1 00:00:11,730 --> 00:00:17,320 At this point we finally understand how to fully build a convolution or neural network in python. 2 00:00:17,580 --> 00:00:21,830 Now we're ready to outline the upcoming code as usual. 3 00:00:21,840 --> 00:00:25,920 I'm going to repeat the steps of our analysis just to be sure. 4 00:00:25,950 --> 00:00:30,260 Step number one is to load in the data for the following examples. 5 00:00:30,270 --> 00:00:37,320 We'll be using more difficult datasets in particular fashion amnesty and see 410 fashioned amnesty was 6 00:00:37,320 --> 00:00:43,620 designed to be a drop in replacement for the amnesty data set which today is considered to be too easy 7 00:00:43,620 --> 00:00:49,790 for modern machine learning algorithms one might consider it to be a solved problem. 8 00:00:50,170 --> 00:00:56,560 Thus fashion amnesty was born and it was designed to be a standard benchmark just like amnesty but with 9 00:00:56,560 --> 00:01:04,640 the goalpost further ahead just like amnesty it's a data set of 28 by 28 grayscale images and it even 10 00:01:04,640 --> 00:01:07,220 has the same number of samples. 11 00:01:07,250 --> 00:01:13,020 The difference is that instead of simply handwritten digits we have images of different types of clothing. 12 00:01:13,190 --> 00:01:17,240 For example t shirts shoes pants and so forth. 13 00:01:17,240 --> 00:01:21,450 Your job is to classify these images into the correct category. 14 00:01:21,560 --> 00:01:25,580 The CFR 10 data set is older but more difficult. 15 00:01:25,580 --> 00:01:31,670 This dataset contains color images so the data takes up more ram and the images are also a bit bigger 16 00:01:31,820 --> 00:01:33,690 32 by 32. 17 00:01:33,800 --> 00:01:38,930 It contains classes such as automobile frog horse cat and dog. 18 00:01:38,930 --> 00:01:44,900 You can already imagine that with such tiny images it might be very easy to mistake a cat for a dog 19 00:01:44,900 --> 00:01:46,130 or a dog for a horse 20 00:01:51,340 --> 00:01:53,410 step number two is to build a model. 21 00:01:54,040 --> 00:01:58,660 Luckily we just discussed that in painstaking detail over the past few lectures. 22 00:01:58,750 --> 00:02:02,320 In case you forgot we'll be using convolution or neural networks 23 00:02:07,500 --> 00:02:09,690 step number three is to train the model. 24 00:02:09,690 --> 00:02:15,240 Step number four is to evaluate the model and step number five is to use the model to make predictions 25 00:02:16,580 --> 00:02:17,120 just like the. 26 00:02:17,120 --> 00:02:18,050 And then section. 27 00:02:18,050 --> 00:02:19,470 Nothing here changes. 28 00:02:19,490 --> 00:02:22,270 These steps are model agnostic. 29 00:02:22,520 --> 00:02:28,970 As per my rule all machine learning interfaces are the same one new thing you'll learn about later on 30 00:02:28,970 --> 00:02:34,730 in this section is what to do if your data set is too large to fit into memory or if you want to augment 31 00:02:34,730 --> 00:02:35,990 your data. 32 00:02:35,990 --> 00:02:42,180 For example we know that a dog facing left is a dog but a dog facing right is also a dog. 33 00:02:42,260 --> 00:02:48,140 So should our CNN not be able to recognize both of these orientations as a dog. 34 00:02:48,140 --> 00:02:49,910 The answer is yes. 35 00:02:49,910 --> 00:02:52,100 We also know that deep learning loves data. 36 00:02:52,100 --> 00:02:53,810 The more data the better. 37 00:02:53,810 --> 00:03:00,620 This is why deep learning has come to prominence in recent years so data augmentation is a method of 38 00:03:00,620 --> 00:03:04,460 virtually adding more data without actually having it. 39 00:03:04,490 --> 00:03:05,950 We'll discuss this more later on. 40 00:03:05,980 --> 00:03:09,140 But for our initial example we're going to keep things simple 41 00:03:14,290 --> 00:03:17,490 so let's elaborate on step number one loading in the data. 42 00:03:18,220 --> 00:03:23,710 Luckily both the data sets we'll be using for these examples are also included in torch vision. 43 00:03:23,890 --> 00:03:28,870 It's generally going to follow the same pattern as before you can even probably try to guess the function 44 00:03:28,870 --> 00:03:32,110 names and you might be right for a fashion M.A.. 45 00:03:32,110 --> 00:03:39,440 We load in the data by calling torture vision that data sets that fashion feminist for a CFR 10 and 46 00:03:39,440 --> 00:03:44,020 we load in the data by calling torch vision datasets that see 410. 47 00:03:44,140 --> 00:03:45,370 Nothing too surprising there. 48 00:03:48,310 --> 00:03:53,830 Inside the training loop we can loop over the data sets in batches by using the data loader which you 49 00:03:53,830 --> 00:04:01,460 guessed it looks exactly the same as before. 50 00:04:01,470 --> 00:04:02,790 So what changes. 51 00:04:03,060 --> 00:04:05,920 Other than changing the model and changing the data functions. 52 00:04:05,940 --> 00:04:07,290 Nothing really. 53 00:04:07,320 --> 00:04:10,140 So this is a great time to remind you of my two rules. 54 00:04:10,260 --> 00:04:14,220 All data is the same and all machine learning interfaces are the same 55 00:04:19,330 --> 00:04:20,410 so just as a quiz. 56 00:04:20,410 --> 00:04:25,390 Let's think about which of these steps will be the same and which will be different. 57 00:04:25,390 --> 00:04:29,910 First the training loop as you know we'll be doing Batch gradient descent. 58 00:04:30,070 --> 00:04:32,640 Will this be the same or different. 59 00:04:32,680 --> 00:04:35,710 The answer is the same next question. 60 00:04:36,220 --> 00:04:39,430 Let's consider evaluating the accuracy of the model. 61 00:04:39,460 --> 00:04:41,900 Will this be the same or different. 62 00:04:41,920 --> 00:04:45,700 The answer is the same next question. 63 00:04:45,700 --> 00:04:50,900 Let's consider plotting the confusion matrix given the model predictions and model targets. 64 00:04:50,920 --> 00:04:53,380 Will this be the same or different. 65 00:04:53,380 --> 00:04:54,610 The answer is the same 66 00:04:59,690 --> 00:05:05,000 the only minor non-trivial thing I want to highlight is that the classes are going to be encoded as 67 00:05:05,000 --> 00:05:07,700 integers from 0 up to 9 inclusive. 68 00:05:08,180 --> 00:05:11,660 But how do we know which integer corresponds to which class. 69 00:05:11,660 --> 00:05:14,960 For example what if zero is dress and one is teacher. 70 00:05:14,960 --> 00:05:16,100 Is that obvious. 71 00:05:16,100 --> 00:05:17,400 Of course not. 72 00:05:17,420 --> 00:05:24,030 I just did the easy thing and look them up on the Internet and hardcoded them as strings this way when 73 00:05:24,030 --> 00:05:26,050 we look at the misclassified samples. 74 00:05:26,100 --> 00:05:31,500 It won't say something like 0 is misclassified as one which doesn't really mean anything but rather 75 00:05:31,640 --> 00:05:35,100 t shirt as misclassified as Drescher which is more interpretable.