1 00:00:00,180 --> 00:00:06,750 Hi and welcome back in this lesson, we'll take a look at using cameras to train our one CNN model on 2 00:00:06,750 --> 00:00:12,310 the Catzavelos Dogs dataset, which is basically something you can apply to your own datasets as well. 3 00:00:12,330 --> 00:00:15,330 So open this notebook and we'll get started in this lesson. 4 00:00:15,990 --> 00:00:17,640 So there's a lot going on in this lesson. 5 00:00:17,640 --> 00:00:21,030 So I may be a bit a bit quick with some things, however. 6 00:00:21,330 --> 00:00:25,920 I'll try my best to be slow and explain these things line by line and block by block. 7 00:00:26,580 --> 00:00:31,320 So firstly, in this first section, we're loading all of the necessary functions we're going to be 8 00:00:31,320 --> 00:00:31,740 using. 9 00:00:32,220 --> 00:00:36,480 Take note that we're using the skills tree and test split function. 10 00:00:36,600 --> 00:00:40,890 So it's no that now that's downloadable images and unzip them. 11 00:00:40,920 --> 00:00:42,840 This takes about 30 seconds, roughly. 12 00:00:43,410 --> 00:00:46,800 So while that's how well that works, well, that runs. 13 00:00:46,800 --> 00:00:48,710 I'll talk about this. 14 00:00:48,720 --> 00:00:53,040 This is where we just define our image width and height and image size, which takes more to do this 15 00:00:53,040 --> 00:00:54,990 width and height as inputs. 16 00:00:55,710 --> 00:01:00,090 And then we just set the number of channels because it's a color image we're using. 17 00:01:00,090 --> 00:01:03,360 So it's 60 by 60 by trees, the dimensionality of this. 18 00:01:04,140 --> 00:01:06,480 And next, we'll wait for that to finish. 19 00:01:07,680 --> 00:01:08,490 Yeah, here we go. 20 00:01:08,910 --> 00:01:10,500 So next, forget the error below. 21 00:01:10,530 --> 00:01:12,000 Let's run this and get rid of it. 22 00:01:12,960 --> 00:01:18,110 What we're doing here, we're going to load our data and we're going to store the file name and its 23 00:01:18,120 --> 00:01:19,500 class in a data frame. 24 00:01:19,680 --> 00:01:26,250 And the reason for that for us, because Keros has an automatic function that allows you to pull information 25 00:01:26,250 --> 00:01:31,830 like this out of the data frames, which are which is convenient because data scientists like myself 26 00:01:31,830 --> 00:01:35,590 and I'm assuming many of you would be familiar with pandas. 27 00:01:35,670 --> 00:01:37,320 And this is a very, very good tool. 28 00:01:37,470 --> 00:01:39,930 So Keros integrates directly with pandas. 29 00:01:40,380 --> 00:01:45,450 So once you have your file name and your class here, it's a very good way to separate it, in my opinion. 30 00:01:46,140 --> 00:01:47,310 You can actually do that. 31 00:01:47,340 --> 00:01:54,960 So to do this or to build this, all we have to do is actually just get the file names for all of the 32 00:01:54,960 --> 00:01:56,070 images in the directory. 33 00:01:56,520 --> 00:02:03,690 Split them by the DOT because if it's but that's how the clusters are named, it's not always going 34 00:02:03,690 --> 00:02:04,240 to be this way. 35 00:02:04,260 --> 00:02:09,060 Just remember that this is specific to this dataset, which is let's actually look at the test data 36 00:02:09,240 --> 00:02:10,260 actually again. 37 00:02:10,260 --> 00:02:16,110 And remember, you won't see it, but the images are named something like, Let's do this here. 38 00:02:17,280 --> 00:02:20,190 So if there's a dog in the image, it's going to be dog dot. 39 00:02:20,730 --> 00:02:22,710 Some no jpg. 40 00:02:23,280 --> 00:02:28,860 So if you split on the dot here and if a category, if you get if you grab that first element and split 41 00:02:29,130 --> 00:02:34,860 the fishery, you will probably get you can get whether it's a cat or dog in that sense. 42 00:02:35,760 --> 00:02:43,440 And then you can just use pandas after the reframe here to create the use. 43 00:02:43,440 --> 00:02:44,070 Those are easier. 44 00:02:44,070 --> 00:02:49,020 So you get the file names here and the categories, and you just create this data frame out of this. 45 00:02:49,140 --> 00:02:50,430 So that's quite convenient. 46 00:02:50,940 --> 00:02:56,350 And then you can do things like value counts and plot and bob plot to get the codes that each category. 