torchsparse icon indicating copy to clipboard operation
torchsparse copied to clipboard

Published 20 hours ago • verified publisher icon mit-han-lab

Reproducing results in the paper

Open ahmedius2 opened this issue 2 years ago • 4 comments

Hello,

How can we reproduce the results in the TorchSparse paper at Figure 11? Do you provide the implementation of MinkUNet and CenterPoint that utilizes TorchSparse?

Also, when I look at the CUDA sparse convolution code, it appears that your adaptive grouping strategy is not implemented. Could you tell me where should I look for it, is there a specific branch?

ahmedius2 avatar Aug 04 '22 17:08 ahmedius2

Hi @ahmedius2, thanks for your interest in our work. We have a development branch for the features related to MLSys 2022, which we are currently cleaning up and verifying the correctness. Stay tuned!

zhijian-liu avatar Aug 05 '22 05:08 zhijian-liu

@Xiuyu-Li Any progress of adaptive matrix multiplication grouping yet?

zeng-hello-world avatar Aug 19 '22 08:08 zeng-hello-world

Mark!

haiduo avatar Sep 08 '22 07:09 haiduo

@ahmedius2 @zeyu-hello @haiduo The latest main branch contains our adaptive matrix multiplication grouping strategy now. You can take a look at https://github.com/mit-han-lab/torchsparse/blob/36ca82457a053207ecb31578badb887e8356a482/examples/example.py#L98-L108 for a basic example usage. More detailed documentations and an official package will be released soon.

Xiuyu-Li avatar Sep 08 '22 23:09 Xiuyu-Li