# -*- coding: utf-8 -*-
__author__ = "Ngoc Huynh Bao"
__email__ = "ngoc.huynh.bao@nmbu.no"
import numpy as np
import tensorflow as tf
from tensorflow.keras.losses import Loss, deserialize
from ..utils import Singleton
[docs]class BinaryFbetaLoss(Loss):
def __init__(self, reduction='auto', name="binary_fbeta", beta=1):
super().__init__(reduction, name)
self.beta = beta
[docs] def call(self, target, prediction):
size = len(prediction.get_shape().as_list())
reduce_ax = list(range(1, size))
eps = 1e-8
true_positive = tf.reduce_sum(prediction * target, axis=reduce_ax)
target_positive = tf.reduce_sum(tf.square(target), axis=reduce_ax)
predicted_positive = tf.reduce_sum(
tf.square(prediction), axis=reduce_ax)
fb_numerator = (1 + self.beta ** 2) * true_positive + eps
fb_denominator = (
(self.beta ** 2) * target_positive + predicted_positive + eps
)
return 1 - fb_numerator / fb_denominator
[docs]class ModifiedDiceLoss(Loss):
def __init__(self, reduction='auto', name="modified_dice_loss", beta=1):
super().__init__(reduction, name)
self.beta = beta
[docs] def call(self, target, prediction):
size = len(prediction.get_shape().as_list())
reduce_ax = list(range(1, size))
eps = 1e-8
true_positive = tf.reduce_sum(prediction * target, axis=reduce_ax)
target_positive = tf.reduce_sum(target, axis=reduce_ax)
predicted_positive = tf.reduce_sum(
prediction, axis=reduce_ax)
fb_numerator = (1 + self.beta ** 2) * true_positive + eps
fb_denominator = (
(self.beta ** 2) * target_positive + predicted_positive + eps
)
return 1 - fb_numerator / fb_denominator
[docs]class Losses(metaclass=Singleton):
"""
A singleton that contains all the registered customized losses
"""
def __init__(self):
self._losses = {
'BinaryFbetaLoss': BinaryFbetaLoss,
'ModifiedDiceLoss': ModifiedDiceLoss
}
[docs] def register(self, key, loss):
if not issubclass(loss, Loss):
raise ValueError(
"The customized loss has to be a subclass"
+ " of keras.losses.Loss"
)
if key in self._losses:
raise KeyError(
"Duplicated key, please use another key for this loss"
)
else:
self._losses[key] = loss
[docs] def unregister(self, key):
if key in self._losses:
del self._losses[key]
@property
def losses(self):
return self._losses
[docs]def register_loss(key, loss):
"""
Register the customized loss.
If the key name is already registered, it will raise a KeyError exception
Parameters
----------
key : str
The unique key-name of the loss
loss : tensorflow.keras.losses.Loss
The customized loss class
"""
Losses().register(key, loss)
[docs]def unregister_loss(key):
"""
Remove the registered loss with the key-name
Parameters
----------
key : str
The key-name of the loss to be removed
"""
Losses().unregister(key)
[docs]def loss_from_config(config):
if type(config) == dict:
if 'class_name' not in config:
raise ValueError('class_name is needed to define loss')
if 'config' not in config:
# auto add empty config for loss with only class_name
config['config'] = {}
return deserialize(
config,
custom_objects=Losses().losses)
return deserialize(config, custom_objects=Losses().losses)