[Bug] Incompatible pointops version – furthestsampling() missing 'new_offset'

[DoctorHaibara]

Describe the Bug

I'm trying to run the point cloud training script from compressai based on examples/train_pointcloud.py. Specifically, I adapted it for the KITTI dataset (my script: train_pointcloud_kitti.py). When running the script, I encountered multiple issues related to the missing or incompatible pointops dependency.

---

❗ Key Errors

1. furthestsampling signature mismatch

TypeError: forward() missing 1 required positional argument: 'new_offset'

This error originates from:

indices = pointops.furthestsampling(xyz_tr, npoint).long()

Apparently, the pointops library currently available at repos like [Silverster98/pointops] expects an extra argument (new_offset). But the version used in this repo does not.

---

2. No version or repo for pointops is specified

There is no documentation, requirements.txt, or submodule pointing to the expected pointops version. This makes it impossible to guarantee reproducibility or compatibility.

🔍 Context

• OS: Ubuntu 20.04
• PyTorch: 2.7.0
• CUDA: 12.4
• pointops: tried with [Silverster98/pointops]

📌 Request

Please consider:

• Indicating the specific GitHub repository and commit hash for the pointops version used.
• Packaging the correct version of pointops as a submodule or providing installation instructions.
• Adding minimal documentation for running point cloud models under compressai.

Thank you!

[YodaEmbedding]

Try:

pip install pointops-yoda
pip install pyntcloud-yoda

CompressAI/setup.py

Lines 109 to 112 in d997a40

	POINTCLOUD_REQUIRES = [
	"pointops-yoda",
	"pyntcloud-yoda", # Patched version of pyntcloud.
	]

---

Notes:

• pyntcloud-yoda includes performance improvements and other fixes, detailed in perf: read_ply replace pandas.read_csv engine=python with c; improve read_off header-parsing robustness daavoo/pyntcloud#352.

[DoctorHaibara]

Thanks for the support! The issue is resolved.

The final training error was due to a BatchNorm1d layer receiving input with shape [1, C, 1], which caused a crash. I fixed it by replacing BatchNorm1d with GroupNorm, which works better for small batch sizes or limited spatial dimensions.

Closing this now. Appreciate your work on the project!