Darknet is an open source neural network framework written in C and CUDA. It is fast, easy to install, and supports CPU and GPU computation.

For more information see the Darknet project website.

For questions or issues please use the Google Group.

The author documents quite well how to use YOLOv3 here https://pjreddie.com/darknet/yolo/. Darknet is his own equivalent to TensorFlow or Pytorch written completely in C with no dependencies (CUDA is optional). I’m amazed that (1) one person was able to do this and (2) the code is as simple and beautiful as it is. Some of the code is difficult to read, but the overall structure is very clear.

The common neural network operations each have their own .c file in the /src directory e.g. connected_layer.c and convolutional_layer.c. The general structure of the codebase is that the “library” part of Darknet that contains common neural network operations is in /src and the “applications” part of Darknet that contains different specific uses of the core library is in /examples. The different applications are the YOLO detector and the author’s other neural network projects.

1. The main() method for Darknet is in examples/darknet.c link

   a. main() contains the functions calls for each of the command line options explained in the YOLO website: detect, detector test, detector demo, and detector train.

   b. We will pursue detector train because it should execute the meat of the YOLO paper. If the main() function is passed the word detector it calls the run_detector(int argc, char **argv) function.

   c. run_detector(int argc, char **argv) is defined in examples/detector.c link This function calls many instances of find_char_arg(int argc, char **argv, char *arg, char *def) defined in src/utils.c. find_char_arg() searches command line arguments for expected char sequences such as -prefix and returns the argument values.

   d. After parsing command line arguments, run_detector() will call train_detector() if the train argument is given. The important arguments for this function are (1) the location of the data configuration file e.g. cfg/voc.data which contains training and validation labels and (2) the network configuration file e.g cfg/yolov3-voc.cfg which contains the definition of the network architecture.

2. The training operation for YOLOv3 is in train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, int ngpus, int clear) link

   a. read_data_cfg(char *filename) in src/option_list.c parses the data configuration file. Options read from this configuration file are stored in a linked list and parsed by option_find_str() defined in src/option_list.c.

   b. The following commonly used "classes" (there are no classes in C) appear in train_detector() and are defined in include/darknet.h: A network is a struct containing dozens of parameters such as learning rate and momentum. A matrix is a struct that contains height and width and a pointer to cell contents. A data object is a struct containing x and y matrices. A layer is a massive struct containing hundreds of parameters such as nbiases, nweights, and inputs.

   c. An array of networks is allocated using calloc() based on the number of GPUs available. For every available gpu, load_network() is called to convert the network definition in the configuration file into a network in memory. The core heavy lifting inside load_network() is network *parse_network_cfg(char *filename) in src/parser.c. It's worth looking at cfg/yolov3-voc.cfg to compare the network definition in that configuration file to the network definition in the YOLOv3 paper shown below: