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?