how do you decide loss weight?

[timemao]

Hi, you did good work!

But I have trouble with loss weight, the L2-loss(reconstruction loss) finally converge to about 1000 for 16512512*3 [0,1],but the TV-loss calculated from groundtruth illumination map (get by input/gt image) and input is about 10^6, so we cannot place 1 for L2-loss and 2 for Tv-loss.

Any advise?

[wangruixing]

Did you mean L2 loss is around 1000? tv loss is around 10^6? In tv loss(local smooth loss) function, the input is origin image, the output is the illumination map. Actually, during my training process, the loss is around 0.033. I suggest you can only use the input image pixel divide illumination map at first, then calculate the l2 loss between the predicted image and ground truth. Only use l2 loss and make sure the loss is right at first.

[timemao]

Could you make sure metrics.py tv-loss code?

[wangruixing]

It is the local smoothness code, not standard tv loss.

[wangruixing]

The second parameter is not a predicted image, it is the illumination map.

[timemao]

Actually I calculate illumination map with

---

S=min(1.0,max(im1d,im1d./(im2d+0.0001)));

---

You mean the second parameter not like this? In metrics.py, you use tf.image.rbg_to_grayscale, this function get 1-channal output

[wangruixing]

You forgot to calculate the mean. tf.reduce_mean. and during processing image, the scale is 0-1.