WEBVTT - generated by wenglor-media

1
00:00:00.040 --> 00:00:01.440
In this video, we will learn

2
00:00:01.440 --> 00:00:04.080
how to review a trained model in AI Lab.

3
00:00:04.120 --> 00:00:06.640
We will use one of the example datasets,

4
00:00:06.640 --> 00:00:09.720
since they already include pre-trained models.

5
00:00:09.880 --> 00:00:12.960
Go to the “Models” page by clicking on
“All Models”

6
00:00:12.960 --> 00:00:15.400
and select a model you would like to review.

7
00:00:18.480 --> 00:00:19.920
On the “Model Review” page,

8
00:00:19.920 --> 00:00:23.040
you will find detailed information about its
performance —

9
00:00:23.040 --> 00:00:25.560
including balanced accuracy in percent,

10
00:00:25.560 --> 00:00:27.280
the number of wrong predictions,

11
00:00:27.280 --> 00:00:29.160
the inference time in milliseconds

12
00:00:29.160 --> 00:00:31.680
and the recall and precision for each class.

13
00:00:40.200 --> 00:00:41.920
The “Confusion Matrix” shows

14
00:00:41.920 --> 00:00:45.360
how the actual ground truth compares to the
predicted results,

15
00:00:45.360 --> 00:00:48.240
and how accurately each class was identified.

16
00:00:52.960 --> 00:00:54.960
You can also view examples of correct

17
00:00:54.960 --> 00:00:56.960
and wrong predictions.

18
00:01:01.000 --> 00:01:02.000
In the “Info” tab,

19
00:01:02.000 --> 00:01:05.000
you will see details such as image resolution,

20
00:01:05.000 --> 00:01:08.000
dataset size and the number of classes.

21
00:01:08.000 --> 00:01:10.000
Click “Review Wrong Predictions”

22
00:01:10.000 --> 00:01:13.000
to inspect the misclassified images in full
size.

23
00:01:20.000 --> 00:01:22.000
Here, you can enable the “Heatmap” to understand

24
00:01:22.000 --> 00:01:25.000
which parts of the image influenced the prediction.

25
00:01:33.000 --> 00:01:36.000
If needed, you can relabel these images

26
00:01:36.000 --> 00:01:38.000
and retrain the model to improve accuracy.

27
00:01:54.000 --> 00:01:56.000
Finally, all model performance details

28
00:01:56.000 --> 00:01:58.000
can be viewed as a printable report —

29
00:01:58.000 --> 00:02:01.000
perfect for sharing or documentation.