problem about training SOS-Net

[Lu7z-727]

Before I trained SOS-Net, I trained my AECoder model based on the ReadMe file, where I took the geometric center of the small organ tag and cropped it to an 8x64x64 image as the center point. The input is the small organ label, and the output is the reconstructed small organ label diagram. The loss function is MSE. But I can't do the anti-habitual training properly.