search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.63k stars 7.95k forks source link

yolov4 training #5483

Open wdxybhb opened 4 years ago

wdxybhb commented 4 years ago

Hi there,

I would like to experiment to deploy yolov4 on an embedded linux with a npu on it.

Unfortunately, the model conversion tool does not support mish activation function and turns it into ReLU automatically, which generates wrong output.

May I ask is it okay to re-train the network via changing the mish to Relu while keeping yolov4.conv.137 as initial weight file?

Thank you so much!

AlexeyAB commented 4 years ago

This is yolov4 without mish:

DrXuQian commented 3 years ago

Hi AlexeyAB, we are currently training a Yolov4 without both the mish and leaky relu. In our config, we replaced both activation functions to Relu. But after following the Train Detector on MS COCO (trainvalno5k 2014) dataset, the mAp dropped a lot (< 20mAp @0.5IOU). I am wondering do I need to retrain the classifier on Imagenet first before training the detector? Thank you so much.