Weighting Classes in a Binary Classification Neural Network

5 Ansichten (letzte 30 Tage)
Arjun Desai
Arjun Desai am 25 Mai 2018
I am building a binary classification neural network. The last 3 layers of my CNN architecture are the following:
fullyConnectedLayer(2, 'Name', 'fc1');
softmaxLayer
classificationLayer
Currently, the classificationLayer uses a crossentropyex loss function, but this loss function weights the binary classes (0, 1) the same. Unfortunately, in my total data is have substantially less information about the 0 class than about the 1 class.
As a result, I want to weight the loss function to penalize misclassifying the 0 class more, with classWeights proportional to 1/(class frequency).
I noted that there is a way to weight classes in the pixelClassificationLayer but not the general classificationLayer, which I would be using as I am working on a classification problem.
How can I add class weights to my loss function for training?
  3 Kommentare
Arjun Desai
Arjun Desai am 24 Mär. 2019
Hi Samreen,
Unfortunately i could not find a good work around. might be something to pitch for future development.
Arjun
Eugene Alexander
Eugene Alexander am 28 Mai 2019
Please take a look at Define Custom Weighted Classification Layer and the example on Speech Command Recognition using Deep Learning. I am trying it right now on a binary classification problem.

Melden Sie sich an, um zu kommentieren.

Antworten (0)

Kategorien

Mehr zu Deep Learning Toolbox finden Sie in Help Center und File Exchange

Produkte


Version

R2018a

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by