Meaning of constant K

[gerkulesov]

Hello!

You wrote that parameter k is the number of classes and last fc layer before loss layer must have 2k outputs. But in the paper about Ordinal Regression authors says that if k is the number of classes, than last layer must have 2(k-1) outputs -- two for each of k-1 binary classificators.

Where is the truth?

Update: I think I'll understand better if you answer one more question: is your model predicts age in diapason [0;99] or [1;100]?

[luoyetx]

If classifier of C0 ~ Ci gives label 1, then the output label is i, we predict age of [0, 99]. It's different from original paper. I recommend you can implement this layer using PyTorch. Caffe is too old to use.

[gerkulesov]

Ok. It is means that C0 is always 1, right? Did you draw the ROC curve for your model? How did you use a threshold for prediction?