Three line by line explanations of the source code of anchor free series network yolox (a total of ten articles, which are guaranteed to be explained line by line. After reading it, you can change the

The whole series includes ：Demo Explain the source code line by line ->train The script source code is explained line by line ->backbone Explain the source code line by line ->FPN Explain the source code line by line ->Head Explain the source code line by line ->loss Explain the source code line by line -> The source code of data loading is explained line by line -> The source code of data enhancement is explained line by line ->simOTA Explain the source code line by line . Ensure line by line , Note that line by line , Include python grammar ,tensor Role and application of dimension and line by line code . In fact, there is no mystery about the network structure , It's a stack of modules , There is no reason why you can modify any module . After reading this series, I can do anything about any network structure at will , It's not just limited to one caller .

This article is about YOLOX Medium decoupling Head Build , Before that, you must read the previous blog posts , Otherwise, there will be a lot of content in this article that you can't understand .

---

Above, YOLOX Network structure diagram of the whole decoupling head . The yellow font is the intermediate variable in the code , Easy to understand . Dimension is when yolox_s Network is the output characteristic dimension of the whole network , And the default detection categories are 8 individual . First of all, we must understand what is decoupling head ： Decoupling head is to separate classification and detection tasks . Classification task and regression task are completely different in principle , Forecast them separately , The accuracy of the results is improved , Real time proof is true