PointNet++ two questions

[dokato]

I follow along the PointNet++ example:
https://github.com/pyg-team/pytorch_geometric/blob/master/examples/pointnet2_classification.py
Looking great, but I have two questions around the specific problem I face:

1. Looks like all the pointclouds are reduced to only 1024 points T.SamplePoints(1024). Can PointNett++ be applied also to data with variable number of points?

2. After classification, is there a way to extract from the model some sort of embedding (i.e., lower-dimensional representation of my pointcloud) with a fixed size, let's say 32?

Thanks

[rusty1s]

1. Yes, our implementation works seamlessly with a dynamic number of points across examples.
2. Of course. You can consider returning the graph-level embedding x of this line in your forward function as well.

[dokato]

Great, thanks @rusty1s . BTW since we're on that, does pytorch_geometric implement any unsupervised methods to learn embeddings from pointclouds too?

[rusty1s]

Sadly, no. Our point cloud examples are currently supervised only, sorry :( Any pointers/links highly appreciated though :)