MeanAbsolutePercentageError

Calcula o erro percentual médio absoluto entre rótulos e previsões.

loss = 100 * abs(labels - predictions) / labels

Uso autônomo:

    Operand<TFloat32> labels =
        tf.constant(new float[][] { {2.f, 1.f}, {2.f, 3.f} });
    Operand<TFloat32> predictions =
        tf.constant(new float[][] { {1.f, 1.f}, {1.f, 0.f} });
    MeanAbsolutePercentageError mape = new MeanAbsolutePercentageError(tf);
    Operand<TFloat32> result = mape.call(labels, predictions);
    // produces 50f
 

Ligando com peso da amostra:

    Operand<TFloat32> sampleWeight = tf.constant(new float[] {0.7f, 0.3f});
    Operand<TFloat32> result = mape.call(labels, predictions, sampleWeight);
    // produces 20f
 

Usando o tipo de redução SUM :

    MeanAbsolutePercentageError mape = new MeanAbsolutePercentageError(tf, Reduction.SUM);
    Operand<TFloat32> result = mape.call(labels, predictions);
    // produces 100.0f
 

Usando o tipo de redução NONE :

    MeanAbsolutePercentageError mape = new MeanAbsolutePercentageError(tf, Reduction.NONE);
    Operand<TFloat32> result = mape.call(labels, predictions);
    // produces [25f, 75f]