47 00:02:57,120 --> 00:03:00,790 And you can even view a sample image here using the lower image function. 48 00:03:00,820 --> 00:03:02,820 This is not related to the pandas part of it. 49 00:03:02,820 --> 00:03:06,720 This is just a way for us to get a random image is random. 50 00:03:06,720 --> 00:03:11,880 The choice in the file names here we get our sample image name because it's in this. 51 00:03:12,030 --> 00:03:13,170 This is our actual pet. 52 00:03:13,740 --> 00:03:16,380 And we just use my plot lib to show the image. 53 00:03:18,180 --> 00:03:22,350 So it's quite simple and you get a different random image every time you run this code. 54 00:03:23,280 --> 00:03:28,830 So now this will create a model which we've have done many times before, and this is the same model 55 00:03:28,830 --> 00:03:30,330 we used to pay to listen to. 56 00:03:30,630 --> 00:03:37,650 It's a very simple model 16 filters, 3D filters, 64 filters, three battery filters, really activation, 57 00:03:38,340 --> 00:03:39,660 max pooling here. 58 00:03:40,260 --> 00:03:47,010 Then we just have to fully connected layers, 256 nodes, 128 nodes up to finally two nodes, which 59 00:03:47,010 --> 00:03:48,560 are where we have to solve Max. 60 00:03:48,570 --> 00:03:54,570 This gives us the probability we use Adam as the optimizer with a very low learning rate that first 61 00:03:54,720 --> 00:03:59,580 point zero zero five, you're probably going to have to go that low, but that's just keep it that way 62 00:03:59,580 --> 00:04:00,060 for now. 63 00:04:00,750 --> 00:04:06,090 And then we just find our model has four hundred and sixty six thousand parameters, not that much. 64 00:04:06,750 --> 00:04:11,040 Now what we're going to do is we're going to create our data generators. 65 00:04:11,040 --> 00:04:16,410 So let's take a look at how we do that to firstly, we need to replace all the classrooms, the ones 66 00:04:16,410 --> 00:04:21,240 and zeros that we saw previously in the data frame with the cat or dog. 67 00:04:21,480 --> 00:04:22,480 So let's do that. 68 00:04:22,920 --> 00:04:24,030 Let's move our convenience. 69 00:04:24,030 --> 00:04:28,920 We could have left it at one and zero, but we'll change it to cat and dog because it's nicer to look 70 00:04:28,920 --> 00:04:31,200 out for us, for human to and to it. 71 00:04:31,200 --> 00:04:37,530 If we see the grass next, we'll use the scale and train test bed function under the F dataset here 72 00:04:37,980 --> 00:04:43,260 to create training data sets and validation data frames are in here and data frame here. 73 00:04:43,920 --> 00:04:48,960 And then we just reset the index and drop trou just to get back a nice data frame like this. 74 00:04:49,320 --> 00:04:52,170 So let's run this and then we go. 75 00:04:52,170 --> 00:04:58,050 So we have that one day, let's take a look at our validation dataset, which is going to be. 76 00:04:59,940 --> 00:05:02,130 Oops, didn't start properly. 77 00:05:06,830 --> 00:05:09,860 So there we go, so you can see those are the frames here. 78 00:05:10,610 --> 00:05:14,630 This this data frame contains all the image names with the validation data. 79 00:05:15,140 --> 00:05:17,360 This one contains all the ones with the training data. 80 00:05:18,050 --> 00:05:20,870 Next, this is what is cool about Keros. 81 00:05:21,380 --> 00:05:27,740 There is this flow from data frame function that exists in the train to generate the function that we 82 00:05:27,740 --> 00:05:29,840 create with the image data generated here. 83 00:05:30,380 --> 00:05:33,050 This is similar to PyTorch, where we have our transmissions. 84 00:05:33,050 --> 00:05:36,680 We can reskill the image here, so played at risk of transformation here. 85 00:05:37,040 --> 00:05:39,590 It's not exactly the same thing, but it's quite similar. 86 00:05:40,310 --> 00:05:46,640 However, this is the cool part here with trained data gender flow from data frame, we can set the 87 00:05:46,640 --> 00:05:48,830 part to the data frame. 88 00:05:49,070 --> 00:05:49,910 That's the train. 89 00:05:49,910 --> 00:05:54,560 The data frame here said the part with the images where they're located. 90 00:05:55,220 --> 00:06:01,010 Then we just put the column names here with the X column name is found in what the white labels name 91 00:06:01,010 --> 00:06:02,240 is going to be class. 