Model on hardware doesn't seem to run like on the website

I’m getting good results with the test data on my model on the website.

When I run the model on the hardware (arduino ble nano) it sometimes works, but not with the same reliability as the website.

So I get 0.99 a lot for the different categories, but it doesn’t seem to tip over into a prediction (stays uncertain).

Any ideas why this is? I still have very small amount of data - about 30s.
Can I reduce the threshold somehow?


It looks like the prediction has to stay on one label for some amount of time before the output prediction actually changes.
So if I spin for a few seconds then it will trigger. I want it to trigger after seeing the spin for just a few hundred ms.

You can control these with the smooth` structure. E.g.:

    // This is a structure that smoothens the output result
    // With the default settings 70% of readings should be the same before classifying.
    ei_classifier_smooth_t smooth;
    ei_classifier_smooth_init(&smooth, 10 /* no. of readings */, 7 /* min. readings the same */, 0.8 /* min. confidence */, 0.3 /* max anomaly */);

Just set no. of readings and readings the same to 1.

making some progress:

@mattvenn \o/

You could lower the window size if you need even faster response times by the way.

1 Like