To address #124, the suggestion is to use resized convolution (upscale -> convolve) instead of transposed convolution.
It turns out that we're already using resized convolution. To make everything consistent with SsnL's recommendation, I've just changed padding to reflection padding (shouldn't make a big difference).
To address #124, the suggestion is to use resized convolution (upscale -> convolve) instead of transposed convolution.
It turns out that we're already using resized convolution. To make everything consistent with
SsnL's recommendation, I've just changed padding to reflection padding (shouldn't make a big difference).