The implementation of DiceLoss

[hwei-hw]

Thanks for your share. When I run this version (40ae1aa) code and I get the following error: `/userfolder/relaynet_pytorch/relaynet_pytorch/net_api/losses.py in forward(self, output, target, weights, ignore_index) 60 encoded_target[mask] = 0 61 else: ---> 62 encodedtarget.scatter(1, target.unsqueeze(1), 1) 63 64 if weights is None:

RuntimeError: invalid argument 3: Index tensor must have same dimensions as output tensor at /pytorch/aten/src/THC/generic/THCTensorScatterGather.cu:289 `

And the input data's size is [NumData, 1, rows, cols], the corresponding label's size is [NumData, 1，rows, cols]. The target of our is to segment 10 class.
I read the code and don't understand these code:

` relaynet_pytorch/relaynet_pytorch/net_api/losses.py

lin51：output = output.exp()

lin52：encoded_target = output.detach() * 0

line61：encodedtarget.scatter(1, target.unsqueeze(1), 1)

`

In the older versions, the implementation of DiceLoss is different from this version code. And the older vision as follows:

` class DiceLoss(_Loss):

def forward(self, input, target):

    return 1 - dice_coeff(input, target)

`

So which is better ? or which is the right version for the "ReLayNet" paper ? Thanks a lot!

[abhi4ssj]

Hello,

The previous version of the code required passing of the binarized (one-hot encoded) ground truth for the Dice Loss, whereas the new version takes in the Ground Truth as [1,..,NumClass-1], and one-hot encode them within the function. This saves some RAM space when dealing with large number of classes. Functionally, both should perform the same.

Best Abhijit

[hwei-hw]

Thanks for your reply ! I haven met some error when run the new version code. Is there some error in the input data and corresponding labels?

[abhi4ssj]

I made some changes to the DataLoader file. Not sure if you are using the latest version. I do not think the error is in the input data.

[hwei-hw]

Thanks very a lot ! I will check these possible error result from the changes in the new version code.

[MustafaArikan]

Hi @Atomwh and @abhi4ssj , im getting the same error RuntimeError: invalid argument 3: Index tensor must have same dimensions as output tensor at /pytorch/aten/src/THC/generic/THCTensorScatterGather.cu:289

The shape of my data is: Tr_Dat = np.zeros((16, 1, 512, 512), dtype="float32"); Tr_Label = np.zeros((16, 1, 512, 512), dtype="uint8"); Tr_weights = np.zeros((16, 1, 512, 512), dtype="float32");

train_data = ImdbData(Tr_Dat, Tr_Label, Tr_weights);

Which pytorch version should I use? And what is the solution to the error above?

Thanks