1

00:00:01,050 --> 00:00:06,840

Had you got to answer the question, how do we evaluate our models while we make predictions with them



2

00:00:07,320 --> 00:00:11,060

and then we compare the predictions to what the model should have actually done.



3

00:00:11,610 --> 00:00:20,130

So we need to get the labels from the test data set and compare them to these labels and models predictions.



4

00:00:20,640 --> 00:00:21,420

So let's do that.



5

00:00:21,690 --> 00:00:22,850

Well, right now, there's a little note.



6

00:00:23,340 --> 00:00:35,030

Now we've got predictions, array of all of our models, predictions to evaluate them.



7

00:00:35,490 --> 00:00:42,750

We need to compare them to the original test data set levels.



8

00:00:43,500 --> 00:00:50,670

So to do that, to get that we've currently our test data is wrapped up, an intensive low batch data



9

00:00:50,670 --> 00:00:50,970

set.



10

00:00:51,970 --> 00:00:57,820

So we're going to have to unpatched this data set and pull out the labels and get them in a format that's



11

00:00:57,820 --> 00:01:02,170

like this, because what is one of the most important things in machine learning and deep learning is



12

00:01:02,170 --> 00:01:07,720

comparing apples to apples, making sure you tenses are in the right format, and making sure your array



13

00:01:07,720 --> 00:01:09,580

of data in the right shape.



14

00:01:09,580 --> 00:01:16,490

Because if we try to compare this array here to our test data as it is now, we'll just get errors galore.



15

00:01:16,900 --> 00:01:18,090

So that's what we're going to do.



16

00:01:18,100 --> 00:01:19,380

We need to unravel this.



17

00:01:20,920 --> 00:01:31,050

So to get our test labels, we need to unravel our test data batch data set.



18

00:01:31,600 --> 00:01:38,710

And to do so, we're going to create a list called Y Labels is going to be for our test labels and then



19

00:01:38,710 --> 00:01:45,130

we can go for images, labels, because right now our test data might just put it down here just in



20

00:01:45,130 --> 00:01:46,030

case it disappears.



21

00:01:46,180 --> 00:01:51,970

Test data is in a batch of image label tuples.



22

00:01:52,090 --> 00:01:56,680

So we go for images, labels in test data.



23

00:01:57,130 --> 00:02:02,210

Now there's a method for batch data sets intensive low called one batch.



24

00:02:02,830 --> 00:02:07,350

So as you might imagine, this is going to one batch, our data set.



25

00:02:07,630 --> 00:02:15,640

So go from our test data is currently we've seen this before, 790 batches.



26

00:02:16,690 --> 00:02:24,390

Of size 32 of image label Tupperware's, but what we want is we want to unpack all of the test data



27

00:02:24,720 --> 00:02:30,090

and then retrieve just the labels from our test data set.



28

00:02:30,660 --> 00:02:34,170

And now I'm going to do this code before I start jumping around.



29

00:02:34,170 --> 00:02:35,200

So let's do this together.



30

00:02:35,700 --> 00:02:43,920

We're going to go why labels don't append labels don't no high ARG, Max.



31

00:02:45,660 --> 00:02:51,240

Beautiful, and then we're going to have a look at what Y labels the first 10.



32

00:02:52,670 --> 00:03:03,070

Look at the first 10, by the way, this is a Tenzer or an array, the labels of one hot encoded labels.



33

00:03:03,320 --> 00:03:13,190

So the reason why we get Amax is currently test labels look like zero zero zero.



34

00:03:13,400 --> 00:03:19,580

And where the target label is, say, for the class doughnut, it's going to have a one there and then



35

00:03:19,580 --> 00:03:23,780

it'll be dot, dot, dot, and then it'll be zeros for the rest of it for one hundred and one.



36

00:03:24,740 --> 00:03:27,050

We want the index value.



37

00:03:27,060 --> 00:03:27,620

That's what we do.



38

00:03:27,620 --> 00:03:31,850

Amax where the one occurs.



