Improve convergence behavior with groupnorm

[michaelmckinsey1]

Replaced BatchNorm3d in the 3D U-Net with GroupNorm, and made the number of groups configurable via group_norm_groups. This improves training stability for low batch sizes.

1 minibatch size

batchnorm

groupnorm

• need to check if this is worse at higher batch sizes than batch norm

groupnorm does converge in up to 2x less epochs, even at higher batch sizes

8 minibatch size

batchnorm

groupnorm

• should we give an option for either group norm or batch norm?
  • I think this is only necessary if groupnorm is significantly worse at higher batch sizes, which it isn't.