Semantic segmentation with large images and U-Net

Hello,
someone nice could please adapt this code : Dropbox - train.py - Simplify your life to exploit 3500x2500 pixels images at input of neural network :
I have a 24GB RAM GPU, if 3500x2500 pixels is too much large to fit the RAM, please take a smaller dimension.
Thank you
Best regards

other files needed :

noone could do that please ?

As with your previous posts I don’t think you are giving enough information about your issue, where you are stuck, and especially what you’ve tried so far to solve it.
I don’t think just asking users to implement your request is too compelling.