I am doing 4-class semantic segmentation with U-net using generalised dice loss as loss function.

General approach to save best model during training is to monitor validation loss at each epoch and save the model if val loss decreases than previous minimum.

But, I am interested in the "model which gives best the average dice score of 4 classes".

During training in my case, using validation loss as criteria doesn't lead to best avg dice score. So what should I consider the best model the one with less validation loss or highest validation dice score?

Below is my validation loss and avg dice score after each epoch. Out of this which epoch gives the best model?

epoch  1/10     validation loss: 0.95     avg dice score: 0.17
epoch  2/10     validation loss: 0.86     avg dice score: 0.23
epoch  3/10     validation loss: 0.77     avg dice score: 0.34
epoch  4/10     validation loss: 0.74     avg dice score: 0.40
epoch  5/10     validation loss: 0.71     avg dice score: 0.45
epoch  6/10     validation loss: 0.69     avg dice score: 0.34
epoch  7/10     validation loss: 0.79     avg dice score: 0.45
epoch  8/10     validation loss: 0.75     avg dice score: 0.51
epoch  9/10     validation loss: 0.76     avg dice score: 0.36
epoch 10/10     validation loss: 0.75     avg dice score: 0.38

If I go by val loss as criteria epoch-6 gives the best model, if i choose avg dice score as critieria epoch-8 gives the best model? how to choose?

  • $\begingroup$ Why not just measure the average DICE score on the validation set? $\endgroup$
    – Sycorax
    May 10, 2021 at 15:49
  • $\begingroup$ Actually average DICE score is measured on the validation set itself $\endgroup$
    – spb
    May 10, 2021 at 15:50
  • 1
    $\begingroup$ Then what's the problem here? $\endgroup$ May 10, 2021 at 15:52
  • $\begingroup$ My question is, what should be the criteria to choose the best model? with min validation loss or highest dice score? $\endgroup$
    – spb
    May 10, 2021 at 15:54
  • 2
    $\begingroup$ If DICE is the sole criterion for the best model, why does it matter if it corresponds to the minimum val loss or not? $\endgroup$
    – Sycorax
    May 10, 2021 at 16:32


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.