1 00:00:00,060 --> 00:00:00,570 Hi, guys. 2 00:00:00,600 --> 00:00:01,950 Welcome back to the course. 3 00:00:02,460 --> 00:00:06,540 This is going to be our first object detection lesson. 4 00:00:07,230 --> 00:00:08,400 So let's get started. 5 00:00:08,490 --> 00:00:15,270 So the first model we're going to trend this object detector is a skilled YOLO v4, and this one is 6 00:00:15,270 --> 00:00:16,040 a very cool one. 7 00:00:16,240 --> 00:00:18,510 I started with a pretty cool project, in my opinion. 8 00:00:19,020 --> 00:00:22,440 This one is a gun and pistol detection detector. 9 00:00:23,070 --> 00:00:24,660 So let's take a look. 10 00:00:24,870 --> 00:00:30,900 So open up Notebook 42, which is highlighted here, and that'll bring up this window. 11 00:00:31,560 --> 00:00:33,230 So this is a lesson here. 12 00:00:33,240 --> 00:00:41,280 So firstly, like I mean, I've mentioned before all of these notebooks adopted from rebel floor, they 13 00:00:41,280 --> 00:00:46,140 may not be the latest notebooks that are available right now in rubber floor site because they're constantly 14 00:00:46,470 --> 00:00:49,350 tweaking and making little changes to these notebooks. 15 00:00:49,800 --> 00:00:54,480 So what I did this was back in 2021, believe November. 16 00:00:55,290 --> 00:01:01,020 I captured those notebooks and I may have made some minor changes because some of those things, some 17 00:01:01,020 --> 00:01:03,440 of their notebooks are broken. 18 00:01:03,450 --> 00:01:07,110 Unfortunately, they do get around to fixing it quite quick. 19 00:01:07,140 --> 00:01:12,450 So if you do encounter one of the newer ones that are broken, you can just send them a message and 20 00:01:12,450 --> 00:01:14,220 they'll probably update it quite quickly. 21 00:01:15,600 --> 00:01:24,000 So this one here is a skilled healer before, and it's going to be trained on the Pistol Pistols dataset. 22 00:01:24,660 --> 00:01:29,790 As you can see here, this is an example of how the dataset looks, and we have pictures of guns where 23 00:01:29,790 --> 00:01:32,580 we have the bounding boxes annotated around it. 24 00:01:32,670 --> 00:01:33,820 This is our training data. 25 00:01:33,840 --> 00:01:38,910 This is what we'll use to create our very own gun and pistol detector. 26 00:01:39,540 --> 00:01:44,340 So you can see how all of these things, these the guns here in this footage just as a robbery. 27 00:01:45,090 --> 00:01:48,990 Most of these look like studio or movie shots or stock images. 28 00:01:49,500 --> 00:01:54,570 Nevertheless, that's still useful because they all have guns in them, and you're not going to encounter 29 00:01:54,570 --> 00:01:59,040 a real life gun dataset unless someone put in the works to build one. 30 00:01:59,160 --> 00:02:00,030 It's possible. 31 00:02:00,030 --> 00:02:07,410 It's a lot of footage on YouTube of like CCTV robberies, especially in the US, where there is no gun 32 00:02:07,410 --> 00:02:07,980 control. 33 00:02:09,150 --> 00:02:10,890 That's not a political opinion. 34 00:02:10,890 --> 00:02:12,180 That's just a fact. 35 00:02:12,810 --> 00:02:14,640 So let's start this notebook. 36 00:02:15,660 --> 00:02:17,820 So one of them, the first is not books. 37 00:02:18,330 --> 00:02:20,110 They they're a bit heavy. 38 00:02:20,130 --> 00:02:26,820 There's a lot of setup going on because using your look for you'll five all of these abductors. 39 00:02:27,300 --> 00:02:32,160 They're not actually meant to be run in notebooks because they got a lot of things going on in the background. 40 00:02:32,640 --> 00:02:39,180 A lot of parts that need to be build a lot of different configurations, so it looks a bit messy when 41 00:02:39,180 --> 00:02:40,110 it's in a notebook. 42 00:02:40,200 --> 00:02:46,890 But nevertheless, this is perhaps the easiest way to get started with this object detector notebooks. 43 00:02:47,520 --> 00:02:51,720 So I'm not going to run this code here because right now I've run these before. 44 00:02:52,320 --> 00:02:57,900 I will just say this takes these few blocks cells out of code. 45 00:02:58,770 --> 00:03:00,960 Some of them do take maybe a couple minutes to run. 46 00:03:01,650 --> 00:03:02,610 Shouldn't be that long. 47 00:03:02,620 --> 00:03:05,100 The longest one might be this one, I believe. 