39

00:03:32,510 --> 00:03:33,710

So they're one encoded.



40

00:03:33,900 --> 00:03:39,170

That's why we do amax and we do some pie so that it turns into something like this in array.



41

00:03:40,270 --> 00:03:45,550

Beautiful, but we can see what this does by running, it might take a minute or so because on batching



42

00:03:45,550 --> 00:03:49,360

data does take a fair while, depending on how many batches of data you have.



43

00:03:49,660 --> 00:03:50,560

So I'll let this run.



44

00:03:55,770 --> 00:03:56,400

Excellent.



45

00:03:56,700 --> 00:04:02,940

Now, you might be wondering why we get some labels in all zeros or actually, let's just quickly just



46

00:04:02,940 --> 00:04:07,040

want to know how many why labels are there.



47

00:04:07,170 --> 00:04:11,780

So this is the original test data set labels to Gaylan y labels.



48

00:04:11,970 --> 00:04:13,050

There should be how many?



49

00:04:13,620 --> 00:04:14,990

One for each image.



50

00:04:15,000 --> 00:04:18,690

And we had twenty five thousand two hundred and fifty images.



51

00:04:18,720 --> 00:04:19,370

Beautiful.



52

00:04:19,890 --> 00:04:22,240

So you might be wondering why they're all zero.



53

00:04:22,830 --> 00:04:27,360

Well, it's because when we originally loaded come up here.



54

00:04:28,890 --> 00:04:34,710

When we originally loaded will go to where we loaded our test data, do you remember we set Shuffle



55

00:04:34,710 --> 00:04:35,650

Eagle to false.



56

00:04:36,300 --> 00:04:37,890

So what that means is going to happen.



57

00:04:37,920 --> 00:04:41,050

This is where I told you we're going to come back to where shuffle equals false.



58

00:04:41,220 --> 00:04:46,290

This is so we can make evaluations on our test data set because otherwise, every time we loaded test



59

00:04:46,290 --> 00:04:48,890

data, the order of them would be shuffled.



60

00:04:48,900 --> 00:04:54,900

And so if we made predictions on our test data set and then we unraveled it, it would shuffle at every



61

00:04:54,900 --> 00:04:55,200

time.



62

00:04:55,200 --> 00:04:57,360

And so we couldn't compare apples to apples.



63

00:04:57,900 --> 00:05:03,270

So if we come here because we've turned shuffle off, it means that the test labels that we've now got,



64

00:05:03,540 --> 00:05:11,730

the Y labels array will be in the order of apple pie, first baby back ribs, baklava the governance



65

00:05:11,730 --> 00:05:12,470

of these foods.



66

00:05:12,840 --> 00:05:13,590

So if we come back.



67

00:05:14,640 --> 00:05:17,400

To making predictions with that train model, we've now got to.



68

00:05:18,270 --> 00:05:23,820

A rise of twenty five thousand two hundred fifty, so we've got Y labels, these are original labels



69

00:05:24,060 --> 00:05:25,650

and these are our model's predictions.



70

00:05:25,800 --> 00:05:32,250

So ideally, our model would have the same first ten samples as the true labels.



71

00:05:32,250 --> 00:05:34,710

But we can see there's already some errors here.



72

00:05:34,710 --> 00:05:36,900

So predicting 50 to where it should have been zero.



73

00:05:37,410 --> 00:05:42,000

It's predicting 80, where it should have been zero and so on and so on.



74

00:05:42,130 --> 00:05:48,020

But the important thing is because we don't shuffle off, these are now in the same order.



75

00:05:48,240 --> 00:05:54,570

So when we compare them later on, we can be sure that the index zero, this one lines up with index



76

00:05:54,570 --> 00:05:56,190

zero of our predicted class.



77

00:05:56,460 --> 00:05:57,840

But let's not take my word for it.



78

00:05:58,200 --> 00:06:02,850

Let's write some code in the next few videos to further evaluate and visualize our predictions.



79

00:06:03,360 --> 00:06:03,930

I'll see you there.