Understanding ei_classifier_smooth_update()

ShawnHymel · April 28, 2021, 12:00am

I’m looking through the ei_classifier_smooth_update() definition (in ei_classifier_smooth.h), as a student is asking about this function. I’d like to make sure I understand what it’s doing.

From what I can tell, the function looks at the previous 10 inferences. If a label has the highest score over classifier_confidence (e.g. 0.8) min_readings_same (e.g. 7) out of n_readings (e.g. 10), it will be returned as the chosen label (otherwise, you’ll get “unknown”).

Am I on the right track?

louis · April 28, 2021, 3:39pm

Hi @ShawnHymel,

I haven’t tried it yet myself but I do understand the same from the code than what you’ve just explained.
As a return you will get either the label, uncertain or anomaly:

github.com

edgeimpulse/inferencing-sdk-cpp/blob/master/classifier/ei_classifier_smooth.h#L69


    smooth->min_readings_same = min_readings_same;
    smooth->classifier_confidence = classifier_confidence;
    smooth->anomaly_confidence = anomaly_confidence;
    smooth->count_size = EI_CLASSIFIER_LABEL_COUNT + 2;
}
/**
 * Call when a new reading comes in.
 * @param smooth Pointer to an initialized ei_classifier_smooth_t struct
 * @param result Pointer to a result structure (after calling ei_run_classifier)
 * @returns Label, either 'uncertain', 'anomaly', or a label from the result struct
 */
const char* ei_classifier_smooth_update(ei_classifier_smooth_t *smooth, ei_impulse_result_t *result) {
    // clear out the count array
    memset(smooth->count, 0, EI_CLASSIFIER_LABEL_COUNT + 2);
    // roll through the last_readings buffer
    numpy::roll(smooth->last_readings, smooth->last_readings_size, -1);
    int reading = -1; // uncertain

Best regards,

Louis

janjongboom · April 28, 2021, 4:24pm

Correct. We also have a moving-average filter in the code which can be used to smooth out results.

ShawnHymel · April 29, 2021, 2:44am

Thanks all!

Is the moving average filter a separate part of the same ei_classifier_smooth_update() function? I’m not seeing it if so. Or are you referring to the part where it requires e.g. 7 out of the previous 10 readings to be above a threshold?

janjongboom · April 29, 2021, 9:32am

No, it’s actually here: https://github.com/edgeimpulse/inferencing-sdk-cpp/blob/6b402ac113af8e602246fbdcd49bcdaf06b826ce/classifier/ei_run_classifier.h#L133

We use it for smoothing out audio in continuous classification.

ShawnHymel · April 29, 2021, 2:18pm

Got it, thank you! I remember seeing the MAF in the audio example…I just didn’t see it in the motion example