1
00:00:00,450 --> 00:00:03,390
Okay everyone I hope you're as excited as I am.

2
00:00:03,390 --> 00:00:04,350
Why.

3
00:00:04,350 --> 00:00:06,500
Well because we're on the homestretch.

4
00:00:06,570 --> 00:00:10,640
We're up to number six save and load a trained model.

5
00:00:10,950 --> 00:00:12,910
So we've gone through all these steps here.

6
00:00:12,990 --> 00:00:17,430
We've seen how to improve our models with hyper parameter tuning and let's say we're at a point where

7
00:00:17,430 --> 00:00:21,780
we've improved our model to the point that we're happy to either share that with one of our colleagues

8
00:00:21,780 --> 00:00:27,300
or something like that or export it and save it and put it in one of our applications so that our customers

9
00:00:27,300 --> 00:00:30,430
might want to use it might want to access its predictions.

10
00:00:30,540 --> 00:00:34,750
Something along those lines essentially we want to take our model that we've trained.

11
00:00:34,800 --> 00:00:36,210
We've got these results.

12
00:00:36,480 --> 00:00:42,360
Export it to file and so that we can import it and use it somewhere else rather than it just being within

13
00:00:42,360 --> 00:00:46,910
our Jupiter notebook so let's create a heading.

14
00:00:46,910 --> 00:00:54,260
Six guys saving and loading trained machine learning models.

15
00:00:55,520 --> 00:00:56,580
Beautiful.

16
00:00:56,840 --> 00:00:59,300
And there's two ways we can do this.

17
00:00:59,300 --> 00:01:05,930
So two ways to save and load chain learning models.

18
00:01:05,930 --> 00:01:18,740
First I'm gonna cover is with Python's pickle module and the second one is with the job leave module.

19
00:01:18,740 --> 00:01:27,610
So let's check out how we might do it with pickle so the first thing is to import pickle and then and

20
00:01:27,990 --> 00:01:33,270
if you're wondering what pickle is you can go Python pickle as always if if we're not sure what something

21
00:01:33,270 --> 00:01:38,550
is we can search it up so python object serialization.

22
00:01:38,550 --> 00:01:44,730
So if you read this the pickle module implements binary protocols for serialization decentralizing a

23
00:01:44,730 --> 00:01:52,650
python object structure well in our case if you wondering what a python object structure is in our case

24
00:01:52,930 --> 00:01:55,860
a python object is our model.

25
00:01:55,860 --> 00:02:03,060
So if we want to say for example save G.S. classifiers who say we're most happy with our grid search

26
00:02:03,060 --> 00:02:04,260
parameters model.

27
00:02:04,380 --> 00:02:07,500
So the random forest classifier we've trained with grid search.

28
00:02:07,680 --> 00:02:13,120
If we wanted to export that this is our python object and that's what people can help us do.

29
00:02:13,200 --> 00:02:13,860
So let's do that.

30
00:02:13,860 --> 00:02:20,460
Let's say an example of how we might export our GCF which is a python object to a pickle file.

31
00:02:20,640 --> 00:02:25,950
So save an existing model to file.

32
00:02:26,390 --> 00:02:28,880
Yes please.

33
00:02:28,880 --> 00:02:35,030
And to do so we're going to use pickle dot dump we're going to pass it the object we would like to save

34
00:02:35,030 --> 00:02:41,660
in our case G.S. CnF our grid searched classifier and then you can name it something here.

35
00:02:41,660 --> 00:02:47,540
So we need to use the open function so Python's open function for dealing with input and output of files

36
00:02:48,290 --> 00:02:53,900
and in our case we're just going to I guess random forest model 1.

37
00:02:53,900 --> 00:02:56,720
Now you can get as creative as you like with these names.

38
00:02:56,720 --> 00:03:00,650
There might be some nomenclature that you decide upon with your team or something like that.

39
00:03:00,650 --> 00:03:02,110
That's what you use for your models.

40
00:03:02,240 --> 00:03:09,050
But now case I've just gone something nice and simple and Dot P K L stands for pickle file and we also

41
00:03:09,050 --> 00:03:17,240
pass because we're writing a file we pass the parameter WB for write binary so if we hit shift and enter

42
00:03:18,430 --> 00:03:27,120
beautiful this is going to save our grid search model with this file name to our current working directory.

43
00:03:27,120 --> 00:03:32,580
So if we go up here if we go guess random what we've got to randoms in there.

44
00:03:32,760 --> 00:03:34,780
So he s random random forest.

45
00:03:34,820 --> 00:03:40,170
So this is a random random forest really random one dot scale and you can see there's some other model

46
00:03:40,170 --> 00:03:40,970
files here.

47
00:03:41,040 --> 00:03:46,500
So dot P K L This is one I tried just before and this is another one we we created right at the start

48
00:03:46,500 --> 00:03:51,960
where we were looking at a cyclone workflow and now if you've exported a fair few of these models you

49
00:03:51,960 --> 00:03:56,600
might want to put them in like a model file like we've got our data in a data file.

