1 00:00:00,360 --> 00:00:07,440 Hi and welcome back in this lesson, we'll take a look at using PyTorch to implement a Siamese network 2 00:00:07,440 --> 00:00:11,940 that can compare faces, so it's a very cool project, so let's get started. 3 00:00:12,630 --> 00:00:17,730 So open notebook too late, which I've already done here, and we can begin to listen. 4 00:00:18,360 --> 00:00:25,530 So this lesson has a lot of small building blocks, so I'll try to go for them as detailed as I can. 5 00:00:25,620 --> 00:00:31,710 So firstly, we're going to import our modules and then we're going to import our data and some utility 6 00:00:31,710 --> 00:00:32,430 functions. 7 00:00:32,730 --> 00:00:35,570 So let's just run this first block of code at Hustle. 8 00:00:35,640 --> 00:00:38,250 Follow the inputs that we'll be using. 9 00:00:40,170 --> 00:00:43,080 Next, we're going to create some utility functions here. 10 00:00:43,200 --> 00:00:45,570 One is an IM show just to show the image. 11 00:00:46,000 --> 00:00:47,810 Another one is to show the plot. 12 00:00:47,820 --> 00:00:53,400 That's a plot for doing training to represent, to show a training loss and accuracy loss. 13 00:00:53,910 --> 00:00:55,690 Next will download our data. 14 00:00:55,770 --> 00:01:00,090 This is a source of the data here, so you can just click this link if you want to check out the source. 15 00:01:00,260 --> 00:01:06,090 The AT&T database of faces and you can expand this and read all about it here. 16 00:01:06,480 --> 00:01:13,860 So the images 1982 by a hundred and twelve pixels and it's a grayscale image, so we can just take a 17 00:01:13,860 --> 00:01:15,480 look at that after. 18 00:01:15,570 --> 00:01:17,750 So let's look at our data here. 19 00:01:22,910 --> 00:01:25,400 OK, that was pretty quick, just took about five seconds. 20 00:01:26,030 --> 00:01:32,780 Next, we create a config class here, so it is still a training directory, training tested directory 21 00:01:33,380 --> 00:01:38,390 of that size and the number of epochs with the training for which is 100, which is quite a lot. 22 00:01:39,350 --> 00:01:41,990 Now we have a Siamese network dataset. 23 00:01:41,990 --> 00:01:47,660 This is what creates our image pairs here, so we just need 50 percent of the images to be in the same 24 00:01:47,660 --> 00:01:48,320 class. 25 00:01:48,650 --> 00:01:51,500 So this creates the image pins right there. 26 00:01:53,000 --> 00:01:57,620 Next, we set out image folders here using image for that right now. 27 00:01:58,340 --> 00:02:04,580 Next, we create or transform sort of transforms will be using or just resizing it to one hundred by 28 00:02:04,580 --> 00:02:06,650 100 and converting it to 10. 29 00:02:06,650 --> 00:02:09,470 So and no one getting here. 30 00:02:10,040 --> 00:02:11,420 So let's go. 31 00:02:12,260 --> 00:02:17,450 So now we can create our data loaders here from the same dataset that we've created right here that 32 00:02:17,450 --> 00:02:19,670 transforms will be created as well. 33 00:02:20,360 --> 00:02:26,930 So we shuffle them, said the number of workers to it, although that might be a bit high for collab. 34 00:02:26,940 --> 00:02:30,650 So maybe we can try for, although it should default to forfeit. 35 00:02:30,650 --> 00:02:33,790 Can't get it to believe that's how it should working with Potlatch. 36 00:02:34,520 --> 00:02:42,950 And then we have a terrible data here so we can take a look at the plots we use to watch visions, utils 37 00:02:42,950 --> 00:02:49,220 and make grid to just create a simple plot of some of the some of the sample images so you can see them 38 00:02:49,220 --> 00:02:49,550 there. 39 00:02:50,210 --> 00:02:52,760 These are all three sides to 100 by 100. 40 00:02:55,340 --> 00:02:58,100 Now we can start building out Siamese network. 41 00:02:58,670 --> 00:03:00,560 So let's take a look at this network. 42 00:03:00,920 --> 00:03:07,340 Basically, we use something called reflection, but to this pads the tensor using the reflection of 43 00:03:07,340 --> 00:03:08,480 the input boundary. 44 00:03:08,930 --> 00:03:14,630 Then we have a convoluted velu, then a batch norm, then another reflection pattern. 45 00:03:15,230 --> 00:03:18,490 Then we have another con's would descend with eight filters. 46 00:03:18,500 --> 00:03:23,120 Initially, it was four kernel sizes poetry and then we have. 47 00:03:23,240 --> 00:03:27,860 So we have tree convalesce here of increasing amount. 48 00:03:27,950 --> 00:03:31,400 And so we actually know it's about the same hand to see it. 49 00:03:32,090 --> 00:03:33,820 And then we have our sequentially. 50 00:03:34,010 --> 00:03:36,860 This produces our embedding. 