search

traveller59 / spconv

Spatial Sparse Convolution Library
Apache License 2.0
1.87k stars 365 forks source link

Does SparseConv3d & SubMConv3d always produce exactly the same outputs? #584

Open zodiac000 opened 1 year ago

zodiac000 commented 1 year ago

I found out by given the same set of inputs always receiving slightly different outputs. The diff is roughly about 10^-5. This is awkward. Is there any randomness involved in the implementation of 3DConv in spconv2.X or spconv1.X? If so, is there any way to avoid the randomness occuring.

jisoo0-0 commented 11 months ago

same here. .. Did you solved the problem..??

harrydobbs commented 6 months ago

Same here - did either of you find a solution?

x1y9 commented 2 months ago

same here. I have set torch.manual_seed and random.seed, still get changed outputs

jisoo0-0 commented 2 months ago

Nah... I have not found the solution for this problem. So i have repeated experiment for multiple times.

lysukhin commented 2 weeks ago

+1. Seems like the source of the randomness is somewhere inside the implementation of the sparse convolutions