1
00:00:00,960 --> 00:00:10,500
So the model is now fit, but when I posed the video, I had earlier done this for 20 epochs, but I

2
00:00:10,500 --> 00:00:13,110
saw that the solution was not converged.

3
00:00:13,650 --> 00:00:15,930
So I had landed for another 30.

4
00:00:16,020 --> 00:00:17,540
Still, the solution was not in words.

5
00:00:17,610 --> 00:00:19,890
So I have done it for no 50 bucks.

6
00:00:21,540 --> 00:00:24,300
So overall, I have run this model.

7
00:00:24,370 --> 00:00:32,790
Four hundred epochs and after 100 epochs, I can see that in some places, in some epochs, we have

8
00:00:32,820 --> 00:00:38,500
achieved a model giving a validation accuracy of nearly 84 percent also.

9
00:00:39,870 --> 00:00:47,910
So if you are using callbacks, you can store the best model using a callback and you can use that model

10
00:00:48,390 --> 00:00:50,850
to test your performance on your test dataset.

11
00:00:51,450 --> 00:00:54,090
For now, we have this model.

12
00:00:54,570 --> 00:01:01,800
This last model, which is giving us validation accuracy of eighty three point three six percent, which

13
00:01:01,800 --> 00:01:02,520
is also good enough.

14
00:01:02,640 --> 00:01:11,940
You can see that we have achieved over eight nine percent accuracy improvement just by doing the documentation.

15
00:01:13,200 --> 00:01:22,200
So even if you have fewer data observations to bring in CNN, you have the option of creating artificial

16
00:01:22,200 --> 00:01:30,120
data by using the data augmentation for which we used this function image data generator.

17
00:01:32,310 --> 00:01:34,860
So this was the second part of our project.

18
00:01:36,090 --> 00:01:41,400
We have seen considerable improvement in accuracy after using data augmentation.

19
00:01:42,780 --> 00:01:48,950
And the last part we are going to use up retained CNN model.

20
00:01:50,310 --> 00:01:56,610
There are certain architectures which have been trained on large datasets and these networks are huge

21
00:01:56,760 --> 00:02:00,840
and they have performed really well in certain competitions.

22
00:02:01,530 --> 00:02:05,480
And some of these architectures are part of a decade as Labadie.

23
00:02:05,820 --> 00:02:09,940
So we can straightaway use the weight of these Bre-X train models.

24
00:02:10,800 --> 00:02:19,370
So in the next video, we are going to use such a retrain model and applied on our God vs. Dogs dataset.

25
00:02:19,870 --> 00:02:22,660
And let us see if we can improve the accuracy for the.