1 00:00:01,230 --> 00:00:02,280 Hi and welcome back. 2 00:00:02,640 --> 00:00:07,500 In this section, we'll take a look at using PyTorch to implement Google Deep Streams algorithm. 3 00:00:08,070 --> 00:00:09,060 So let's get started. 4 00:00:09,090 --> 00:00:12,930 So open notebook tricks here which are already have opened. 5 00:00:13,590 --> 00:00:14,910 And let's take a look at this. 6 00:00:14,920 --> 00:00:19,310 So in this in this vision, we're not going to use the inception network. 7 00:00:19,320 --> 00:00:20,790 We're going to try different networks. 8 00:00:20,790 --> 00:00:22,560 So let's use Figure 16. 9 00:00:23,070 --> 00:00:30,630 So let's load our model and load or libraries, as well as download our test sample image and we'll 10 00:00:30,630 --> 00:00:34,410 wait for that to be completed should be finished very shortly. 11 00:00:43,580 --> 00:00:47,810 OK, so it took a little while because video is quite a quite big network. 12 00:00:48,410 --> 00:00:49,760 So let's see. 13 00:00:49,790 --> 00:00:52,400 So test image to see one we looked at than the previous one. 14 00:00:52,700 --> 00:00:54,050 That picture in Tobago. 15 00:00:54,950 --> 00:00:58,710 So now what we do, we're going to get the outputs of Aleah. 16 00:00:58,730 --> 00:01:04,700 So to do that, we create this classical hook and then we create these functions called get gradients. 17 00:01:05,180 --> 00:01:09,980 This function makes brilliant calculations from the output channels of the target layer. 18 00:01:10,490 --> 00:01:15,980 Then we also have the D normalized normalization of the image transform into a transformer to be using, 19 00:01:16,610 --> 00:01:23,240 as well as we have the dream, which basically combines all the functions above and executes Google's 20 00:01:23,420 --> 00:01:24,860 deep dream algorithm on them. 21 00:01:25,130 --> 00:01:30,410 So like I said in the previous lesson, I'm not going to go into this code in detail because it's quite 22 00:01:30,890 --> 00:01:34,130 confusing sometimes for beginners, and it's a lot going on. 23 00:01:34,520 --> 00:01:40,250 So for now, let's just run this and remember it at a high level how it operated. 24 00:01:40,850 --> 00:01:44,810 Now we're going to run an execute of this Google deep dream here. 25 00:01:44,930 --> 00:01:45,890 So let's take a look. 26 00:01:47,000 --> 00:01:52,730 We're using the dream function that we created up here, and you can see we just generate this image 27 00:01:52,730 --> 00:01:53,540 quite quickly. 28 00:01:54,020 --> 00:01:56,090 And you get some pretty trippy effects already. 29 00:01:56,090 --> 00:02:00,290 You can see some with spirals added, some spikes, it added. 30 00:02:00,290 --> 00:02:01,250 Like, what is that? 31 00:02:01,250 --> 00:02:02,840 Like a tutorial to this rock? 32 00:02:04,550 --> 00:02:05,460 This is quite trippy. 33 00:02:05,690 --> 00:02:06,890 And then that was quite quick. 34 00:02:07,640 --> 00:02:11,120 So this is a pretty nice and quick implementation of the dream. 35 00:02:11,540 --> 00:02:16,530 It's not a high res image by any means, but you can adjust the scaling parameters as well as layer 36 00:02:16,550 --> 00:02:19,370 parameters and octaves to tweak that. 37 00:02:19,550 --> 00:02:21,860 Now let's make some improvements. 38 00:02:22,010 --> 00:02:27,470 So as you can see where the patterns we have to see, we have this, we can see the patterns have the 39 00:02:27,470 --> 00:02:28,250 same skill. 40 00:02:28,880 --> 00:02:32,180 And that's a problem because we want to get different skills in image. 41 00:02:32,600 --> 00:02:38,750 So to do this with updated prior could, but every time we resize image to a different scale. 42 00:02:38,900 --> 00:02:40,850 So that's what this court is doing here. 43 00:02:41,270 --> 00:02:44,780 It's a lot more functions and I mean a lot more lines of code going on to do this. 44 00:02:45,320 --> 00:02:48,350 So let's run this and take a look at the output. 45 00:02:50,820 --> 00:02:54,390 This will take longer, though, because we're doing different skills now. 46 00:03:03,310 --> 00:03:06,130 All right, so that's finished, and now we can visualize it. 47 00:03:06,160 --> 00:03:08,830 Let's take a look at the different octave skills. 48 00:03:23,340 --> 00:03:24,840 All right, finally, here we go. 49 00:03:24,990 --> 00:03:29,400 And you can see we've created a very, very cool, but we're trippy image. 50 00:03:29,880 --> 00:03:36,420 You can definitely see some, some reptile reptilian looking skills, something that looks like octopus 51 00:03:36,420 --> 00:03:37,950 tentacles coming out of it. 52 00:03:38,580 --> 00:03:39,300 These are weird. 53 00:03:39,300 --> 00:03:40,840 Eyes the sky. 54 00:03:40,860 --> 00:03:42,030 Let's take a look at the sky. 55 00:03:42,690 --> 00:03:45,450 Wow, you can see this little face in the background here. 56 00:03:46,290 --> 00:03:47,240 Wow, this really. 57 00:03:47,250 --> 00:03:52,320 This really did create some weird, trippy effects, so I would encourage you to mess with the parameters 58 00:03:52,320 --> 00:03:57,690 here, as well as like the octave scaling and some of the other parameters defined above. 59 00:03:58,620 --> 00:03:58,890 Yeah. 60 00:03:59,250 --> 00:04:02,520 So take a look and try and experiment on your own. 61 00:04:03,330 --> 00:04:04,950 All right, we'll stop there for now. 62 00:04:05,370 --> 00:04:09,210 I did have the code for a targeted deep dream algorithm here. 63 00:04:09,600 --> 00:04:14,340 However, it broke, it doesn't no longer works in color because of some visioning issue. 64 00:04:14,760 --> 00:04:19,830 So maybe I'll leave this code here for you guys, but maybe it would work if you toy with it or mess 65 00:04:19,830 --> 00:04:20,160 with it. 66 00:04:20,790 --> 00:04:24,000 So for now, that's it for this lesson in PI touch with deep dream. 67 00:04:24,000 --> 00:04:29,970 I hope you enjoyed it, and I hope you have fun experimenting on your images and feel free to share 68 00:04:29,970 --> 00:04:33,510 them with the rest of the students in the Udemy forum. 69 00:04:33,930 --> 00:04:34,340 Thank you.