Nonsensical Classification Result after inference using different version of pgeof

[jako5]

✅ Code of conduct checklist

• 🌱 I am using the latest version of the code.
• 👩‍💻 I made sure the bug concerns the official project's codebase and not something I coded myself on top of it. (We only provide support for code we wrote and released ourselves).
• 🔎 I took appropriate time to investigate the problem before filing an issue, but am unable to solve it myself.
• 📙 I thoroughly went through the README, but could not find the solution there.
• 📘 I thoroughly went through the tutorial slides, notebook, and video, but could not find the solution there.
• 📗 I thoroughly went through the documentation, but could not find the solution there.
• 📜 I went through the docstrings and comments in the source code parts relevant to my problem, but could not find the solution there.
• 👩‍🔧 I searched for similar issues, but could not find the solution there.
• 🔎 I made sure my bug is related to the project's codebase and is not in fact a sub-dependency issue (e.g.: FRNN installation issue, PyTorch installation issue, PyTorch Geometric not installed with GPU support, ...).
• ⭐ Since I am showing interest in the project, I took the time to give the repo a ⭐ to show support. Please do, it means a lot to us !

🐛 Describe the bug

Hi Damien,

first of all thank you and congratulations for your amazing work.

It has come to my attention after a longer bug search in the last days that the latest update on point_geometric_features has rendered my latest trained model unusable, because the calculated features differ. I do not (yet) consider this a problem of point_geometric_features, as I have not retrained using the latest version of pgeof yet. Its just not backwards compatible. Therefore, if a checkpoint is to be reused after reinstallation, it seems wise to me to switch to a fixed version install, in my case

pip install git+https://github.com/drprojects/point_geometric_features.git@4102aa9

instead of using the latest repo version.

Best regards

Jakob

📜 Log output

🤖 Steps to reproduce the bug

Train using commit 4102aa9 of https://github.com/drprojects/point_geometric_features, then infer/test using the latest version of the same repo

📚 Additional information

No response

🖥️ Environment

python3 collect_env.py doesn't work:

AttributeError: module 'torch' has no attribute 'xpu'

[drprojects]

Hi @jako5 thanks for raising this issue and for the pre-investigation.

From what I understand, there must have been a change in the order or in the normalization of the features produced by pgeof between commit 4102aa9 and the current 9c2e7b9. @rjanvier you might be more up to date than me in that regard, does that sound familiar to you ?

[rjanvier]

yes, I just noticed a bug I introduced in 0.3.3, sorry: The E values and E vector are sorted with the use of selfadjoint matrix, but not in the order we expect, so it leads to wrong results... and In the micro benchmark I use to validate it does not show up clearly. please use 0.3.2 and I will propose a fix in the coming hours/days in the code. Sorry again and thanks.

[drprojects]

Super ! Thanks a lot @rjanvier for the fast reply 🙏 🙏 🙏

[drprojects]

This issue was addressed by @rjanvier in the pgeof PR: drprojects/point_geometric_features#20

Thanks @rjanvier 🙏