92 00:06:02,870 --> 00:06:09,050 Then we have an image size, which we defined before class melodies categorical that size here and there 93 00:06:09,050 --> 00:06:09,350 we go. 94 00:06:09,770 --> 00:06:15,140 So and now we do the same for the validation data generated exactly the same thing, actually. 95 00:06:15,920 --> 00:06:19,250 The only difference here is that we're using a different data frame. 96 00:06:19,370 --> 00:06:21,170 Otherwise, everything else remains the same. 97 00:06:22,190 --> 00:06:24,230 And here's an example of how it works. 98 00:06:24,770 --> 00:06:26,930 So this is just an example, this generator here. 99 00:06:28,010 --> 00:06:29,120 So let's take a look at this. 100 00:06:29,120 --> 00:06:34,340 So example the f here is just this is just a random line from this. 101 00:06:34,340 --> 00:06:35,810 Here's one image, basically. 102 00:06:36,530 --> 00:06:42,220 And what we do is if you just run this here, it's found one image one in this class. 103 00:06:42,230 --> 00:06:46,580 That's because it's only one image was set here, but we can then preview that image. 104 00:06:47,030 --> 00:06:48,440 So how do we use it out there? 105 00:06:48,860 --> 00:06:54,470 Well, for example, generate the returns, put the image and label. 106 00:06:54,950 --> 00:06:57,470 So you can actually get the image out of it here. 107 00:06:57,500 --> 00:06:59,480 Images stored in a multidimensional array. 108 00:06:59,480 --> 00:07:01,220 So you just index and the first one. 109 00:07:01,880 --> 00:07:04,000 And you can display that random image. 110 00:07:04,010 --> 00:07:07,370 For that example, image had its first image in our training dataset. 111 00:07:07,970 --> 00:07:09,580 So that's how this works. 112 00:07:09,680 --> 00:07:16,880 So you can see now that all you have to do now in fitting the model is to model that fit set pointed 113 00:07:16,880 --> 00:07:22,130 to a train that they generate through the generator, which contains our training images set to part 114 00:07:22,130 --> 00:07:28,010 four, set to a validation data generator to be equal to the validation data part of it. 115 00:07:28,490 --> 00:07:32,630 Remember, these are the things we created above here validation generator. 116 00:07:33,320 --> 00:07:39,470 And this is just a function that points basically to where the data is using the pandas data frame we 117 00:07:39,470 --> 00:07:43,650 created and we can train. 118 00:07:43,670 --> 00:07:47,150 Also remember, we have to do the validation steps and steps we book. 119 00:07:47,600 --> 00:07:48,860 This is a validation site. 120 00:07:48,860 --> 00:07:55,070 Abide by the VAT status of the training administered data size divided by the body size. 121 00:07:55,070 --> 00:07:56,540 So let's run this. 122 00:07:57,500 --> 00:08:00,530 It's going to take a little while to train roughly about a minute for each epoch. 123 00:08:01,100 --> 00:08:03,780 So while the trends continue with the lesson. 124 00:08:04,760 --> 00:08:09,740 And since most of this code has been run before the outputs exist here, so I can we can. 125 00:08:09,890 --> 00:08:15,660 I can teach you from those so you can see this model is being trained to see immediately. 126 00:08:15,680 --> 00:08:23,420 It's a 55 percent accuracy, which is a good sign because in a binary classification problem, 50, 127 00:08:23,420 --> 00:08:27,410 if it starts at 50 percent and it continues at 50 percent, it means it's not learning. 128 00:08:27,410 --> 00:08:32,000 But now we can see it's learning, so it's getting better, even though it's not perfect yet. 129 00:08:32,000 --> 00:08:33,380 Obviously, it's just started. 130 00:08:34,010 --> 00:08:36,950 So let's continue with this lesson to the next block. 131 00:08:36,950 --> 00:08:41,320 Here's where we saved the model, and here's where we thought the results here. 132 00:08:41,330 --> 00:08:47,810 So you can see in the first graph we have treating this with the legend, isn't he actually not sure 133 00:08:47,820 --> 00:08:49,960 why it's not OK? 134 00:08:49,970 --> 00:08:51,230 And take a look at occlude? 135 00:08:51,230 --> 00:08:57,050 Yeah, we need to have another legend above, but that's OK, so we can see this as the training accuracy 136 00:08:57,170 --> 00:08:58,400 which is going up here. 137 00:08:58,880 --> 00:09:01,880 Validation accuracy is going around 80 percent. 