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

About implementation of MLP #84

Closed root116688 closed 1 year ago

root116688 commented 1 year ago

Thanks for sharing code I have some questions.

  1. why use Conv1d to implement MLP?
  2. Why Didn't use Linear to implement MLP?
  3. https://github.com/ma-xu/pointMLP-pytorch/blob/3e3d80cff5c23a631fe5ba4ca97db3f452893ed2/classification_ModelNet40/models/pointmlp.py#L209 When will it be greater than 1, and what does this parameter mean?

Thanks!

ma-xu commented 1 year ago

Hi Thanks for your interest.

For Q1 and Q2, they are mathmatically equal. For Q3, we didn't use groups and the parameter should always be 1. We implement it for additional experiments.

root116688 commented 1 year ago

Thanks a lot, Linear and Conv1d(kernel size =1) Seems like the features value is equal, but seems the weight and bias initial or backward is different? Or totally same?

ma-xu commented 1 year ago

They may be different but make almost no difference to final results.

root116688 commented 1 year ago

Thaks for your reply!