Hello. I have trouble in training that the loss hardly decreases at all. Since the model.compile are not allowed to tweak, I could only think of that the mistake should be either in the model structure or in the model.fit. I used almost the same structure as the structure in the Fashion Mnist CNN autoencoder except for using relu activation function in all conv2d layers. Could you give me some suggestions? Thank you.
Yeah probably the it is in the model, keep trying to change a bit the architecture. Maybe increasing or decreasing the conv2d layers.