search

mit-han-lab / torchsparse

[MICRO'23, MLSys'22] TorchSparse: Efficient Training and Inference Framework for Sparse Convolution on GPUs.
https://torchsparse.mit.edu
MIT License
1.16k stars 132 forks source link

[BUG] Only support subm mode? #249

Closed matri123 closed 8 months ago

matri123 commented 9 months ago

Is there an existing issue for this?

Current Behavior

In rule_book generation process, only has subm mode's rule_book. so, it doesn't have normal mode? And in spconv process, only subm rule_book can be used? or, normal rule_book also works?

Expected Behavior

No response

Environment

- GCC:
- NVCC:
- PyTorch:
- PyTorch CUDA:
- TorchSparse:

Anything else?

No response

zhijian-liu commented 9 months ago

@ys-2020, could you please take a look at this issue when you have time? Thanks!

ys-2020 commented 9 months ago

Hi @matri123 , a normal convolution mode is supported in our framework. Could you further clarify which part makes you confused? Maybe with a link to the doc or a sample code. Thanks

matri123 commented 8 months ago

https://github.com/mit-han-lab/torchsparse/blob/master/torchsparse/nn/functional/build_kmap.py#L32 https://github.com/mit-han-lab/torchsparse/blob/master/torchsparse/nn/functional/build_kmap.py#L54 Only has subm and downsample, and doesn't has normal mode.

ys-2020 commented 8 months ago

Hi. In subm mode, the input feature map will not be downsampled. (The output resolution and point coordinates will be kept the same as input.) I guess this might be the normal mode you want?

ys-2020 commented 8 months ago

Close this issue as complete. If you have further questions, feel free to reopen it.