Some time ago, I had a look at the Scaled YOLOv4 architecture. This architecture utilizes the “Mish” activation function. Since I consider darknet as be broken beyond repair, I was happy to find a pytorch implementation. Well, the code quality is similarly questionable but I simply wanted to try things out and not use in a production environment. Unfortunately, it requires the installation of mish_cuda. After a few attempts of compiling and installing it with useless error messages, I simply replaced mish_cuda with a pure python/pytorch version. However, if we want to load some of the pre-trained models provided, then pickle requires that mish_cuda.MishCuda exists, so a quick code change wouldn’t do the job. You’ll find mish-cuda-dummy here. I want to point out that this is nothing I would recommend for production use as it is slower than mish_cuda. It simply provides a compatible replacement if you want to test something that requires mish_cuda but you’re having trouble installing it in < 2 mins.