Problem loading Yolo Version 11 on Android Studio using Interpreter api and GPU delegate

[AndrewQianNorthernVue]

I have converted a YoloV11 model using Ultralytics' export function into a .tflite format. I then attempted to load the interpreter using Interpreter(model, options). The options had a GPU delegate set:

        val delegate = GpuDelegate()
     
        val options = Interpreter.Options().addDelegate(delegate)
      
        interpreter = Interpreter(model, options)

This DID NOT work for YoloV11 but DID work for Yolov8, which I converted the exact same way. I also attempted to convert to onnx and then tflite, but the Interpreter(model, options) would also stall (just never finish) for yolov11. I believe this might be due to unsupported operations, and have looked at Netron, but it is quite tedious.

I would like to know if this is a known issue and if there is a fix. Here is some additional context:

1. I attempted to just use the CPU for my YoloV11 using options.numThreads(4) instead of the delegate code above. It worked just fine.

2. I attempted to use the GPU for YoloV8 which was converted the exact same way, it worked without issue.

3. compatList.isDelegateSupportedOnThisDevice somehow returns false no matter which model I am loading, but this doesn't seem to stop my V8 model from loading and running.

4. I am running a demensity chip on my device.

[gamirou]

I am currently encountering the same problem on Yolov10

[AndrewQianNorthernVue]

I am currently encountering the same problem on Yolov10

The solution I found was

1. Directly convert to TFLite instead of going through --> Onnx --> TFLite
2. More important: Keep image sizes around 1280. I believe this is close to the cutoff where the number of channels becomes too large for the graph to support, causing the GPU to stall.

From what I can understand, I do not think compatlist as mentioned in my above post is that reliable of an indicator on whether or not GPU can be used.