✅ Deploy Preview for frigate-docs canceled.

If there are commonalities then you could put that in the installation section of the docs then have the hardware and object detector link to that as a prerequisite for specific setup

as far as post processing goes, it could make sense to have a util.py in detectors that handles this

as far as post processing goes, it could make sense to have a util.py in detectors that handles this

I agree. I will try to implement post-processing for all detectors. I will use the mode: model_type: config option that is already implemented to handle this. But I will also include a way to return the pure detections (similar to how its already done) to allow the detectors to handle post-processing differently (e.g. using opencl on GPU).

This is now ready to merge

I think it's best to add some information about the license on your repo where you published the models. Also, probably best to note that here as well. We wouldn't want someone using Frigate for commercial purposes to assume that these models are free for commercial use.

Lastly, how painful would it be to update this so the models are downloaded at runtime? That prevents frigate from actually redistributing the models itself.

Good idea. I will add notes in my repo and frigate and I will exclude the models from the image. Maybe I will add some other models (maybe yolox) to the image later.

I added a note to my repo, the docs and to the logger on each start, so that the user knows about the license situation.

I also excluded the models from the image.

I see that Frigate 0.14 is coming out soon. Are there any plans for a beta phase? When should I finish my bigger changes so that they make it in that version?

I don't know that I would call it "soon". There should be a beta within a week or two, and I am sure there will be at least a few betas to follow that.

This PR is ready to merge from my side.

@blakeblackshear would you be open to shipping a yolo-nas model in the image if it was trained using no pre-trained weights? According to many of their github responses you aren't bound by the hostile license if you train your own model using their architecture only.

I ask because I have a super-gradients python environment set up already with an rtx 4090 and I wouldn't mind throwing a few days of compute at training a model for Frigate because it seems like a cool project :)
I'm already training a yolo-nas pose model up currently with no pre-trained weights off the coco dataset for a VR full body tracker because of the licensing.

Apache 2.0 could apply if a court determines the architecture as an input to the process instead of part of the tool: Deci-AI/super-gradients#1993 (comment)

..or, it could be any license if the architecture is deemed as part of the tool to create the model and not an input to the process. It's a large legal gray area but the worst outcome here is the license is Apache 2.0 which is pretty unrestrictive.

I would be open to it. I have a 4090 as well and plan to look at supporting yolo-nas for frigate+ too.

Sweet, I'll see if I can train up a model in a few days when mine is done :p

What size do you think would be the best to use in Frigate?

Also, FYI, if you're wanting to train up models with your computer as well, when you install super-gradients, you need to use master, stable PyPi and github both have broken dependencies.
pip install git+https://github.com/Deci-AI/super-gradients.git@master or clone the git repo and pip install -e . inside dir

I haven't looked at what's possible yet, but I was hoping to try a quantisized 320x320 version of the small model. Since frigate crops the image ahead of time, smaller objects aren't as important. Hard to say without comparing real world performance of a few options.

Hey @tenten8401,
this would be really great!
I don't know if you are specifically interested in yolo-nas on rockchip devices or if you just posted your idea here. I use the pre-trained image from super-gradients but I'm worried about the license. I would be interested in using your model for the rknn detector if you can provide one.

You might consider hosting a repo and upload your models + weights there. I think other people might be interested in using a pre-trained yolo-nas model without these license terms as well.

I have a 320x320 small model training now based on coco2017, but I did spot one thing in the github repo that has me a bit worried:
https://github.com/Deci-AI/super-gradients/blob/master/LICENSE.YOLONAS.md

or any components comprising the model and the associated documentation

@tenten8401 awesome!

I am not a lawyer, but this license applies only to yolo-nas with pre-trained weights. Based on the answers to similiar questions in the super gradients repo (see here: https://github.com/Deci-AI/super-gradients/issues?q=is%3Aissue++License++Yolo-NAS), I think it is save to use your model.

Here a collaborator clearly confirms this: Deci-AI/super-gradients#894 (comment)

Here is also the answer to my question from a deci employee that also says this: Deci-AI/super-gradients#1993 (comment)