train speed

[easyfcn]

dear ,why the POINT+GCN train speed is so slow??
I want to know your train speed.
thank you.

[chendaichao]

Thanks for your questions. The low training speed is mainly caused by the construction of input graph, where we need to compute all pairwise distances, determine the weights and pack these weights to a sparse adjacency matrix. I admit that this is really time-consuming; furthermore, since the data points are augmented by random rotation, the graph has to be reconstructed in every epoch.
In terms of the training speed, the training took about half an hour per epoch when n=1024 on a single GTX 1080Ti. When n=128 or 256 it run much faster, so indeed I first train the model with n=256, and tune it for larger n. I have also tried not using data augmentation and constructing the graph in the preprocessing stage. This really made it faster but the model seemed to overfit.
I think probably there would be a more efficient and elegant solution. If you find one, please share it with me.