Keras: averaging different loss functions

Question

I am using Keras with TensorFlow backend, and want to define a custom loss function like this:

It computes the euclidean distance between the first 2 entries of y_true and y_pred
It computes the absolute error between the rest of the entries

Then it takes the average.

I do this:

def custom_objective(y_true, y_pred):
    y_true = backend.get_value(y_true)
    y_pred = backend.get_value(y_pred)

    a = np.sqrt(np.mean(np.square(y_pred[:2] - y_true[:2]), axis=-1))
    b = np.sum(np.abs(y_true[2:] - y_pred[2:]))
    return (a + b) / 5

I get an InvalidArgumentError when I compile the model:

model.compile(loss=custom_objective, optimizer='adam')

nuric nuric · Accepted Answer · 2018-05-20T18:15:54

You are using NumPy on Keras tensors, unfortunately that is a deadly combination. What you are looking for is something along the lines of:

def custom_objective(y_true, y_pred):
  a = K.sqrt(K.mean(K.square(y_pred[:2] - y_true[:2]), axis=-1))
  b = K.sum(K.abs(y_true[2:] - y_pred[2:]))
  return (a + b) / 5 # these operators work on tensors

Keras: averaging different loss functions

1 Answers