Srinivas Kaza

@twtygqyy Awesome. I just pushed a few minor tweaks and fixes (changed the cache size, disabled shuffle). I mostly wrote this because my desktop doesn't have enough memory to run...

The image augmentation now uses PyTorch's API rather than OpenCV. `torch.flip` is unreasonably slow, but otherwise I'm noticing a substantial performance improvement over the old version.

On a different note, I finished training on DIV2K using this method and I was unable to replicate your results. I haven't looked much in `eval.py`, so maybe I'm wrong,...

Ah, good point. I'll try generating some of my own bicubic-downsampled images and seeing if the results get any better. I'd rather use PyTorch's bicubic downsample than Matlab's (or the...

I wrote an implementation of the inference half in a OpenGL [compute shader](https://github.com/AnimatedRNG/rusty-raisr). Performance is okay but could probably be improved; something like 8ms to upscale from 1080p to 4K...

I came across the same problem. I think this fork doesn't work with LLVM 8 -- I've had success with LLVM/Clang 7.0.1 (although I think LLVM 7.1.0 should also work)