ppogg / ncnn-yolov4-int8

NCNN+Int8+YOLOv4 quantitative modeling and real-time inference

Hi there! I'm Xiangrong Chen.

Hi, I’m Xiangrong Chen, a student who likes to record life, and here is my Garden of Eden~

• 🔭 When I was in undergraduate, I teamed up with my classmates to attend competition. At that time I was exposed to machine learning and was fascinated by it.🤖
• 🌱 During graduate school, I chose deep learning as my research direction. I believe that there will be many interesting people and things on this way.🤣
• 👯 During the second year of graduate school, I have practiced in Guangzhou and recorded my own interesting experiments during the period.
• 💡 I like to program by python and C++
• 🤣 Hope to meet more friends. Nice to meet you~ 😁

Things I code with

Open source projects

🎁 Projects	⭐ Stars	📚 Forks	🛎 Issues	📬 Pull requests
YOLOv5-Lite
YOLO-Streaming
Retinaface_Ghost
ncnn-yolov4-int8

I like listening to music, playing ball, and like to adventurous more. 🎼🎸🎶 → Spotify Playing 🎧👀

---

At the same time, I also have other blogs, welcome to pay attention!💻📌

• ➡️ Xiangrong Chen's home page in zhihu💛
• ➡️ Xiangrong Chen's home page in CSDN💛

Connect with me and Grab a Coffee! 🙋‍♀️🥤

This README file is generated every 3 hours!
Last refresh: Monday, 14:07 CEST
Create your own here!

ncnn-yolov4-int8's Issues

量化以后，最后计算得到的位置xywh为定点数，怎么确定在图片中的位置，需要反量化吗

需要反量化吗，因为确定位置是需要浮点数的吧，类别可以比较定点数的概率

int8 fp16速度比较

“对比之前的fp16模型，明显在input_size相同的情况下快了40%-70%，且精度几乎没有什么损耗”

但是我看到表中int8反而更慢，是这样吗？有分析过为什么吗？

ncnn如何使用

pogg，你好！我是一个纯纯的小白，想问问你，我该如何使用ncnn进行推理，是新建一个C++项目，然后添加使用吗？我并没有看明白，可否提供一个教程，谢谢！

quantization/ncnn2table params

thanks for your project, I don't know how you generate table file. I have read your article on zhihu, but I don't know about ncnn2table's params: mean=[104,117,123], norm=[0.017,0.017,0.017] and pixel=BGR (while yolov4 uses PIXEL_BGR2RGB). Thanks a lot!