Skip to main content

YOLOv5

YOLOv5​

This tutorial explains how to run the YOLOv5 model on the RevyOS system using either the CPU or the NPU.

Initial Environment Configuration

Before proceeding with this tutorial, please ensure you have completed the Environment Configuration section.

Obtaining Example Code​

The example code accompanying this tutorial is available on GitHub. Clone it to your local machine using the git command.

$ git clone https://github.com/zhangwm-pt/lpi4a-example.git

The code relevant to this tutorial is located in the detection/yolov5 directory.

Model Acquisition​

The model used in this tutorial originates from the YOLOv5 repository and can be exported via its export.py script.

$ git clone https://github.com/ultralytics/yolov5.git
$ cd yolov5
$ pip3 install ultralytics
$ python3 export.py --weights yolov5n.pt --include onnx --imgsz 384 640
About Github Network Proxy

If you experience network issues when accessing GitHub from mainland China, consider using a network proxy tool to accelerate access.

Model Conversion and Compilation​

After completing the environment configuration, you can use HHB to compile the model into an executable program for the C920.

This tutorial uses the YOLOv5n detection model. For YOLOv5n, the HHB command truncates the graph at the final convolutional layer, and the subsequent post-processing is delegated to the provided yolov5n.c file.

Navigate to the detection/yolov5 directory and execute the following command:

$ hhb -D --model-file yolov5n.onnx --data-scale-div 255 \
    --board th1520 --input-name "images" --output-name \
    "/model.24/m.0/Conv_output_0;/model.24/m.1/Conv_output_0;/model.24/m.2/Conv_output_0" \
    --input-shape "1 3 384 640" --calibrate-dataset kite.jpg  \
    --quantization-scheme "int8_asym"
About Parameters
  • -D: Directs HHB to stop after generating the executable file
  • --model-file: Specifies the input model file
  • --data-mean: Specifies the mean value
  • --data-scale: Specifies the scaling value
  • --board: Specifies the target platform as C920 (CPU) or TH1520 (NPU)
  • --input-name: Names the model input tensor
  • --output-name: Names the model output tensor
  • --input-shape: Specifies the model input tensor shape
  • --postprocess: Configures the post-processing behavior for HHB-generated glue code; save_and_top5 saves the outputs and prints the top-5 results
  • --quantization-scheme: Specifies the quantization type

You can run hhb --help to view all available parameters and options.

About HHB Generated Files

After the command finishes, the current directory contains a newly created hhb_out subdirectory with files such as:

  • hhb.bm: HHB model file containing quantized weight data
  • hhb_runtime: Executable file for the development board, compiled from the C sources in this directory
  • main.c: Reference entry point for the HHB-generated example program
  • model.c: HHB model structure representation file
  • model.params: Model weight file
  • io.c: HHB-generated example program providing file I/O helper functions
  • io.h: HHB-generated example program providing file I/O function declarations
  • process.c: HHB-generated example program providing image preprocessing functions
  • process.h: HHB-generated example program providing image preprocessing function declarations

GCC Post-processing Compilation​

This tutorial implements the latter portion of the model and NMS in C so that post-processing yields detection results for the output image.

$ riscv64-unknown-linux-gnu-gcc yolov5n.c -o yolov5n_example hhb_out/io.c \
    hhb_out/model.c -Wl,--gc-sections -O2 -g -mabi=lp64d -I hhb_out/ -L \
    /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/th1520/lib/ \
    -lshl -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/decode/install/lib/rv \
    -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/runtime/riscv_linux \
    -lprebuilt_runtime -ljpeg -lpng -lz -lstdc++ -lm -I \
    /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/th1520/include/ -mabi=lp64d \
    -march=rv64gcv0p7_zfh_xtheadc -Wl,-unresolved-symbols=ignore-in-shared-libs -I \
    /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/th1520/include/shl_public/ \
    -I /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/th1520/include/csinn/

In the example code, the SHL library is linked; for example, its installation directory can be /usr/local/lib/python3.8/dist-packages/shl.

Option Descriptions
  • -Ihhb_out -I/usr/local/lib/python3.8/dist-packages/shl/install_nn2/c920/include/: Header search paths that include the SHL headers
  • -L/usr/local/lib/python3.8/dist-packages/shl/install_nn2/c920/lib: Library search path that points to the precompiled SHL binaries
  • -static: Enables static linking
  • -o yolov5n_example: Specifies the name of the generated executable

After successful compilation, a yolov5n_example file will be generated in the example directory.

Execution​

After cross-compilation, copy the required files to the development board directory.

For example, with board IP 10.63.x.x using /demo directory, copy the example program directory using scp:

scp -r yolov5n th1520@10.63.x.x:/demo/

When executing from the Linux command line on the board's terminal, the terminal output reports the following stages:

  1. Preprocessing: Scale and pad the original image to 384 × 640
  2. Model execution and post-processing: Perform model inference and apply NMS
  3. Bounding box drawing: Render the detection results on the 384 × 640 image
Required Execution Files
  • kite.jpg: Input image
  • image_preprocessed.bin: Preprocessing intermediate result
  • yolov5n_example: Model execution file compiled on the x86 host with GCC
  • hhb_out/hhb.bm: Model execution file generated on the x86 host by HHB
  • detect.txt: Post-processing output file containing 4 detected objects
  • kite_result.jpg: Output image with detection boxes added

Reference Results​

The input image shows a family of three flying a kite. YOLOv5 is expected to detect three people and one kite. Input Image Detection Image

When the example runs successfully, the terminal prints output similar to the following:

$ python3 inference.py
 ********** preprocess image **********
 ******* run yolov5 and postprocess *******
INFO: NNA clock:1001624 [kHz]
INFO: Heap :anonymous (0x2)
INFO: Heap :dmabuf (0x2)
INFO: Heap :unified (0x5)
WARNING: Mapping to the on chip ram failed (128 > 0), continuing...
FATAL: Importing 737280 bytes of CPU memory has failed (Invalid argument)
Run graph execution time: 11.96299ms, FPS=83.59
detect num: 4
id:     label   score           x1              y1              x2              y2
[0]:    0       0.895277        273.492188      161.245056      359.559814      330.644257
[1]:    0       0.887368        79.860062       179.181244      190.755692      354.304474
[2]:    0       0.815214        222.054565      224.477600      279.828979      333.717285
[3]:    33      0.563324        67.625580       173.948883      201.687988      219.065765
 ********** draw bbox **********
[273.492188, 161.245056, 359.559814, 330.644257, 0.895277, 0]
[79.860062, 179.181244, 190.755692, 354.304474, 0.887368, 0]
[222.054565, 224.4776, 279.828979, 333.717285, 0.815214, 0]
[67.62558, 173.948883, 201.687988, 219.065765, 0.563324, 33]