search

laoreja / HPLFlowNet

Code for our CVPR 2019 paper, HPLFlowNet: Hierarchical Permutohedral Lattice FlowNet for Scene Flow Estimation on Large-scale Point Clouds.
GNU General Public License v3.0
101 stars 9 forks source link

The training speed is very slow #6

Closed pxiangwu closed 5 years ago

pxiangwu commented 5 years ago

Hi, I just found that the training speed is very slow. I wonder if there is any problem with my configuration? I used Pytorch0.4.0 and cuda 9.0, on a single RTX Titan.

How long did it take to train the model for 800 epochs?

Thank you so much.

laoreja commented 5 years ago

Yes, the training is slow. Some CUDNN optimization doesn’t work for varying input size. My model is trained for about 500~600 epochs. I didn’t remember the exact training time, roughly several days.