1 00:00:00,910 --> 00:00:10,930 Now, what we will do is we will take the, um, well, this example and we will apply it using PyTorch. 2 00:00:10,930 --> 00:00:14,560 We will see how simple it is to do so. 3 00:00:14,560 --> 00:00:18,520 So we first start with this, um. 4 00:00:20,250 --> 00:00:26,250 Just, uh, Jupyter Notebook and we call it, let's say basic. 5 00:00:27,950 --> 00:00:28,310 Three. 6 00:00:28,310 --> 00:00:31,760 This is the third or I don't know, we can call it actually back. 7 00:00:32,660 --> 00:00:35,960 For back propagation. 8 00:00:36,760 --> 00:00:38,440 And main. 9 00:00:39,260 --> 00:00:42,410 So we start with X. 10 00:00:43,260 --> 00:00:44,130 X and y. 11 00:00:44,160 --> 00:00:48,180 Of course x or not actually import import. 12 00:00:50,040 --> 00:00:51,030 Torch. 13 00:00:52,830 --> 00:00:57,660 And we start with defining x x equals. 14 00:00:58,960 --> 00:01:00,040 Torch. 15 00:01:00,760 --> 00:01:02,290 Dot tensor. 16 00:01:03,410 --> 00:01:05,630 Of the value. 17 00:01:06,640 --> 00:01:09,580 Of the value of x, which is. 18 00:01:10,980 --> 00:01:12,690 Which is one. 19 00:01:14,070 --> 00:01:17,910 And why is going to be. 20 00:01:19,800 --> 00:01:20,880 Equals. 21 00:01:22,910 --> 00:01:26,420 The tensor of value of two. 22 00:01:29,000 --> 00:01:34,520 On the other hand, w w is a little bit. 23 00:01:35,580 --> 00:01:36,870 A different. 24 00:01:37,050 --> 00:01:38,910 Let's let's see how we will identify. 25 00:01:39,120 --> 00:01:41,880 It will be torch dot. 26 00:01:45,690 --> 00:01:46,530 Tensor. 27 00:01:46,560 --> 00:01:47,550 Tensor. 28 00:01:48,030 --> 00:01:50,970 And then we will give it an initial value. 29 00:01:50,970 --> 00:01:52,860 In our case we give it one. 30 00:01:53,220 --> 00:01:58,710 And then we will say require requires gradient. 31 00:01:58,710 --> 00:02:00,870 Gradient equals true. 32 00:02:02,310 --> 00:02:07,110 And we shift enter if you print these values. 33 00:02:07,110 --> 00:02:12,150 So x print y print. 34 00:02:13,660 --> 00:02:15,760 W we see how they look. 35 00:02:16,180 --> 00:02:22,030 X and y is just a tensor and w is a tensor, but it says required gradient. 36 00:02:22,030 --> 00:02:29,350 So in the operations we will do to the w it will compute the d w over. 37 00:02:29,350 --> 00:02:33,490 You know, the thing we are connecting it with in our case is the loss. 38 00:02:34,610 --> 00:02:38,150 Now what we need to do is to do the operations. 39 00:02:38,150 --> 00:02:39,830 We said that. 40 00:02:40,950 --> 00:02:42,090 Operations is. 41 00:02:42,090 --> 00:02:51,000 The first one is w multiplied x will equal y hat y hat minus y will equal s, and s squared will equal 42 00:02:51,000 --> 00:02:51,840 loss. 43 00:02:51,840 --> 00:02:53,760 So we do the same thing. 44 00:02:54,240 --> 00:03:06,000 Y hat y hat equals w multiplied x s equals w or y hat. 45 00:03:06,950 --> 00:03:13,910 Minus y and loss will equal s. 46 00:03:15,810 --> 00:03:16,470 Squared. 47 00:03:17,170 --> 00:03:20,230 So these are just the operations we have done. 48 00:03:22,110 --> 00:03:27,600 And then we put shift enter and then we have the actual loss. 49 00:03:27,660 --> 00:03:35,760 So we said the first we go forward uh operation and then we can calculate the loss. 50 00:03:35,760 --> 00:03:38,460 So how much is the loss in this case. 51 00:03:38,460 --> 00:03:43,830 Well actually we already did the computation and the already computed this value. 52 00:03:43,830 --> 00:03:48,390 So we just, uh, print and we just print the loss. 53 00:03:48,630 --> 00:03:49,830 What do we have? 54 00:03:49,830 --> 00:03:52,800 We have the loss equals one. 55 00:03:52,830 --> 00:04:00,330 Of course, if we want to just show the value item and we just put like this and we will get it. 56 00:04:00,480 --> 00:04:01,350 Uh, like this. 57 00:04:01,350 --> 00:04:04,050 Of course the uh, loss is also a tensor. 58 00:04:04,560 --> 00:04:07,080 So here is we put it an item. 59 00:04:07,080 --> 00:04:11,100 And now we are having the forward pass value. 60 00:04:11,100 --> 00:04:12,870 How about the backward pass. 61 00:04:12,870 --> 00:04:14,850 Well, in PyTorch is very simple. 62 00:04:14,850 --> 00:04:16,860 We just one command. 63 00:04:16,860 --> 00:04:24,810 We just say loss dot back forward and that's it. 64 00:04:25,260 --> 00:04:26,850 And shift enter. 65 00:04:27,540 --> 00:04:28,470 That's it. 66 00:04:29,120 --> 00:04:32,750 But what does it what did it do it actually? 67 00:04:32,750 --> 00:04:34,460 Or how did the. 68 00:04:35,130 --> 00:04:37,620 Loss and W got calculated. 69 00:04:37,620 --> 00:04:43,290 Actually, we use the loss to make a backward, uh, propagation. 70 00:04:43,290 --> 00:04:51,630 And then he used these operations to go back to uh w which has the gradient equals to true. 71 00:04:52,250 --> 00:04:58,550 If we have more gradients, then it will calculate the loss and how everything is, uh, how the loss 72 00:04:58,550 --> 00:05:00,830 will change with every W we have. 73 00:05:00,830 --> 00:05:05,090 So print let's, let's see the results w. 74 00:05:05,090 --> 00:05:08,210 But this time we need to say gradient. 75 00:05:08,950 --> 00:05:10,690 And then we shift into. 76 00:05:10,930 --> 00:05:17,290 Of course we can just also print w like this, but we will see only the value of w. 77 00:05:17,320 --> 00:05:22,060 In order to get the gradient value we just put dot gradient. 78 00:05:22,210 --> 00:05:27,160 So with this way of course we have this value. 79 00:05:27,160 --> 00:05:29,740 Of course as a tensor we can also item. 80 00:05:30,520 --> 00:05:32,890 And do like this, but I prefer. 81 00:05:32,920 --> 00:05:34,450 Okay, let's let's do. 82 00:05:35,570 --> 00:05:37,190 The Or, let's put it. 83 00:05:38,030 --> 00:05:42,950 Two print statements because usually it's better just this way. 84 00:05:43,280 --> 00:05:53,810 So minus two, which is d loss over d w is what we calculated by hand which is this one minus two.