mit-han-lab / torchsparse

[MICRO'23, MLSys'22] TorchSparse: Efficient Training and Inference Framework for Sparse Convolution on GPUs.

https://torchsparse.mit.edu

Does this repo support Deconv or other upsample layer?

kellenf opened this issue 8 months ago · comments

kellenf commented 8 months ago

Thanks for this work, but how can I use a deconv or other upsample layer? It seems not supported?

Zhijian Liu commented 8 months ago

You may use the transposed sparse convolution to upsample the data: https://github.com/mit-han-lab/torchsparse/blob/master/torchsparse/backbones/modules/blocks.py#L40-L47.