The smoothness loss is also different: I kept getting nan with the logarithmic operation so I deleted that part, the left part could be seen as total-variation loss (tv-loss)

SystemErrorWang / deepupe_tensorflow

An unofficial implementation of Underexposed Photo Enhancement using Deep Illumination Estimation

MIT License

19 stars 5 forks source link

The smoothness loss is also different: I kept getting nan with the logarithmic operation so I deleted that part, the left part could be seen as total-variation loss (tv-loss) #4

Open lilin-git opened 5 years ago

lilin-git commented 5 years ago

You normalize the data in (-1,1) and the parameter of tf. log () is greater than 0, so loss is nan.

SystemErrorWang commented 5 years ago

sorry for late reply, i also found this problem in my training. I suggest you to try deep guided filter as it only use tensorflow default ops and no additional self-defined layers. I got good result with it