Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,690 --> 00:00:06,930 Hello and welcome to this patented Hoyo today we're going to create the SSD neural network. 2 00:00:06,930 --> 00:00:08,890 So we already defined our detect function. 3 00:00:08,890 --> 00:00:15,330 So now the main thing that we have to do left is to indeed create this as is the neural network. 4 00:00:15,360 --> 00:00:21,540 But the good news is that we already have the weights of a pre-trained as is the neural network. 5 00:00:21,570 --> 00:00:26,550 So what will simply do is create an object that will represent the neural network itself. 6 00:00:26,700 --> 00:00:34,260 Thanks to the build SSD the function that we import from this SSD pipe and fallar recommend to have 7 00:00:34,260 --> 00:00:35,390 a look at this. 8 00:00:35,550 --> 00:00:41,520 And then after we create this neural network object we will get the weights by loading from an already 9 00:00:41,520 --> 00:00:43,680 pre-trained as is the neural network. 10 00:00:43,860 --> 00:00:45,840 The weights are contained in this file. 11 00:00:45,870 --> 00:00:50,880 We're going to get these weights by using torture load which is a function of torch and that will open 12 00:00:50,880 --> 00:00:57,650 a tensor that will contain these weights and then using another function to load state dict function. 13 00:00:57,660 --> 00:01:04,100 We will attribute these loaded weights to our instance object of our neural network. 14 00:01:04,140 --> 00:01:05,730 So that's exactly the process. 15 00:01:05,730 --> 00:01:06,930 It's not that hard. 16 00:01:06,930 --> 00:01:09,550 We will do it in three or four lines of code. 17 00:01:09,780 --> 00:01:10,780 So let's do this. 18 00:01:10,860 --> 00:01:16,030 The first thing that we need to do is to create our neural network object. 19 00:01:16,200 --> 00:01:21,990 So we're going to call this new one that work not just to align with this variable here but we could 20 00:01:21,990 --> 00:01:26,010 actually choose another name it's just less confusing this way. 21 00:01:26,190 --> 00:01:32,210 And now as we just said we're going to use to build as is the function that is a function from the SSD 22 00:01:32,210 --> 00:01:33,250 pattern file. 23 00:01:33,330 --> 00:01:37,280 So we call this function build underscore SSD. 24 00:01:37,500 --> 00:01:41,430 And in this function we actually have to input only one argument. 25 00:01:41,490 --> 00:01:45,470 This argument is the face you have two possible phases. 26 00:01:45,510 --> 00:01:47,590 The train phase and the test phase. 27 00:01:47,890 --> 00:01:54,240 But here since we're going to use an already pre-trained model bellowing its weight well we're not going 28 00:01:54,240 --> 00:01:55,520 to train anything. 29 00:01:55,560 --> 00:02:01,020 We're just going to test actually because we're going to test as is the model on our Video of the funny 30 00:02:01,020 --> 00:02:01,570 dog. 31 00:02:01,800 --> 00:02:08,820 So the face we have to choose here is test and we just put it this way in quotes test just like that 32 00:02:09,270 --> 00:02:10,230 and that's it. 33 00:02:10,440 --> 00:02:16,130 Our neural network our SSD neural network is created with this single line of code. 34 00:02:16,410 --> 00:02:17,630 Awesome. 35 00:02:17,640 --> 00:02:24,060 Then next step as we said the next step is to load the weights of an already pre-trained SSD neural 36 00:02:24,060 --> 00:02:30,990 network and the name of this already pre-trained neural network is exactly SSD 300 and this core map 37 00:02:31,230 --> 00:02:37,270 underscores seventy seven point forty three underscore V-2 that's the name of the neural network. 38 00:02:37,380 --> 00:02:38,530 That's a powerful one. 39 00:02:38,560 --> 00:02:44,220 It was pre-trained to detect between 30 to 40 objects and you're going to see that on the video that 40 00:02:44,220 --> 00:02:46,150 it's not only going to detect the drug. 41 00:02:46,350 --> 00:02:47,670 It's going to detect more. 42 00:02:47,820 --> 00:02:53,520 So actually this model and this will be part of the homework you will have to test this model on other 43 00:02:53,520 --> 00:02:55,650 videos containing other objects. 44 00:02:55,830 --> 00:02:58,670 So that's the best pre-trained model we could find. 45 00:02:58,770 --> 00:03:05,250 And believe me or believe the paper is actually more powerful than some other great object detection 46 00:03:05,250 --> 00:03:08,810 Morell's like faster or CNN or yellow. 47 00:03:09,120 --> 00:03:14,820 According to the paper that is according to the cases tested by the paper the SSD is the most powerful. 48 00:03:14,830 --> 00:03:19,120 A mom says the first of our CNN and yellow. 49 00:03:19,170 --> 00:03:19,740 All right. 50 00:03:19,740 --> 00:03:21,670 So let's slow this wait. 51 00:03:21,780 --> 00:03:26,790 And to load these weights we just need to take our new network object and then we're going to use the 52 00:03:26,790 --> 00:03:35,060 load underscore state underscored dict function and inside this function we exact input. 53 00:03:35,460 --> 00:03:49,260 Well in quotes our pre-trained model says the 300 this core and a P underscore 77 dot 43 underscore 54 00:03:49,350 --> 00:03:52,250 the two dot p t h. 55 00:03:52,620 --> 00:03:59,040 But these weights we're going to put them into a tensor and therefore inside the load state dict function 56 00:03:59,400 --> 00:04:06,520 I'm going to call another function which I already mentioned that is the torch that load function. 57 00:04:06,630 --> 00:04:14,100 So the load function from the torture library and this towards stop load function will open a tensor 58 00:04:14,340 --> 00:04:20,460 that will contain these weights and then the use of the load stated function is to attribute these weights 59 00:04:20,580 --> 00:04:23,670 to our esset the neural network object. 60 00:04:23,670 --> 00:04:29,910 All right so that's almost ready we just need to add two more arguments in our torch's upload function. 61 00:04:29,910 --> 00:04:40,320 The first one is map underscore location and that should be equal to lambda storage and a third argument 62 00:04:40,680 --> 00:04:44,610 luck for location storage. 63 00:04:44,610 --> 00:04:48,700 All right so that's just the way to open a center that will contain these weights. 64 00:04:48,840 --> 00:04:53,840 And so now not only do we have a sensor that contains these weights but also these weights or attribute 65 00:04:53,840 --> 00:04:57,200 it to our as is the net object. 66 00:04:57,210 --> 00:04:57,570 All right. 67 00:04:57,570 --> 00:05:02,720 Now the neural network SSD single shot multi-post detection is created. 68 00:05:02,880 --> 00:05:08,790 So that means that we have the frames coming from the video and we have our neural network nets but 69 00:05:08,790 --> 00:05:11,630 remember to apply the detect function on the frames. 70 00:05:11,850 --> 00:05:18,030 We not only need the frames and the net but we also need the transform transformation and that's exactly 71 00:05:18,030 --> 00:05:20,450 what we're going to create in the next tutorial. 72 00:05:20,510 --> 00:05:26,400 We're going to create this transformation that will make sure that the input frames coming from the 73 00:05:26,400 --> 00:05:33,200 video and that will be the input of the function will be compatible with our SSD neural network. 74 00:05:33,210 --> 00:05:35,920 So let's just create this last thing we need. 75 00:05:35,970 --> 00:05:38,260 That is a transformation in the next tutorial. 76 00:05:38,280 --> 00:05:40,110 And until then you can build a vision. 8247