50
00:03:56,670 --> 00:04:01,680
So this is just one way it to tidy up the directory but you can see the one we've just saved here that's

51
00:04:01,680 --> 00:04:05,810
seconds ago now if we come back how would we load that in.

52
00:04:06,060 --> 00:04:11,750
Well we can do it with people as well so load a saved model.

53
00:04:11,880 --> 00:04:18,520
So we're going to go loaded people model Eagles pickle dot load.

54
00:04:18,540 --> 00:04:24,270
Now the advantage here is that because we've saved a model to a pickle file and because it's stored

55
00:04:24,270 --> 00:04:31,380
in our directory here is that someone who wanted to use our model won't have to go through all of the

56
00:04:31,380 --> 00:04:33,150
training steps that we've done above.

57
00:04:33,150 --> 00:04:39,010
They can just access the model and be able to make predictions by loading it in we could copy this file

58
00:04:39,010 --> 00:04:39,480
name here.

59
00:04:39,490 --> 00:04:45,760
But we're just going to type it in Jesus tab we used tab auto complete for that and now because we're

60
00:04:45,760 --> 00:04:51,390
reading in a file we passing the parameter R B for read binary.

61
00:04:51,400 --> 00:04:52,010
Wonderful.

62
00:04:52,090 --> 00:04:54,080
So that's loaded in our save model.

63
00:04:54,220 --> 00:05:00,580
Now the way to check it is we can make some predictions to see if this model if it actually worked if

64
00:05:00,580 --> 00:05:01,330
we saved it.

65
00:05:01,480 --> 00:05:05,090
And we can use it to make predictions while then we're in a good place.

66
00:05:05,120 --> 00:05:11,620
We want to do loaded people model don't predict now similar to how we made predictions before with our

67
00:05:11,830 --> 00:05:13,060
GCF.

68
00:05:13,130 --> 00:05:21,010
We're just going to run this exact same line here but make predictions without pickled model predict

69
00:05:21,190 --> 00:05:22,910
on the test dataset.

70
00:05:23,050 --> 00:05:23,620
Wonderful.

71
00:05:23,650 --> 00:05:32,020
And then we'll call our evaluate parades function and compare it the test labels to the pickle y parades

72
00:05:34,050 --> 00:05:38,310
what have we got here found input variables with inconsistent number of samples.

73
00:05:38,730 --> 00:05:39,170
Mm hmm.

74
00:05:40,970 --> 00:05:42,070
I know what we've done.

75
00:05:42,110 --> 00:05:46,320
We've got a data mismatch here what we'd have to do

76
00:05:49,190 --> 00:05:54,800
is I believe if we just commented out this line and recreate we have to re instantiate our train and

77
00:05:54,800 --> 00:05:56,690
test split that's what we have to do.

78
00:05:56,780 --> 00:06:05,400
So maybe I'll just comment out these lines for the time being run this cell come back down here we load

79
00:06:05,400 --> 00:06:09,360
it in so the mistake here I believe and maybe wrong hit shift enter.

80
00:06:09,690 --> 00:06:10,870
There we go.

81
00:06:11,160 --> 00:06:18,720
Is that our x test had been instantiated somewhere above and we had to re instantiate it so that it

82
00:06:18,720 --> 00:06:20,910
was reset back to the correct shape.

83
00:06:21,060 --> 00:06:26,530
And now the way we can know that this is gone through correctly is if we can pair.

84
00:06:26,610 --> 00:06:30,330
So if we compare our original guess CnF predictions.

85
00:06:30,390 --> 00:06:36,240
So these values here if we copy these is the original grid search classify metrics

86
00:06:38,960 --> 00:06:39,860
so let's have a look.

87
00:06:40,010 --> 00:06:44,550
So this is our loaded models metrics accuracy seventy eight point six nine.

88
00:06:44,600 --> 00:06:53,690
Same precision zero point seven four same recall zero point eight two same and F1 score zero point seven

89
00:06:53,690 --> 00:06:53,900
eight.

90
00:06:53,930 --> 00:06:59,330
So this same that means that our import our saving our model and reporting it.

91
00:06:59,330 --> 00:07:05,150
So loading it in has worked correctly and now again these names here you can change them to whatever

92
00:07:05,150 --> 00:07:05,890
you like.

93
00:07:05,900 --> 00:07:11,300
Same with the loaded variable names but this is just a quick example of how you can use pickle to quickly

94
00:07:11,300 --> 00:07:17,810
save one of your models to file and then load it back in at a later date if need be and then make predictions

95
00:07:17,810 --> 00:07:24,050
using the exact same loaded model that get the exact same results as the model we originally saved in

96
00:07:24,050 --> 00:07:28,550
the next video we'll see how we can do the same thing except this time using job Lib.