experiments/ptv3/README.md

Point Transformer V3 encoder (negative result)

This is the Point Transformer V3 encoder experiment from the write-up. We replaced the Perceiver encoder with a Point Transformer V3 encoder (per-point features, no latent bottleneck) to test whether a stronger encoder would scale past the 8k Perceiver.

Outcome. Trained from scratch at 8k for 200k steps, it plateaued at 0.323 local HSS, below the curriculum-trained Perceiver (0.357), and ran at roughly 6 s/sample on an A5000 — over the two-hour T4 evaluation budget. It was never submitted. This folder is provided as confirmation that the experiment was run.

Contents

checkpoint_ptv3_8k.pt     trained weights (200k steps at 8k)
train_args.json           training configuration
ptv3_code/                the PT v3 encoder (adapted from the Pointcept release)
                          and the [B,T,*] <-> flat adapter (encoder_wrapper.py)
model_with_ptv3.py        EdgeDepthSegmentsModel with the arch="ptv3" branch
                          (drop-in replacement for s23dr_2026_example/model.py)

Notes

• Extra dependencies beyond the main requirements.txt: spconv-cu121, torch-scatter, addict (and optionally flash-attn).
• To run it, place ptv3_code/ as s23dr_2026_example/ptv3/, use model_with_ptv3.py as s23dr_2026_example/model.py, and load the checkpoint with arch="ptv3". Inference must use fp32 (spconv does not tune fp16 kernels reliably on these GPUs).
• The PT v3 encoder is adapted from the authors' Pointcept release.