以下信息均来自官网
------------------------------------------------------------------------------------------------------------
损失函数的使用
损失函数（或称目标函数、优化评分函数）是编译模型时所需的两个参数之一：
model.compile(loss='mean_squared_error', optimizer='sgd') from keras import losses model.compile(loss=losses.mean_squared_error, optimizer='sgd')
你可以传递一个现有的损失函数名，或者一个 TensorFlow/Theano 符号函数。 该符号函数为每个数据点返回一个标量，有以下两个参数:
- y_true: 真实标签。TensorFlow/Theano 张量。
- y_pred: 预测值。TensorFlow/Theano 张量，其 shape 与 y_true 相同。
实际的优化目标是所有数据点的输出数组的平均值。
可用损失函数
mean_squared_error
mean_squared_error(y_true, y_pred)
mean_absolute_error
mean_absolute_error(y_true, y_pred)
mean_absolute_percentage_error
mean_absolute_percentage_error(y_true, y_pred)
mean_squared_logarithmic_error
mean_squared_logarithmic_error(y_true, y_pred)
squared_hinge
squared_hinge(y_true, y_pred)
hinge
hinge(y_true, y_pred)
categorical_hinge
categorical_hinge(y_true, y_pred)
logcosh
logcosh(y_true, y_pred)
预测误差的双曲余弦的对数。
对于小的 x，log(cosh(x)) 近似等于 (x ** 2) / 2。对于大的 x，近似于 abs(x) - log(2)。这表示 'logcosh' 与均方误差大致相同，但是不会受到偶尔疯狂的错误预测的强烈影响。
参数
- y_true: 目标真实值的张量。
- y_pred: 目标预测值的张量。
返回
每个样本都有一个标量损失的张量。
categorical_crossentropy
categorical_crossentropy(y_true, y_pred)
sparse_categorical_crossentropy
sparse_categorical_crossentropy(y_true, y_pred)
binary_crossentropy
binary_crossentropy(y_true, y_pred)
kullback_leibler_divergence
kullback_leibler_divergence(y_true, y_pred)
poisson
poisson(y_true, y_pred)
cosine_proximity
cosine_proximity(y_true, y_pred)
注意: 当使用 categorical_crossentropy 损失时，你的目标值应该是分类格式 (即，如果你有 10 个类，每个样本的目标值应该是一个 10 维的向量，这个向量除了表示类别的那个索引为 1，其他均为 0)。 为了将 整数目标值 转换为 分类目标值，你可以使用 Keras 实用函数 to_categorical：
from keras.utils.np_utils import to_categorical
categorical_labels = to_categorical(int_labels, num_classes=None)
如果还不明白，请看下面的源码
1 """Built-in loss functions.
2 """
3 from __future__ import absolute_import
4 from __future__ import division
5 from __future__ import print_function
6
7 import six
8 from . import backend as K
9 from .utils.generic_utils import deserialize_keras_object
10 from .utils.generic_utils import serialize_keras_object
11
12
13 def mean_squared_error(y_true, y_pred):
14 return K.mean(K.square(y_pred - y_true), axis=-1)
15
16
17 def mean_absolute_error(y_true, y_pred):
18 return K.mean(K.abs(y_pred - y_true), axis=-1)
19
20
21 def mean_absolute_percentage_error(y_true, y_pred):
22 diff = K.abs((y_true - y_pred) / K.clip(K.abs(y_true),
23 K.epsilon(),
24 None))
25 return 100. * K.mean(diff, axis=-1)
26
27
28 def mean_squared_logarithmic_error(y_true, y_pred):
29 first_log = K.log(K.clip(y_pred, K.epsilon(), None) + 1.)
30 second_log = K.log(K.clip(y_true, K.epsilon(), None) + 1.)
31 return K.mean(K.square(first_log - second_log), axis=-1)
32
33
34 def squared_hinge(y_true, y_pred):
35 return K.mean(K.square(K.maximum(1. - y_true * y_pred, 0.)), axis=-1)
36
37
38 def hinge(y_true, y_pred):
39 return K.mean(K.maximum(1. - y_true * y_pred, 0.), axis=-1)
40
41
42 def categorical_hinge(y_true, y_pred):
43 pos = K.sum(y_true * y_pred, axis=-1)
44 neg = K.max((1. - y_true) * y_pred, axis=-1)
45 return K.maximum(0., neg - pos + 1.)
46
47
48 def logcosh(y_true, y_pred):
49 """Logarithm of the hyperbolic cosine of the prediction error.
50 `log(cosh(x))` is approximately equal to `(x ** 2) / 2` for small `x` and
51 to `abs(x) - log(2)` for large `x`. This means that 'logcosh' works mostly
52 like the mean squared error, but will not be so strongly affected by the
53 occasional wildly incorrect prediction.
54 # Arguments
55 y_true: tensor of true targets.
56 y_pred: tensor of predicted targets.
57 # Returns
58 Tensor with one scalar loss entry per sample.
59 """
60 def _logcosh(x):
61 return x + K.softplus(-2. * x) - K.log(2.)
62 return K.mean(_logcosh(y_pred - y_true), axis=-1)
63
64
65 def categorical_crossentropy(y_true, y_pred):
66 return K.categorical_crossentropy(y_true, y_pred)
67
68
69 def sparse_categorical_crossentropy(y_true, y_pred):
70 return K.sparse_categorical_crossentropy(y_true, y_pred)
71
72
73 def binary_crossentropy(y_true, y_pred):
74 return K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
75
76
77 def kullback_leibler_divergence(y_true, y_pred):
78 y_true = K.clip(y_true, K.epsilon(), 1)
79 y_pred = K.clip(y_pred, K.epsilon(), 1)
80 return K.sum(y_true * K.log(y_true / y_pred), axis=-1)
81
82
83 def poisson(y_true, y_pred):
84 return K.mean(y_pred - y_true * K.log(y_pred + K.epsilon()), axis=-1)
85
86
87 def cosine_proximity(y_true, y_pred):
88 y_true = K.l2_normalize(y_true, axis=-1)
89 y_pred = K.l2_normalize(y_pred, axis=-1)
90 return -K.sum(y_true * y_pred, axis=-1)
91
92
93 # Aliases.
94
95 mse = MSE = mean_squared_error
96 mae = MAE = mean_absolute_error
97 mape = MAPE = mean_absolute_percentage_error
98 msle = MSLE = mean_squared_logarithmic_error
99 kld = KLD = kullback_leibler_divergence
100 cosine = cosine_proximity
101
102
103 def serialize(loss):
104 return serialize_keras_object(loss)
105
106
107 def deserialize(name, custom_objects=None):
108 return deserialize_keras_object(name,
109 module_objects=globals(),
110 custom_objects=custom_objects,
111 printable_module_name='loss function')
112
113
114 def get(identifier):
115 """Get the `identifier` loss function.
116 # Arguments
117 identifier: None or str, name of the function.
118 # Returns
119 The loss function or None if `identifier` is None.
120 # Raises
121 ValueError if unknown identifier.
122 """
123 if identifier is None:
124 return None
125 if isinstance(identifier, six.string_types):
126 identifier = str(identifier)
127 return deserialize(identifier)
128 if isinstance(identifier, dict):
129 return deserialize(identifier)
130 elif callable(identifier):
131 return identifier
132 else:
133 raise ValueError('Could not interpret '
134 'loss function identifier:', identifier)