0

For context, I trained two separate autoencoders in Keras: one with a standard MSE loss function and one with a customized MSE loss function. When evaluating both models at a given point in training, they have very similar performance but the losses are very different.

My metric for performance is mean percent error. Both models are recreating the original image with mean error on the order of 3%. However, when these models were saved, the standard Keras MSE model had a loss less than 1.0 while the model with the customized MSE cost function had a loss on the order of 30.

If they perform on such a consistent level, why are the losses so drastically different?

WVJoe
  • 515
  • 7
  • 21

1 Answers1

1

Loss is just a scalar that says to model train which direction take to adjust weights. If you multiply a loss by a scalar, the result will be almost the same. Im not saying that the absolute value doesnt matter, it matters. But it's not the central point. Probably the difference in your case happens because Keras MSE do some normalizations that you dont.

Augusto Maillo
  • 166
  • 1
  • 6
  • That seems to be the cause for it. I'm not very concerned with it as my performance metrics are what is important to me, but I just wanted a better understanding on what was going on. – WVJoe Apr 25 '20 at 16:05
  • 1
    Try to normalize your y_pred data before do the math. – Augusto Maillo Apr 27 '20 at 16:27