https://www.cnblogs.com/smuxiaolei/p/8662177.html
The objective function, or loss function, is the performance function in the network and one of the two parameters necessary to compile a model. Due to the many types of loss functions, the following is an example of the keras official website manual.
In the official keras.io, there are the following information:
-
mean_squared_error或mse
-
mean_absolute_error or mae
-
mean_absolute_percentage_error或mape
-
mean_squared_logarithmic_error或msle
-
squared_hinge
-
hinge
-
binary_crossentropy (also known as log loss, logloss)
-
categorical_crossentropy: also known as multi-class logarithmic loss. Note that when using this objective function, the label needs to be converted into
(nb_samples, nb_classes)
a binary sequence of the form -
sparse_categorical_crossentrop: As above, but accepts sparse tags. Note that when using this function, you still need your label to have the same dimension as the output value. You may need to add a dimension to the label data:
np.expand_dims(y,-1)
-
kullback_leibler_divergence: The information gain from the predicted value probability distribution Q to the true value probability distribution P to measure the difference between the two distributions.
-
cosine_proximity: the opposite of the average value of the cosine distance between the predicted value and the true label