- set 'BatchNormalizationStatistics','moving'
- set 'Shuffle','every-epoch'
- use larger miniBatchSize
The reason for the sudden changes in the amount of accuracy in the MATLAB Deep Learning application?
1 view (last 30 days)
Show older comments
My goal is to detect corona from CT scan using neural networks (squeezenet). i have 2 classes of positive corona or negative corona. In both experiments, only Gaussian noise was added to the training images and in Validation images werenot used.
I used the original CT scan images in one type of experiment (type 1)
![](https://www.mathworks.com/matlabcentral/answers/uploaded_files/587726/image.jpeg)
And in another type of the same photos with the difference that the lungs have been extracted in them (the second type)
![](https://www.mathworks.com/matlabcentral/answers/uploaded_files/587731/image.jpeg)
Exactly one graph and exactly one value is used for all parameters
It should be noted that the size of these images is smaller than the first type of images, but the size of the input layer in both cases is equal to 224 224 1
But when I do the evaluation with 664 training samples and 262 samples, in the second type, the amount of accuracy drops sharply, from 62 to 39.85, while the work process in the first type seems normal.
Please help me what is the cause of this problem
type 1
![](https://www.mathworks.com/matlabcentral/answers/uploaded_files/587736/image.jpeg)
type 2
![](https://www.mathworks.com/matlabcentral/answers/uploaded_files/587741/image.jpeg)
0 Comments
Answers (1)
CT Xu
on 25 Feb 2023
A few practices may help:
In addition, set smaller learning rate may also help stably increasing validation accuracy during training
0 Comments
See Also
Categories
Find more on Image Data Workflows in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!