preprocess scale 0.017 about densenet HOT 6 CLOSED

Actually, the reason of 0.017 is the following:
Originally, they preprocess normalized between [0.0; 1.0] RGB images by subtracting a mean value [0.485, 0.456, 0.406] and dividing each channel by a corresponding std value from [0.229, 0.224, 0.225], see here.
But as @titu1994 says that the weights are ported from here and Felix says that the weights were ported from here, where the preprocessing is

• BGR mean values [103.94,116.78,123.68] are subtracted
• scale: 0.017 is used, instead of the original std values for image preprocessing

and if you compute the inverse of 255 * mean(std), it gives 1.0 / (255.0 * 0.227) = 0.01727563272004837

HTH

from densenet.

Primarily, incorrect classification if scaling is not performed. I referred to the Keras-Densenet repo for this right ? Or did I forget that when I rewrote the call

from densenet.

It's approximately 1/58.8 for some reason. I never could find out why this is the case.

from densenet.

Maybe try with 1/255 and see if it still classifies correctly?

from densenet.

Thank you very much for the explanation @vfdev-5 !

from densenet.

@vfdev-5 good sleuthing, thanks!

from densenet.