1 00:00:01,260 --> 00:00:07,770 Hi and welcome back in this lesson, we'll take a look at the PyTorch implementation of nine old style 2 00:00:07,770 --> 00:00:08,340 transfer. 3 00:00:08,850 --> 00:00:11,730 So open this notebook and we'll begin to listen. 4 00:00:12,150 --> 00:00:15,380 So in this lesson, we're doing it from scratch. 5 00:00:15,390 --> 00:00:19,770 We're not going to be loading a pre-trained model like we did with TensorFlow hub. 6 00:00:20,280 --> 00:00:25,530 And we're going to basically build the components from scratch like we did in the second half of that 7 00:00:25,530 --> 00:00:26,280 last lesson. 8 00:00:26,880 --> 00:00:28,110 So let's take a look. 9 00:00:28,170 --> 00:00:33,750 So I'm not going to read through all of these explanation notes here, but lengthy, and you can do 10 00:00:33,750 --> 00:00:34,920 that on your own time. 11 00:00:35,310 --> 00:00:37,590 So let's import these functions that we'll be using. 12 00:00:38,070 --> 00:00:43,490 You can see we're using the regular watch functions, as well as some towards vision models. 13 00:00:43,500 --> 00:00:48,300 That's that's that gives us the pre-trained Kijiji model that we will we will be using of that pre-treatment 14 00:00:48,300 --> 00:00:48,630 model. 15 00:00:48,640 --> 00:00:53,730 We're not going to be using specific, stylized model like we used in TensorFlow Arab. 16 00:00:54,600 --> 00:01:00,960 Let's double check that we're using our GPU and we just print this out. 17 00:01:04,820 --> 00:01:05,870 We are using energy for you. 18 00:01:05,900 --> 00:01:06,290 Great. 19 00:01:07,010 --> 00:01:13,760 Now let's load our test images here of we're going to take a look at so the content image is going to 20 00:01:13,760 --> 00:01:14,630 be a difference. 21 00:01:15,020 --> 00:01:22,730 In this case, I wanted to show you that we can actually apply neural style transfer to a style, an 22 00:01:22,730 --> 00:01:27,490 auto image so you can copy the artistic style of one image onto another image. 23 00:01:27,500 --> 00:01:29,570 So we'll take a look and I'll show you what I mean. 24 00:01:30,410 --> 00:01:33,020 You can use any of the any images you want here. 25 00:01:33,320 --> 00:01:36,950 Just point a URL to it to upload it to your GitHub and point of view URL here. 26 00:01:37,790 --> 00:01:43,880 So I'm using the one on the PyTorch tutorial site to demonstrate this because I think it's a good example 27 00:01:43,880 --> 00:01:44,810 and it works well. 28 00:01:45,350 --> 00:01:49,610 So this is the image loader here where we convert into a pill image. 29 00:01:50,060 --> 00:01:54,170 This is a plotting function, and let's take a look at those test images. 30 00:01:54,650 --> 00:02:00,260 So this is a content image, and we're going to apply this artistic style onto this image. 31 00:02:00,470 --> 00:02:02,510 So that should lead to a very cool result. 32 00:02:03,140 --> 00:02:06,170 So now we have to define our lost functions. 33 00:02:06,680 --> 00:02:14,090 So, you know, we have our content loss here, which we create does a small little function that calculates 34 00:02:14,090 --> 00:02:16,430 the MSI loss between the input and the target. 35 00:02:17,310 --> 00:02:20,240 Then we have a star loss along with our grand matrix. 36 00:02:20,240 --> 00:02:24,650 So in this one, we are using the grand matrix and you can see and PyTorch. 37 00:02:25,010 --> 00:02:28,020 It's actually fairly easy to implement some of these functions. 38 00:02:28,040 --> 00:02:29,540 Not as confusing as to Carrasco. 39 00:02:29,550 --> 00:02:30,050 No, is it? 40 00:02:31,790 --> 00:02:34,540 So now we can take a look at star loss here. 41 00:02:34,550 --> 00:02:36,560 So this is a Star Wars class. 