Why TensorFlow Loss Function Value is Getting Bigger and Bigger: A Simple Solution – TensorFlow Tutorial

By | December 4, 2019

When you are using tensorflow to training your model,  you may find an error: the value of loss function is getting bigger and bigger when training. How to solve this problem? In this tutorial, we will discuss two ways to fix this error.

tensorflow learning rate settings

Why the value of loss function is getting bigger and bigger

There are two reasons:

(1) The learning rate is high

(2) The value of hyperparameters are too large

For example:

As to loss function

loss = loss_1

There is only a loss function, you should check your learning rate is high or not. If you have set it to be 0.001, which means the loss function may be wrong, you should modify your model. However, if the original learning rate is 0.005, you can set it to be 0.001

As to loss function

loss = a * loss_1 + b * loss_2

There are two hyperparameters a and b, if the learning rate is low, you should consider the value of a and b is too large or not. Decrease them.

Leave a Reply