Examples

This Page gives examples and explanations how this library can be used.

Mask Tensor

A mask tensor is a mono value index matrix. It describes at which pixels in an image the pixel is assigned to a class index.

A mask tensor is an 3D tensor with following dimensions:

\[Shape = [N, H, W]\]

with \(N\) as the number of masks, \(H\) as the height of the image and \(W\) as the width of the image, following the order of the dimensions like Pytorch.

As an Example we use this small example of these three 3 by 3 masks. All three masks are displayed overlapping and since in this example only the zeros overlap we are able to see all three of them, with the \(blue\) mask on layer 0, the \(red\) mask on layer 1 and the \(orange\) mask on layer 2.

with \(idx_{blue} = 1\) , \(idx_{red} = 3\) and \(idx_{orange} = 2\).

So the blue pixel in the upper middle has the indices \(N=0\), \(H=0\), \(W=1\) and the \(value=1\). This is also the most left entry in the code example below. Each mask in the code example is ordered from left to right blue -> red -> orange.

To create a Mask Tensor use the MaskSparseTensor class with an index and a value tensor as additional attributes to the size:

import torch

from pytorch_sparse_special.special.sparse_mask import MaskSparseTensor

indices = torch.tensor(
    [
#      < blue   >  < red    >  < orange>
      [0, 0, 0, 0, 1, 1, 1, 1, 2],
      [0, 1, 1, 2, 0, 0, 2, 2, 1],
      [1, 0, 2, 1, 0, 2, 0, 2, 1],
    ],
)

values = torch.tensor(
#  < blue   >  < red    >  < orange>
  [1, 1, 1, 1, 3, 3, 3, 3, 2]
)

size = (3,3,3)

tensor = MaskSparseTensor(indices, values, size)

indices define the position of the pixel in the tensor. They have to match the 3D space.

values define the class index for each pixel. Shape has to match the Number of pixels of indices.

For additional information about the methods implemented to the class refer modules section.