0

enter image description here

Input size is 1 x 128 x 128. Stride used throughout is 2. Looking at the generator, from left to right, the input/output dimensions with every layer should be:

  1. 128/63
  2. 63/30
  3. 30/14
  4. 14/6
  5. 6/2
  6. ?/?
  7. ?/?

Is it possible that the diagram is incorrect?

user6118986
  • 341
  • 2
  • 15
  • I believe it's 128->64 and so on. 2 ** 7 = 128, so you can do downsampling 7 times max. Which paper this figure from? – mrgloom Nov 16 '19 at 23:07
  • Privacy-Protective-GAN for Face De-identification (Wu, Lang, Ying). The thing is that a stride of 2 =/= divide in half, its the difference of the input dimensionality and the convolutional filter dimensionality which gets divided by the stride (then + 1). So for the first layer, its floor((128-4)/2) + 1 = 63. – user6118986 Nov 16 '19 at 23:13

0 Answers0