Keras自定义指标

Question

如何创建一个自定义指标，该指标在分类问题上接受正确的预测，不仅在预测类别和真实类别相同时，而且在预测类别是真实类别的邻居类别时？“ p>

Answer 1

如果你考虑到这个问题，这个答案是有效的，只有一个班级＆＃34;应该输出。

def neighbourMetric(yTrue,yPred):

    #these make this function not differntiable, but since you asked for "metric" it's ok
    trueIndices = K.argmax(yTrue)
    predIndices = K.argmax(yPred)

    minAccepted = trueIndices - 1
    maxAccepted = trueIndices + 1


    satisfiesMin = K.cast(K.greater_equal(predIndices,minAccepted),K.floatx())
    satisfiesMax = K.cast(K.less_equal(predIndices,maxAccepted),K.floatx())

    satisfiesBoth = satisfiesMin * satisfiesMax

    return K.mean(satisfiesBoth)

Answer 2

以下是我解决它的方法：

def one_off(y_true, y_pred):
    return K.cast(K.abs(K.argmax(y_true, axis=-1) -  K.argmax(y_pred, axis=-1)) < 2, K.floatx())