fadamsyah / pytorch_brain_mri_segmentation Goto Github PK

Thanks for sharing your work. It is really cool.
In my opinion, there are some minor misktakes in your code about DiceLoss and Dicesccore.
As for the DiceLoss, i think the DiceLoss should be calculated by averaging the loss in the batch,which is more consistent with the definition of the DIceloss.
So i code this in the way as follows.
class Diceloss(nn.Module):
def init(self):
super().init()
def forward(self,prediction,target,smooth=1e-12):
assert prediction.shape==target.shape,'the dimensions of prediction and target do not match/diceloss'
batch_size=prediction.shape[0]
intersection=(prediction.view(batch_size,-1) * target.view(batch_size,-1)).sum(dim=1)
union=prediction.view(batch_size,-1).sum(dim=1)+target.view(batch_size,-1).sum(dim=1)
dice=(2*intersection+smooth)/(union+smooth)
dice_loss = 1 - dice
dice_loss=torch.mean(dice_loss)
return dice_loss