Closed fishbotics closed 1 year ago
I'm realizing that these tensors are state-ful (unlike traditional Pytorch tensors). I'll leave this to the developers to close, but this is a non-issue. Perhaps this could go in the README?
Thanks for reporting this issue. Yes, sparse tensors are stateful, which means you need to reset the cache in the second run.
@zhijian-liu : what's the right way to reset the cache?
pc.cmaps.clear()
pc.kmaps.clear()
Is there an existing issue for this?
Current Behavior
Hi,
I am running into a very strange issue that I don't know how to replicate super well. I've made a super simple model and when I call it once on the input, the results make sense. When I call it a second time (on the original input), the length of the coords are wrong, while the length of the feats are not. This means I cannot, for example, use GlobalAveragePool because the coordinates do not match the features.
One thing I noticed is that the
cmaps
andkmaps
of the input variablepc
are being modified (I checked by doing a deep copy and then . My guess is that in the__add__
function, these are being assigned by reference to another tensor and then modified on the other tensor (which in turns modifies them on the input tensor). Just to show that:Expected Behavior
I expect the length of the coords to always match the length of the feats
Environment
Anything else?
No response