Dice Loss is Negative for some steps

[shakeel608]

Hi,

I am trying to follow the 3D segmentation Tutorial,

When I run the code on CPU local machine, the tutorial is working fine, But the moment I start training on GPU cluster (CUDA), the model performance is poor and it give dice loss negative for some trainining steps

[aymuos15]

Which tutorial exactly? And by some training steps, do you mean it is fluctuating between values or it just suddenly turns to negative?