search

GrumpyZhou / pytorch-dgcnn

Pytorch implementation of Dynamic Graph CNN for Learning on Point Clouds (EdgeConv)
41 stars 10 forks source link

Dynamic or Static Graph #3

Open cangumeli opened 3 years ago

cangumeli commented 3 years ago

Hi, thank you for providing this nice implementation.

Here in the classification network, it seems to me that you create the KNN graph using the input pts before every graph convolution layer. Therefore, the same KNN graph is computed at every layer. Shouldn't knn_graph take out as the input instead of pts for having dynamic graph updates?

Best regards.

GrumpyZhou commented 3 years ago

Hi @cangumeli , I think you are right. It should take the output from the previous edge_conv to the next one. If you don't mind, you can make a pull request on this change, then I can merge. Otherwise, I will change it at some point later 😅. I stopped maintaining it for years. In case the outdated environment, i suggest you to check out : https://github.com/rusty1s/pytorch_geometric/blob/master/examples/dgcnn_classification.py 😄😄.

cangumeli commented 3 years ago

Thanks, didn't see the official implementation :). I suppose I can create a pull request if the env is not too outdated.