TensorFlow cost function sorting

Question

I'm pretty new to tensor flow and i'm trying to make my network sort an series of 8 numbers.

the way i would like to evaluate how well it has done is by subtracting x by y, invert the negative numbers and multiply them by a weight if thats necessary. And i was wondering if this is possible in tensorflow. Or if in any case there is better way to do this.

Something that is not tensoflow would look something like:

    w = 20
    s = 0
    print prediction, y
    for i in range(len(prediction.val)):
        s+= (y[i] / prediction[i] - 1) * w
    if s < 0:
        s = s * -1
    return s

In this case, if the total is a negative it will be inverted, however i would preferably do each element individually.

I'm currently stuck after:

prediction = neural_network_model(x)
sub = tf.subtract(prediction,y)

Answer 1

i found a solution that works and looks like :

prediction = neural_network_model(x)
sub = tf.subtract(prediction,y)
sign = tf.sign(sub)
cost = tf.multiply(sub,sign)

this does not yet include the weight but for now its working fine