Return to Deep Learning Network model development history of classic catalog
Previous: depth articles - classic History of Deep Learning Network model (eight) elaborate structure and features DenseNet
In this section, the structure and characteristics DarkNet elaborate and accuracy of the models and
II. Network Classic (Classic Network)
9. DarkNet
. (1) compared to tensorflow it, darknet is not so exaggerated, but it has become a darknet advantages:
. ① darknet entirely by the C language (C frame for open neural network cuda and written), and there is no dependency, OpenCV course be used, but at least with its image is displayed for a better visualization;
②. Darknet support cpu (no gpu so it does not matter, of course, with cuda / cudnn use gpu course faster and better).
③. Darknet faster installation speed, easy to install.
④. Darknet is a lightweight architecture, not like tensorflow is so powerful API, it is more flexible and suitable for studying the underlying can be more convenient for them to improve and extend from the bottom.
⑤. There is realization and implementation of the darknet caffe similar places, familiar with the darknet, help to get started caffe.
. (2) darknet architecture:
①. Cfg folder is the architecture of some models, each cfg file similar to the prototxt caffe files that define the structure of the entire model through the file.
②. Data folder label placed some documents, such as COCO 9k category name, etc., and when (the main folder for the presentation charts, or direct training and other COCO corresponding data sets useful for some sample chart, if you use your own the data for training, something that folder is not required).
③. Src folder full bottom frame definition file defines all layers of all other basic functions in the folder, the folder is to be understood that the source of the frame.
④. Examples folder are some of the more high-level functions, such as checking function, identification and other functions, these functions are called directly as a function of the underlying.
⑤. Include file, the name suggests, a place to store the header file
⑥. Python folder using python to invoke methods model, also you need to use the dynamic library libdarknet.so darknet
⑦. Scripts folder scripts, such as downloading COCO data sets, data sets voc format conversion script to train the required format and so on.
⑧. In addition to the license file, and the rest is the Makefile.
. (3) darknet19 network description:
. (4) darknet53 network description:
. (5) YOLO-V3 darknet53 model network description:
10. The accuracy of each model and 
Return to Deep Learning Network model development history of classic catalog
Previous: depth articles - classic History of Deep Learning Network model (eight) elaborate structure and features DenseNet
