1 00:00:00,300 --> 00:00:06,870 Hello, everyone, and welcome to this new and exciting session in which we are going to build our own 2 00:00:06,870 --> 00:00:08,820 YOLO like model. 3 00:00:08,850 --> 00:00:17,250 So from the paper we had already seen that we have this initial conv layers which are pre-trained on 4 00:00:17,250 --> 00:00:27,210 the image net data set such that they could be used to extract very useful features from our input images. 5 00:00:27,780 --> 00:00:38,010 And then this conv layers were followed by two fully connected layers which were designed in order to 6 00:00:38,010 --> 00:00:40,920 adapt to our problem of object detection. 7 00:00:41,340 --> 00:00:49,350 Now, given that we do not want to train this backbone year from scratch on the image net data set, 8 00:00:49,400 --> 00:00:54,210 we'll use an already pre-trained backbone, which is the rest net 50. 9 00:00:54,630 --> 00:01:03,130 Again, we have the output dimension defined as a number of classes plus five times B from the paper. 10 00:01:03,150 --> 00:01:04,500 B is given to B two. 11 00:01:04,530 --> 00:01:11,310 We've seen this already and then five because we have the probability of obtaining an object. 12 00:01:11,310 --> 00:01:16,440 And then for the remaining four we have the bound and box. 13 00:01:16,890 --> 00:01:22,370 So we have two of this bounding box predictions. 14 00:01:22,380 --> 00:01:29,100 That's why here we have two times 510 plus the number of classes which in our case is equal 20. 15 00:01:29,550 --> 00:01:32,930 Now we define this number of filters to be 512. 16 00:01:32,940 --> 00:01:34,440 So that's it. 17 00:01:34,440 --> 00:01:38,700 From year we have our full, complete model. 18 00:01:38,730 --> 00:01:40,620 You could take this off. 19 00:01:40,620 --> 00:01:47,580 You see that we have our pre trained race net, which is our backbone or our base model. 20 00:01:47,580 --> 00:01:54,090 And then we followed this up with several conflicts similar to what we have here in the paper. 21 00:01:54,090 --> 00:02:01,410 And then we have the global average pulling, which is what is given here in the paper. 22 00:02:01,680 --> 00:02:09,060 Now, one thing we should note about the average pulling is the fact that when we have inputs, let's 23 00:02:09,060 --> 00:02:23,340 say we have this seven by seven, then by let's say five, so we have one channel to three, four and 24 00:02:23,340 --> 00:02:23,970 then five. 25 00:02:23,970 --> 00:02:26,940 So let's suppose we have some of my seven, my five input. 26 00:02:26,970 --> 00:02:33,810 Now after going into the global average pulling, what we will have here is the averaging of each and 27 00:02:33,810 --> 00:02:39,870 every value or let's say pixel in each and every channel. 28 00:02:39,870 --> 00:02:46,440 So for this channel, for example, we will have one representative value for this channel will have 29 00:02:46,440 --> 00:02:48,440 one representative value, which is the average. 30 00:02:48,450 --> 00:02:53,700 So for this we would have the average, this would have the average and this will have the average. 31 00:02:53,700 --> 00:02:56,790 So we average all these values here. 32 00:02:57,870 --> 00:03:04,200 And the problem with this is information about object position is lost. 33 00:03:04,200 --> 00:03:10,770 And so instead of using this average pulling is preferable to use the flattening. 34 00:03:11,400 --> 00:03:17,770 And so what we do is we'll take this off from here and then we'll have flatten. 35 00:03:18,090 --> 00:03:18,720 Okay. 36 00:03:18,720 --> 00:03:26,190 So once we have that, just as with the paper you see here, we have the fully connected layer, which 37 00:03:26,190 --> 00:03:30,540 is this dense layer right here and then this other fully connected layer. 38 00:03:30,540 --> 00:03:33,090 So we have that and then we reshape. 39 00:03:33,900 --> 00:03:36,080 Now, this should be actually split. 40 00:03:36,090 --> 00:03:40,620 Let's take all this year copy and paste. 41 00:03:40,620 --> 00:03:48,750 So we're split by split by split or split by split by output dimension. 42 00:03:48,750 --> 00:03:50,550 So it's seven by seven by 30. 43 00:03:50,550 --> 00:03:53,220 So this is now our model. 44 00:03:53,760 --> 00:04:00,240 You can see we have a total of 53 million parameters, 30 trainable and 23 non trainable. 45 00:04:00,240 --> 00:04:02,950 That's from our rest net. 46 00:04:03,090 --> 00:04:09,730 We are then going to define our model checkpoint where our file path is this year. 47 00:04:09,750 --> 00:04:13,830 Then we're going to save only the weights, we're going to monitor the validation loss. 48 00:04:14,580 --> 00:04:22,170 We're going to obviously save the model which produces the minimum or the smallest validation loss, 49 00:04:22,170 --> 00:04:23,190 and that's it. 50 00:04:23,190 --> 00:04:26,010 We save the base, we save the best weights only. 51 00:04:26,010 --> 00:04:33,750 So we run that and then now we move to the scheduling here if the number of epochs is less than 40. 52 00:04:33,750 --> 00:04:40,080 So the first 40 epochs, we use a learning rate of one times ten to a negative three between 40 and 53 00:04:40,080 --> 00:04:40,470 80. 54 00:04:40,470 --> 00:04:45,630 We use a learning rate of five times 10 to -4, and then after that we use a learning rate of one times 55 00:04:45,630 --> 00:04:46,950 ten to the negative four. 56 00:04:46,950 --> 00:04:48,030 So that's it. 57 00:04:48,270 --> 00:04:53,190 We compile our model and then we start with the training. 58 00:04:53,970 --> 00:04:58,770 Now, after training for a few epochs, you will notice that the model starts to over fit. 59 00:04:59,020 --> 00:05:06,370 So in the next section, we are going to treat or use several techniques to help solve this or resolve 60 00:05:06,370 --> 00:05:08,080 the problem of overfitting. 61 00:05:08,380 --> 00:05:15,340 Now, we've been training for over 20 epochs and you could see clearly from the loss and the validation 62 00:05:15,340 --> 00:05:21,850 of the train loss and the validation loss that our model starts performing well and at some point starts 63 00:05:21,850 --> 00:05:22,750 overfitting. 64 00:05:22,780 --> 00:05:29,590 As you could see here, we have the training loss, which keeps dropping right here, and then the validation 65 00:05:29,590 --> 00:05:34,160 loss drops and then at some point starts increasing. 66 00:05:34,180 --> 00:05:37,480 So clearly our model is overfitting.