ResourceExhaustedError about kpconv

[lhyfst]

You can reduce the batch size to solve the problem.

[HuguesTHOMAS]

Hi @fenglupeter,

The parameters are set for a 11GB GPU, like the one used in the paper. Like @lhyfst said, the batch size is the first thing you can reduce if you have a smaller GPU. however, you should try to keep it above 3 or 4, to avoid any problem with our varying batch strategy.

If it is not enough, you can increase the parameter first_subsampling_dl or decrease the parameter in_radius. The actual ratio is in_radius = 50 * first_subsampling_dl. You can reduce to 40 without affecting the performances drastically.

Best,
Hugues

[KangchengLiu]

I wonder how to change batch size in your config.py setting for the training of S3DIS?

[HuguesTHOMAS]

You can change the value ''batch_num'' of the Config class.
Remember that this value is an average as the batch size vary depending on the number of points in each input pointcloud