keras model.compile Loss function and optimizer

Loss function

summary

Loss function is the goal of model optimization , So it's also called objective function 、 Optimize the scoring function , stay keras in , Parameters for model compilation loss Class of loss function specified , There are two ways of specifying ：

model.compile(loss='mean_squared_error', optimizer='sgd')

perhaps

from keras import losses
model.compile(loss=losses.mean_squared_error, optimizer='sgd')

Available loss function

Available loss objective function ：

mean_squared_error or mse
mean_absolute_error or mae
mean_absolute_percentage_error or mape
mean_squared_logarithmic_error or msle
squared_hinge
hinge
categorical_hinge
binary_crossentropy（ Also called logarithmic loss ,logloss）
logcosh

categorical_crossentropy： Also known as multi class logarithmic loss , Note when using this objective function , The label needs to be transformed into a shape like (nb_samples, nb_classes) Binary sequence of 
sparse_categorical_crossentrop： Above , But accept sparse tags . Be careful , When using this function, you still need to have the same dimension as the output value , You may need to add a dimension to the tag data ：np.expand_dims(y,-1)

kullback_leibler_divergence: From the probability distribution of predicted values Q To the truth probability distribution P Information gain of , To measure the difference between two distributions .

poisson： namely (predictions - targets * log(predictions)) The average of 

cosine_proximity： That is, the inverse number between the predicted value and the average cosine distance of the real label 

Loss function formula

https://zhuanlan.zhihu.com/p/34667893

Two classification - Report errors

On the loss function of the report error ：
use Keras Do text classification , I always have mistakes like this ,

My category is 0 or 1, But the mistake told me it couldn't be 1.

See ：Received a label value of 1 which is outside the valid range of [0, 1) - Python, Keras

loss function The problem of .

It used to be sparse_categorical_crossentropy,

Change it to binary_crossentropy Problem solving .

Optimizer

https://www.cnblogs.com/xiaobingqianrui/p/10756046.html