Skip to content

Latest commit

 

History

History
196 lines (121 loc) · 6.08 KB

README.md

File metadata and controls

196 lines (121 loc) · 6.08 KB
Raw

yolov8

Table of contents

1. Description

The model used in this example comes from the following open source projects:

https://github.com/airockchip/ultralytics_yolov8

2. Current Support Platform

RK3562, RK3566, RK3568, RK3576, RK3588, RK1808, RV1109, RV1126

3. Pretrained Model

Download link:

./yolov8n.onnx
./yolov8s.onnx
./yolov8m.onnx

Download with shell command:

cd model
./download_model.sh

Note: The model provided here is an optimized model, which is different from the official original model. Take yolov8n.onnx as an example to show the difference between them.

  1. The comparison of their output information is as follows. The left is the official original model, and the right is the optimized model. As shown in the figure, the original one output is divided into three groups. For example, in the set of outputs ([1,64,80,80],[1,80,80,80],[1,1,80,80]), [1,64,80,80] is the coordinate of the box, [1,80,80,80] is the confidence of the box corresponding to the 80 categories, and [1,1,80,80] is the sum of the confidence of the 80 categories.
Image
  1. Taking the the set of outputs ([1,64,80,80],[1,80,80,80],[1,1,80,80]) as an example, we remove the subgraphs behind the two convolution nodes in the model, keep the outputs of these two convolutions ([1,64,80,80],[1,80,80,80]), and add a reducesum+clip branch for calculating the sum of the confidence of the 80 categories ([1,1,80,80]).
Image

4. Convert to RKNN

Usage:

cd python
python convert.py <onnx_model> <TARGET_PLATFORM> <dtype(optional)> <output_rknn_path(optional)>

# such as: 
python convert.py ../model/yolov8n.onnx rk3588
# output model will be saved as ../model/yolov8.rknn

Description:

  • <onnx_model>: Specify ONNX model path.
  • <TARGET_PLATFORM>: Specify NPU platform name. Such as 'rk3588'.
  • <dtype>(optional): Specify as i8, u8 or fp. i8/u8 for doing quantization, fp for no quantization. Default is i8.
  • <output_rknn_path>(optional): Specify save path for the RKNN model, default save in the same directory as ONNX model with name yolov8.rknn

5. Python Demo

Usage:

cd python
# Inference with PyTorch model or ONNX model
python yolov8.py --model_path <pt_model/onnx_model> --img_show

# Inference with RKNN model
python yolov8.py --model_path <rknn_model> --target <TARGET_PLATFORM> --img_show

Description:

  • <TARGET_PLATFORM>: Specify NPU platform name. Such as 'rk3588'.

  • <pt_model / onnx_model / rknn_model>: Specify the model path.

6. Android Demo

Note: RK1808, RV1109, RV1126 does not support Android.

6.1 Compile and Build

Please refer to the Compilation_Environment_Setup_Guide document to setup a cross-compilation environment and complete the compilation of C/C++ Demo.
Note: Please replace the model name with yolov8.

6.2 Push demo files to device

With device connected via USB port, push demo files to devices:

adb root
adb remount
adb push install/<TARGET_PLATFORM>_android_<ARCH>/rknn_yolov8_demo/ /data/

6.3 Run demo

adb shell
cd /data/rknn_yolov8_demo

export LD_LIBRARY_PATH=./lib
./rknn_yolov8_demo model/yolov8.rknn model/bus.jpg

  • After running, the result was saved as out.png. To check the result on host PC, pull back result referring to the following command:

    adb pull /data/rknn_yolov8_demo/out.png

  • Output result refer Expected Results.

7. Linux Demo

7.1 Compile and Build

Please refer to the Compilation_Environment_Setup_Guide document to setup a cross-compilation environment and complete the compilation of C/C++ Demo. Note: Please replace the model name with yolov8.

7.2 Push demo files to device

  • If device connected via USB port, push demo files to devices:
adb push install/<TARGET_PLATFORM>_linux_<ARCH>/rknn_yolov8_demo/ /userdata/
  • For other boards, use scp or other approaches to push all files under install/<TARGET_PLATFORM>_linux_<ARCH>/rknn_yolov8_demo/ to userdata.

7.3 Run demo

adb shell
cd /userdata/rknn_yolov8_demo

export LD_LIBRARY_PATH=./lib
./rknn_yolov8_demo model/yolov8.rknn model/bus.jpg

  • After running, the result was saved as out.png. To check the result on host PC, pull back result referring to the following command:

    adb pull /userdata/rknn_yolov8_demo/out.png

  • Output result refer Expected Results.

8. Expected Results

This example will print the labels and corresponding scores of the test image detect results, as follows:

person @ (211 241 283 507) 0.873
person @ (109 235 225 536) 0.866
person @ (476 222 560 521) 0.863
bus @ (99 136 550 456) 0.859
person @ (80 326 116 513) 0.311

  • Note: Different platforms, different versions of tools and drivers may have slightly different results.