AlexeyAB opened a new issue #5765: URL: https://github.com/apache/incubator-tvm/issues/5765
Feature-request: State-of-art **Yolo v4** Detector * paper: https://arxiv.org/abs/2004.10934 * source code: https://github.com/AlexeyAB/darknet * cfg-file Yolov4: https://github.com/AlexeyAB/darknet/blob/master/cfg/yolov4.cfg * weights-file Yolov4: https://drive.google.com/open?id=1cewMfusmPjYWbrnuJRuKhPMwRe_b9PaT ---- Required features for YOLOv4: 1. Mish-activation: ```cpp __device__ float mish_yashas(float x) { float e = __expf(x); if (x <= -18.0f) return x * e; float n = e * e + 2 * e; if (x <= -5.0f) return x * __fdividef(n, n + 2); return x - 2 * __fdividef(x, n + 2); } ``` 2. Eliminate grid sensitivity - by using scale_x_y= parameter from [yolo] layer in cfg-file (by default use scale_x_y=1.0 if the value is not set in cfg file) we should use: ```cpp const float x_tmp = logistic_activate(srcData[box_index + 0]) * scale_x_y - (scale_x_y - 1) / 2; const float y_tmp = logistic_activate(srcData[box_index + 1]) * scale_x_y - (scale_x_y - 1) / 2; dstData[box_index + 0] = (x + x_tmp)) / cols; dstData[box_index + 1] = (y + y_tmp)) / rows; ``` instead of **(this code isn't from TVM)**: ```cpp dstData[box_index + 0] = (x + logistic_activate(srcData[box_index + 0])) / cols; dstData[box_index + 1] = (y + logistic_activate(srcData[box_index + 1])) / rows; ``` ---- Accuracy/speed comparison with other detectors on MS COCO dataset. ![comparison_gpus](https://user-images.githubusercontent.com/4096485/84280311-cd11b300-ab3f-11ea-9262-f59f7f12f359.png) ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org