YOLO_PYNQ

Based on yolov2_xilinx_fpga by 陈辰, this project implements the deployment of yolov2-tiny on the Zhengdian Yuanzhi Linghangzhe ZYNQ_7020 board. Features such as photo capture, weight switching, and auto start have been added. (Due to limited resources on the board, the photo capture and weight switching features cannot be implemented for the full YOLOv2, and are instead supported using yolov2_tiny.)

Environment

Development board: Zhengdian Yuanzhi Linghangzhe ZYNQ_7020
Project part: xc7z020clg400-2
Vivado version: 2018.3

Instructions

Development Board Model

• Other development boards such as PYNQ_Z2 may be used, but it's recommended to choose a pynq series board with sufficient resources.
  The camera used is ov5640.
• If a different board model is used, modifications to the XDC file in the Vivado project and the device addresses in the board files are required.
• For PYNQ_Z2, you can refer to the project PYNQ-YOLOv2 by qing-2, which uses the YOLOv2 model.

Training and Quantization

• Training is done using darknet, see software_version, using the int16 version.

• If the network architecture needs to be modified, specify the desired structure during training and update the definitions accordingly during quantization.

  Quantization for yolov2_tiny can be found in software\software_version_yolov2_tiny.

HLS Simulation

HLS files are located at src_int16, using the int16 version. A float32 version can also be used for faster simulation, as int16 simulation is time-consuming.

Vivado Design

The project files are located at Vivado\yolov2_ov5640_hdmi_zdyz7020.
Two physical buttons are added to control photo capture and weight switching.

Block_Design

Board Deployment

Board-related files are based on PYNQ. Two weight models are used: one trained on the COCO dataset for common object detection, and another trained for skin disease classification (with poor performance, included for demonstration purposes only).

• Weight Switching
  Controlled via two PL buttons, which update the model weights during runtime.

• Auto Start
  The PYNQ series runs a Linux kernel, so you only need to add a startup service.

  • command.txt — commands to enable and disable the startup service
  • pynq_yolo.service — the systemd service file
  • yolov2_tiny.py — the Python task executed on the board
  • start_yolov2_tiny.sh — the startup script

TODO

• Dynamic Network Structure Switching
  The hardware structure is not hardcoded in the bitstream, so the network structure can also be dynamically switched. However, the YOLOv2 network, when quantized to int16, is difficult to fit on the ZYNQ_7020 board. Therefore, this feature is only experimental.