How to use label_smooth in Tensorflow object detection API when I train Faster rcnn?

I try to put

label_smoothing = 0.1
factor = tf.constant(label_smoothing)
target_tensor *= (1 - factor)
target_tensor += (factor / target_tensor.shape[2].value)

in losses.py class WeightedSoftmaxClassificationLoss(Loss):

But it has no effect. Even deleting the entire class will not affect the training.

Where should I put the loss?

Thanks