This project aims to provide a modern C++ implementation of the Darknet neural network framework, focusing on efficient training and inference with CUDA support.

• CUDA Support: Utilize the power of NVIDIA GPUs for accelerated neural network training and inference.
• Model Compatibility: Successfully tested with YOLOv1-tiny, YOLOv3-tiny, and YOLOv3 models and ResNet18.
• Performance: Significantly reduced inference time on CPU compared to the original Darknet implementation.
• Modern C++ Features: Leveraging modern C++ standards for improved maintainability.

• Boost (>= 1.74)
• CUDA Toolkit (for GPU support)
• CUDNN8 (for GPU support)
• Cairo (>= 1.16.0) (optional)
• GLib
• LibTIFF
• OpenBLAS library
• OpenCV (>= 4.5.4)
• Pango (optional)
• nlohmann_json (>= 3.10.5)
• yaml-cpp
• Protobuf (>= 3.12.4)

# Clone the repository
git clone https://github.com/trieck/pixienn

# Build the project
cd pixienn
mkdir build
cd build
cmake ..
make

In the bin directory, you will find the pixienn executable. You can use it to run inference on a single image. The following example shows how to run inference on a single image using the YOLOv3-tiny model:

1.) Download the trained darknet weights file and place it in the resources/weights directory below the project root.

./pixienn ../../resources/cfg/yolov3-tiny-cfg.yml ../../resources/images/dog.jpg

If all went well, this will produce an image named predictions.jpg and a JSON file named predictions.geojson in the current directory. The image will have bounding boxes drawn around the detected objects. The JSON file will contain the bounding box coordinates and class labels for each detected object. You should be able to see bounding boxes for a dog and a bicycle and truck in the image.

We will train a YOLOv3-tiny model on the COCO dataset. The easiest thing to do is to follow the instructions on the darknet site for downloading the COCO dataset and creating annotations for the training and validation sets.

You must configure the directory paths for the training and validation images and annotations in the cfg/yolov3-tiny-cfg.yml.

You may wish to edit parameters in the model file under resources/models/yolov3-tiny.yml.

Once you have configured the paths, you can train your model as follows:

./pixienn-train ../../resources/cfg/yolov3-tiny-cfg.yml yolov3-tiny.weights

PixieNN uses the Tensorboard tool for monitoring model training. You can start Tensorboard as follows:

tensorboard --logdir=.