lncc_loss gives too negative value back

[mikami520]

Issue

I am working on Deep Atlas and using the local normalized cross correlation loss function (return -LNCC instead of 1-LNCC) to compute the similarity loss between warped moving image and fixed image. However, the loss is too big (too negative shown in the screenshot) and makes me confused.

Result

The loss of each epoch is also saved as text file attached below.
similarity_loss_reg.txt

Need Help

Could anyone help me work on this confusion? I am appreciated your help, thanks!

[ebrahimebrahim]

A version of this discussion is here