Darknet-cpp project is a bug-fixed and C++ compilable version of darknet, an open source neural network framework written in C and CUDA. Darknet-cpp builds on Linux, Windows and also tested on Mac by users.
Features
-
Uses same code-base as original darknet (ie same .c files are used). Modification is done only for runtime bug-fixes, compile time fixes for c++, and the build system itself. For list of bugs fixed, refer to this thread - https://groups.google.com/forum/#!topic/darknet/4Hb159aZBbA, and https://github.com/prabindh/darknet/issues
-
The Linux build system supports 3 targets -
- original darknet (with gcc compiler),
- darknet-cpp (with g++ compiler), and
- Shared library (libdarknet-cpp-shared.so)
-
Can use bounding boxes directly from Euclid object labeller (https://github.com/prabindh/euclid)
-
C++ API - arapaho, that works in conjunction with libdarknet-cpp-shared.so
-
darknet-cpp supports OpenCV3. Tested on Ubuntu 16.04 and CUDA 8.x
Usage
Using the Makefile in the root directory of the darknet source repository,
make darknet
- only darknet (original code), with OPENCV=0make darknet-cpp
- only the CPP version, with OPENCV=1make darknet-cpp-shared
- build the shared-lib version (without darknet.c calling wrapper), OPENCV=1
Steps to train (Yolov2)
- Download latest commit of darknet-cpp, ex
git clone https://github.com/prabindh/darknet
- Create Yolo compatible training data-set. I use this to create Yolo compatible bounding box format file, and training list file.
https://github.com/prabindh/euclid
This creates a training list file that will be needed in next step.
-
Change the files per below:
- yolo-voc.cfg - change line classes=20 to suit desired number of classes
- yolo-voc.cfg - change the number of filters in the CONV layer above the region layer - (#classes + 4 + 1)*(5)
- voc.data - change line classes=20, and paths to training image list file
- voc.names - number of lines must be equal the number of classes
-
Place label-images corresponding to name of classes in data/labels, ex - data/labels/myclassname1.png
-
Download http://pjreddie.com/media/files/darknet19_448.conv.23
-
Train as below
./darknet-cpp detector train ./cfg/voc-myclasses.data ./cfg/yolo-myconfig.cfg darknet19_448.conv.23
- Atleast for the few initial iterations, observe the log output, and ensure all images are found and being used. After convergence, detection can be performed using standard steps.
If there is a need to report an issue with the darknet-cpp port, use the link - https://github.com/prabindh/darknet/issues.
Information required for filing an issue:
-
Output of
git log --format="%H" -n 1
-
Options enabled in Makefile (GPU,CUDNN)
-
If using Arapaho C++ wrapper, what options were used to build
-
Platform being used (OS version, GPU type, CUDA version, and OpenCV version)
Currently tested with VS2013, CUDA8.0 on Win10.
The solution file requires the below repository.
https://github.com/prabindh/darknet-cpp-windows
The Windows port does not require any additional downloads (like pthreads), and builds the same darknet code-base for Windows, to generate the darknet.dll. Building the Arapaho C++ API and test wrapper, creates arapaho.exe, that works exactly the same way as arapaho on Linux.
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.