search

masszhou / spconv_lite

sparse convolution lib. derived from spconv
Apache License 2.0
50 stars 11 forks source link

Why sparse convolutional inference time is longer #1

Open jiaerwang0328 opened 2 years ago

jiaerwang0328 commented 2 years ago

run python test_timeit.py: compare time with nn.Conv3d() nn.Conv3D() forward time is :0.000066 SparseConv3d() forward time is :0.001239 nn.Conv3D() backward time is :0.000870 SparseConv3d() backward time is :0.001930

ColdCodeCool commented 2 years ago

maybe your sparsity of test case is small, like below 50%, as far as I concerned, the sparse version of Conv performs better when the sparsity of feature is big like above 70%. Hope the author will explain this