metrics.py

import numpy as np


def leaky_accuracy(y_pred, y_true, span=3, threshold=0.05):
    assert span > 1
    #assert epsilon >= 0.0
    #assert y_pred.shape.__len__() == 1
    #assert y_true.shape.__len__() == 1
    y_true = y_true.flatten()
    y_pred = y_pred.flatten()

    #difference = np.abs(y_pred - y_true).__le__(epsilon).reshape(-1, 1)
    #  output = [joystick_erp[:joy_idx + 1].max(keepdims=True) for joy_idx in range(joystick_erp.shape[0])]
    #target = y_true.__ge__(epsilon).flatten()
    pred = y_pred > threshold
    true = y_true > threshold

    FP =  pred & ~true
    #TP =  pred &  true
    FN = ~pred &  true
    TN = ~pred & ~true

    rshift = [np.concatenate([np.zeros(i, dtype=bool), pred[:-i]]) for i in range(1, min(span, len(true)) + 1)]
    lshift = [np.concatenate([pred[i:], np.zeros(i, dtype=bool)]) for i in range(1, min(span, len(true)) + 1)]
    
    TP = (np.stack([*lshift, pred, *rshift]) & true.reshape((1, -1))).any(0)

    acc = (TP | TN).astype(float).mean()
    prec = TP.astype(float).mean() / (TP | FP).astype(float).mean()
    recl = TP.astype(float).mean() / (TP | FN).astype(float).mean()
    f1 = 2 * (prec * recl) / (prec + recl)

    return acc, prec, recl, f1


if __name__ == '__main__':
    output = [[0.0, 0, 0, 0, 0.0, 0.0, 0.0, 0.0, 0.4, 0.0]]
    target = [[ .0, 0, 0, 0, 0.4, 0.4, 0.4, 0.4, 0.4, 0.4]]
    print('accuracy', leaky_accuracy(np.array(output), np.array(target)))


    output = [2.840776687662583e-05, 2.7496724214870483e-05, 2.8048794774804264e-05, 2.794088504742831e-05, 2.693711758183781e-05, 2.976002542709466e-05, 2.9570142942247912e-05, 2.818175380525645e-05, 2.7603786293184385e-05, 2.698144453461282e-05, 2.7740432415157557e-05, 2.82325654552551e-05, 3.1146588298724964e-05, 3.0597115255659446e-05, 2.865607530111447e-05, 3.0131674066069536e-05, 3.2570085750194266e-05, 3.056521381950006e-05, 3.083764386246912e-05, 2.7935528123634867e-05, 3.466392809059471e-05, 3.901164745911956e-05, 3.83157967007719e-05, 4.658328543882817e-05, 4.7162808186840266e-05, 3.4100114135071635e-05, 3.576645758585073e-05, 3.296512659289874e-05, 3.796697637881152e-05, 3.324683348182589e-05, 3.030839070561342e-05, 2.7797528673545457e-05, 2.7219777621212415e-05, 2.948915607703384e-05, 2.815532025124412e-05, 2.7741938538383693e-05, 2.7851576305693015e-05, 2.8814603865612298e-05, 3.3615891879890114e-05, 3.706240386236459e-05, 3.4644595871213824e-05, 2.7978478101431392e-05, 2.6613299269229174e-05, 2.6914496629615314e-05, 2.6453077225596644e-05, 2.6251980671077035e-05, 2.683341153897345e-05, 2.44370039581554e-05]
    target = [0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

    output = (np.array(output) > 0.1).astype(np.float64)
    target = np.array(target)

    print('accuracy', leaky_accuracy(output, target))