1
00:00:00,930 --> 00:00:02,220
Model fitting is complete.

2
00:00:03,120 --> 00:00:10,620
It has run 430 epochs and we can see that we have achieved validation, accuracy of nearly 89 to 90

3
00:00:10,620 --> 00:00:14,990
percent in the twenty seventy Borg.

4
00:00:15,200 --> 00:00:17,440
You can see that it is ninety point five.

5
00:00:18,690 --> 00:00:23,790
So at the 30th Epoch is giving us a validation accuracy of eighty nine point eight percent.

6
00:00:24,750 --> 00:00:28,950
You can see on the graph that our validation accuracy is stagnant.

7
00:00:29,280 --> 00:00:31,170
It has stopped increasing.

8
00:00:33,360 --> 00:00:39,300
Also, you can observe that validation accuracy is a little bit on the higher side, then turning accuracy.

9
00:00:40,920 --> 00:00:47,280
This is a little counterintuitive because usually we see that training accuracy are always higher than

10
00:00:47,280 --> 00:00:48,390
validation equities.

11
00:00:49,710 --> 00:00:55,170
This is mostly because the input data during training is being transformed.

12
00:00:55,470 --> 00:00:56,040
It has.

13
00:00:57,150 --> 00:00:59,130
Zoom out effect it has here.

14
00:00:59,310 --> 00:01:02,670
It has linear transformations and rotation transformations.

15
00:01:03,630 --> 00:01:08,070
So that is why training accuracy is a little bit less than the validation accuracy.

16
00:01:09,810 --> 00:01:11,140
You can save the last model.

17
00:01:11,220 --> 00:01:14,850
That is the more latitude epoch using this same model.

18
00:01:15,060 --> 00:01:19,770
I just got STF five function and you can retrieve this model.

19
00:01:20,460 --> 00:01:26,250
That is, you can create a new model and save these same architecture and read into that.

20
00:01:26,400 --> 00:01:29,740
Using this load model just is DFI function.

21
00:01:31,500 --> 00:01:36,180
Now the last part of our analysis is checking the performance on test data.

22
00:01:37,890 --> 00:01:44,570
Although we saw the performance on validation data and we did not modify anything in the model basis,

23
00:01:44,600 --> 00:01:45,990
the validation deserved.

24
00:01:47,070 --> 00:01:52,270
So the result on validation data will be very similar to the design on that data.

25
00:01:53,520 --> 00:02:01,290
But just to see how it is done, we have to create a test in data that is generated, will have test

26
00:02:01,290 --> 00:02:08,100
territory and test data engine as the input parameters we specified test data.

27
00:02:08,180 --> 00:02:15,630
Again, in this line it was only rescaling the images, the data of pixels by dividing it by 255.

28
00:02:18,120 --> 00:02:20,820
That goes as a parameter and this function.

29
00:02:22,980 --> 00:02:24,360
The other barometers are seen.

30
00:02:25,590 --> 00:02:33,600
We run this and then we can use the evaluate underscored generator function to evaluate the performance

31
00:02:33,600 --> 00:02:36,120
of this model on our deficit.

32
00:02:38,640 --> 00:02:44,090
So do you value it in later function as run and on the test set?

33
00:02:44,490 --> 00:02:50,340
I'm getting an accuracy of eighty nine point nine percent on the validation test.

34
00:02:50,370 --> 00:02:52,500
Also, I was getting eighty nine point eight percent.

35
00:02:53,010 --> 00:02:59,550
Both of these are giving the same level of accuracy for us because for our model validation, data is

36
00:02:59,550 --> 00:03:02,250
also equally unseen as the test data.

37
00:03:05,060 --> 00:03:12,860
Last time, when we only used data augmentation, we were able to get accuracy's in the range of 83

38
00:03:12,860 --> 00:03:13,850
to 84 percent.

39
00:03:14,900 --> 00:03:20,060
Now we have increment of nearly six percent in the accuracy range.

40
00:03:20,690 --> 00:03:25,290
And this increase is because of the convolutional base that we have used.

41
00:03:25,740 --> 00:03:32,780
The convolutional base was an already train set of convolutional and pooling layers, and those trained

42
00:03:32,780 --> 00:03:38,060
LEDs have given us an incremental accuracy of nearly six to seven percent.

43
00:03:38,300 --> 00:03:47,210
So we'll achieve nearly 90 percent, not to further increase the accuracy to 96, 97 percent.

44
00:03:48,470 --> 00:03:51,920
You need to do only one small change to the standard model.

