Paddle-YOLOv4

English | 简体中文

1、Introduction

This project is based on the paddlepaddle_V2.1 framework to reproduce YOLOv4. YOLOv4 is the fourth generation model of YOLO series. While retaining the encoding and decoding mode of YOLOv3 detection head, YOLOv4 uses stronger backbone network, stronger feature fusion module and more data enhancement modes, so that the model performance is significantly improved compared with YOLOv3, and the inference speed is still very fast.

Paper:

• [1] Bochkovskiy A, Wang C Y, Liao H. YOLOv4: Optimal Speed and Accuracy of Object Detection[J].2020.

Reference project：

• https://github.com/AlexeyAB/darknet

The link of aistudio：

• notebook：https://aistudio.baidu.com/aistudio/projectdetail/2479219

2、Accuracy

The model is trained on COCO2017's train set and tested on COCO2017's testdev set (the training can be verified on val set first to evaluate the training of the model).

The AP on testdev set given in the paper when the input size is 416x416 is 41.2%, and the AP obtained in this project is 41.2%.

3、Dataset

COCO Dataset

• Dataset size:
  • train set: 118,287
  • val set: 4952
  • testdev set: 20288
• Data format: standard COCO format, marked with rectangular boxes

4、Requirements

• Hardware：CPU、GPU（a machine with four Tesla V100-32G is recommended）

• Framework：

  • PaddlePaddle >= 2.1.2

5、Quick Start

step1: clone

# clone this repo
git clone https://github.com/nuaaceieyty/Paddle-YOLOv4.git
cd Paddle-YOLOv4
export PYTHONPATH=./

Install dependencies

pip install -r requestments.txt

step2: Training

1. Create an Output folder in the top level directory and download CSPDarkNet backbone pre-training weights (I have converted the official CSPDarkNet weight to .pdparams) here: https://aistudio.baidu.com/aistudio/datasetdetail/103994.
2. This project can be trained using four card Tesla V100-32G. Note: COCO dataset should be prepared in advance, and decompressed into the data directory under the top-level directory (data set address is: https://aistudio.baidu.com/aistudio/datasetdetail/7122). If the dataset address is incorrect, change the corresponding address to the absolute path in the configs/datasets/coco_detection.yml file.

python -m paddle.distributed.launch --gpus 0,1,2,3 train.py -c configs/yolov4/yolov4_coco.yml --eval

step3: Evaluating

Note: Make sure the best_model.pdparams file is in the output directory.

python3 eval.py -c configs/yolov4/yolov4_coco.yml

step4: Generate results of testdev set for submission

Note: Make sure the best_model.pdparams file is in the output directory.

python3 eval.py -c configs/yolov4/yolov4_coco_test.yml

Zip the bbox. Json file generated in the live directory and send it to the evaluation server

Prediction using pre training model

Put the images to be tested in the data directory, run the following command, and save the output images in the Output directory; If there is a GPU in the machine environment, delete -o use_gpu=False from the command

python3 predict.py -c configs/yolov4/yolov4_coco.yml --infer_img data/1.jpg -o use_gpu=False

The result is shown as follows：

六、Code structure

6.1 Structure

├─config                          
├─model                                                     
├─data                            
├─output                          
│  eval.py                        
│  predict.py                     
│  README.md                      
│  README_CN.md                   
│  requirements.txt               
│  train.py                       

6.2 Parameter description

Parameters related to training and evaluation can be set in train.py, as follows:

Parameters	default	description	other
config	None, Mandatory	Configuration file path
--eval	None, Optional	Evaluate after an epoch	If you don't select this, you might have trouble finding the best_model
--fp16	None, Optional	Semi-precision training	If this option is not selected, 32GB of video memory may not be sufficient
--resume	None, Optional	Recovery training	For example: --resume output/yolov2_voc/66

6.3 Training process

See 5、Quick Start

7、Model information

For other information about the model, please refer to the following table:

information	description
Author	YU Tianyang(EICAS)
Date	2021.10
Framework version	Paddle 2.1.2
Application scenarios	Object detection
Support hardware	GPU、CPU
Download link	Pre training model
Online operation	notebook