2

In Keras I'm trying to figure out how to calculate a custom metric or loss that filters out or masks some values so that they don't contribute to the returned value. I'm stuck on how to get a tensor slice or how to iterate with an if: over the values in the tensor to select the values of interest.

I happen to be using the Tensorflow backend but would like to do something that is portable.

Attached is a rough outline of what I'm trying to do but it throws the error: TypeError: 'Tensor' object does not support item assignment

    def my_filtered_mse(y_true, y_pred):
        #Return Mean Squared Error for a subset of values
        error = y_pred - y_true
        error[y_true == 0.0]  =  0 #Don't include errors when y_true is zero
        # The previous like throws the error : TypeError: 'Tensor' object does not support item assignment
        return K.mean(K.square(error))

#...other stuff ...

    model.compile(optimizer=optimizers.adam(), 
        loss='mean_squared_error',
        metrics=[my_filtered_mse])
user3459022
  • 41
  • 2

1 Answers1

2

The failure happens in this line:

error[y_true == 0.0]  =  0 #Don't include errors when y_true is zero

Because error is a tensor, which doesn't support item assignment. You can change this to:

error = tf.gather(error, tf.where(tf.not_equal(y_true, 0.0)))
Max
  • 1,014
  • 6
  • 7