Closed lovehuanhuan closed 3 years ago
Hi @lovehuanhuan No, the one showed in the README and at the end of the demo consider only the inference. The reason behind this choice is that tkDNN optimize inference only, while pre and post-processing are almost the same as in darknet and could be further optimized (with CUDA kernels maybe). However, if interested, you can check our paper in which we consider everything or also only our benchmarks results.
Does the FPS Results include the pre process time and the post process time?