Can I reconstruct an image from the graph?

[yc-cui]

@Somedaywilldo visualizes the mnist_superpixel: https//github.com/pyg-team/pytorch_geometric/issues/320#issuecomment-828382760

My questions is whether the mnist_superpixel can be reconstructed to the origial RGB image?

[rusty1s]

Thanks for the issue. The superpixel graph is a reduced/sparse representation of the original RGB image (the image is reduced from 28x28 pixel to 75 pixels), so there is no way to reconstruct it in a lossless fashion.

For MNIST, you could create a Voronoi diagram from the given point cloud, and color each cell with the color of the original node. This should give you a pretty good visualization of the original MNIST image.

[rusty1s]

We do not have an official example for this but I once created these two gists:

• https://gist.github.com/rusty1s/74b1efe973482e070c50f650ece82b70
• https://gist.github.com/rusty1s/866ca8b830f22aa34923a0f8164f6e64

Hopefully, they are helpful to you.

[yc-cui]

Many thanks！

[yc-cui]

We do not have an official example for this but I once created these two gists:

• https://gist.github.com/rusty1s/74b1efe973482e070c50f650ece82b70
• https://gist.github.com/rusty1s/866ca8b830f22aa34923a0f8164f6e64

Hopefully, they are helpful to you.

Hello, @rusty1s . Sorry to bother you again. I have seen these two gists. May I ask how could I combine Dataloaders in PyTorch and in PyG? The batch from Dataloader in PyG is different from Dataloader in PyTorch. In PyTorch, it's like [B, C, H, W]. I have implemented a part of my network (which deals with images using CNN) using PyTorch, and I want to insert some GNN layers into the network. However, the input of the GNN layers is the output of the CNN which is [B, C, H, W]. How could I deal with it in this situation?

[rusty1s]

For a GNN layer in PyG, it means that B, C and H denote the nodes in your graph (for which you want to share parameters). As such, if x is given by [B, C, H, W] you need to convert it to [B * H * W, C] via x.permute(0, 3, 1, 2).view(-1, num_features).

[yc-cui]

Oh! I see. Many thanks again :)