Pytorch 扩展类标签张量

Question

I am working on a semantic segmentation project in pytorch and I have class maps in the following shapes: [H,W] where each element is an integer between 0-n where n is the number of classes, H the height of the image and W the width of the image.

Here is an example:

test_label = torch.zeros([10,10])
test_label[:5,:5] = 1
test_label[5:,:5] = 2
test_label[:5,5:] = 3
test_label

Output:

tensor([[1., 1., 1., 1., 1., 3., 3., 3., 3., 3.],
    [1., 1., 1., 1., 1., 3., 3., 3., 3., 3.],
    [1., 1., 1., 1., 1., 3., 3., 3., 3., 3.],
    [1., 1., 1., 1., 1., 3., 3., 3., 3., 3.],
    [1., 1., 1., 1., 1., 3., 3., 3., 3., 3.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.]])

Now, what I want is something of the shape [n,C,H] where [1,C,H] would be eg:

tensor([[1., 1., 1., 1., 1., 0., 0., 0., 0., 0.],
    [1., 1., 1., 1., 1., 0., 0., 0., 0., 0.],
    [1., 1., 1., 1., 1., 0., 0., 0., 0., 0.],
    [1., 1., 1., 1., 1., 0., 0., 0., 0., 0.],
    [1., 1., 1., 1., 1., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])

And [2,H,W] would be:

tensor([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.],
    [2., 2., 2., 2., 2., 0., 0., 0., 0., 0.]])

Is there a pytorch function which does this? My current approach would be iteratively masking over each unique element in the original tensor and insert them into a tensor of the shape [n,H,W] initially filled with all zeros. But that doesn't seem to be the best way to do it. I tried to look it up but it seems like I am not able to find the right name for this operation.

Thank you very much for your time.

Answer 1

You could apply nn.functional.one_hot to convert the dense format into one-hot encoding and then multiple with the label value to get the desired result:

>>> C = int(x.max()) + 1
>>> ohe = F.one_hot(x.long(), num_classes=C)

Then multiple by the label values:

>>> res = ohe*torch.arange(C)
>>> res.permute(2,0,1)
tensor([[[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],

        [[1, 1, 1, 1, 1, 0, 0, 0, 0, 0],
         [1, 1, 1, 1, 1, 0, 0, 0, 0, 0],
         [1, 1, 1, 1, 1, 0, 0, 0, 0, 0],
         [1, 1, 1, 1, 1, 0, 0, 0, 0, 0],
         [1, 1, 1, 1, 1, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],

        [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [2, 2, 2, 2, 2, 0, 0, 0, 0, 0],
         [2, 2, 2, 2, 2, 0, 0, 0, 0, 0],
         [2, 2, 2, 2, 2, 0, 0, 0, 0, 0],
         [2, 2, 2, 2, 2, 0, 0, 0, 0, 0],
         [2, 2, 2, 2, 2, 0, 0, 0, 0, 0]],

        [[0, 0, 0, 0, 0, 3, 3, 3, 3, 3],
         [0, 0, 0, 0, 0, 3, 3, 3, 3, 3],
         [0, 0, 0, 0, 0, 3, 3, 3, 3, 3],
         [0, 0, 0, 0, 0, 3, 3, 3, 3, 3],
         [0, 0, 0, 0, 0, 3, 3, 3, 3, 3],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]])