[BUG] what 2.0.0b0 diffrence about 2.1.0? use 2.1.0 and SparseResUNet42, at foward encode at last layer, coord.shape is [0,4],but use 2.0.0.b0,it is ok.
yeyan00 opened this issue · comments
Is there an existing issue for this?
- I have searched the existing issues
Current Behavior
use 2.1.0 and SparseResUNet42 as backbone, same data,after encoder last layer of at foward downsample , the coord.shape is [0,4], but switch 2.0.0.b0,nothing modifed,it work well.
for pypi server is blocked, i install 2.1.0 from source code by use
python -m pip install -e .
when use 2.1.0,i set coordinate order is [batch, x, y, z],when use 2.0.0b0,the order is[x, y, z, batch]
is there other parameters should i set?
unet.py-->_unet_forward
# len(encoders)=1
# downsample
xd = encoders[0](x)
# set coords order
if ts_version>='2.1.0':
x.coords = torch.cat((batch.reshape(-1,1),x.coords), dim=1).int()
else:
# version<=2.0.0
x.coords = torch.cat((x.coords, batch.reshape(-1,1)), dim=1).int()
Expected Behavior
No response
Environment
- GCC:9.4.0
- NVCC:11.6
- PyTorch:1.13
- PyTorch CUDA:11.6
- TorchSparse:2.1.0
Anything else?
No response
We have just fixed our PyPI server. Could you please try installing TorchSparse with PyPI? Thanks!
Hi @yeyan00 , could you please provide a runnable snippet code for us to reproduce the error if it still exists?
Hi @yeyan00 , could you please provide a runnable snippet code for us to reproduce the error if it still exists?
ok,i will install use pypi server first, if it still has this error, i will provide runnable code
We have just fixed our PyPI server. Could you please try installing TorchSparse with PyPI? Thanks!
i install it from download whl file,train 3 epoch, it seems ok.
is there any difference between compiled version and source code?
i install from latest source code like this:
export CUDA_HOME=/home/szlt/anaconda3/envs/torch_113
python -m pip install -e .
I think there is no major difference between the pre-compiled version and the source code. Please feel free to use the pre-compiled version if you cannot build it from source successfully.
Close this issue as completed. Feel free to reopen it if you have any further questions.