Loss_ss is 0.000

[Wangyulin-user]

Hello chaofeng, thank you for your wonderful work. I would like to ask why the Loss_ss is 0 when I trained with my own dataset? The masks I used are not generated by the net_P, and there are only a part of the images have the label masks, for example, my dataset is about brain, and only the tumor part of the brain has label. So could you please explain what's the Loss_ss equals to 0 means? And also the Loss_pcp is 0.

[chaofengc]

Thank you for interest.

• For your first question, please make sure you correctly modify RegionStyleLoss for your data. There is a region_num parameter for this loss. At least two kinds of labels are required. Loss_pcp is not used in the paper, so we set its weight --lambda_pcp to zero by default.
• For the second question, please noted that the mask label should be converted to one hot format.

BTW: If you only want to calculate style loss for specific region, you need to make further modifications to the RegionStyleLoss class, for example, only use gram matrix for specific label in this function https://github.com/chaofengc/PSFRGAN/blob/c882315c9f1fc91c43d13a709016dd411f73d713/models/loss.py#L192

[Wangyulin-user]

Thank you for your kindly explanation, now I set the region_num as 1 and I got Loss_SS even though it still so small like 0.006. So is there an ideal value for the Loss_ss? In view of it is MSE loss, is 0 the training goal?

[chaofengc]

The value of Loss_SS depends on your data. Similar to other networks, it is usually impossible to reach 0 if you use mini batch stochastic gradient descent.

[zhangyunming]

if use Loss_pcp , the result will be better? and lambda_pcp should be?

[chaofengc]

I tried with perceptual loss, but did not find improvement over visual quality. Perhaps will improve quantitative results. I did not test that.