How to use the mean_relative_error function as loss in keras

Question

I set up a neural network and I am trying to give it the inbuilt mean_relative_error as a loss function. I set it up as follows

def customLoss(yTrue,yPred):
    err, loss_value = mean_relative_error(yTrue, yPred, yTrue)
    return loss_value

def model(inp_size):
    inp = Input(shape=(inp_size,))
    x1 = Dense(100, activation='relu')((inp))
    for i in range (6):
        x1 = Dense(100, activation='relu')(x1)
    x1 = Dense(1, activation = 'linear')(x1)

    x2 = Dense(100, activation='relu')(inp)
    for i in range (6):
        x2 = Dense(100, activation='relu')(x2)
    x2 = Dense(1, activation = 'linear')(x2)

    x3 = Dense(100, activation='relu')(inp)
    for i in range (6):
        x3 = Dense(100, activation='relu')(x3)
    x3 = Dense(1, activation = 'linear')(x3)

    x4 = Dense(100, activation='relu')(inp)
    for i in range (6):
        x4 = Dense(100, activation='relu')(x4)
    x4 = Dense(1, activation = 'linear')(x4)



    x1 = Lambda(lambda x: x * baseline[0])(x1)
    x2 = Lambda(lambda x: x * baseline[1])(x2)
    x3 = Lambda(lambda x: x * baseline[2])(x3)
    x4 = Lambda(lambda x: x * baseline[3])(x4)

    out = Add()([x1, x2, x3, x4])

    return Model(inputs = inp, outputs = out)
y_train=y_train.astype('float32')
y_test=y_test.astype('float32')

NN_model = model(X_train.shape[1])
NN_model.compile(loss=customLoss, optimizer='Adamax', metrics=[customLoss])
NN_model.build(X_train.shape)

#NN_model.summary()
NN_model.fit(X_train, y_train, epochs=2,verbose = 1)
train_predictions = NN_model.predict(X_train)


predictions = NN_model.predict(X_test)

However, I get the following error

ValueError: An operation has None for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.

Does anyone have any idea? Thanks!

mujjiga mujjiga · Accepted Answer · 2019-05-10T08:19:41

Use

def customLoss(yTrue,yPred):
    return tf.reduce_mean(tf.abs(yPred-yTrue))

mean_absolute_error is intended for evaluation and it does not have a gradient and so it cannot be used in the back propagation.

Reference

How to use the mean_relative_error function as loss in keras

1 Answers