weight parameter of DiceCELoss

[NastaranVB]

Hello,
I am doing multi class segmentation which contains 3 classes including background. I want to consider weight parameter of "DiceCEloss", so have defined loss function as below:
loss_function=DiceCELoss(include_background=False, to_onehot_y=True, softmax=True, weight=torch.tensor([1, 10])).
I get the below error:
ValueError: the length of the weight sequence should be the same as the number of classes.
If include_background=False, the weight should not include
the background category class 0.

Thank you in advance.

[KumoLiu]

Hi @NastaranVB, please set the weight contain background; this will remove the background class internal. Because weight must also be used in 'CrossEntropyLoss'.

MONAI/monai/losses/dice.py

Line 714 in e9e2738

dice_weight = weight[1:]

Thanks.

[NastaranVB]

@KumoLiu thanks for coming back to me so quickly.
So, it should be as below, am I right?
loss_function=DiceCELoss(include_background=False, to_onehot_y=True, softmax=True, weight=torch.tensor([1., 2., 10.])).
And now the weight is applied for both 'DiceLoss' and 'CrossEntropyLoss'?

As an another question, let me ask what is the parameter "ce_weight=None" in definition of class "DiceCELoss"? Is it weight related to 'CrossEntropyLoss'? How should it be set?
I just got a bit confused, I would be thankful if you could please explain a bit how I should set "ce_weight' and "weight".

Thank you in advance.

[KumoLiu]

Hi @NastaranVB, it's a deprecated argument which will be removed in 1.4.

MONAI/monai/losses/dice.py

Line 650 in e9e2738

"ce_weight", since="1.2", removed="1.4", new_name="weight", msg_suffix="please use `weight` instead."

Thanks.