Regression with pytorch geometric
See original GitHub issueHi @rusty1s,
I am interested to use pytorch_geometric for a regression problem and I wanted to ask you whether you think it would be possible. To give you an understanding of my dataset I have a set of point clouds of different sizes and for which I have available the vertices n
, faces f
(quad meshed) and a set of features vector fx8
which include the normals, the centroid, the face area, etc. Now for each face I would like to regress an intensity value for the corresponding target that I want to give as an input. Due to the spatial arrangement of my dataset I would have a lot of occlusions and a kind of long tail distribution of these intensity values.
Considering that I am quite new to pytorch_geometric/GCNs/GNNs do you think that this would be feasible. If yes do you have any specific model which you think that it would be best to use and possibly point me to point me to an example if there is any.
Thanks a lot.
Issue Analytics
- State:
- Created 2 years ago
- Comments:6 (1 by maintainers)
Thanks a lot, quite helpful. I think I could modify the
from_scratch()
function to load my data.Skipping through the
from_scratch()
function I have a few more questions though:Is it possible to load any custom extra features other than the ones extracted here.
My meshes point coordinates are in meters which could vary from 0 to 10/15m on each mesh, do I need to normalize them to [0,1] or it is not necessary. If yes, each mesh should be normalized individually based on each own
min
,max
dimensionality or based on themin
,max
dimensions of the whole dataset.Hi @ttsesm, MeshCNN works with .obj files (you can see an example of different datasets in the link you shared here). If you want to understand fully how it loads the data, you should go to get function in MeshCnnClassificationDataset class here. The actual loading is in MeshPrepare class here - go to “from_scratch” function. Hope it helps.