138 00:09:01,950 --> 00:09:02,600 Not too bad. 139 00:09:03,620 --> 00:09:10,400 And similarly, we can see this is the validation loss here, which you can see it's sort of a Fitbit 140 00:09:10,400 --> 00:09:13,550 and training dataset because the validation loss is going up already. 141 00:09:14,120 --> 00:09:17,300 And that's because it's probably just a very small network anyway. 142 00:09:17,330 --> 00:09:19,960 Even the small networks tend to generalize better. 143 00:09:20,540 --> 00:09:26,210 They don't always work too well, especially if you will between them, because you can see the training 144 00:09:26,390 --> 00:09:28,580 loss this blue line went down. 145 00:09:29,940 --> 00:09:31,460 Now let's take a look at this. 146 00:09:31,850 --> 00:09:34,880 Let's get the predictions for validation images here. 147 00:09:35,420 --> 00:09:41,990 So we just use the validation data generated here, and we use models that predict generator and steps 148 00:09:41,990 --> 00:09:43,360 is what we have before. 149 00:09:43,550 --> 00:09:45,680 Five divided that's about size. 150 00:09:46,160 --> 00:09:50,450 We use ceiling just to make sure it's not to make sure it's an integer that point. 151 00:09:51,080 --> 00:09:53,930 So we get the model's predictions out of it here. 152 00:09:54,980 --> 00:09:57,920 So then we can actually concatenated with our data frame here. 153 00:09:58,640 --> 00:10:03,770 This could I'm not going to go into too much depth with this code, but just look at the output to see 154 00:10:03,770 --> 00:10:04,380 what it does. 155 00:10:04,380 --> 00:10:06,060 So know what a. 156 00:10:06,160 --> 00:10:11,770 Both in know, you can see we have to ground your class here in the middle column and then the predicted 157 00:10:11,770 --> 00:10:15,940 class for my model, so you can see for all of these, it got rough. 158 00:10:16,120 --> 00:10:17,410 No, it didn't get that much, right? 159 00:10:17,770 --> 00:10:20,050 This one, in particular, the dog, when it's really a cat. 160 00:10:20,560 --> 00:10:21,970 This is not a mistake. 161 00:10:22,000 --> 00:10:25,030 I believe as silly of these two here and this one here. 162 00:10:25,510 --> 00:10:26,860 So it got a few them wrong. 163 00:10:26,980 --> 00:10:31,750 So at 80 percent accuracy, that's what you would expect generally. 164 00:10:32,410 --> 00:10:37,510 So no, we can inference and a bunch of images here and get the probabilities out of it. 165 00:10:38,050 --> 00:10:39,550 So you can see this. 166 00:10:39,550 --> 00:10:44,320 Here's a cat image here when to cut the probability in the brackets. 167 00:10:45,070 --> 00:10:47,410 That's what the title is giving us. 168 00:10:47,410 --> 00:10:50,890 We're getting a category out of it category is predicted. 169 00:10:50,890 --> 00:10:52,090 Finally, but this rule? 170 00:10:52,540 --> 00:10:58,960 So we're printing category here into the image as long as well as the file name, so you can see the 171 00:10:58,960 --> 00:11:00,260 file name as a ground shirt. 172 00:11:00,280 --> 00:11:03,060 And this is a prediction ground truth prediction. 173 00:11:03,070 --> 00:11:03,840 Good so far. 174 00:11:03,850 --> 00:11:07,630 Ground truth prediction or good or these? 175 00:11:07,630 --> 00:11:10,510 This one isn't good, though it's a cat, but it predicted a dog. 176 00:11:11,650 --> 00:11:14,110 And let's see if there are any more mistakes. 177 00:11:16,210 --> 00:11:18,280 No, but this is a mistake. 178 00:11:18,280 --> 00:11:19,510 Very obvious mistake here. 179 00:11:19,510 --> 00:11:20,920 This is definitely not a cat. 180 00:11:22,150 --> 00:11:24,880 So is this one definitely not a cat as well? 181 00:11:26,020 --> 00:11:28,490 This one is definitely not a cat. 182 00:11:28,510 --> 00:11:30,160 It's actually that it's a cat. 183 00:11:30,870 --> 00:11:35,290 And yeah, so you can see at 80 percent accuracy was still making quite a few mistakes. 184 00:11:35,290 --> 00:11:38,620 But generally, this cat, this dog classifier works fairly well. 185 00:11:39,080 --> 00:11:42,100 What's with such a simple model that's actually pretty good? 186 00:11:42,670 --> 00:11:48,520 So now let's take a look at implementing checkpoints and Keros checkpoints are actually super simple 187 00:11:48,520 --> 00:11:49,090 to implement. 