I have a custom esp32+BNO085+Arduino board, capable of running tflite models. I wanted to do a simple “pinch” detection using the accelerometer+gyroscope data.
I was able to make a model using the Spectral Features block, but when converting to a tflite model I wasn’t able to run it on my board because it just inputs the raw imu data (accelerometer xyz + gyroscope xyz = 6 values total).
If I made a model using the Raw Data block, I can run the model since the number of inputs are expected (numberOfSamples * valuesPerSample), e.g. a 500ms window with a 20ms interval would be ((500/20) * 6) = 150 inputs, but when looking at the Spectral Analysis exported Arduino model it says it assumes 78 inputs.
This is an example of it working in the browser using the javascript+webAssembly deployment of the spectral analysis model (https://www.youtube.com/shorts/jo-db-tsj6g), where I stream the sensor data from my pcb to the browser and run the model in the browser, but I’d like to be able to run that model on the pcb itself (and using the Raw Data block just isn’t as good as the Spectral Analysis block)
I don’t use the exported Arduino code, I just get the .tflite file and upload it to my pcb via bluetooth (I’m able to upload any tflite model to the pcb and specify which sensors it uses, and at what sample rate). Is there any reference to see exactly what values I need to put into the model to classify?