51 00:03:36,980 --> 00:03:43,040 So the embedding, the final embedding is a one dimensional vector that's five hundreds of one by 500. 52 00:03:43,730 --> 00:03:45,380 And then we create this loop here. 53 00:03:45,590 --> 00:03:48,440 Well, listen, not this loop is going to create a forward pass here. 54 00:03:48,770 --> 00:03:50,330 So this is called forward once. 55 00:03:50,750 --> 00:03:55,280 And the reason we have forward ones is because we use it to create output one output two, which is 56 00:03:55,280 --> 00:03:58,610 basically the two same system networks. 57 00:03:58,880 --> 00:04:00,200 So let's run this. 58 00:04:01,160 --> 00:04:06,860 Next, we can define our contrast of most function or class two effectively, and then the function 59 00:04:06,860 --> 00:04:07,580 isn't so different. 60 00:04:07,580 --> 00:04:10,370 That creates two that gives us the contrast of loss. 61 00:04:11,300 --> 00:04:17,270 Now we create our training data loader and then set about contrast to the loss criterion, as well as 62 00:04:17,270 --> 00:04:19,880 a optimizer which pleasing Adam. 63 00:04:20,900 --> 00:04:25,790 And once that's done, we can actually start training in the training loop is actually quite simple 64 00:04:25,790 --> 00:04:32,540 compared to the usual PyTorch training grids, training plot loops that you see. 65 00:04:33,590 --> 00:04:34,270 You can see this. 66 00:04:34,540 --> 00:04:36,440 This is here where we run everything. 67 00:04:37,010 --> 00:04:44,150 So we send over two images here, two batches to the network here, and then we just get the loss out 68 00:04:44,150 --> 00:04:47,260 of it and an update gradients as well. 69 00:04:47,270 --> 00:04:52,220 And then we just plot print some results every 10 iterations. 70 00:04:52,430 --> 00:04:54,020 So let's load this. 71 00:04:55,130 --> 00:04:59,840 We'll run this, I should say, and you can see we're training for 100 epochs and it's running quite 72 00:04:59,840 --> 00:05:00,350 fast. 73 00:05:00,950 --> 00:05:03,200 So let's just wait for this to be complete. 74 00:05:23,840 --> 00:05:27,560 All right, so we're finished training and we can see a nice training plot here. 75 00:05:27,980 --> 00:05:33,200 And you can see loss went down to quite low point zero zero six and that's a reasonably good score. 76 00:05:33,770 --> 00:05:38,660 Now let's see how this network performs when we actually compare some fierce, fierce images. 77 00:05:39,140 --> 00:05:43,700 So we'll load some of the testing set here, and then it will create that. 78 00:05:43,700 --> 00:05:48,350 We created the same dataset again right here, and then we create the test data loader. 79 00:05:48,920 --> 00:05:52,790 We just pass to get a batch and then we just get zero here. 80 00:05:53,330 --> 00:05:56,570 So we just put that batch into it here to compare. 81 00:05:56,600 --> 00:06:02,990 So it concatenated, and then we just get the pairwise distance out of it, and then we may use torture 82 00:06:02,990 --> 00:06:08,030 machines to make the plot, the grid, the visual display. 83 00:06:08,690 --> 00:06:10,610 And we just use everything here. 84 00:06:10,610 --> 00:06:14,570 So we get particularly Euclidean distance to similarity score in it. 85 00:06:15,140 --> 00:06:19,280 So let's run that and see how it performance is a pretty cool idea. 86 00:06:20,090 --> 00:06:24,050 So let's take a look at this so you can see the first one. 87 00:06:24,620 --> 00:06:30,930 We get a high dissimilarity score of three point zero six and there are two different people. 88 00:06:31,470 --> 00:06:35,360 However, in the next one, we get a lower dissimilarity school under the same person. 89 00:06:35,570 --> 00:06:36,740 It's quite cool, isn't it? 90 00:06:37,340 --> 00:06:40,930 Next expected high school the similar. 91 00:06:41,780 --> 00:06:43,040 This one is in the similar. 92 00:06:43,040 --> 00:06:48,140 And again, even though they do appear to look similar, they are different people, obviously, and 93 00:06:48,140 --> 00:06:49,460 that work is picking up on that. 94 00:06:50,390 --> 00:06:52,910 Again, good good results. 95 00:06:53,750 --> 00:06:54,340 Good result. 96 00:06:54,350 --> 00:06:55,880 Again, no mistake. 97 00:06:55,890 --> 00:07:01,610 So far, it's all pretty good results that we're comparing this guy with a multitude of guys, including 98 00:07:01,610 --> 00:07:06,710 himself at different angles, because you can see definitely is a different angle for this and this, 99 00:07:06,710 --> 00:07:10,040 and he gets a very low dissimilarity score. 100 00:07:10,640 --> 00:07:11,720 So that's it. 101 00:07:11,930 --> 00:07:15,380 That concludes this lesson and hope you enjoyed it. 102 00:07:16,220 --> 00:07:17,450 Thanks for watching. 103 00:07:17,450 --> 00:07:17,750 But.