48 00:03:05,730 --> 00:03:08,050 But you can see this is the output here. 49 00:03:08,070 --> 00:03:09,510 This is what you should be seeing. 50 00:03:10,050 --> 00:03:11,570 We have to install PI. 51 00:03:11,760 --> 00:03:12,150 No. 52 00:03:12,870 --> 00:03:15,280 Then we get to navigate to the directory. 53 00:03:15,840 --> 00:03:17,550 Then we have to download the data set. 54 00:03:17,560 --> 00:03:19,590 So we have the instructions here. 55 00:03:20,040 --> 00:03:25,920 So the format we're going to be using is the YOLO v five PI torch export format, which I've already 56 00:03:25,920 --> 00:03:26,670 done here. 57 00:03:27,000 --> 00:03:31,470 This is the well, this is the this part doesn't really work actually anymore. 58 00:03:31,740 --> 00:03:33,040 Let's just to live this. 59 00:03:33,780 --> 00:03:39,150 So I've actually uploaded the dataset to Google Drive in some cases in a sets of go down from river 60 00:03:39,150 --> 00:03:42,000 forward to change the format of their hosting. 61 00:03:42,510 --> 00:03:46,860 So we get the unzip sort of the pistols here and we unzip it. 62 00:03:47,580 --> 00:03:56,640 Then we just set up the windmill file and then we just inspect the architecture by doing cat to. 63 00:03:56,930 --> 00:04:01,170 This is displaced the architecture of the yellow freeform scaled model. 64 00:04:01,710 --> 00:04:03,390 You can see it's here. 65 00:04:03,420 --> 00:04:05,610 You can see the backbone of the anchors. 66 00:04:06,180 --> 00:04:09,060 This is a number of classes that we're retraining on here. 67 00:04:09,090 --> 00:04:09,270 Well. 68 00:04:09,510 --> 00:04:11,250 That's what it was pre-trained on that. 69 00:04:11,250 --> 00:04:15,270 Actually, we're going to be training on one class in a way that students dataset. 70 00:04:15,960 --> 00:04:17,700 And this is the head of Uber. 71 00:04:17,700 --> 00:04:23,130 For her, this is a useful tool in the CSP darknet backbone, I believe. 72 00:04:23,310 --> 00:04:29,310 And now here's where we train the other model, so you can see if we can pass a number of these input 73 00:04:29,310 --> 00:04:33,740 arguments here using the OK path style. 74 00:04:34,290 --> 00:04:40,920 So you can see we run the Python file like this, but exclamation so that we can run a Python file. 75 00:04:41,880 --> 00:04:43,740 Time actually just times it's performance. 76 00:04:43,740 --> 00:04:47,040 Pretty useful metric to look at when you're training models. 77 00:04:47,580 --> 00:04:52,560 So we set the image size, how many batches of images we're going to use per batch. 78 00:04:52,560 --> 00:04:55,130 So sixteen number of epochs. 79 00:04:55,140 --> 00:04:59,520 Then we set the directory of the Y.A. data file. 80 00:05:00,300 --> 00:05:03,930 Then we set the model, the sort of the model configuration. 81 00:05:04,620 --> 00:05:08,310 Then we said to which we're not going to use between weights, we're training from scratch. 82 00:05:08,850 --> 00:05:14,720 And then we just put two results here, and cash means that we're going to put the data in your. 83 00:05:15,810 --> 00:05:18,870 So this is a GPO will be recruiting it on. 84 00:05:20,290 --> 00:05:22,250 And now here we go. 85 00:05:22,260 --> 00:05:26,880 So it's training, so you can see it takes quite some time. 86 00:05:27,150 --> 00:05:35,130 I've been training this for maybe 15, 20 minutes now and you can see it's already reached at AT&T Park 87 00:05:35,130 --> 00:05:38,520 and that is with using the GPU and club. 88 00:05:38,520 --> 00:05:44,790 And I have a cooler approach which may give me slightly faster GPU compared to what you might be using, 89 00:05:44,790 --> 00:05:46,080 although I can't say for sure. 90 00:05:46,770 --> 00:05:49,560 So you can see it's taking roughly a minute to beatbox. 91 00:05:49,560 --> 00:05:50,250 Not that bad. 92 00:05:51,660 --> 00:05:56,520 So this is what you can expect after you finish training so you can actually launch a tensor board. 93 00:05:56,970 --> 00:06:03,810 And measures to its training performance is quite useful to use if you want to make it persistent so 94 00:06:03,810 --> 00:06:05,310 that you will receive the results. 95 00:06:05,760 --> 00:06:11,400 I would encourage you to use something like weights and biases, which is a really good library and 96 00:06:11,400 --> 00:06:13,950 tool to keep storing your training results. 