1 00:00:00,940 --> 00:00:02,050 Hi and welcome back. 2 00:00:02,800 --> 00:00:08,110 We're about to take a look at denseness, which basically effectively resonates, but better. 3 00:00:08,590 --> 00:00:09,460 So let's get started. 4 00:00:09,550 --> 00:00:18,760 So dense that was introduced in 2016 by a joint effort by the Cornwall University, Shinhwa University 5 00:00:18,760 --> 00:00:20,500 and Facebook's EIA research team. 6 00:00:20,920 --> 00:00:28,090 And in fact, denseness won best people in 2017 CPR Conference, which is the Premier Computer Vision 7 00:00:28,090 --> 00:00:29,440 Conference for publishing. 8 00:00:29,440 --> 00:00:30,490 So that's a big deal. 9 00:00:31,090 --> 00:00:36,340 And it was able to obtain an even higher accuracy than resonate with fewer parameters. 10 00:00:36,880 --> 00:00:40,060 This is just a preview of the architecture we'll we'll be discussing shortly. 11 00:00:40,540 --> 00:00:41,770 So how did it do that? 12 00:00:41,860 --> 00:00:47,170 Well, basically, it solves the same problem that residents of which is a vanishing gradient problem. 13 00:00:47,650 --> 00:00:53,950 And as we know, just to do a recap when networks get very, very big, having a very small gradient 14 00:00:54,250 --> 00:01:00,010 in that network can basically cause the gradients along that chin to become zero, which is not good. 15 00:01:00,160 --> 00:01:06,010 It basically means that you're kind of wasting parameters, wasting computation time at that point. 16 00:01:06,430 --> 00:01:12,700 So then, since we're able to solve this concept by introducing something called collective knowledge, 17 00:01:13,060 --> 00:01:16,690 where each layer receives information from two previous layers. 18 00:01:16,690 --> 00:01:18,850 So let's take a look and see how that's implemented. 19 00:01:19,330 --> 00:01:23,230 So this here is all classical scene in architecture. 20 00:01:23,800 --> 00:01:24,520 You remember this? 21 00:01:24,520 --> 00:01:29,020 This was the feature maps, feature map outputs and blah blah blah. 22 00:01:29,350 --> 00:01:33,250 We have okunnu layers to reload and all of that stuff. 23 00:01:33,640 --> 00:01:35,560 So you should be familiar with this. 24 00:01:37,180 --> 00:01:43,090 Now, this is a resonant architecture resonance basically introduced at Short-Circuit here, so you 25 00:01:43,090 --> 00:01:44,770 don't always get vanishing gradients. 26 00:01:45,100 --> 00:01:47,950 So what did dense density do differently? 27 00:01:48,220 --> 00:01:49,240 Well, let's take a look. 28 00:01:49,990 --> 00:01:53,140 Dense that's connect to have multiple connections so you can see. 29 00:01:53,140 --> 00:01:54,760 Let's follow the path of this read. 30 00:01:54,760 --> 00:01:57,880 One can see it connects here just like a resonant short-circuit. 31 00:01:58,360 --> 00:02:03,250 Then there's a yellow pad that goes all the way to here, and then there's a green one that goes all 32 00:02:03,250 --> 00:02:06,070 the way to here and a pink one that goes all the way to here. 33 00:02:06,490 --> 00:02:12,100 So you can see this input is received at all the intermediate CNN layers. 34 00:02:12,880 --> 00:02:15,580 So each layer receives information from the previous layers. 35 00:02:15,580 --> 00:02:20,410 And the growth rate, which is the additional number of channels for each layer, is controlled by a 36 00:02:20,410 --> 00:02:21,580 parameter called K. 37 00:02:23,050 --> 00:02:24,610 So that's not all there is to then. 38 00:02:24,670 --> 00:02:29,020 That's the basic dense and composition layer has this foundation here. 39 00:02:29,440 --> 00:02:31,600 They have a pre activation batch norm. 40 00:02:33,070 --> 00:02:37,240 And then the second layer with Crabtree filters, that's basically it. 41 00:02:37,600 --> 00:02:40,580 However, they also introduce something called bottleneck layer. 42 00:02:40,600 --> 00:02:42,340 So what are these bottleneck layers? 43 00:02:42,940 --> 00:02:48,310 Well, remember we have batch known Mary-Lou, and then it's followed by a one by one layer here. 44 00:02:48,790 --> 00:02:51,100 And then that's a bottleneck point right there. 45 00:02:51,100 --> 00:02:54,610 And then we can have a batch known and a tree by tree columns here. 46 00:02:55,090 --> 00:03:00,640 So you can see this sort of like a squeeze or the file module similar to the file modulo, squeezing 47 00:03:00,650 --> 00:03:03,580 it sort of a squeeze and a squeeze expand layer. 48 00:03:04,060 --> 00:03:05,200 It's actually quite similar. 49 00:03:05,740 --> 00:03:12,220 So this was the architecture diagram used to buddy that Transnet researchers and the people, and you 50 00:03:12,220 --> 00:03:14,230 could kind of get a feeling of what it is here. 51 00:03:14,230 --> 00:03:18,820 It's not that descriptive, but basically, here's a note I put in. 52 00:03:19,180 --> 00:03:23,080 There's one by one converters, followed by a two by two averaging pool. 53 00:03:23,470 --> 00:03:29,170 And these are used as transition layers between two contiguous, contiguous dense blocks. 