Usage questions about pytorch_spline_conv HOT 6 CLOSED

Hi,
no this is not correct. We scale the pseudo coordinates in preprocessing to avoid performing the same operation over and over again. Note that the scaling is automatically done when using the Cartesian() transform (when using pytorch_geometric).

from pytorch_spline_conv.

Thanks for the quick response! The cartesian transform is very helpful.

I noticed there is no mention of local rotation invariance of the filter in the paper. I'm assuming it is left to the user to apply the filter with different rotations of the pseudo coordinates and to do angular max pooling (as in GCNN) to achieve this. Is this correct?

from pytorch_spline_conv.

That is correct. Filters of SplineCNN are not rotation invariant, but the rotation invariance can be enforced by the method you mentioned.

from pytorch_spline_conv.

Good to know. I haven't implemented the angular pooling but my experiments seem to be going well without it. Thanks again!

from pytorch_spline_conv.

That is correct. Filters of SplineCNN are not rotation invariant, but the rotation invariance can be enforced by the method you mentioned.

Couldn't you also just use rotationally invariant pseudo coordinates? For example using edge lengths and angles defined by the principal curvature directions, which should be both transnationally and rotationally invariant?

from pytorch_spline_conv.

Sure, something like Point Pair Features.

from pytorch_spline_conv.