Hardware components | ||||||
 |
| × | 1 | |||
Software apps and online services | ||||||
 |
| |||||
 |
| |||||
Ship Detection based on YOLOv3 and KV260
OverviewThis is the entry project of the Xilinx Adaptive Computing Challenge 2021. It uses YOLOv3 for ship target detection in optical remote sensing images, and deploys DPU on the KV260 platform to achieve hardware acceleration.
First, I trained the YOLOv3 model for ship detection using the darknet framework. Secondly, referring to the darknet_yolov4 tutorial of Xilinx, I quantified and compiled the network model and evaluated the quantized model. Finally, I deployed the compiled xmodel to the KV260 platform with PYNQ framework to implement the hardware-accelerated YOLOv3 network for ship detection tasks. The model with quantized and compiled performs well on KV260 and achieves 7.63 FPS. In the follow-up work, I will apply methods such as compression pruning to the model to improve the running speed of the model.
Things used in this projectHardware components- AMD-Xilinx Kria KV260 Vision AI Starter Kit
- A computer with a high-performance GPU such as an RTX3090 to accelerate model training
- AMD-Xilinx PYNQ Framework
- AMD-Xilinx Xilinx Vitis-AI
Follow the steps below to reproduce it.
1. Prepare KV260 and PYNQRefer to Kria PYNQ Github Repository.
Get the Ubuntu SD Card ImageFollow the steps to Get Started with Kria KV260 Vision AI Starter Kit until you complete the Booting your Starter Kit section.
Install PYNQThen install PYNQ on your Kria KV260 Vision AI Starter Kit. Simply clone this repository from the KV260 and run the install.sh script.
git clone https://github.com/Xilinx/Kria-PYNQ.git
git clone https://github.com/Xilinx/Kria-PYNQ.git
cd Kria-PYNQ/
cd Kria-PYNQ/
sudo bash install.sh
sudo bash install.sh
This script will install the required Debian packages, create Python virtual environment and configure a Jupyter portal. This process takes around 25 minutes.
Open JupyterJupyterLab can now be accessed via a web browser <ip_address>:9090/lab or kria:9090/lab. The password is xilinx.
Clone this Repository on Kria-PYNQ directory and run yolov3_dpu/dpu_yolov3_voc.ipynb.
Image test:
Clone AlexeyAB Darknet, and train your own YOLOv3 model.
For ship detection, my model have 93.1% mAP with 0.5 IoU thresh.
Then launch Vitis-AI-CPU-1.4 docker environment, run scripts in folder to quantize and compile xmodel.
conda activate vitis-ai-caffe
conda activate vitis-ai-caffe
bash scripts/darknet_convert.sh # convert darknet model to caffe model
bash scripts/darknet_convert.sh # convert darknet model to caffe model
bash scripts/run_vai_q.sh # quantize model
bash scripts/run_vai_q.sh # quantize model
bash run_vai_c_kv260.sh # compile caffe model to xmodel
bash run_vai_c_kv260.sh # compile caffe model to xmodel
Comments