0

I am working on semantic segmentation using CNNs.

I have normalized the values of images to range [0,1]. I have trained my network many times, learning curve seems is learning well, however, the output is always black image. My question is does really scaling affects the learning or should the range of pixel values remains in 0-255 range?

Thanks a lot.

S.EB
  • 1,966
  • 4
  • 29
  • 54
  • 3
    Have you re-scaled the output to the range 0-255? If not, then you're displaying values in the range [0, 1) on a 0-255 scale, which is going to look pretty black. – Prune May 30 '17 at 23:45

1 Answers1

1

In general, it doesn't affect the training. In many situations, [0 1] is actually better than [0 255] for training a DNN since the latter may cause some numerical issues in learning the first layer weights if you don't use lower learning rates in the first layer since w' = y' * x, where ' denotes derivative.

DataHungry
  • 351
  • 2
  • 9
  • Thank you very much for ensuring me whether I am in right track or not. Thanks a lot :) – S.EB Jun 07 '17 at 17:23