Hello, I know that many people like my DarkNet on OpenCL. Recently I also made an improved version of the original DarkNet on CUDA, and I called it DarkNet-vNext, and it has all improvements that you were probably looking for. All Yolo models are supported from the 1st to the 4th version. It supports OpenCV 4 and, for example, benchmark possibility. The good news is it is high-speed. As you probably know, CUDA performs faster than OpenCL, but OpenCL is more system-ready. For example, I recently helped someone run an Intel CPU with Intel HD 620 GPU to run OpenCL DarkNet as Convolutional Neural Network, which worked flawlessly. On the training capability, the CUDA version is the best and the fastest I know. Now on DarkNet-vNext, you may train Yolo4 models, which was impossible before on the original version. And below, you may see changes in statistics.
Enjoy!
p ;).