Combined loss

As far as I know, the ncc loss returns a negative value. How we combined it with another loss, such as segmentation loss (cross-entropy, dice…)–that is positive value?

I suggest we can use a upper bound value and plus it with ncc to convert it to the positive range, such as

L_total = (5+ncc_loss) + dice_loss

where 5 is the maximum loss value that NCC can achieve. But the problem is how we can define the upper bound value, it really depends on the dataset.

[adalca]

the Dice loss is available in the losses.py file.

Or do you mean the whole semi-supervision part of the network? if so, This is available in the redesign branch under VxmDenseSegSemiSupervised – this branch will be merged into the master soon, but can still be used if you just switch the branch in your current repository clone.

[happybcloud]

Would you please share the code that combined loss with segmentation (cross-entropy, dice…)-