YOLO

A pytorch implementation of YOLOv1-v3.  
Project only supports python3.x.

To do

• [ ] Test the model of YOLOv1.
• [ ] Train the model of YOLOv1.
• [x] Test the model of YOLOv2.
• [x] Train the model of YOLOv2.
• [x] Test the model of YOLOv3.
• [x] Train the model of YOLOv3.
• [x] Data augmentation.
• [x] Using k-means to generate the priors.
• [x] Evaluate the model including mAP, precision, recall, etc.
• [x] Data of VOC format -> YOLO format.

Dependency

• torch 0.3.1
• opencv-python
• torchvision
• numpy
• pillow
• argparse

Train

Prepare

VOC -> YOLO

Data of VOC format(lxml) -> YOLO format(txt).
You can use the script in ./TOOL/voc2yolo.py to complete the work of conversion.

• Usage
  • modify Line<13~15> according to your needs
  • run "python3 voc2yolo.py"

Get good priors

Run k-means clustering on the dimensions of bounding boxes to get good priors for our model.
You can use the script in ./TOOL/genPriors/genPriors.py to get the good priors.

• Usage
  • modify the options.json according to your needs
  • run "python3 genPriors.py"

YOLOV1

preparing...

YOLOV2

Step1

pip install -r requirements.txt

Step2

modify the config.py-yolo2_options

• set mode -> train
• set weightfile -> darknet19_448.conv.23
• set clsnamesfile -> coco.names, voc.names, etc.
• set trainSet, testSet, cfgfile, gpus, ngpus, etc.

Step3

run "python3 train.py --version yolo2"

YOLOV3

Step1

pip install -r requirements.txt

Step2

modify the config.py-yolo3_options

• set mode -> train
• set weightfile -> darknet53.conv.74
• set clsnamesfile -> coco.names, voc.names, etc.
• set trainSet, testSet, cfgfile, gpus, ngpus, etc.

Step3

run "python3 train.py --version yolo3"

Test

YOLOV1

preparing

YOLOV2

Step1

modify the config.py-yolo2_options

• set mode - test
• set weightfile -> yolov2.weights
• set clsnamesfile -> coco.names, voc.names, etc.

Step2

run "python3 detector.py --version yolo2"

YOLOV3

Step1

modify the config.py-yolo3_options

• set mode - test
• set weightfile -> yolov3.weights
• set clsnamesfile -> coco.names, voc.names, etc.

Step2

run "python3 detector.py --version yolo3"

Eval

preparing

Reference

• marvis
• yolov1
• yolov2
• yolov3