The flops may not decrease?

[Alexanzhuo]

Thank you for providing this implementation!
As one earlier issue mentioned, I found that you computed the whole convolution and then did interpolation with the mask.
I wonder that if the x here in Convolution 1 is original? That is, you have sampled x before?

[zdaxie]

Thanks for your attention!

In fact, even in test mode, the implementation does not involve real-time acceleration. We just make sure that only sampling points are used in the test. As you can see in the function self.reconstruct_feature, we only use feature * mask to reconstruct the complete feature map, which means that we don't actually need to calculate the complete convolution 1.
https://github.com/zdaxie/SpatiallyAdaptiveInference-Detection/blob/master/mmdet/models/backbones/sparse_resnet.py#L190