A. Use function loss
Loss function [also known as optimization objective function or scoring function] is one of the two parameters required for the compilation model.
model.compile(loss='mean_squared_error', optimizer='sgd')
or
from keras import losses
model.compile(loss=losses.mean_squared_error, optimizer='sgd')
You can pass a current loss function name or a TensorFlow / Theano signum function. Each data point returns a scalar, has a lower symbol of the two parameters as a function of:
1.y_true
The real label, TensorFlow / Theano tensor.
2.y_pred
Prediction value, / Theano tensor TensorFlow, which is the same shape and y_true.
The actual optimization goal is an average of all the output array data points.
II. Available loss function
1.mean_squared_error (y_true, y_pred) [MSE, mean square error}
The formula is:
Source:
2.mean_absolute_error (y_true, y_pred) [MAE, the average absolute error]
MAE mentioned can not say significant target detection, target the so-called significant, for example, when we look at a picture, we will first focus on those colorful, eye-catching content. As we will first of all look at the same time to see Transformers Optimus Prime, this is absolutely the C bits. So we define Transformers Optimus Prime is a significant goal.
The evaluation index calculating significance target detection, the detection algorithms have used mean absolute error, which is calculated as follows:
Source:
3.mean_absolute_percentage_error [MAPE, the mean absolute percentage error]
Similarly the average absolute error, mean absolute percentage error between the predicted results and the variation ratio of the true value. Calculated as follows:
Source:
Remarks:
1.clip
The number of element by element, will exceed the specified range into a number of border enforcement.
2.epsilon
Fixed parameters, the default value 1 * e-7.
[4.mean_squared_logarithmic_error MSLE, logarithmic mean square error}
Before calculating the mean square error to logarithmic data, recalculated.
The formula is:
Source:
5.squared_hinage [not used]
The formula is:
Source:
6.hinage [not used]
The formula is:
Source:
7.categorical_hinge [not used]
Source:
8.logcosh [not used]
Prediction error of the logarithm of the hyperbolic cosine. The results mean square error with roughly the same, but not strongly influenced by the occasional crazy wrong prediction.
Source:
9.categorical_crossentropy [not used]
When using categorical_crossentropy loss, the target should be classified [i.e., if the format type 10, then the target value of each sample should be a 10-dimensional vector, this vector represents the index categories in addition to 1, others are 0} . To convert the integer target classification target, utility function can use keras to_categorical.
from keras.utils.np_utils import to_categorical
categorical_labels = to_categorical(int_labels, num_classes=None)
Source:
10.sparse_categorical_crossentropy [not used]
Source:
11.binary_crossentropy [not used]
Source:
12.kullback_leibler_divergence [not used]
Source:
13.poisson [not used]
The formula is:
Source:
14.cosine_proximity [not used]
The formula is:
Source:
III. Other types of loss function
1.ctc_batch_cost [high performance]
Source:
CTC loss algorithm running on each batch element.
parameter:
1.y_true
It contains the true value of the tag tensor. Type (samples, max_string_length).
2.y_pred
包含预测值或softmax输出的张量。类型(samples, time_steps, num_categories)。
3.input_length
张量(samples, 1)，包含y_pred中每个批处理项的序列长度。
4.label_length
张量(samples, 1)， 包含y_true中每个批处理项的序列长度。
返回shape为(samples, 1)的张量，包含每一个元素的CTC损失。