How to achieve 'multi-step end-to-end training' in this code?

[SATURN2021]

Hello, it's very nice for me to appreciate your approach! I get some problems when studying your code... I've learnt that the multi-step end-to-end training was used, but I didn't find how this is done in the code. As we can see in your paper, 'every upsampling unit has two stages except the first one', so I guess there should be something in your code to control the training process, and I didn't find them however. I guess that the 'detach()' controls the training process, but it doesn't appear at the right position. Maybe my thought was too simple because I'm new in deep learning. Thank you for your reading and help! This has confused me for some days...