Convert Superpixel to Mesh data

[hkim716]

I'm trying to convert MNIST superpixel dataset to Mesh data, which has pos and faces.
I know that converting graph to mesh is now an easy task, but is there a way (or previous experience) that I could try?
Thanks!

[rusty1s]

You could make use of our Delaunay transformation.

[hkim716]

Thanks Matt. It worked!

Is there a way to assign both edge_attr and face to the data object?
When I printed

transform_graph = T.Cartesian(cat=False)
transform_mesh = T.Delaunay()

mnist_dataset_graph = MNISTSuperpixels(path, True, transform=transform_graph)[:10] # True means it's train dataset
mnist_dataset_mesh = MNISTSuperpixels(path, True, transform=transform_mesh)[:10] # True means it's train dataset

print(mnist_dataset_graph[0])
print(mnist_dataset_mesh[0])

I saw

Data(edge_attr=[1399, 2], edge_index=[2, 1399], pos=[75, 2], x=[75, 1], y=[1])
Data(edge_index=[2, 1399], face=[3, 138], pos=[75, 2], x=[75, 1], y=[1])

How could I get datasets
Data(edge_attr=[1399, 2], edge_index=[2, 1399], face=[3, 138], pos=[75, 2], x=[75, 1], y=[1]) like this efficiently?

[rusty1s]

You can compose transformations via T.Compose:

transform = T.Compose([T.Delaunay(), T.Cartesian(cat=False)])

[hkim716]

When I use Delaunay transformation, then it generates new triangular meshes which has faces.
However, it looks edge_index [2, 1399] is still edges for original graphs.
How could I get and transform edge_index based on the newly generated triangular meshes?

[rusty1s]

You can call the FaceToEdge transform afterwards :)