loss values are not changed due to normalization of the input image

[crazy-zxx]

I tried to apply this code to a small object semantic segmentation task with only two categories, but encountered some problems.

My loss function values (loss_adv1, loss_adv2 , loss_D1 , loss_D2 ) are almost always 0.693 .

iter =    0/250000, loss_seg1 = 0.625 loss_seg2 = 0.570 loss_adv1 = 0.695, loss_adv2 = 0.689 loss_D1 = 0.693 loss_D2 = 0.693
iter =    1/250000, loss_seg1 = 0.327 loss_seg2 = 0.462 loss_adv1 = 0.610, loss_adv2 = 0.785 loss_D1 = 0.697 loss_D2 = 0.697
iter =    2/250000, loss_seg1 = 0.503 loss_seg2 = 0.581 loss_adv1 = 0.686, loss_adv2 = 0.708 loss_D1 = 0.693 loss_D2 = 0.693
iter =    3/250000, loss_seg1 = 0.544 loss_seg2 = 0.604 loss_adv1 = 0.724, loss_adv2 = 0.674 loss_D1 = 0.694 loss_D2 = 0.693
iter =    4/250000, loss_seg1 = 0.541 loss_seg2 = 0.625 loss_adv1 = 0.704, loss_adv2 = 0.678 loss_D1 = 0.693 loss_D2 = 0.693
iter =    5/250000, loss_seg1 = 0.347 loss_seg2 = 0.492 loss_adv1 = 0.672, loss_adv2 = 0.694 loss_D1 = 0.693 loss_D2 = 0.693
.....

The above problem occurs when I normalize the image to the range [0.1]. When I get the output below without normalizing the image, it seems the problem is solved.

iter =    0/500000, loss_seg1 = 0.745 loss_seg2 = 0.515 loss_adv1 = 0.690, loss_adv2 = 0.697 loss_D1 = 0.693 loss_D2 = 0.693
iter =    1/500000, loss_seg1 = 0.764 loss_seg2 = 0.526 loss_adv1 = 0.709, loss_adv2 = 0.680 loss_D1 = 0.692 loss_D2 = 0.692
iter =    2/500000, loss_seg1 = 0.735 loss_seg2 = 0.519 loss_adv1 = 0.676, loss_adv2 = 0.703 loss_D1 = 0.692 loss_D2 = 0.691
iter =    3/500000, loss_seg1 = 0.705 loss_seg2 = 0.513 loss_adv1 = 0.693, loss_adv2 = 0.699 loss_D1 = 0.693 loss_D2 = 0.691
iter =    4/500000, loss_seg1 = 0.692 loss_seg2 = 0.505 loss_adv1 = 0.714, loss_adv2 = 0.676 loss_D1 = 0.690 loss_D2 = 0.689
iter =    5/500000, loss_seg1 = 0.689 loss_seg2 = 0.500 loss_adv1 = 0.707, loss_adv2 = 0.700 loss_D1 = 0.693 loss_D2 = 0.687
iter =    6/500000, loss_seg1 = 0.638 loss_seg2 = 0.494 loss_adv1 = 0.678, loss_adv2 = 0.705 loss_D1 = 0.689 loss_D2 = 0.684
iter =    7/500000, loss_seg1 = 0.611 loss_seg2 = 0.501 loss_adv1 = 0.684, loss_adv2 = 0.678 loss_D1 = 0.684 loss_D2 = 0.680
iter =    8/500000, loss_seg1 = 0.610 loss_seg2 = 0.493 loss_adv1 = 0.718, loss_adv2 = 0.750 loss_D1 = 0.685 loss_D2 = 0.679
iter =    9/500000, loss_seg1 = 0.586 loss_seg2 = 0.481 loss_adv1 = 0.730, loss_adv2 = 0.627 loss_D1 = 0.691 loss_D2 = 0.675
iter =   10/500000, loss_seg1 = 0.580 loss_seg2 = 0.479 loss_adv1 = 0.697, loss_adv2 = 0.811 loss_D1 = 0.681 loss_D2 = 0.653
iter =   11/500000, loss_seg1 = 0.510 loss_seg2 = 0.473 loss_adv1 = 0.701, loss_adv2 = 0.754 loss_D1 = 0.676 loss_D2 = 0.646
iter =   12/500000, loss_seg1 = 0.488 loss_seg2 = 0.468 loss_adv1 = 0.696, loss_adv2 = 0.650 loss_D1 = 0.664 loss_D2 = 0.633
iter =   13/500000, loss_seg1 = 0.480 loss_seg2 = 0.474 loss_adv1 = 0.755, loss_adv2 = 0.955 loss_D1 = 0.649 loss_D2 = 0.618
iter =   14/500000, loss_seg1 = 0.456 loss_seg2 = 0.465 loss_adv1 = 0.752, loss_adv2 = 0.707 loss_D1 = 0.650 loss_D2 = 0.587
iter =   15/500000, loss_seg1 = 0.438 loss_seg2 = 0.464 loss_adv1 = 0.732, loss_adv2 = 0.805 loss_D1 = 0.687 loss_D2 = 0.560
iter =   16/500000, loss_seg1 = 0.406 loss_seg2 = 0.440 loss_adv1 = 0.542, loss_adv2 = 1.086 loss_D1 = 0.703 loss_D2 = 0.518
...

This seems to be due to the normalization I'm using on the input image? So that means I shouldn't process the input image? Can you help me ? Thanks.