1 00:00:11,690 --> 00:00:16,670 In this lecture we are going to look at a code lab notebook that does transfer learning without data 2 00:00:16,700 --> 00:00:18,410 augmentation. 3 00:00:18,410 --> 00:00:21,470 This lecture is going to walk you through a prepared code lab notebook. 4 00:00:21,530 --> 00:00:27,230 Although a very good exercise which I always recommend is once you know how this is done to try and 5 00:00:27,230 --> 00:00:33,380 recreate it yourself with as few references as possible as usual you can look at the title of the notebook 6 00:00:33,620 --> 00:00:36,010 to determine what notebook we are currently looking at. 7 00:00:37,250 --> 00:00:44,580 And this notebook is called Pi torch transfer learning fast so our main goal in this script is to see 8 00:00:44,970 --> 00:00:52,140 if we pre compute the features a Will training B faster and B Well we get an unacceptable reduction 9 00:00:52,170 --> 00:00:57,960 in accuracy I'll assume that in this lecture you've already seen the code that does transfer learning 10 00:00:58,380 --> 00:00:59,610 with data augmentation. 11 00:00:59,640 --> 00:01:05,700 So I'm not going to explain any of the lines that I've already explained so at the top everything is 12 00:01:05,700 --> 00:01:07,310 the same as before. 13 00:01:07,560 --> 00:01:14,340 Same imports same data same process to move the data into the folders which are structured the right 14 00:01:14,340 --> 00:01:17,130 way for the image folder data set object 15 00:01:23,520 --> 00:01:24,160 OK. 16 00:01:24,200 --> 00:01:28,260 So when we get to the transforms you'll notice that we only have one. 17 00:01:28,310 --> 00:01:30,410 Why do we only need one. 18 00:01:30,410 --> 00:01:35,300 Well since we're not going to do any data augmentation we can apply the same transformations to both 19 00:01:35,300 --> 00:01:41,460 the train and test that and these are all non data augmentation and transformations 20 00:01:44,380 --> 00:01:48,290 so next we create our image folder objects. 21 00:01:48,350 --> 00:01:52,390 Next we create our data load our objects. 22 00:01:52,500 --> 00:01:55,560 Next I grab the pre train model with prearranged equals true 23 00:01:58,840 --> 00:02:00,970 so the next part is also new 24 00:02:05,210 --> 00:02:08,980 since the V-J operation is separate from the rest of the network. 25 00:02:09,080 --> 00:02:16,360 We can define a separate model for it so I'll define a class called Veggie features in the constructor. 26 00:02:16,370 --> 00:02:20,890 The only thing I need to do is save the VB model in the forward function. 27 00:02:20,900 --> 00:02:27,020 I take X and I pass it through the different parts of the e.g. one at a time until right before the 28 00:02:27,020 --> 00:02:28,530 final classifier. 29 00:02:28,730 --> 00:02:30,860 At that point I just flatten and then return 30 00:02:33,600 --> 00:02:39,370 next I instantiate this class and I call it VEGF next. 31 00:02:39,370 --> 00:02:45,830 As a sanity check just to make sure this works we pass in a random input output a shape one by twenty 32 00:02:45,830 --> 00:02:46,960 five thousand eighty eight. 33 00:02:47,090 --> 00:02:50,780 So it works. 34 00:02:50,800 --> 00:02:55,530 Next we're going to transform our data using the feature transformer that we just instantiated. 35 00:02:56,140 --> 00:02:58,490 So first we need to find the length of each dataset. 36 00:02:59,560 --> 00:03:02,830 Luckily we can use the lend function on the data set objects 37 00:03:07,440 --> 00:03:11,740 so we have entrained three thousand and test one as. 38 00:03:11,790 --> 00:03:15,330 Next we're gonna grab D the feature dimensionality. 39 00:03:15,330 --> 00:03:20,370 This is the size of the output of the VEGF transformer which is equal to twenty five thousand eighty 40 00:03:20,370 --> 00:03:26,050 eight. 41 00:03:26,050 --> 00:03:31,750 Next we're going to work on creating our tabular data set consisting of transform the features from 42 00:03:31,750 --> 00:03:33,320 the original data. 43 00:03:33,430 --> 00:03:37,510 Once we have this we no longer have to work with the original dataset. 