Darknet-cpp

Darknet-cpp project is a bug-fixed and C++ compilable version of darknet (including Yolov3 and v2), an open source neural network framework written in C and CUDA. Darknet-cpp builds on Linux, Windows and also tested on Mac by users.

Prebuilt binaries for evaluation now available

• Prebuilt binaries are provided for evaluation purposes.

  • Tegra TK1 (CPU, CUDA, CUDA + CUDNN) - Yolov2 only
  • Jetson Nano - CUDA + CUDNN - Yolov3 - tiny only
  • Windows x64 (CUDA + CUDNN) - Yolov3
  • Linux x64 (CUDA + CUDNN) - Yolov3
  • Darwin Mac x64 (CUDA + CUDNN) - Yolov3

• Download the binaries from yolo-bins

  • https://github.com/prabindh/yolo-bins

Features of darknet-cpp

• Uses same source 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 Visual Studio 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, and a test wrapper that can read images or video files, and show detected regions in a complete C++ application.

• darknet-cpp supports OpenCV3 and OpenCV4. Tested on Ubuntu 16.04, 18.04 and windows, with CUDA 8.x to 10.x

• Note: darknet-cpp requires a C++11 compiler for darknet-cpp, and arapaho builds.

• Note: CMake wrapper for Linux is available at - https://gist.github.com/prabindh/6bb3c429b8e8347edcf9b5e0b2f54a20 This uses existing Makefile and is just a wrapper.

Usage

Using the Makefile in the root directory of the darknet source repository,

• make darknet - only darknet (original code), with OPENCV=0
• make darknet-cpp - only the CPP version, with OPENCV=1
• make darknet-cpp-shared - build the shared-lib version (without darknet.c calling wrapper), OPENCV=1
• make arapaho - build arapaho and its test wrapper (from within arapaho folder)

Steps to test (Yolov3)

After performing `make darknet-cpp`, the executable `darknet-cpp` is generated. Run the below command to recognise the provided data objects.

