The model does not work so well.
I combined the forward projections (using torch repeat) using orthogonal vectors by transposing the projection matrices (torch transpose) and then adding them. I think that the best would be to make a logical and operation or using a better activation function.
There is a problem too with conv2d. It detects edges, but do not make the inside of the shape made by edges full (a positive grayscale value), and so it makes a projection that is not valid.
If anyone has an idea on how to detect shapes and assign a positive grayscale value inside of the enclosed face created from the edges, that would be nice.
On the other hand, using a vertices + faces reconstruction after conv2d could be more precise, but I still haven’t figured how to do it.