4

I'm building a model for Time series classification. The data is very unbalanced so I've decided to use a weighted cross entropy function as my loss.

Tensorflow provides tf.nn.weighted_cross_entropy_with_logits but I'm not sure how to use it in TF 2.0. Because my model is build using tf.keras API I was thinking about creating my custom loss function like this:

pos_weight=10
def weighted_cross_entropy_with_logits(y_true,y_pred):
  return tf.nn.weighted_cross_entropy_with_logits(y_true,y_pred,pos_weight)

# .....
model.compile(loss=weighted_cross_entropy_with_logits,optimizer="adam",metrics=["acc"])

My question is: is there a way to use tf.nn.weighted_cross_entropy_with_logits with tf.keras API directly?

username0101
  • 43
  • 1
  • 5

1 Answers1

5

You can pass the class weights directly to the model.fit function.

class_weight: Optional dictionary mapping class indices (integers) to a weight (float) value, used for weighting the loss function (during training only). This can be useful to tell the model to "pay more attention" to samples from an under-represented class.

Such as:

{
    0: 0.31, 
    1: 0.33, 
    2: 0.36, 
    3: 0.42, 
    4: 0.48
}

Source

Edit: JL Meunier answer explains well how to multiply the logits with class weights.

Hichame Yessou
  • 2,658
  • 2
  • 18
  • 30
  • 4
    Thank you for the response. But could I get the same result by using my custom function instead of passing the weights to the `model.fit` function? – username0101 Aug 12 '19 at 11:18