Keras backend (tensorflow) vs Keras

Question

I would like to custom a Keras loss function but I do not really understand something.

If I use tensorflow as a backend for Keras, do I need to use functions from keras.backend or can I use functions directly from tensorflow.

I only see posts where people are using functions from keras.backend but not from tensorflow (even if tensorflow has much more functions). Are there reasons to do so?

For a toy example :

from keras import backend as K 
import tensorflow as tf

def loss_keras(y_true, y_pred):

    square_error = K.square(y_pred - y_true)
    loss = K.mean(square_error)

    return loss

def loss_tf(y_true, y_pred):

    square_error = tf.squared_difference(y_pred, y_true)
    loss = tf.reduce_mean(square_error)

    return loss

Both of these functions work well but one is using directly tensorflow and the other is using keras.backend functions.

I know that this is a silly example but when you want to do more complicated stuff, I thought that using tensorflow would be easier than keras functions as there are more functions available

Answer 1

如评论中所指出和在此答案中所述： “在以下情况下，必须使用Keras后端函数（即keras.backend。*）：1）需要预处理或扩充传递给实际函数的参数Tensorflow或Theano后端或对返回的结果进行后处理，或2）您想编写一个可在所有Keras支持的后端上使用的模型。”