PPYOLO-Tiny for Raspberry 4B

PPYOLO-Tiny

PP-YOLO Tiny is more suitable for mobile devices, which implement MobileNetV3 and Depthwise Seqarable Convolution.

The structure of PP-YOLO Tiny is shown at the end of this file.

Paddle Lite

paddlepaddle is a deep learning framework, but it do not support Arm devices such as Raspberry. So, the official support another tool called paddle-lite, which can be deployed on much more mobile devices.

Usage

About

This project is based on a detection mission which will be deployed on Raspberry 4B. The dataset is originally from berkeley-BDD100K dataset. The official annotations are two json files and I translate them into 79000 xml file (see bdd100k_voc_labels), so that we can easily train our model by using PaddleX.

Requirement

# Window/Linux
python 				3.7
CUDA(Optional) 		11.0
cudnn(Optional) 	8.0
paddlepaddle-gpu    2.1.2.post110
paddlex             2.0.0rc3

# Raspberry OS 		
raspios_full_armhf 2020-05-28-05:28
# https://mirrors.tuna.tsinghua.edu.cn/raspberry-pi-os-images/raspios_full_armhf/archive/

Step

1. Build Paddle Lite and install it or install the package I have already built

   The package is in the same respository called paddlelite-a0e14603f-cp37-cp37m-linux_armv7l.whl

   using pip3 install can install paddle-lite.

   OR:

   Strongly suggest this blog 【超详细】树莓派4B 安装Paddle-Lite 2.8.0

   **note: **When build paddle lite, use ./lite/tools/build_linux.sh --arch=armv7hf --with_python=ON --with_extra=ON --python_version=3.7 --with_cv=ON instead of ./lite/tools/build_linux.sh --arch=armv7hf --with_python=ON

2. Export your model

   By using paddlex, paddlex --export_inference --model_dir=./output/ppyoloTiny/best_model --save_dir=./inference_model is very simple.

3. Optimize your model

   This step can only be done on Raspberry, because the paddle-list version on PC may be different from Raspberry and the final version of building is a strange string (mine is a0e14603f)

   Using opt.py can translate model into paddle-lite form.

4. Run

   run python3 predict.py in your terminal is OK

   possible result:

   

   The model is trained from 1% of BDD100K dataset for 250 epochs

   If you want to predict from a camera, change if True into if False

Structure of PPYOLO-Tiny