Keras自定义损失函数输出负值，不明白为什么？

Question

Dear StackOverflow Community, I have the following loss function in keras:

return K.mean((y_true+K.epsilon()) * K.square(y_pred - y_true), axis=-1)

When I try to train my network (y normalized to 0 - 1) with it, the loss appears to get to an negative value, which I just can't understand. I calculated the same thing with numpy, and everything worked fine and as intended.

I would be really delighted If someone knows the cause for this weird negative solutions, so thank you for your help.

Answer 1

If y_true is really normalized to 0-1 that only possible cause that I see is K.epsilon() . As this page suggest epsilon can be changed by user and this can cause a problem.

Try to hardcode epsilon value or just throw it away.