Question about cfgs file
Seong-min-Park opened this issue · 4 comments
Thank you for the excellent research.
I have a question about the cfgs file provided in the code. When I tried to reproduce using the cbgs_voxel0075_voxelnext.yaml, I noticed that the FLOPs calculation shows a model with a spatial pruning ratio of 0 (83.8 GFLOPs), as seen in Table 3. Is there a model config file available with a spatial pruning ratio of 0.5, or is there a way to separately configure the pruning ratio?
You can change the backbone name to be "VoxelResBackBone8xVoxelNeXtSPS" in the config file.
Thank you for your response. I have an additional question. To obtain a model with 33.6 GFLOPs, similar to the result with a ratio of 0.5 in table3, should I set the pruning ratio of the SPS module to 0.5? (When I tried reproducing it, setting the pruning ratio of SPS module to 0.5 didn't result in the expected reduction in computational cost.)
Yes. 0.5 is the right setting. Would you please show me how you calculate FLOPs?
I used the method provided in this work.
https://github.com/CVMI-Lab/SparseKD
I measured the computation cost of the model by directly calculating the number of active points in each layer using input.features.shape[0] in the conv forward of the spconv library. The flops of each layer were computed as input_channel * output_channel * kernel_size^dim * #active_points, and then I summed up the flops of all layers.