The image output from the training is fine, but using the retrained model for testing, the effect is very bad. What is the reason for this?  