StephanST

/

WALDO30

like 222

also @StephanST would you like to build a demo for this? we can provide zero A100 grant

Hey Merve, here I am distributing the trained models and not any of the training code used to train these models. The models themselves are an output of the code and don't fall under any license other than what I put on them (MIT here).
If people wanted to get access to the training code from Ultralytics, which is under AGPL, it's public here https://github.com/ultralytics/ultralytics/ 😄 But once again I'm not claiming ownership of that code base and nothing in this release includes code from there.

As for setting up a demo: for sure that would be cool! I'd be happy to build a little app for that so people ca play with the models.

@StephanST AFAIK if you're using an AGPL3.0 backbone (weights) your model has to be AGPL3.0 because it's a derivative work, see here the owner of YOLO is arguing that the weights are derivative work. do not get me wrong, I'm a strong advocate for open-source, it's just we wouldn't like people using this model in production to get in any trouble, hence this discussion to fix the license.

let me know when you build the demo under this thread, I'll keep an eye on it. follow the instructions here https://huggingface.co/zero-gpu-explorers (you just have to wrap a function and that's it) and I'll assign you an A100. thank you!

Oh right yeah, but I train 100% from random init and don't use any of the backbones from anyone else for this reason. Actually I have my own pre-training systems with some other datasets as well, but all 100% my work :)

I guess it's a gray area then (I saw in the model card you use YOLO backbone hence I opened this). Thanks for clarification!

As I've mentioned here, under AGPLv3, the model would be an object code of the AGPLv3 training source code, but the "Corresponding Source" would also include the dataset. Being the viral license AGPLv3 is, it would also make your own work AGPLv3, because that's the whole point of AGPLv3.

You can't use an AGPL library in your project and say "hey, that's my project, I just used this library". All that is specifically made clear in the license.

If you wish to keep your dataset private, you shouldn't be making a model from an AGPLv3 training code. YOLO does have non-Ultralytics implementations like DarkNet/YOLO, YOLO-NAS, YOLOX and the one here. Don't let the version numbers fool you, they depend on the publishers, and Ultralytics just makes more releases.

Even now, anyone who downloaded the model could make a lawsuit and ask for you to give your dataset and win the case. Businesses stay away from GPL and AGPL for that very reason.