The reason for Loss can't be decreased in deep learning

Ref: https://blog.csdn.net/u010159842/article/details/83614791

https://blog.csdn.net/u011534057/article/details/51452564

train loss 不断下降，test loss趋于不变，说明网络过拟合;

Thus I will redesign the network .

 

To go througth the data of the classes, I use keyboard from matlab to view them

matlab keyboard

i=1;
while(1)
[~,~,button]=ginput(1);
button
      switch button
          case 28 %left
              i=i-1
          case 29 %right
              i=i+1
              
          case 27
              break;
      end
end

 

Although the three classes are ovbvious different, but the deep model is still not sensitve to classify them.

deep learning with time-frequency features

CWT and its frequecyfilter have been taken to improve its accuracy, the major problem is the mis-classify with joint and moisture...

 

Use Resenet 18 network

Accuracy of Validation is 89.4% for time-frequency features.

Although it has high accuracy on test dataset, the result is very bad on field dataset. 

 

Try use image ,not data.