pornspective
patreon
Great news, progress on the AI modelling ...
🕑 Added 2025-10-23 07:58:20 +0000 UTCWow I now remember why I've left the original model for so long, that was a huge task back then and this has been no different. It's took hours of patience and meant I've only had time to put my head in this task for a few weeks now, but I won't be beaten and that has eventually led to success.
Since my last update I've: -
1. Trained the model again, but failed to convert to ONNX
2. Use CoPilot for hours trying to tweak third party library code on each failure on the ONNX convert scripts. Each time getting further away from the actual released library version. Obviously doesn't help my ambition to make this a neat pipeline that can be repeated and new versions downloaded without breaking things.
3. Threw everything in the bin, and started again.
4. Found a lifeline where someone suggested latest Python was problematic. Removed Python and installed a slightly older version.
5. Still failed to convert to ONNX, same errors.
6. Decided for one last time to start everything again on the older Python.
7. Built a new model which took 12 hours.
8. Converted to ONNX successfully, although wasn't able to get any results from it. Turns out me trying to fix the number of labels issue was causing the problem.
9. Ignored the number of labels issues and last night found it worked perfectly in Python!
10. Couple of hours then porting to .NET and trying with a basic console application and finally made the breakthrough! All working in .NET too.
So I'm now 100% confident I have what I need to bring this into the AI application, I proven it works and simply need to now create a plugin for it. My plan is to release a beta version of this soon so people can try out the model and in the meantime I'll look to make something more configurable so new models can be developed and users choose which they prefer.
So in summary I've cracked it guys, RF-DETR models will now be a repeatable modelling pipeline that will hopefully improve the AI further and allow others to contribute if they wish with their own trained models.
I'd envisage a release of the AI Beta in a week or two. Drop me a comment or message if you'd like to be testers for me :). Ideally those with both CPU and GPU setups would be great!
Comments
jej79bob
I'm willing to participate in the beta test, but only for the CPU version; my old GPU (Nvidia GeForce GTX 1060) isn't powerful enough.
David Jackson
Im totally down for the beta test