97 00:06:13,950 --> 00:06:17,700 It works seamlessly when you'll have five for multiple attacks. 98 00:06:18,240 --> 00:06:20,340 So I would encourage you to use that. 99 00:06:20,430 --> 00:06:24,450 This because this tensor mode is not going to remain after we exit the notebook. 100 00:06:24,460 --> 00:06:28,230 I mean, it will be there, but the actual explicit results wouldn't be there. 101 00:06:28,680 --> 00:06:32,940 And then it'll be difficult for you to compare experiments when you have two different runs. 102 00:06:33,100 --> 00:06:33,390 Yeah. 103 00:06:35,480 --> 00:06:37,460 So you can see this is another metric. 104 00:06:37,490 --> 00:06:40,630 This is from the actually all four results directory. 105 00:06:40,630 --> 00:06:45,470 You can look at the IOU, the object in this, which I'm not even sure what that is. 106 00:06:45,980 --> 00:06:48,170 You can look at precision and recall always good. 107 00:06:48,290 --> 00:06:54,410 The map scores are very good to look at as well, because these take into consideration bounding boxes 108 00:06:54,410 --> 00:06:56,120 and classification scores. 109 00:06:56,540 --> 00:07:01,910 Precision and recall again, she didn't mention, but they're always very useful to look at. 110 00:07:02,060 --> 00:07:08,810 So the four boxes on the right here are perhaps the most useful metrics to look at when treating these 111 00:07:08,810 --> 00:07:09,290 models. 112 00:07:09,920 --> 00:07:15,500 And I agree, because you can directly compare map, especially these to map scores with other models 113 00:07:15,500 --> 00:07:21,170 as well, because they tend to all use maps two point five and a map of the range point five two point 114 00:07:21,170 --> 00:07:22,550 ninety five as well. 115 00:07:23,720 --> 00:07:25,490 So let's take a look at this. 116 00:07:26,030 --> 00:07:28,490 All right, let's visualize some of our training data here. 117 00:07:28,520 --> 00:07:30,410 So this is the 22 we saw previously. 118 00:07:30,980 --> 00:07:31,400 No. 119 00:07:31,430 --> 00:07:37,670 Let's take a look at some of the actual results from our train from training on model. 120 00:07:38,180 --> 00:07:40,460 You can see it doesn't have a class theme like gun, a pistol. 121 00:07:40,460 --> 00:07:42,260 It just has class zero, which is OK. 122 00:07:42,290 --> 00:07:48,560 You can always rename those things later on and you can see it's getting the gun in almost every photo 123 00:07:48,560 --> 00:07:48,750 here. 124 00:07:48,770 --> 00:07:49,940 This is actually quite good. 125 00:07:49,940 --> 00:07:52,550 I'm actually quite impressed with this performance. 126 00:07:53,330 --> 00:07:59,600 It definitely has learnt the shape of a gun in all of these photos, and the bombing boxes are quite 127 00:07:59,600 --> 00:08:00,170 good as well. 128 00:08:00,200 --> 00:08:02,000 However, actually, this isn't the only one here. 129 00:08:02,360 --> 00:08:04,260 It didn't get these two rifles here. 130 00:08:04,400 --> 00:08:05,800 That's a bit surprising. 131 00:08:05,820 --> 00:08:11,150 Maybe doesn't have enough training data, or maybe it's just trained on pistols and not rifles. 132 00:08:11,500 --> 00:08:12,950 I actually haven't checked. 133 00:08:14,060 --> 00:08:19,810 I believe it missed part of this gun here because this is a cropped image right now, and you can see 134 00:08:19,810 --> 00:08:22,670 it doesn't get the gun here, but that's still not too bad. 135 00:08:23,450 --> 00:08:28,700 So if he wants to run an inference windows string widths, these are the words that we've trained so 136 00:08:28,700 --> 00:08:29,030 far. 137 00:08:29,030 --> 00:08:34,820 You can see it's towards the words after every epoch here, and this basically is probably the last 138 00:08:34,940 --> 00:08:35,500 we had to. 139 00:08:35,810 --> 00:08:42,690 And this is a one who probably want to use the best wits and you can see just the drunk rendered early 140 00:08:42,920 --> 00:08:43,550 detection. 141 00:08:44,000 --> 00:08:47,360 All you have to do is store some images in a test over here. 142 00:08:48,150 --> 00:08:49,220 Set these parameters. 143 00:08:49,220 --> 00:08:50,660 This is a confidence threshold. 