kwjinwoo / YOLOv1_Tensorflow2

YOLOv1 Implementation with Tensorflow2

Introduction

• This repository is implementation code with tensorflow2

Difference from Original Paper

• Using Xception as Backbone network(to get pre-trained weight)
• Following train strategy from this code (Batch size, num_epoch, augmentation...)

Requirement

python == 3.8
tensorflow == 2.8.0
opencv-python == 4.5.5
albumentations == 1.2.1

Dataset

• Using PASCAL VOC 2007, PASCAL VOC 2012
• For train, using PASCAL VOC 2007 trainval, PASCAL VOC 2012 trainval
• For validation, using PASCAL VOC 2007 test
• For test, using PASCAL VOC 2012 test
• Dataset dir structure

  $root/
  ├── VOC2007
  |   ├── Annotations
  |   ├── ImageSets
  |   └── JPEGImages
  ├── VOC2007_test
  ├── VOC2012
  └── VOC2012_test
  

• Make Dataset

  python make_dataset.py --dataset_root --save_dir
  
  args
  --dataset_root : VOC dataset files root directory path
  --save_dir : saved tffrecords files directory
  

• After run

  $save_dir/
  ├── trainval_2007.tfrecord
  ├── trainval_2012.tfrecord
  └── test_2007.tfrecord
  

Train

• Run

  python train.py --img_size --dataset_dir --s --num_class --num_epoch --batch_size
  
  args
  --img_size : model input size
  --dataset_dir : tfrecord file directory
  --s : output grid number (s X s)
  --num_class : the number of class
  --num_epoch : the number of epoch
  --batch_size : Batch size
  

• After run
  • model weights saved at ckpt dir
    yolo_final : saved after all epoch
    yolo_valid_best : saved at lowest valid loss

    $ckpt/
    ├── yolo_final.index
    ├── yolo_final.data-00000-of-00001
    ├── yolo_valid_best.index
    └── yolo_valid_best.data-00000-of-00001
    

  • model asset saved at model_asset dir

    $model_asset/
    ├── yolo_final
    └── yolo_valid_best
    

Test

• Inference

  python test.py --model_asset --images_dir
  
  args
  --model_asset : model asset dir path
  --images_dir : test image dir
  

• After run
  Inference result saved at infer dir

Result

• Train
• Validation
• Test

Reference

[1] You Only Look Once: Unified, Real-Time Object Detection. Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi
[2] devbruce / yolov1-tf2