search

ma-xu / pointMLP-pytorch

[ICLR 2022 poster] Official PyTorch implementation of "Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework"
Apache License 2.0
501 stars 65 forks source link

what takes advance in using 'from pointnet2_ops import pointnet2_utils' #63

Closed hawkinglai closed 2 years ago

hawkinglai commented 2 years ago

I noticed that u have achieved the part of 'farthest_point_sample' by function, but in your repo, u are using pointnet2_utils.furthest_point_sample(xyz, self.groups).long() instead of that function. I am new coming to the research of point cloud, so I have this question. Why u use this lib but that achieved function?

By the way, why I have this question because I ran this code and found that I could not do the fps. out = _ext.furthest_point_sampling(xyz, npoint) RuntimeError: false INTERNAL ASSERT FAILED at "pointnet2_ops/_ext-src/src/sampling.cpp":83, please report a bug to PyTorch. CPU not supported Do u know how to fix it?

It is very enough to answer the first question. If u could answer those, plz update me anytime and I will appreciate u with many thanks!

ma-xu commented 2 years ago
  1. using pointnet2_utils.furthest_point_sample can achieve faster speed since it operations on CUDA.
  2. The reason is that you did not install the cuda lib. Install the lib will fix this problem.

For your development, you can simply uncomment line 160 https://github.com/ma-xu/pointMLP-pytorch/blob/b5dcf4d0ca2085d10fe1dc5e6f2972cddec8399d/classification_ScanObjectNN/models/pointmlp.py#L160 and comment line 161 https://github.com/ma-xu/pointMLP-pytorch/blob/b5dcf4d0ca2085d10fe1dc5e6f2972cddec8399d/classification_ScanObjectNN/models/pointmlp.py#L161 to run codes on cpu (I would not suggest training like this due to slow speed).

hawkinglai commented 2 years ago

Inspire me a lot for your reply with much information. Thank u!

hawkinglai commented 2 years ago

Back to the paper, how to plot out the 'Loss landscape' like figure 4?

ma-xu commented 2 years ago

@Yukimori-GitHub See here: https://github.com/ma-xu/pointMLP-pytorch/issues/42#issuecomment-1153547400

hawkinglai commented 2 years ago

Ok, yes. And now I have a new question, how can I get the information about the train/test speed like table 2?

ma-xu commented 2 years ago

@Yukimori-GitHub See here: https://github.com/ma-xu/pointMLP-pytorch/issues/10 and https://github.com/ma-xu/pointMLP-pytorch/issues/40

hawkinglai commented 2 years ago

thanks!