huber_loss for Tensorflow Keras

loss.py

def huber_loss(y_true, y_pred, clip_delta=1.0):
  error = y_true - y_pred
  cond  = tf.keras.backend.abs(error) < clip_delta

  squared_loss = 0.5 * tf.keras.backend.square(error)
  linear_loss  = clip_delta * (tf.keras.backend.abs(error) - 0.5 * clip_delta)

  loss = tf.where(cond, squared_loss, linear_loss)

  return tf.keras.backend.mean(loss)