I’m wondering if anyone has implemented the above paper, found at [2002.02117] Fixed smooth convolutional layer for avoiding checkerboard artifacts in CNNs that purports to fix checkerboard artifacts in conv and convtranspose layers with stride greater than one.
I have no experience with time-based DSP and I can’t get my head around what their ‘zero-order hold kernel’ actually equates to when using an image dataset …
Thanks in advance for any insight!
This is the first time I’ve seen this paper and they provide 0 information on how the operation was actually implemented in code, so I imagine not? Something like this should really have an accompanying code listing or repo…
Thanks for the reply.
I was surprised I couldn’t find any implementation details or repo - not even a note on which packages they used to do the tests.
I’ll see if I can cobble something together over the weekend, as a test. I have a basic idea how to implement it, but it likely won’t be pretty!