54 00:03:29,650 --> 00:03:34,990 So you can see they have a block here block here that is followed by a two by two average pooling. 55 00:03:35,500 --> 00:03:36,790 And that's the transition layers. 56 00:03:36,910 --> 00:03:45,370 So this here then Senate introduced several networks of different depths this that 121, 169 to one 57 00:03:45,370 --> 00:03:46,500 and 264. 58 00:03:46,510 --> 00:03:47,950 You can see that it's quite deep. 59 00:03:48,700 --> 00:03:52,600 And growth rate is controlled by a but I mentioned for you guys. 60 00:03:53,500 --> 00:03:55,060 So let's take a look at performance now. 61 00:03:55,060 --> 00:04:01,060 You can see the error rate for theinternet was quite good and you can see it going down as network layers 62 00:04:01,060 --> 00:04:01,640 increase. 63 00:04:02,110 --> 00:04:06,370 Similar for the top five or it, you can see it went down to quite low values here. 64 00:04:06,940 --> 00:04:12,340 But let's compare this to Reznik now, given the number of parameters given less parameters here. 65 00:04:12,730 --> 00:04:18,070 You can see denseness performed quite similarly, basically better with less parameters. 66 00:04:18,070 --> 00:04:24,190 Even like look at this one, denseness 121 is getting an error rate of twenty five point five, whereas 67 00:04:24,220 --> 00:04:27,160 resonated for IS getting error rate of twenty six point five. 68 00:04:27,700 --> 00:04:30,910 Accuracy reversed, sensitive using accuracy, we use error rate. 69 00:04:31,360 --> 00:04:38,590 I do tend to prefer to use accuracy, but some researchers keep this standard going forward, so let's 70 00:04:38,590 --> 00:04:40,000 just stay with it anyway. 71 00:04:40,030 --> 00:04:44,560 So basically, you can see the lowest error rate here for theInternet 264. 72 00:04:44,830 --> 00:04:50,530 It still performs better than a resonant, which has twice the amount of parameters. 73 00:04:51,070 --> 00:04:57,070 Similarly, for flops, which is the computational inference number of calculations that it does in 74 00:04:57,120 --> 00:04:57,850 inference. 75 00:04:58,390 --> 00:05:00,250 So you can see denseness do perform. 76 00:05:00,550 --> 00:05:04,300 They're faster than resonates and as well as more accurate. 77 00:05:04,780 --> 00:05:06,610 So that's quite an achievement. 78 00:05:07,420 --> 00:05:12,640 So that's it for a tour of all these CNN architectures. 79 00:05:13,180 --> 00:05:20,590 These this should give you a very good grasp and understanding of complicated state of the art scene 80 00:05:20,590 --> 00:05:23,290 and architecture designs and why they work. 81 00:05:23,830 --> 00:05:29,840 So what we'll do now, we'll stop there, even though there's more to this, which was expanded for 82 00:05:29,920 --> 00:05:30,490 the lessons. 83 00:05:30,940 --> 00:05:34,090 I'll talk about things like So we'll stop here for now. 84 00:05:34,120 --> 00:05:39,130 That basically ends the tour of the state of the art CNN architectures. 85 00:05:39,520 --> 00:05:41,290 Not all of these were state of the art. 86 00:05:41,320 --> 00:05:48,310 However, it is very important to understand the foundation of this the designs of these CNN's because 87 00:05:48,760 --> 00:05:54,400 many of them will be utilized in future networks as well as we take bits of Terry here and there that 88 00:05:54,400 --> 00:05:54,990 we've learned. 89 00:05:55,010 --> 00:05:56,770 Researchers have learned that this works. 90 00:05:57,250 --> 00:06:02,890 So it's very, very good that you understand this will also take a tour of vision transformers. 91 00:06:02,890 --> 00:06:05,110 However, I'll do that later on in the course. 92 00:06:05,620 --> 00:06:11,140 For now, though, what we'll do, we'll dive into the code and actually load some of these pre-trained 93 00:06:11,140 --> 00:06:13,270 networks and explore them. 94 00:06:13,780 --> 00:06:17,080 However, these networks are all trained well. 95 00:06:17,080 --> 00:06:22,000 The ones we can load with cameras of PyTorch are all trained on the image networks. 96 00:06:22,360 --> 00:06:26,500 Which brings me to the next chapter of what is image that we may have mentioned. 97 00:06:26,890 --> 00:06:33,040 It's the benchmark image dataset, computer vision dataset that we used to benchmark these models. 98 00:06:33,640 --> 00:06:36,760 However, we haven't really explored it or understood it. 99 00:06:37,180 --> 00:06:39,400 So that's what we'll do in the next section. 100 00:06:39,400 --> 00:06:44,680 And then after that, we'll go back into the code and start working with pre-trained models. 101 00:06:45,040 --> 00:06:49,780 Many of these pre-trained models like Feig and ception mobile net efficient net. 102 00:06:50,140 --> 00:06:55,780 So we'll take a tour of this and give you some hands on experience with loading and unloading some of 103 00:06:55,780 --> 00:07:01,150 the best networks and implementing them yourself and testing testing them on some images. 104 00:07:01,570 --> 00:07:03,700 So I'll see you in the next section. 105 00:07:03,820 --> 00:07:04,270 Thank you.