search

Guillem96 / efficientdet-tf

Custom implementation of EfficientDet https://arxiv.org/abs/1911.09070
GNU General Public License v3.0
96 stars 23 forks source link

Question regarding huber loss #24

Closed UmarSpa closed 4 years ago

UmarSpa commented 4 years ago

Why the huber loss is mulitplied by 50 ?

   return 50. * tf.divide(self.loss_fn(y_true, y_pred), normalizer)
UmarSpa commented 4 years ago

And also the normalizer is first multiplied by 4. How come ?

   normalizer = tf.multiply(normalizer, tf.constant(4., dtype=tf.float32))
Guillem96 commented 4 years ago

These values should not be here, I was doing some testing and unfortunately, I upload them. Thanks for the comment