Code isuue

[Emimal]

if scale >= 2:
output = Subpixel(64, (3,3), r = 2,padding='same',activation='relu')(output)
if scale >= 4:
output = Subpixel(64, (3,3), r = 2,padding='same',activation='relu')(output)
if scale >= 8:
output = Subpixel(64, (3,3), r = 2,padding='same',activation='relu')(output)

All the three lines are same.

Isn't it

output = Subpixel(64, (3, 3), r=scale, padding='same', activation='relu')(output) ?

[rajatkb]

I think when working on the paper, the paper was not explicitly mentioning what do to when we need to output in scales above 4 or 8. I figured that using the pixel shuffle upscaling a number of times for higher upscales won't hurt. I verified it by changing r = 8 , or using the same operation 3 times. The final results are based on this final variant of code. The code only applies the same operation 3 times if scale is set to 8 , 2 times if 4 and 1 time is 2. I am looking into the paper again , to look at how to do higher level upscaling (progressively or imediately upscale to the scale)

PS: I just checked my mail exchange with the author he they also did the same, you can either put r = scale or scale it up multi stage, they are saying it has same effect. However i remember i verified by repeated experiment which one works. Since I used the same configuration to add perceptual loss to it.