silvanus0930/tool

This document describes rv1109 vm and its tools.

Tools

VirtualBox-7.2.6a-172322-Win.zip

• virtual box version 7.2.6
• vm : ubuntu 18.04, username: server, password: TNK#$104

python-3.9.13-amd64.zip

python installer which can be used to train yolov5-master

wheels_win64_py39_cu117.zip

python wheel files which are yolov5 dependencies. cuda version is 11.7.

yolov5-master.zip

ultralytics yolov5 repository. The models trained by this repository can be converted to rknn which can inference on RV1109 board.

VM Overview

• All source code and toolchain exists in home/server/Work dirctory
• RV1109 and RV1126 has the same spec except for its NPU capacity is different.

深圳RV1126开发资料

This is a toolchain of 深圳RV1126开发板 and it might mismatched with the actual RV1109 board. The toolchain is exactly installed in /home/thomas/rv1126/repo_new directory and verified for compile succed. This toolchain is used as reference purpose.

RKNN-TOOLKIT

• the last version of rknn-toolkit for RV1109 is 1.7.1. see rknn-toolkit17.tar.gz.
  It contains sample code to convert and test several CNN models on RV1109 and RV1126 including onnx and darknet. I provided the another sample code to convert ultralytics yolov5 into rknn.
• rknn-toolkit-v1.7.1-packages: It contains python packages of RKNN-toolkit, and it is already installed on system python env(3.6.9). So no need to install again.
• yolov5 This is a repository to convert the ultralytics yolo model into rknn.
  The source code which can train ultralytics on windows11-GPU env will be shared in either way. yolov5 folder only can be used to convert to rknn model. The dedicated conda env(conda activate yolov5) is installed to convert the model. To reference details steps for model conversion, see ~/Work/yolov5/ultralytics/yolov5-master/script/convert.md file. This repository contains the python code to inference yolov5 rknn model on RV1109. C++ inferencing code should be modified based on yolov3 inference code.