42 00:02:36,560 --> 00:02:40,700 I should say that it keeps track of both of these things with the ground matrix. 43 00:02:41,330 --> 00:02:45,140 And now we just import or model set it to evaluation mode. 44 00:02:45,650 --> 00:02:52,670 So we download a model from Dubai to its repo and then now we just have a normalization means a standard 45 00:02:52,670 --> 00:02:59,810 deviation for the image net pre-processing, and we create another normalization function here. 46 00:03:00,680 --> 00:03:04,310 And then, no, this is a big class that this is what it's doing. 47 00:03:04,790 --> 00:03:06,290 Most of the heavy lifting here. 48 00:03:06,920 --> 00:03:10,880 We basically run the the neural cell transfer algorithm right here. 49 00:03:11,390 --> 00:03:15,680 So this is going to step true for all of the different layers we want to specify. 50 00:03:16,250 --> 00:03:21,650 And this is going to basically get the content layers, get the star layers and then trim off the layers 51 00:03:21,650 --> 00:03:25,430 of last content and style and return our model. 52 00:03:25,520 --> 00:03:28,610 So this basically is the core of our network. 53 00:03:29,300 --> 00:03:30,260 So let's run that. 54 00:03:30,950 --> 00:03:38,060 And no and no, we can just take a look at our test image just again did with image, just so content 55 00:03:38,060 --> 00:03:38,390 image. 56 00:03:39,020 --> 00:03:40,760 And here we have a optimizer. 57 00:03:40,760 --> 00:03:44,580 We're using the LBC's algorithm, so optimizer. 58 00:03:44,580 --> 00:03:48,230 It's recommended for neural cell transfer to get these people. 59 00:03:49,400 --> 00:03:55,430 And then we can now finally step through the network because we had the class above that basically extracted 60 00:03:55,430 --> 00:03:59,150 all the layers that we're working with and now and and losses. 61 00:03:59,150 --> 00:04:03,380 And no, we can actually compute everything with our training steps. 62 00:04:03,380 --> 00:04:09,770 So that's run that we have everything in our PyTorch setup optimizes. 63 00:04:09,770 --> 00:04:10,370 You agreed. 64 00:04:11,090 --> 00:04:12,710 So everything should be fine. 65 00:04:12,980 --> 00:04:21,870 So let's create this and now we can do to run out and run style transfer function and implement our 66 00:04:21,890 --> 00:04:23,870 new rules to transfer using by touch. 67 00:04:24,030 --> 00:04:24,890 So let's take a look. 68 00:04:26,910 --> 00:04:28,500 You can see it's running fine. 69 00:04:29,310 --> 00:04:30,750 We don't see an output yet. 70 00:04:31,320 --> 00:04:37,590 So we're running for the default of 200 steps here, and these are some of the content with some weight 71 00:04:37,590 --> 00:04:39,600 ratios and you can change as well. 72 00:04:40,140 --> 00:04:45,370 So you can see after running for those 200 iterations, we get this nicely generated image here. 73 00:04:45,900 --> 00:04:46,890 So that's pretty cool. 74 00:04:47,130 --> 00:04:49,410 And you can experiment with different weights to try it. 75 00:04:50,190 --> 00:04:52,270 But this one actually looks quite good, in my opinion. 76 00:04:52,290 --> 00:04:56,910 It looks like it actually was painted in that style that autistic. 77 00:04:56,910 --> 00:04:57,240 So. 78 00:04:57,840 --> 00:05:04,380 So that's it for the PyTorch implementation of neural cell transfer and the next lesson, we'll start 79 00:05:04,380 --> 00:05:10,140 taking a look at auto incluidas and then micho it again, which is a really cool part of the course. 80 00:05:10,680 --> 00:05:12,000 So I'll see you then. 81 00:05:12,210 --> 00:05:12,480 But.