144 00:08:51,020 --> 00:08:55,760 If you want to lose your confidence threshold like point one, you're probably going to get a lot of 145 00:08:55,760 --> 00:08:56,720 false positives. 146 00:08:56,720 --> 00:09:02,720 However, you set it too high at twenty seven point six, you're going to miss a lot of guns and hence 147 00:09:02,720 --> 00:09:03,890 have a bad recoil. 148 00:09:04,520 --> 00:09:06,560 So let's give and take point four. 149 00:09:06,920 --> 00:09:11,390 I would say 0.3 to point six is usually a good value to use. 150 00:09:11,900 --> 00:09:18,020 More often than not, I tend to use lower confidence thresholds because it depends on the application. 151 00:09:18,020 --> 00:09:23,630 But sometimes you just want to try to get as much of the detectors going, and you don't care too much 152 00:09:23,630 --> 00:09:24,740 about false positives. 153 00:09:24,740 --> 00:09:28,190 But that really does depend on your application. 154 00:09:29,390 --> 00:09:36,170 And you can see we are loading the best model with its here and we're running detect so detected as 155 00:09:36,170 --> 00:09:42,230 a python file in the area of reform model that runs detections on these images here. 156 00:09:42,680 --> 00:09:46,220 So we can see it runs on all of these images in test folder. 157 00:09:46,220 --> 00:09:50,360 Here you can see it's quite a bit of a tiff that runs pretty quick, to be honest. 158 00:09:50,660 --> 00:09:53,810 When zero, two or three seconds, that's almost real time, in my opinion. 159 00:09:53,990 --> 00:09:56,930 Almost 20 frames a second believe actually, it would be. 160 00:09:58,130 --> 00:10:03,290 And then you can see you can display some of the outputs here, so you can take a look at some of the 161 00:10:03,290 --> 00:10:04,000 images here. 162 00:10:04,090 --> 00:10:08,570 You can see it auto getting the guns correctly, which is quite good. 163 00:10:11,450 --> 00:10:15,680 The French are going to sequence pictures that don't look too difficult. 164 00:10:16,400 --> 00:10:22,130 Maybe there will be some more difficult ones lower down, hopefully because it will look a bit repetitive 165 00:10:22,130 --> 00:10:22,520 here. 166 00:10:24,770 --> 00:10:28,910 But nevertheless, we are getting the guns down great. 167 00:10:29,600 --> 00:10:37,700 So that concludes this lesson on the yellow v4 skilled object detector trained on the guns and arms 168 00:10:37,700 --> 00:10:40,440 data sets if you wanted to export your words to download. 169 00:10:41,000 --> 00:10:43,220 You can just run this line here as well. 170 00:10:44,420 --> 00:10:45,730 Always good to keep track. 171 00:10:45,740 --> 00:10:50,990 Keep track of your training day to hear your training progress, I should say so you can see what's 172 00:10:50,990 --> 00:10:54,970 going on to know if it's 17 epochs and things to look for. 173 00:10:54,980 --> 00:10:57,990 You can see the scores different metrics here. 174 00:10:58,010 --> 00:11:06,140 My point, Dave, my point five point nine five point five Precision Recall targets, classes. 175 00:11:06,660 --> 00:11:07,970 All this good stuff here. 176 00:11:08,060 --> 00:11:12,920 All the metrics actually are quite here right now, and you can measure it so you can just take a look 177 00:11:12,920 --> 00:11:15,050 and see if these metrics are getting better all the time. 178 00:11:15,530 --> 00:11:18,230 And you can see you can look at the Map Point five score. 179 00:11:18,740 --> 00:11:20,420 You can see it's two point fifty five now. 180 00:11:20,900 --> 00:11:26,510 And initially we started at point where it's a point five zero two, I guess. 181 00:11:26,660 --> 00:11:28,760 Yeah, it looks like it could be boxier. 182 00:11:29,570 --> 00:11:31,370 So that's it for this lesson. 183 00:11:31,370 --> 00:11:32,270 I hope you enjoyed it. 184 00:11:32,690 --> 00:11:39,170 In the next lesson, we'll take a look at mass detection using TensorFlow Object Detection Library. 185 00:11:39,800 --> 00:11:46,400 So my selection meeting the Alison King machines initially in my head but is actually to mass like those 186 00:11:46,760 --> 00:11:48,830 face masks to wear because of the pandemic. 187 00:11:49,430 --> 00:11:53,120 Well, we're going to do a mass detector next, so stay tuned for that. 188 00:11:53,300 --> 00:11:53,720 Thank you.