Uploading a custom model to Limelight

[marty]

I have created a custom model using the Limelight documentation instructions on neural networks and they say that in 2023.1 custom models will be able to be uploaded. However I am on 2023.1 OS and I am unable to upload my custom model and I don’t see anywhere on the Limelight website to upload the model. Has anybody else been able to get this working?

[Phlogiston]

I’m not very familiar with the limelight’s new classifier features, but I believe it requires a google coral accelerator plugged into the limelight USB port to be enabled.

[UnofficialForth]

This. Plus you need a limelight 3, as 2+ and older version did not have this feature.

[KaseyB]

This works on the 2+ with the current dataset that was preloaded in the 2023 image. I believe the website docs are not up to date, as the custom model feature has been pushed to at least the next image.

This is written at the bottom of the 2023.1 announcement thread.
Nerual Net upload is being pushed to 2023.2!

[UnofficialForth]

Interesting, my understand was that you needed the google coral. Thanks for the correction!

[KaseyB]

You do need the google coral as well, but it’s not limited to the 3 only.

[Joe_Ross]

Introducing Limelight 3

@Jack_1221 @fedorlspp26 @jdao @BoxOfScraps @Kalimbus @Joel_Singh AprilTags and Coral-accelerated models work on LL1, LL2, and LL2+. Object detection models will run at 20-30fps on older hardware as opposed to 45-60fps on LL3. Classification models run at 90fps on LL3, and I believe 60fps on older hardware. H.264 streaming is the only feature exclusive to LL3.

[veg]

For teams without a Coral, it is possible to evaluate ML models on an unaccelerated Raspberry Pi if it is exported in the right format. I’ve been testing Teachable Machines models that run at the same ~7.5 fps on an RPi 4 with or without the Coral attached. Axon object tracking models also work. Not blazing performance, but maybe adequate for a lot of uses.

I am documenting this process, but it is one of about six “should have been an offseason” projects on my desk right now.

[Brandon_Hjelstrom]

@marty

Limelight 2023.2 - Easier 3D, Neural Net Upload

Making 3D easier than ever. WPILib-compatible Botposes Botpose is now even easier to use out-of-the-box. New NetworkTables Key “botpose_wpired” - botpose, but with the origin at the right-hand side of the driverstation on the red side of the field. New NetworkTables Key “botpose_wpiblue” - botpose, but with the origin at the right-hand side of the driverstation on the blue side of the field. New Json arrays - botpose_wpired, and botpose_wpiblue All of the above botposes are compatible with …

[marty]

Hi, I’m trying to upload the neural network and it says the model and labels has been uploaded but it still is using the built-in model and labels.

[Brandon_Hjelstrom]

Did you unzip the model before uploading the two files separately?

Can you try switching pipelines and switching back in the web UI?

You can send the files to me for testing w/ google drive if something seems wrong!

[marty]

Yes I tried both of these and the upload still is not working.

[marty]

This is the link to my labels
This is the link to my model
You can test these out, maybe I’m just doing something wrong

[Brandon_Hjelstrom]

It works and seems to perform pretty well in my limited testing! Can you try uploading the labels file first, and then the model file?
Edit - Maybe you are using the detection pipeline instead of the classifier pipeline? Make sure you’re set to “classifier.”
Thanks

[joseppi]

Is detector not supported yet? We are getting the same results, only classifier works with the model and labels.

[ashwinshrivastav8728]

You have to update the Limelight Software to 2023.3 or higher to be able to upload a custom neural network. As for uploading, you have to open the Limelight Dashboard once the Limelight is connected to the RoboRIO and you have to specify the type to a Classifier or Detector which then gives you the option to upload the model and a label map.