SparseResUNet42 should have LAYERS = (2, 3, 4, 6, 2, 2, 2, 2) instead of LAYERS = (2, 2, 2, 2, 2, 2, 2, 2)

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

SparseResUNet42 should have LAYERS = (2, 3, 4, 6, 2, 2, 2, 2) instead of LAYERS = (2, 2, 2, 2, 2, 2, 2, 2) #301

Closed YilmazKadir closed 2 months ago

YilmazKadir commented 3 months ago

From my understanding, in the original Spatiotemporal Segmentation paper MinkowskiNet42 (Res16UNet34C) has LAYERS = (2, 3, 4, 6, 2, 2, 2, 2). However, your SparseResUNet42 implementation has LAYERS = (2, 2, 2, 2, 2, 2, 2, 2). Is there a mistake in the naming or am I missing a detail?

zhijian-liu commented 2 months ago

Our implementation serves as a reference point. We didn't adhere to the exact settings of MinkowskiNet.