What is our story around streamed data? about rune

[meelislootus]

@Michael-F-Bryan yes - that sounds about right! The case of an image model and an audio/accelerometer model are slightly different:

For image, we can just take one sample (one snap of the camera), at say 2 Hz (twice per second) to begin with, for the simplest case (though later we can use multiple samples as @kthakore mentioned).

CAPABILITY<f32[96,96]> image IMAGE --sampling_frequency_hz=2

For audio/accelerometer: as you explain above. Depending on the model/application, we can decide / a data scientist can decide what length window, in seconds, is appropriate to capture the signal. E.g. in gesture, a suitable window size might be 2 seconds. Then we need to consider what size the model requires as input: in the case of our gesture model, this is 128 samples. Therefore we ask the accelerometer for 128 samples over 2 seconds. And let's say we provide take the window of 128 samples twice per second (at 2Hz).

CAPABILITY<f32[128,3]> accelerometer ACCEL --window_size_in_milliseconds 2000 --sampling_frequency_hz=2