DarkTorch

Designed as a drop in replacement for the YOLO Darknet framework written in idiomatic PyTorch.

This project was written both as a learning experience and to make hacking on YOLO easier than the alternative C framework.

Good luck!

Supports YOLOv2 with YOLOv3 support coming soon!

Setup

Use conda for easiest setup

conda env create -f environment.yml
conda activate darktorch

Required packages if not using conda:

• pytorch
• torchvision
• opencv-python

Optional packages:

• visdom
• matplotlib
• pytest

Detection

Detection with COCO

# Download weights
wget https://pjreddie.com/media/files/yolov2.weights

# Run detect script
python3 detect.py --weights=yolov2.weights --cfg=cfg/yolov2.cfg --image=data/dog.jpg

Detection with VOC

# Downloads weights
wget https://pjreddie.com/media/files/yolov3-voc.weights

# Or for CPU use VOC-TINY
wget https://pjreddie.com/media/files/yolov3-tiny.weights

# Webcam demo

## CPU
python3 webcam_demo.py --no-cuda

## IF you have Nvidia CUDA
python3 webcam_demo.py



# Run detect script
python3 detect.py --weights=yolov2-voc.weights --cfg=cfg/yolov2-voc.cfg --image=data/dog.jpg

Training

Training on VOC

# Download VOC dataset
cd data
cp ../scripts/get_voc_dataset.sh ./
bash get_voc_dataset.sh

# Label VOC dataset
cp ../../scripts/voc_label.py
python3 voc_label.py
cd ../..

python3 train.py

Training on COCO

# Download COCO dataset
cd data
cp ../scripts/get_coco_dataset.sh ./
bash get_coco_dataset.sh

cd ..

python3 train.py --cfg=cfg/yolov2-coco.cfg --weights=darknet19_448.conv.23

Training arguments

• --no-cuda
• --num-workers
• --clipping-norm
• --cfg
• --data
• --weights
• --no-shuffle
• --non-random
• --fintune
• --once

Running the test suite

Download both the COCO and VOC datasets

# From the root directory run:
cd data
cp ../scripts/get_voc_dataset.sh ./
cp ../scripts/get_coco_dataset.sh ./
bash get_voc_dataset.sh
cd ..
bash get_coco_dataset.sh
cd ..

# Run the label VOC script
cd voc
cp ../../scripts/voc_label.py
python3 voc_label.py
cd ../..

# Run the test suite
python3 -m pytest

Citations

@misc{darknet13,
  author =   {Joseph Redmon},
  title =    {Darknet: Open Source Neural Networks in C},
  howpublished = {\url{http://pjreddie.com/darknet/}},
  year = {2013--2016}
}