Holy duck, I lost two days on a convolutional autoencoder splitted in two separate neural networks to encode and decode separately, it reconstruction had some strange behaviours. I was giving as input an image and then saving the encoded compressed representation in a new image, in this way I could decode it with the decoder whenever I want saving space.
How much retarded am I?
The internal layer's weights hadn't constraints so in learning phase the convolutional filters can contain any number, positive > 255 or even negative and I cannot save it in a new image as they are so they were clipped automatically between 0 and 255 with an huge information loss.
It's so frustrating when you rewrite the code in any possible way, you obtain the same wrong result and then you realize that was a borderline behaviour of a third part library.