search

NVIDIA / MinkowskiEngine

Minkowski Engine is an auto-diff neural network library for high-dimensional sparse tensors
https://nvidia.github.io/MinkowskiEngine
Other
2.48k stars 367 forks source link

Create layer with ME.RegionType.CUSTOM #303

Closed dragonlong closed 3 years ago

dragonlong commented 3 years ago

Hi Chris,

Thanks for keep updating this library. I tried to run the Synthia 4D experiment based on ME v0.5 pn a remote CentOS server. Everything works well, except that when I try to create a ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS convolution layer, which will be mapped to ME.RegionType.CUSTOM automatically. But it seems region_offsets is not generated in the convert_conv_type() function, which is needed by ME.KernelGenerator() . A naive NxD list doesn't seem to work here. If it doesn't bother too much, may I know what would be a quick fix to that? I noticed that it seems to be similar to HYBRID in previous versions.

Thanks!

Best, Xiaolong

dragonlong commented 3 years ago

Just a follow-up, I am able to use 'ConvType.SPATIO_TEMPORAL_HYPERCROSS' or 'ConvType.SPATIO_TEMPORAL_HYPERCUBE', the training seems to go well. So I will close this issue report now. Thanks!

Pexure commented 3 years ago

Hi @dragonlong I have the same problem with ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS. May I ask how ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS is mapped to ME.RegionType.CUSTOM automatically? And is there any compensation for missing ME.RegionType.HYBRID in 0.5.x?

sandeepnmenon commented 3 years ago

+1 I have the same issue with 0.5.x version of ME. Not sure how to map ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS with ME.RegionType.CUSTOM

@dragonlong Where you able to reproduce the results by replacing this with 'ConvType.SPATIO_TEMPORAL_HYPERCROSS' or 'ConvType.SPATIO_TEMPORAL_HYPERCUBE' ? Did you change the CONV_TYPE for STRes16UNetBase ?