Inconsistent convolution result between WarpConvNet and MinkowskiEngine

[Xiashangning]

Hi, thanks for open-sourcing yet another sparse pytorch library, excellent job!

However, when I comparing the results from SparseConv3d with that from MinkowskiConvolution, I found that even if the input and their weight/bias are strictly equal, their outputs are very different.

Therefore, I would like to know if there are any difference in the implementation of the sparse 3d convolution algorithm.

Besides, it seems that SparseConv3d is much slower than MinkowskiConvolution, with 120% more computational time used. Is this expected? How can we improve the speed of WarpConvNet?

Thanks,

[chrischoy]

The order of kernel is inversed. Try setting weights from MinkowskiEngine to weights[::-1, :, :] for Warpconvnet.

With the new automatic kernel selection, it should be faster or at least as fast as MinkowskiEngine.

[Xiashangning]

Thanks for your response! I will try it asap.