SmoothL1BoundingBoxLoss #112
Description
This is used in RetinaNet:
Basic prototype from keras.io example:
class SmoothL1BoundingBoxLoss(tf.losses.Loss):
"""Implements Smooth L1 loss"""
def __init__(self, delta):
super().__init__()
self._delta = delta
def call(self, y_true, y_pred):
difference = y_true - y_pred
absolute_difference = tf.abs(difference)
squared_difference = difference ** 2
loss = tf.where(
tf.less(absolute_difference, self._delta),
0.5 * squared_difference,
absolute_difference - 0.5,
)
return tf.reduce_sum(loss, axis=-1)