Mixed precision (16-bit) training error

[asadabbas09]

I'm trying to use 16 bit precision in pytorch lightning to save some gpu memory, but I'm getting this error:

RuntimeError:MinkowskiEngine/src/convolution_gpu.cu:69, assertion (in_feat.scalar_type() == kernel.scalar_type()) failed. type mismatch

Is there a way to fix this error?

[jh-chung1]

same problem..

[Ltwicke]

I also get that error and i dont understand, how im supposed to influence the data type of the kernel, as im using standard ME convs

EDIT: I fixed the problem by adjusting dtypes in the creation of my sparse tensors. dtype=torch.int16 did it for me

[houyongkuo]

same problem

[houyongkuo]

I carefully compared my input with the input of FCGF, Minkowskiengine/example/reconstruction.py&completion.py, and found that the dtype in my input is different from the others, mine is float64, the others are float32, guess this should be the reason error caused by

[luoao-kddi]

Hi all, I met this problem too, I found it occurs after removing nn.Sequential in my code.

work well: self.module = nn.Sequential(ME.MinkowskiConvolution(3, 128, 3, dimension=3))

error: self.module = ME.MinkowskiConvolution(3, 128, 3, dimension=3)

[ZiliangMiao]

same problem, you have to make sure the tensor type of features is torch.float32.