Open seanco-hash opened 4 years ago
I encountered the same issue. I am using PointNet++ instead of PointNet as a backbone in another network and the training speed is very slow. Is this expected? I also have batch size reduced to 10 because GPU memory allocation errors. GPU utilization is 35-40%
Hi, My teammates and I are trying to use PointNet++ with our data. While on PointNet both the results and the running time are good (100 epochs in 25 minutes) PointNet++ training is very slow (3 epochs in 25 minutes) and we use google colab to run it so the time we have is not enough to run enough epochs to get good results.
Is PointNet++ really supposed to be so slow in comparison to PointNet? Is there something we can do to make it faster? thanks!