HausdorffERLoss is not differentiable

[shijianjian]

I wrote a doctest for the erosion based hausdorff loss and I found it is not differentiable:

    Examples:
        >>> image_preds = torch.tensor([[
        ...    [0, 0.1, 0.1, 0],
        ...    [0, 0.3, 0.6, 0],
        ...    [0, 0.4, 0.8, 0],
        ...    [0, 0, 0, 0],
        ... ]], requires_grad=True)[None]
        >>> image_gt = torch.tensor([[
        ...    [0, 1, 0, 0],
        ...    [1, 1, 1, 0],
        ...    [1, 1, 1, 0],
        ...    [0, 0, 0, 0],
        ... ]])[None]
        >>> hdloss = HausdorffERLoss(from_logits=False)
        >>> hdloss(image_preds, image_gt)
        tensor(0.0625)
        >>> hdloss(image_preds, image_gt).backward()

I would suspect something is missing from your implementation.

[JunMa11]

Hi, @shijianjian
I do not implement the erosion-based HD loss.
I only use the DTM-based HD loss in my repo.

[shijianjian]

Isn't this the erosion based HD loss?

https://github.com/JunMa11/SegLoss/blob/master/losses_pytorch/hausdorff.py#L83-L173

[JunMa11]

Oh, I apologize for the inconvenience.

The erosion-based HD loss was copied from https://github.com/SilmarilBearer/HausdorffLoss and I have not tested it.

https://github.com/JunMa11/SegLoss/blob/71b14900e91ea9405d9705c95b451fc819f24c70/losses_pytorch/hausdorff.py#L15

I would recommend raising an issue in the original repo.

[lucafei]

is the HausdorffDLoss possible for multi class segmentation？

[shijianjian]

@JunMa11 Sure.

@lucafei It worked in multi-label segmentation from BCHW to BCHW.

[CheeseInTrap]

Hi, I wonder why there is no normalize operation when calculating edt in HausdorffDTLoss, the loss value may be too large