Unstable convergence PSNR

[stc1995]

Thanks for your reimplementation.
However, in my 5 trials of running the original training code of Lego, the results are unstable. In detail, 3 trials results are ended with PSNR in 9~10, the other two trials ended with PSNR above 30.
Did you encounter similar issues? Thanks for your response!

[kwea123]

It highly depends on the network initialization and the first training samples. Since there is a large portion of white background, if the network overfits to this background at the beginning, it makes the result very bad. You can see a solution here, or simply increase the batch size or trying other optimizers such as radam or ranger might help.