 39 00:02:25,320 --> 00:02:27,270 Now let's take a look at our test image to test. 40 00:02:27,270 --> 00:02:33,580 Image we've been using for this lesson is a beautiful picture of a beach in the island of Tobago. 41 00:02:34,050 --> 00:02:36,790 That's like home country of Trinidad and Tobago. 42 00:02:37,320 --> 00:02:38,310 It's a star. 43 00:02:38,490 --> 00:02:42,360 It's actually a heavenly bay, which is a bay next to costar a beach. 44 00:02:43,110 --> 00:02:45,780 So it's a very, very beautiful spot that I love. 45 00:02:46,560 --> 00:02:49,520 So now let's take a look at the deep dream process. 46 00:02:49,560 --> 00:02:52,770 So remember, we're first loading our original image. 47 00:02:53,370 --> 00:02:58,620 Then we define the number of processing skills which you have seen above in those parameters from small 48 00:02:58,620 --> 00:02:59,340 astrologers. 49 00:02:59,820 --> 00:03:03,450 Then we recites the original image to to the smallest skill. 50 00:03:03,900 --> 00:03:08,400 And then for every skill, starting with the smallest one, we run gradient descent. 51 00:03:08,850 --> 00:03:14,250 We upskill the image to the next skill and then re-injected the detail that was lost to the upskilling 52 00:03:14,280 --> 00:03:14,670 time. 53 00:03:15,240 --> 00:03:16,290 And then we stop. 54 00:03:16,470 --> 00:03:18,660 When we cook, we are back to the original skill. 55 00:03:19,170 --> 00:03:22,620 So now add them to obtain the detail lost during the upsampling. 56 00:03:23,040 --> 00:03:24,990 We simply take the original image. 57 00:03:25,140 --> 00:03:30,750 Shrink it down and then upskill it again and then compare the results of the original resized image. 58 00:03:31,260 --> 00:03:34,380 So that's basically the process in the overview. 59 00:03:34,860 --> 00:03:40,320 You can go back to the slides if you want to get a recap, more detailed recap with more, more visual 60 00:03:40,590 --> 00:03:41,280 explanation. 61 00:03:42,000 --> 00:03:47,970 So here we create a pre processing function for the image that we load, and this uses the Inception 62 00:03:47,980 --> 00:03:51,070 version three pre processes because that's the model will be loading. 63 00:03:51,090 --> 00:03:53,880 That's a pre-trained model that we've been using. 64 00:03:54,840 --> 00:03:59,730 And then we also have a deep process function image, which is what we need to do when we do the mixing 65 00:03:59,730 --> 00:04:04,950 of the images, and that's specific to the inception of each processing. 66 00:04:04,950 --> 00:04:06,540 That's where these values come from. 67 00:04:07,110 --> 00:04:09,270 So don't worry about these too much. 68 00:04:09,780 --> 00:04:13,490 And no, we compute the gibbering loss. 69 00:04:13,500 --> 00:04:15,000 So we load on that to work here. 70 00:04:15,600 --> 00:04:22,530 We get upwards to orleaneans here in this and then we just create a feature extract extractor here using 71 00:04:22,530 --> 00:04:26,250 those layers that we've loaded because we don't include the top layer. 72 00:04:26,430 --> 00:04:28,200 That means that's a fully connected layers. 73 00:04:28,620 --> 00:04:33,150 So it's basically a feature extract, a big feature extractor network. 74 00:04:33,810 --> 00:04:36,900 So that's how we create this model, this feature to model here. 75 00:04:37,860 --> 00:04:41,040 And now the most common computation is actually very simple. 76 00:04:41,580 --> 00:04:45,000 We just promises how we run the gradient ascent part of it here. 77 00:04:45,930 --> 00:04:47,520 So let's create that function. 78 00:04:48,240 --> 00:04:51,730 And then we set up the gradient descent loop for one octave here. 79 00:04:51,750 --> 00:04:53,490 So this is the gradient descent setup. 80 00:04:54,000 --> 00:04:59,880 We're using TensorFlow as gradient to appear to run to compute these gradients at some. 81 00:04:59,960 --> 00:05:06,020 Butterfly torches, Auchterarder, and then we just get to gradients normalized to gradients as we said 82 00:05:06,020 --> 00:05:07,100 we would have to do. 83 00:05:07,730 --> 00:05:14,150 And then this is the gradient descent loop that executes everything here next. 84 00:05:14,300 --> 00:05:18,830 Now we can actually start the training loop and iterating over different octaves. 85 00:05:19,280 --> 00:05:25,130 So this takes a little while to run when when it's finished running, we can display all final deep 86 00:05:25,130 --> 00:05:25,880 dream output. 87 00:05:26,480 --> 00:05:29,730 So let's go ahead and run this feature. 88 00:05:29,740 --> 00:05:34,520 Extract is not defined within run this block of code on top here. 89 00:05:34,820 --> 00:05:36,020 Let's do that now. 90 00:05:36,890 --> 00:05:39,590 OK, so now we can run this. 91 00:05:41,860 --> 00:05:47,440 So it's going to take a while, as they said, so we'll sit back and relax, go have a cup of tea or 92 00:05:47,440 --> 00:05:49,120 coffee for the time being. 93 00:05:50,020 --> 00:05:53,470 And then maybe in about a few minutes, it should be finished. 94 00:05:56,490 --> 00:05:59,820 OK, actually, that took a lot shorter than expected 22 seconds. 95 00:06:00,390 --> 00:06:07,320 Maybe we can actually use the parameters and tweak them to get the more fancy results, so let's take 96 00:06:07,320 --> 00:06:08,010 a look at the output. 97 00:06:08,760 --> 00:06:09,750 This should be exciting. 98 00:06:11,760 --> 00:06:14,130 Let's see our first Google deep dream. 99 00:06:14,220 --> 00:06:14,790 Here we go. 100 00:06:15,570 --> 00:06:17,870 So you can see this is actually quite a way continues. 101 00:06:17,880 --> 00:06:23,700 There's a weird little ES pattern that it created here. 102 00:06:23,750 --> 00:06:26,100 I'm not entirely sure what it looks like. 103 00:06:26,880 --> 00:06:28,620 It looks like a fish head, isn't it? 104 00:06:29,610 --> 00:06:31,890 And then just some generic patterns here. 105 00:06:32,370 --> 00:06:36,090 These patterns are due to the filters that we've we looked at. 106 00:06:36,510 --> 00:06:42,000 You can you can feel free to change the mixing change a lot of those different parameters in the beginning 107 00:06:42,000 --> 00:06:46,170 of the notebook and see if you get any other cool results. 108 00:06:46,380 --> 00:06:47,070 Let's see. 109 00:06:48,330 --> 00:06:50,040 Yeah, some of the cool patterns here. 110 00:06:50,040 --> 00:06:55,560 Something trippy, not super trippy, but you can experiment and run this for longer with more iterations 111 00:06:55,560 --> 00:06:58,320 as well to get even more color effects. 112 00:06:58,950 --> 00:07:02,940 So that's it for the cameras, Google's deep dream algorithm. 113 00:07:03,780 --> 00:07:07,920 Next, we'll take a look at how we do this and PyTorch, so I'll see you there. 114 00:07:08,220 --> 00:07:08,490 But.