0

Hello I'm implementing a CycleGAN and most of the other implementations I've seen on the internet use Convolution with stride 2 instead of a Maxpoolinglayer for downsample.

On to my question, why should we dismiss Maxpooling and instead add stride 2 to the Convolutions in the U-Net for CycleGANs. Is it because to much information gets lost in the Maxpoolingoperation? Or is there a different reason?

Zitrus
  • 1

1 Answers1

0

In general, you loose information with max-pooling yes, as you only choose some certain amount of feature from the previous features. Because of this reason, people may prefer to use CNNs with more strides to shrink the spatial size. Unless there is another explanation in the paper you referred, this should be the main reason. Nevertheless, you are right with your explanation.

ai-py
  • 1
  • Umm, does using strides preserve more information than with maxpooling? I'd argue the opposite is true, since with maxpooling you are calculating the convolution on more datapoints. – NikoNyrh Apr 28 '23 at 08:36