关于sparse的计算量

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

1.15k stars 131 forks source link

关于sparse的计算量 #277

Closed nancyhxn closed 6 months ago

nancyhxn commented 7 months ago

请问将sparse运用在自己的算法中，该如何计算sparse所占用的计算量？我目前的发现是直接调用“from thop import profile”的profile函数，不能计算到sparse的计算量，会被直接忽略掉。请问这个问题该如何解决？

zhijian-liu commented 6 months ago

Yes, you need to write a customized hook for sparse convolution.

ys-2020 commented 6 months ago

Hi @nancyhxn , I think the kmap for sparse convolution might be helpful. You can refer to the out_in_map to calculate the theoretically minimal flops for sparse convolution. You may also need the reduced_mask to compute the actual macs in TorchSparse v2.1.0. Please refer to our paper for more details.

ys-2020 commented 6 months ago

Close this issue due to inactivity. Please feel free to reopen it if you have any further questions.