1 00:00:00,270 --> 00:00:07,830 So we're now ready to train our CNN to CNN that we just constructed a pill, let's begin treating it. 2 00:00:07,920 --> 00:00:15,410 So an increase it's actually quite simple to train CNN's well, neural networks compared to PyTorch. 3 00:00:16,140 --> 00:00:18,810 That's a given by batting with PyTorch, to be honest. 4 00:00:19,140 --> 00:00:21,570 It allows you some nitty-gritty control. 5 00:00:22,140 --> 00:00:26,910 Some good python type style language that we can actually interact with. 6 00:00:26,910 --> 00:00:28,380 Get the opposite inputs. 7 00:00:28,470 --> 00:00:30,750 Perform operations in real time and loop. 8 00:00:31,260 --> 00:00:31,980 It's quite nice. 9 00:00:32,580 --> 00:00:37,290 However, wood carries a lot of that is abstracted away, and you can see what I mean here. 10 00:00:37,860 --> 00:00:43,710 Once we have a model that's compiled here, all we have to do is no specified about a specified number 11 00:00:43,710 --> 00:00:47,160 of epochs and we just do model dot fit. 12 00:00:47,790 --> 00:00:54,600 We have history equals a model that fit only because the model not fit after every iteration of every 13 00:00:54,600 --> 00:00:54,960 epoch. 14 00:00:54,960 --> 00:01:01,380 I should say it outputs its towards tick accuracy to validation, lost validation, accuracy and training 15 00:01:01,380 --> 00:01:03,720 loss us so we can plot this afterward. 16 00:01:04,200 --> 00:01:05,220 So it's a good thing. 17 00:01:05,880 --> 00:01:06,960 So don't don't worry about that. 18 00:01:06,990 --> 00:01:14,020 So in the model that fit function, we simply input our training data or treating labels, specify what 19 00:01:14,040 --> 00:01:20,520 that size, which is 128 we're going to use specify epochs verbose, which is a mode where we display 20 00:01:20,520 --> 00:01:21,330 this information. 21 00:01:21,750 --> 00:01:23,640 If you said that to zero, you get less information. 22 00:01:23,640 --> 00:01:27,150 Instead of two, you get even more information during the training process. 23 00:01:27,780 --> 00:01:29,440 I think it's just three modes right now. 24 00:01:29,460 --> 00:01:30,240 Zero one two. 25 00:01:30,840 --> 00:01:34,710 And then we specify our validation data in this tuple here. 26 00:01:35,100 --> 00:01:37,500 This is the testino, and this is a test levels. 27 00:01:38,280 --> 00:01:44,010 And then in the end, we can actually simply just do Model three, evaluate give it to tested the test 28 00:01:44,010 --> 00:01:44,490 labels. 29 00:01:44,970 --> 00:01:50,760 So the verbose to zero by, I think it's by default zero anyway, and we can get the score out of it. 30 00:01:51,150 --> 00:01:54,930 The score gives you in scoring first index gives us a test loss. 31 00:01:55,350 --> 00:01:57,720 Second index gives us the test accuracy. 32 00:01:58,230 --> 00:02:00,150 So this is quite simple. 33 00:02:00,570 --> 00:02:02,550 So now let's begin training this model. 34 00:02:04,110 --> 00:02:08,520 Oops model wasn't defined, and that was because we didn't run this block of code previously. 35 00:02:10,080 --> 00:02:12,750 And that's also because we didn't run the code above it. 36 00:02:13,500 --> 00:02:18,510 Sorry about that, but it's a good practice to keep running these blocks of good. 37 00:02:19,560 --> 00:02:22,830 Make sure you ran sequentially because they depend on each other. 38 00:02:24,390 --> 00:02:26,970 And now we do need to run that. 39 00:02:27,420 --> 00:02:30,660 Now we can start training our model, and there we go. 40 00:02:31,110 --> 00:02:32,430 So it's beginning to first book. 41 00:02:32,820 --> 00:02:37,170 Pretty soon you'll start seeing the output in real time as it goes through each batch. 42 00:02:37,680 --> 00:02:38,190 There we go. 43 00:02:38,580 --> 00:02:40,110 So what are some of his here? 44 00:02:40,170 --> 00:02:42,780 This number is a number of batches in our data. 45 00:02:43,620 --> 00:02:51,990 Remember, we set it at 128 well, 5th 60000, divided by 128 is 469. 46 00:02:52,560 --> 00:02:57,120 So that's why you have four hundred and sixty nine as total number of batches here. 47 00:02:57,420 --> 00:03:01,440 So you can see how quickly it's going through all the batteries in the data, taking the three seconds 48 00:03:01,440 --> 00:03:04,560 of each one, which is actually much faster than our PyTorch model. 49 00:03:05,010 --> 00:03:08,100 And I will see the PyTorch model in that. 50 00:03:08,100 --> 00:03:12,930 In that proof, his example was slower, even though we're using the same hardware and GBU just because 51 00:03:12,930 --> 00:03:14,430 of how we were doing things in the loop. 52 00:03:14,970 --> 00:03:17,790 There's other ways with Pi too we can make things much faster. 53 00:03:18,150 --> 00:03:20,370 But for now, that was just a basic example. 54 00:03:20,640 --> 00:03:22,830 So I didn't want to complicate things too much for you guys. 55 00:03:25,680 --> 00:03:33,030 One thing you wouldn't notice still is that after 10 epochs, the Python model was actually more accurate. 56 00:03:33,610 --> 00:03:39,330 We're at 92 percent here with Keros, but if I, it was at ninety seven point eight percent accuracy, 57 00:03:39,330 --> 00:03:45,420 and that probably has to do with the normalization that we with zero scented invito, which you can 58 00:03:45,420 --> 00:03:51,720 see how it actually does make a difference where the convergence and at the end of the twenty five epochs 59 00:03:51,720 --> 00:03:58,320 would still finished and quite in a very short space of time, we get ninety five percent accuracy in 60 00:03:58,320 --> 00:03:58,620 the end. 61 00:03:59,340 --> 00:04:06,770 And you can see the output, the Tesla says accuracy that came from the score, this variable and that 62 00:04:06,870 --> 00:04:08,670 came from a model that evaluates here. 63 00:04:09,810 --> 00:04:11,370 So we'll stop there for now. 64 00:04:11,760 --> 00:04:15,810 And in the next section, we'll take a look at plotting our loss and accuracy charts. 65 00:04:16,830 --> 00:04:17,820 See you in the next section. 66 00:04:17,970 --> 00:04:18,420 Thank you.