当我没有真实值时，如何编写自己的自定义损失函数？

Question

I'm trying to figure out a way to create my own loss function. I'm using keras_model_sequential() model on R.

custom_loss <- function(x){
    post <- second_model(x)   #the current model
    pri <- first_model()      #another already defined model
    LOSS <- sum((pri-post)^2)
    return(LOSS)
}

The problem is that I do not have the basic y_pred and y_true variables (which keras's default loss functions require), as I do not have labeled examples. I just have a random model with defined input. And my goal is to shape the model by minimizing the cost of my loss function. In other words, I want my network to learn itself the good values for the output (y).

Edit:

x_train <- model1 %>% predict(input_vector)

--code defining the model2 --

y_true <- matrix(c(1),100,16)  #dummy, because no target values

delta <- model2 %>% predict(x_train)
adjusted_input <- input_vector + delta
adjusted_y <- model1 %>% predict(adjusted_input)
y_pred <- adjusted_y #(just to have the same variable names as argument)

custom_loss <- function(y_pred, y_true){

    LOSS <- sum((10-y_pred)^2)
    return(LOSS)
}

And now comes the problem...

model2 %>% compile(
  loss = custom_loss(y_pred, y_true),
  optimizer = optimizer_nadam(),
  metrics = c("mae")
)

Answer 1

Ok, now I get it. Since you use tensorflow backend you have to pass tensor objects in the loss functions and import the backend of tensorflow for calculations.

So

1. Import tensorflow backend
2. Do these calculations in the loss function directly with tensorflow operations

 delta <- model2 %>% predict(x_train) adjusted_input <- input_vector + delta adjusted_y <- model1 %>% predict(adjusted_input) y_pred <- adjusted_ 

3. To input the result of the first model you can use the keras Input layer:

   layer_input() %>%

Here is a guide for it.