README.md

This repository implements YOLOv1 with Pytorch. Paper Link

I implemented it for studying purpose, and I used Pytorch functions as much as possible for simplicity, such as calculating IoU, non-maximum suppression operations. I didn't encapsulate the function for some same operation, so it may have some code overlap, but it is more clearer without many function calls.

What's more, I tried my best to reproduce the paper as much as possible(network architecture, loss function, data augmentation and so on). But I didn't pretrain on the Imagenet.

Finally, I hope this repsitory will help you to understand original YOLO algorithms. You can follow the step to train your own YOLOv1 model. I will publish my training results when my program finishes running.

Clone this repository

git clone https://github.com/li624120638/yolov1

[optional] Creating a Virtual Environment

conda create -n yolov1 python=3.8 -y
conda activate yolov1

Install Pytorch Link

The version I installed is pytorch-1.12.1, torchaudio-0.12.1, with cudatoolkit-11.6.0

conda install pytorch torchvision torchaudio cudatoolkit=11.6 -c pytorch -c conda-forge

Install other package

pip install -r requirements.txt

Train model

If you haven't downloaded the VOC dataset yet, please set download to true, it will download VOC2007 trainval, test and VOC2012 trainval in dataset_root. VOC2007 test used as validate set, while other used as train set.

cd train_model
python train_yolo.py --config ./configs/VOC2007_detection_yolov1.yaml

Test model

Set model_args['weights'] to be the path of model you want to test.

python train_yolo.py --config ./configs/VOC2007_detection_yolov1_vis.yaml

The results will be saved in ${work_dir}/visulization/test_predict