cropped the training images to 256*256, i met some errors

[wzr0108]

I cropped the training images so that the length and width are both 256, and the original image will become 4 small images, but the result of this training is completely wrong. I took the valid set to train again, and performed well on the test set, which proves that my training code is correct. The strangest thing is that I took the upper left corner of the training set picture for training, and the test result was good, but I took the upper right corner for training, but the result was very poor

[wzr0108]

[wzr0108]

I found a solution. I set the norm_method to "std" before and changed it to "whole_image". I would like to ask what is the difference between the two.