Dice Score Becomes `nan` After Epoch 130 in xLSTM-uNet 3d_fullres Model Training

[matteodelucchi]

I am experiencing an issue with the xLSTM-uNet model, where the Dice score becomes nan after epoch 130. Up until this point, the Dice score has increased, and the losses have decreased as expected. However, at epoch 130, the validation Dice score suddenly drops to nan.

Possible Causes: I suspect issues with exploding gradients or divisions by zero but not sure...

• Is there a way to extract the gradients?
• What are suggested regularization techniques?
  I would appreciate any guidance or suggestions on how to resolve this issue. Thank you!