Confusion about the function _make_enc() and _make_dec()

[auniquesun]

@chrockey @junha-l @hmyang0727 @Kwonyoung-Ryu Thanks for sharing the code.

One point I am confused with is the PointTransformerSeg architecture re-implemented in this repo.

Specifically, in model/pointtransformer/pointtransformer_seg.py, the class PointTransformerSeg defines the function _make_enc(), according to the code, each one in self.enc1, self.enc2, self.enc3, self.enc4, self.enc5 contains multiple Point Transformer Blocks.

However, according to the model architecture given by the paper, there is nothing explicitly indicating that each encoder should consist of multiple Point Transformer Blocks. Multiple point transformer blocks in each encoder make the model unnecessarily complicated, which also applies to the function _make_dec() in same class.

Instead, I mean each encoder should contain one TransitionDown Layer and one Point Transformer Block, and the re-implementations by others also follow this, e.g., Point-Transformers.

[chrockey]

Hi @auniquesun,

Instead, I mean each encoder should contain one TransitionDown Layer and one Point Transformer Block, and the re-implementations by others also follow this, e.g., Point-Transformers.

We also have tried a lightweight model like model = PointTransformerSeg(PointTransformerBlock, [2, 2, 2, 2, 2], **kwargs), but failed to reproduce mIoU 70.4.
There could be many reasons like data augmentations (e.g., Elastic distortion), and we didn't try all of the possibilities.
If you can find such a setting, please feel free to open a PR :).

Hope this helps your understanding.