Keras Python custom loss function to give the maximum value of (absolute difference of) Tensor?

Question

I have as my output a 1-D array of 5 elements, it looks like this in the model:

out_vector = Dense(out_count, activation='relu', name='out_vector')(network_layer_3)

where out_count is 5 (not that it really matters). When I compare it to true_out_vector , another 5D array, I want the loss to be the "maximum of the absolute differences of the elements".

Simple example what I mean:

v1 = [94, 1000, 50, 85, 23]

v2 = [100, 430, 88, 12, 90]

I want my loss to equal the biggest absolute difference, which is |1000 - 430| = 570 |1000 - 430| = 570 clearly, since element 2 has this biggest difference. I'm having trouble making this happen in Keras. Here's what I've tried:

def customLoss(yTrue,yPred):
        return K.maximum(K.abs(yTrue - yPred))  

But I get the error:

File "C:\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py", line 609, in ndim
    dims = x.get_shape()._dims
AttributeError: 'int' object has no attribute 'get_shape'

I'm sure there should be a simple way to do what I'm going after.

Answer 1

How about this:

from keras import backend as K
v1 = K.constant([94, 1000, 50, 85, 23])
v2 = K.constant([100, 430, 88, 12, 90])


def customLoss(yTrue, yPred):
    return K.max(K.abs(yTrue - yPred))

result = customLoss(v1, v2)

sess = K.get_session()
print(sess.run(result))

and the output is:

570.0

NB: The method K.maximum takes two tensors while you are passing only one