Invalid index at scatter?

1

I’m trying to implement scatter_ in my custom loss function:

encoded_target = input.data.clone().zero_()
encoded_target.scatter_(1, target.unsqueeze(1), 1)

input is the output from my CNN network and is of the shape [1,1,32,64,64] and target is the ground truth used to calculate the loss, it has the shape [1,32,64,64], hence the unsqueeze(1) to bring it to the same shape as input.

input is of type float32 and target is of type long.

I get the following error at scatter_:

RuntimeError: Invalid index in scatter at c:\new-builder_2\win-wheel\pytorch\aten\src\th\generic/THTensorMath.cpp:703

Ive checked the documentation on this and my implementation is correct, Is there any reason this is not working?

P.S. I noticed that if I change the dimension in scatter_ it works:

encoded_target.scatter_(2, target.unsqueeze(1), 1)

But this is not what i want, why is working for dimension 2 and failing for 1?

Thanks

2 Likes
2

I couldnt get the scatter_ to work. As a work around I managed to create my target using a thesholding method:

where_ones = torch.ones(target.shape, dtype=torch.long) 
where_zeros = torch.zeros(target.shape, dtype=torch.long)
target_ = torch.where(target > 0.1, where_ones, where_zeros)
target = torch.unsqueeze(target_, 0).float()

This works and gives the loss output but I’m not sure if this is the preferred method.
Is what what I am doing equivalent to the output produced by scatter_?
I read in one of the docs that for target is is advised to use the scatter_ method as this was the tensor is differentiable and autograd can be utilized. Or was this the case for previous versions of pytorch?

3

I’m not sure, how you would like to apply scatter in this case, as dim1 only is of size 1.
You would get this operation:

self[i][index[i][j][k]][k] = src[i][j][k] # if dim == 1

, which means that your scatter index (target in your case` can only hold zeros.
Other values will yield an invalid index.

Could you post a small sample with values and how the result should look like?

1 Like
4

Sorry, I found this code online for DiceLoss and wanted to use it in my code, below is the full code for DiceLoss:

    def forward(self, input, target):
        weights = 1
        encoded_target = input.data.clone().zero_()
        encoded_target.scatter_(1, target.unsqueeze(1), 1)
        # encoded_target = Variable(encoded_target)
        assert input.size() == encoded_target.size(), "Input sizes must be equal."
        assert input.dim() == 5, "Input must be a 4D Tensor."
        num = (input * encoded_target).sum(dim=4).sum(dim=3).sum(dim=2)
        den1 = input.pow(2).sum(dim=4).sum(dim=3).sum(dim=2)
        den2 = encoded_target.pow(2).sum(dim=4).sum(dim=3).sum(dim=2)
        dice = (2 * num / (den1 + den2)) * weights
        return ((1-dice.sum()) / dice.size(1)) / 5

I want to create a target tensor to calculate the loss against. Thats why I thought if I create it in the same was as in my second post will it be the same thing as scatter?

I assume scatter here is trying to do the same thing…

5

@farazk86
Hi, I am working on a code and I have run to the same error derivating form the loss function. This link explains my error https://github.com/michuanhaohao/reid-strong-baseline/issues/120. Did you find a solution for fixing the error?