44 00:03:37,630 --> 00:03:45,990 So we're going to start by instantiating empty arrays to store X train y train x test and Y test. 45 00:03:46,120 --> 00:03:51,850 Next we're going to set the device to the GP you and move the feeder transformer to the GP you 46 00:03:56,860 --> 00:03:59,670 next we're going to populate X train and Y training. 47 00:04:00,070 --> 00:04:02,470 We can do this by looping over the train at dataset. 48 00:04:03,370 --> 00:04:09,340 So on each iteration we start by moving the inputs to the GP you it's not necessary to move the targets 49 00:04:09,340 --> 00:04:13,540 that the GP you because we only want to make a prediction with the model. 50 00:04:13,840 --> 00:04:20,840 Next we get the output by passing the inputs through VEGF. 51 00:04:20,870 --> 00:04:23,340 Next we get the size of the output. 52 00:04:23,360 --> 00:04:28,700 Note that this is not equal to the back size since the number of samples may not divide evenly into 53 00:04:28,700 --> 00:04:30,070 the batch size. 54 00:04:30,110 --> 00:04:37,510 So the final batch is less than or equal to the batch size next we assign the features and the targets 55 00:04:37,540 --> 00:04:41,920 to X train and Y train at the current index ie. 56 00:04:42,160 --> 00:04:49,700 Then we increment I by the current that size and go to the next iteration of the loop so since we print 57 00:04:49,730 --> 00:04:55,730 ie on each iteration we can confirm that the number of samples in our new dataset is equal to the original 58 00:04:55,730 --> 00:04:56,800 number of images. 59 00:04:58,970 --> 00:05:03,570 All right so we can see that the final value of AI is three thousand which is the correct number of 60 00:05:03,570 --> 00:05:08,800 samples. 61 00:05:08,810 --> 00:05:16,520 Next we do the same thing for X test and Y test. 62 00:05:16,730 --> 00:05:22,460 So one thing that's interesting to check is the men in the max of the feature values. 63 00:05:22,460 --> 00:05:28,580 As you can see the max in X train is about sixty two which should seem weird to you because you should 64 00:05:28,580 --> 00:05:32,390 expect it values in a neuron that work to be somewhat normalized. 65 00:05:32,390 --> 00:05:35,870 Since we're using a real you activation at the minimum value with zero 66 00:05:39,330 --> 00:05:44,580 so anyway since the data is at this weird scale we're going to use a standard scalar object to transform 67 00:05:44,580 --> 00:05:49,150 it so we'll call the scale data X train to an X test to 68 00:05:53,600 --> 00:05:56,670 now because the next step is to do logistic regression. 69 00:05:56,720 --> 00:05:59,690 There's no actual reason we need to use pi torch. 70 00:05:59,740 --> 00:06:05,250 In fact we can just use the logistic regression that's built into socket learn so let's do that. 71 00:06:06,200 --> 00:06:12,500 As usual it's just fit and predict which I've taught you for free in the past so everybody knows this. 72 00:06:12,500 --> 00:06:18,950 And as you can see we do very well just as well as when we did use data augmentation in fact. 73 00:06:18,950 --> 00:06:30,470 So maybe data augmentation wasn't so useful in this instance. 74 00:06:30,480 --> 00:06:36,000 Next we're going to complete this script by doing transfer learning the usual way which is to use the 75 00:06:36,000 --> 00:06:38,130 same library for the classifier. 76 00:06:38,400 --> 00:06:44,700 So let's start by creating a linear model and the rest of these steps you've seen before so we move 77 00:06:44,700 --> 00:06:53,750 the model to the GP you create the loss and optimizer make some tensor data set objects create the data 78 00:06:53,750 --> 00:07:05,890 loaders define the training function call the training function blood loss preparation and calculate 79 00:07:05,890 --> 00:07:06,580 the accuracy 80 00:07:10,410 --> 00:07:12,450 and so we end up with pretty much the same result 81 00:07:15,440 --> 00:07:16,740 also important. 82 00:07:16,820 --> 00:07:20,390 Notice how much faster this is with data augmentation. 83 00:07:20,420 --> 00:07:27,050 Each epoch takes about 40 seconds without data augmentation each epoch is nearly instant about zero 84 00:07:27,050 --> 00:07:28,460 point two seconds. 85 00:07:28,580 --> 00:07:33,950 So that's a two hundred x speed up by not using data augmentation and not having to go through the pre 86 00:07:33,950 --> 00:07:35,630 train network on each pass.