45
00:03:54,380 --> 00:03:59,160
And that small changes to remove this freeze roots line.

46
00:04:00,050 --> 00:04:07,820
As you remember, we used this line to freeze the weight of our convolutional base, which means that

47
00:04:08,430 --> 00:04:10,760
the convolutional base will not be retrained.

48
00:04:12,740 --> 00:04:17,510
Now, one way is to unfreeze the last few layers.

49
00:04:18,560 --> 00:04:22,300
And another way is to unfreeze the entire convolutional base.

50
00:04:24,530 --> 00:04:28,760
Both of these will give you significant increase in the accuracy.

51
00:04:30,800 --> 00:04:37,880
The concept is, since the features that we were finding out from this convolutional base were targeted

52
00:04:38,000 --> 00:04:39,610
at the image data.

53
00:04:41,120 --> 00:04:45,420
It may need some tuning in, finding out the features.

54
00:04:45,460 --> 00:04:47,450
Also for our particular problem.

55
00:04:48,800 --> 00:04:56,330
So if you do not freeze the rates of last few conversion layers or you do not freeze the of the entire

56
00:04:56,330 --> 00:05:00,800
convolutional base, then you can further increase the accuracy.

57
00:05:01,110 --> 00:05:07,520
And with BDD model, without freezing the rates, you should be able to get accuracy's in the range

58
00:05:07,520 --> 00:05:08,660
of 97 percent.

59
00:05:10,100 --> 00:05:18,440
But keep in mind that since this is a very big network, it will take a lot of computation power of

60
00:05:18,440 --> 00:05:19,100
your computer.

61
00:05:19,430 --> 00:05:23,000
So it will take a lot of memory from your DRAM and Dipu.

62
00:05:24,410 --> 00:05:29,030
Also it will take one or two hours extra if you're running a very big model.

63
00:05:31,310 --> 00:05:38,840
So it may not be feasible to run this on every computer, but just to tell you that this is a possibility.

64
00:05:39,110 --> 00:05:44,480
And you can get 97 percent accuracy in image recognition problems.

65
00:05:45,680 --> 00:05:52,520
This is the only change you have to make and you will be able to get 97 percent accuracy for such problems.

66
00:05:54,650 --> 00:05:56,030
So that's the entire project.

67
00:05:56,750 --> 00:06:04,940
We started this project by importing the data and building a normal, convolutional neural network,

68
00:06:05,630 --> 00:06:09,770
which was giving us accuracy's in the range of 70 to 75 percent.

69
00:06:11,780 --> 00:06:15,440
Then we improve the accuracy using the documentation.

70
00:06:16,370 --> 00:06:20,780
We updated our image data generator with few more parameters.

71
00:06:21,200 --> 00:06:28,310
And this data generator was giving us artificial images with some transformations on the original images

72
00:06:28,310 --> 00:06:28,630
only.

73
00:06:30,050 --> 00:06:37,250
So instead of inputting 2000 images in one epoch, we were inputting 30, 200 images, and all these

74
00:06:37,250 --> 00:06:40,340
images had some transformations done on them.

75
00:06:42,020 --> 00:06:50,090
This improved the accuracy of our model from nearly 73 74 percent to 83 84 percent.

76
00:06:52,420 --> 00:06:53,950
Then we made another small change.

77
00:06:54,490 --> 00:07:00,850
We removed our own convolutional base and used the convolutional base of a pre train model.

78
00:07:03,190 --> 00:07:04,720
We used we did you 16.

79
00:07:06,220 --> 00:07:10,210
You can try out on other models also with this change.

80
00:07:10,360 --> 00:07:14,080
We were able to get accuracy's in the range of 90 percent.

81
00:07:15,630 --> 00:07:22,420
And the last part, which I'm not running right now, but you can try on your only if you have a very

82
00:07:22,420 --> 00:07:31,990
good system with large arm and large Deepu, then you can probably run that part also in which you unpretty

83
00:07:31,990 --> 00:07:32,620
delayers.

84
00:07:35,630 --> 00:07:36,650
You removed this line

85
00:07:39,370 --> 00:07:47,390
from the entire model compiler for data game and then take out the desert accuracy and that will be

86
00:07:47,390 --> 00:07:49,660
coming out at nearly 97 percent.

87
00:07:52,040 --> 00:07:58,220
And with this project, I believe that you'll be pretty comfortable in building of a major recognition

88
00:07:58,220 --> 00:08:01,940
model using convolutional neural networks on your system.
