Skip to main content

YOLOX for Object Detection

Author: Meng Chang

This topic presents an example about how to deploy a YOLOX model on a RISC-V development board for object detection. 

The content of this topic is derived from the tutorial of LicheePi Module 4A (LPi4A) development board provided by Sipeed. A LPi4A development board uses a TH1520 chip and is equipped with a Xuantie C920 CPU. The software is HHB-onnxruntime customized by T-Head. Based on the onnxruntime project, HHB-onnxruntime supports vector instructions on C920 to optimize execution performance.

In the use case described in this topic, only RISC-V development boards are used, and no additional x86 hosts and other devices are required.

The following sections describe how to:

  • Install the Python environment on LPi4A.
  • Execute the source code from YOLOX repository on C920.
  • Execute ONNX models with HHB-onnxruntime.

The normal open-source model deployment process:

  1. Configure the basic Python environment on LPi4A.
  2. Get the YOLOX source code and model.
  3. Install the Python package that YOLOX depends on.
  4. Run a demo using HHB-onnxruntime on LPi4A

Configure the basic Python environment

Configure the basic hardware and software

Install the development board and enter the certificate information of the root user after power-on. For more information, see Unboxing.

Make sure that you are connected to the Internet and update the apt source.

apt update

Install the software required for running the demo.

apt install wget git vim

Install the SHL library. SHL provides optimized code for RISC-V vector extension instructions on C920.

tar xf c920.tar.gz
cp c920/lib/* /usr/lib/riscv64-linux-gnu/ -rf

Configure the Python environment

By default, Python V3.11 is installed in the system burned by LPi4A. You can run the following command to check the Python version.

python3 --version

In the following example, Python V3.11 is used. If you want to use another version of Python, specify the version that you want to use in the command when installing dependencies.

Most of the software packages that Python program software depends on can be installed by pip. You can run the following command to install pip.

apt install python3-pip

Before installing other Python packages, install the venv package that can be used to create a Python virtual environment.

apt install python3.11-venv

Create a Python virtual environment and activate it.

cd /root
python3 -m venv ort
source /root/ort/bin/activate

So far, the basic Python environment is created. Similar to other architectures, pure Python packages can be installed by the pip install command.

Get the YOLOX model

YOLOX is a YOLO-like object detection model with excellent performance.

You can download the source code and model from GitHub.

git clone
cd YOLOX/demo/ONNXRuntime

Modify the source code

To use HHB-onnxruntime to execute the model, switch to HHB-onnxruntime first. In the onnxruntime sample directory in the source code, modify the beginning content in the demo/ONNXRuntime/ file and add the fourth and fifth lines of code as shown below.

#!/usr/bin/env python3
# Copyright (c) Megvii, Inc. and its affiliates.

+import sys
+sys.path.insert(0, "../../")

import argparse
import os

Use sys.path.insert to specify the search path in the code. Therefore, you do not need to install the YOLOX package from the source code.

Install dependencies

The Python ecology of the RISC-V architecture is still lacking. After the update is released in the future, the dependent packages of YOLOX can be installed directly from the requirements.txt file.

The YOLOX demo relies on multiple Python packages. Therefore, you need to download the precompiled Python packages.

git clone -b python3.11
cd prebuilt_whl

You can run the following command in sequence to install the dependencies.

pip install numpy-1.25.0-cp311-cp311-linux_riscv64.whl
pip install opencv_python-4.5.4+4cd224d-cp311-cp311-linux_riscv64.whl
pip install kiwisolver-1.4.4-cp311-cp311-linux_riscv64.whl
pip install Pillow-9.5.0-cp311-cp311-linux_riscv64.whl
pip install matplotlib-3.7.2.dev0+gb3bd929cf0.d20230630-cp311-cp311-linux_riscv64.whl
pip install pycocotools-2.0.6-cp311-cp311-linux_riscv64.whl
pip3 install loguru-0.7.0-py3-none-any.whl
pip3 install torch-2.0.0a0+gitc263bd4-cp311-cp311-linux_riscv64.whl
pip3 install MarkupSafe-2.1.3-cp311-cp311-linux_riscv64.whl
pip3 install torchvision-0.15.1a0-cp311-cp311-linux_riscv64.whl
pip3 install psutil-5.9.5-cp311-abi3-linux_riscv64.whl
pip3 install tqdm-4.65.0-py3-none-any.whl
pip3 install tabulate-0.9.0-py3-none-any.whl

In the installation process, other pure Python dependent packages are required and pip automatically downloads them from the official sources.

Install HHB-onnxruntime

HHB-onnxruntime is a transplanted SHL backend (execution provider), and onnxruntime can be reused in SHL for high-performance optimized code on black iron CPUs.

pip install onnxruntime-1.14.1-cp311-cp311-linux_riscv64.whl

Execute the demo

Execute the demo in the examples directory.

python3 -m yolox_s.onnx -i ../../assets/dog.jpg -o outdir -s 0.7 --input_shape 640,640

Parameter description:

  • -m: specifies the model. In this example, yolox_s that was downloaded is specified as the model.
  • -i: specifies the image. In this example, a sample image that comes with the repository is specified.
  • -o: specifies the output directory as outdir. The outdir directory will be created during the execution.
  • -s: Specifies the detection threshold value as 0.7. Objects with a probability lower than 0.7 will be ignored.
  • –input_shape: Specifies the size of the image used for model detection.

Reference result

In this example, the following figure is used. With a threshold value of 0.7, the expected detection result of yolov5 includes a dog and a bicycle.

After the demo is executed, the result image dog.jpg is generated in the outdir directory. Each detected object is marked with a frame in the picture, and the probability is also marked. The result is as follows:


HHB-onnxruntime customized by T-head is used to execute the YOLOX model to complete the object detection task, and the C920 CPU on the RISC-V development board is used to quickly deploy the model.

Stay Connected With RISC-V

We send occasional news about RISC-V technical progress, news, and events.