188 00:11:49,900 --> 00:11:52,660 You would have seen some of the examples in the slides previously. 189 00:11:53,110 --> 00:11:56,420 So all you do, basically, you just look at the checkpoint you want to use. 190 00:11:56,440 --> 00:11:57,820 We're going to use model checkpoints. 191 00:11:58,300 --> 00:12:05,630 This saves animal waits after every epoch and early stopping, which is another one where we do it. 192 00:12:05,860 --> 00:12:10,900 We use it to monitor the devaluation, loss or validation accuracy, whichever one you want to monitor. 193 00:12:11,440 --> 00:12:13,530 You can specify whatever parameter you want. 194 00:12:13,540 --> 00:12:19,840 Here we are looking at validation, loss both of them and basically we set the parameters here. 195 00:12:19,870 --> 00:12:24,580 This is the value that it's required to change before we stop. 196 00:12:24,970 --> 00:12:27,460 This is no the box we wait before we stop. 197 00:12:28,330 --> 00:12:32,810 This is to open settings and we can restore best weights to be equal to true. 198 00:12:32,830 --> 00:12:39,400 So once we stop it, it saves the model with the best weights, and that specified actually just saves 199 00:12:39,400 --> 00:12:41,020 them the best, which is the model. 200 00:12:41,440 --> 00:12:47,080 So whatever model it keeps, it returns at the end is the best model in this case here. 201 00:12:48,040 --> 00:12:53,320 Similarly, for model check pointing here, you can set a pattern for where you want the model to be 202 00:12:53,320 --> 00:12:53,830 saved. 203 00:12:54,400 --> 00:13:01,300 So I'll just take off this part because this path, as it exists in this directory and this is the saving 204 00:13:01,300 --> 00:13:02,860 it to the best start for here. 205 00:13:03,610 --> 00:13:04,970 Again, that's it. 206 00:13:06,190 --> 00:13:12,960 Also, let's take a look at the lending rate on the tool so we can just set the parameters here as well. 207 00:13:12,970 --> 00:13:15,640 I wouldn't go into details with this because we did have in the slides. 208 00:13:15,640 --> 00:13:23,140 So what we do know, we have those three checkpoints we created here model check point, early stopping 209 00:13:23,830 --> 00:13:25,570 and reduced lending rates. 210 00:13:26,110 --> 00:13:28,990 So we put all of those check points in an array here. 211 00:13:28,990 --> 00:13:30,730 So we have recorded callbacks. 212 00:13:31,300 --> 00:13:36,610 And then all we have to do when we're doing model of fitting tariffs is put callbacks equal to the callbacks 213 00:13:36,610 --> 00:13:37,480 that we set here. 214 00:13:38,050 --> 00:13:38,890 And that's it. 215 00:13:38,980 --> 00:13:44,380 And then you start creating your model, and all of us basically obeys all of those rules. 216 00:13:44,380 --> 00:13:50,200 And you can see here that in this example, it's stopped trading after the sixth iteration because at 217 00:13:50,200 --> 00:13:55,150 that point, violation was, I believe, stopped it and improve from point sixty five. 218 00:13:55,630 --> 00:13:59,680 So it implemented it, really stopping and return the model with the best suites here. 219 00:13:59,950 --> 00:14:05,410 So the model now has those best suites based on the sixty bucks here, which most likely would have 220 00:14:05,410 --> 00:14:06,670 been the last one either. 221 00:14:06,720 --> 00:14:09,100 Actually, this will be this one. 222 00:14:09,220 --> 00:14:13,930 The third one point six point seven four, which is the one with the lowest suites. 223 00:14:15,040 --> 00:14:16,540 So that's it. 224 00:14:16,540 --> 00:14:20,290 For this lesson, I notice a bit lengthy and it was a lot going on in this lesson. 225 00:14:21,430 --> 00:14:23,660 I know things are getting a bit more difficult. 226 00:14:23,680 --> 00:14:29,950 However, I'm hoping you can keep up, and if you don't, please feel free to message me on Udemy support 227 00:14:29,950 --> 00:14:32,440 forums and I'll try to get back to you as soon as possible. 228 00:14:33,070 --> 00:14:37,840 Thank you, and we'll stop there for now and then after we'll take a look at PyTorch lightning. 229 00:14:38,440 --> 00:14:38,740 OK. 230 00:14:39,220 --> 00:14:40,300 So stay tuned for that. 231 00:14:40,450 --> 00:14:40,810 Thank you.