DiceCELoss gets 0 CE component for binary segmentation

[kretes]

I'm using MONAI for binary 3D segmentation.

I've noticed that in my case the 'CE' component in DiceCELoss is 0 and therefore has no effect, however I wouldn't know that if I didn't dig deep and debug the code.

repro script:

import torch
from monai.losses.dice import DiceLoss, DiceCELoss

shape = (2,1,3)

label = torch.randint(2, shape).type(torch.float)
pred = torch.rand(shape, requires_grad=True)

dceloss = DiceCELoss(include_background=True, sigmoid=True, lambda_ce=1)

dice_loss = dceloss.dice(pred, label)
ce_loss = dceloss.ce(pred, label)

loss = dceloss.lambda_dice * dice_loss + dceloss.lambda_ce * ce_loss

print("total", dceloss(pred, label), loss)
print("dice", dice_loss)
print("ce", ce_loss)

This is basically extracted from forward of DCELoss here

MONAI/monai/losses/dice.py

Line 723 in be4e1f5

def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:

I think what's going on here is CELoss is not aimed for binary case.
However - DiceCELoss isn't shouting at me that I'm doing something wrong, and at the same time it gave me confidence I can use it for a single-channel case (e.g. because it gives some warnings about doing single-channel e.g. single channel prediction, include_background=False ignored. .).
Am I right that it should either be:

• shout at the user that that DiceCELoss can't be used in single-channel scenario
• handle this scenario internally using BCE?

[KumoLiu]

Fixed in #6924.