./darknet-cpp detect cfg/yolov3.cfg yolov3.weights data/dog.jpg
layer     filters    size              input                output
0 conv     32  3 x 3 / 1   256 x 256 x   3   ->   256 x 256 x  32  0.113 BFLOPs
1 conv     64  3 x 3 / 2   256 x 256 x  32   ->   128 x 128 x  64  0.604 BFLOPs
2 conv     32  1 x 1 / 1   128 x 128 x  64   ->   128 x 128 x  32  0.067 BFLOPs
3 conv     64  3 x 3 / 1   128 x 128 x  32   ->   128 x 128 x  64  0.604 BFLOPs
4 res    1                 128 x 128 x  64   ->   128 x 128 x  64
5 conv    128  3 x 3 / 2   128 x 128 x  64   ->    64 x  64 x 128  0.604 BFLOPs
6 conv     64  1 x 1 / 1    64 x  64 x 128   ->    64 x  64 x  64  0.067 BFLOPs
7 conv    128  3 x 3 / 1    64 x  64 x  64   ->    64 x  64 x 128  0.604 BFLOPs
8 res    5                  64 x  64 x 128   ->    64 x  64 x 128
9 conv     64  1 x 1 / 1    64 x  64 x 128   ->    64 x  64 x  64  0.067 BFLOPs
10 conv    128  3 x 3 / 1    64 x  64 x  64   ->    64 x  64 x 128  0.604 BFLOPs
11 res    8                  64 x  64 x 128   ->    64 x  64 x 128
12 conv    256  3 x 3 / 2    64 x  64 x 128   ->    32 x  32 x 256  0.604 BFLOPs
13 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
14 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
15 res   12                  32 x  32 x 256   ->    32 x  32 x 256
16 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
17 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
18 res   15                  32 x  32 x 256   ->    32 x  32 x 256
19 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
20 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
21 res   18                  32 x  32 x 256   ->    32 x  32 x 256
22 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
23 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
24 res   21                  32 x  32 x 256   ->    32 x  32 x 256
25 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
26 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
27 res   24                  32 x  32 x 256   ->    32 x  32 x 256
28 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
29 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
30 res   27                  32 x  32 x 256   ->    32 x  32 x 256
31 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
32 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
33 res   30                  32 x  32 x 256   ->    32 x  32 x 256
34 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
35 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
36 res   33                  32 x  32 x 256   ->    32 x  32 x 256
37 conv    512  3 x 3 / 2    32 x  32 x 256   ->    16 x  16 x 512  0.604 BFLOPs
38 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
39 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
40 res   37                  16 x  16 x 512   ->    16 x  16 x 512
41 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
42 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
43 res   40                  16 x  16 x 512   ->    16 x  16 x 512
44 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
45 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
46 res   43                  16 x  16 x 512   ->    16 x  16 x 512
47 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
48 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
49 res   46                  16 x  16 x 512   ->    16 x  16 x 512
50 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
51 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
52 res   49                  16 x  16 x 512   ->    16 x  16 x 512
53 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
54 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
55 res   52                  16 x  16 x 512   ->    16 x  16 x 512
56 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
57 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
58 res   55                  16 x  16 x 512   ->    16 x  16 x 512
59 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
60 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
61 res   58                  16 x  16 x 512   ->    16 x  16 x 512
62 conv   1024  3 x 3 / 2    16 x  16 x 512   ->     8 x   8 x1024  0.604 BFLOPs
63 conv    512  1 x 1 / 1     8 x   8 x1024   ->     8 x   8 x 512  0.067 BFLOPs
64 conv   1024  3 x 3 / 1     8 x   8 x 512   ->     8 x   8 x1024  0.604 BFLOPs
65 res   62                   8 x   8 x1024   ->     8 x   8 x1024
66 conv    512  1 x 1 / 1     8 x   8 x1024   ->     8 x   8 x 512  0.067 BFLOPs
67 conv   1024  3 x 3 / 1     8 x   8 x 512   ->     8 x   8 x1024  0.604 BFLOPs
68 res   65                   8 x   8 x1024   ->     8 x   8 x1024
69 conv    512  1 x 1 / 1     8 x   8 x1024   ->     8 x   8 x 512  0.067 BFLOPs
70 conv   1024  3 x 3 / 1     8 x   8 x 512   ->     8 x   8 x1024  0.604 BFLOPs
71 res   68                   8 x   8 x1024   ->     8 x   8 x1024
72 conv    512  1 x 1 / 1     8 x   8 x1024   ->     8 x   8 x 512  0.067 BFLOPs
73 conv   1024  3 x 3 / 1     8 x   8 x 512   ->     8 x   8 x1024  0.604 BFLOPs
74 res   71                   8 x   8 x1024   ->     8 x   8 x1024
75 conv    512  1 x 1 / 1     8 x   8 x1024   ->     8 x   8 x 512  0.067 BFLOPs
76 conv   1024  3 x 3 / 1     8 x   8 x 512   ->     8 x   8 x1024  0.604 BFLOPs
77 conv    512  1 x 1 / 1     8 x   8 x1024   ->     8 x   8 x 512  0.067 BFLOPs
78 conv   1024  3 x 3 / 1     8 x   8 x 512   ->     8 x   8 x1024  0.604 BFLOPs
79 conv    512  1 x 1 / 1     8 x   8 x1024   ->     8 x   8 x 512  0.067 BFLOPs
80 conv   1024  3 x 3 / 1     8 x   8 x 512   ->     8 x   8 x1024  0.604 BFLOPs
81 conv    255  1 x 1 / 1     8 x   8 x1024   ->     8 x   8 x 255  0.033 BFLOPs
82 detection
83 route  79
84 conv    256  1 x 1 / 1     8 x   8 x 512   ->     8 x   8 x 256  0.017 BFLOPs
85 upsample            2x     8 x   8 x 256   ->    16 x  16 x 256
86 route  85 61
87 conv    256  1 x 1 / 1    16 x  16 x 768   ->    16 x  16 x 256  0.101 BFLOPs
88 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
89 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
90 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
91 conv    256  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 256  0.067 BFLOPs
92 conv    512  3 x 3 / 1    16 x  16 x 256   ->    16 x  16 x 512  0.604 BFLOPs
93 conv    255  1 x 1 / 1    16 x  16 x 512   ->    16 x  16 x 255  0.067 BFLOPs
94 detection
95 route  91
96 conv    128  1 x 1 / 1    16 x  16 x 256   ->    16 x  16 x 128  0.017 BFLOPs
97 upsample            2x    16 x  16 x 128   ->    32 x  32 x 128
98 route  97 36
99 conv    128  1 x 1 / 1    32 x  32 x 384   ->    32 x  32 x 128  0.101 BFLOPs
100 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
101 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
102 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
103 conv    128  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 128  0.067 BFLOPs
104 conv    256  3 x 3 / 1    32 x  32 x 128   ->    32 x  32 x 256  0.604 BFLOPs
105 conv    255  1 x 1 / 1    32 x  32 x 256   ->    32 x  32 x 255  0.134 BFLOPs
106 detection
Loading weights from yolov3.weights...Done!
data/dog.jpg: Predicted in 0.066842 seconds.
dog: 100%
truck: 51%
car: 77%
bicycle: 74%

Yolov3-tiny with Jetson-Nano (Maxwell 128GPU)

In Makefile, update this line

NVCC=/usr/local/cuda/bin/nvcc

./darknet-cpp detect cfg/yolov3-tiny.cfg ../yolov3-tiny.weights data/dog.jpg layer filters size input output 0 conv 16 3 x 3 / 1 416 x 416 x 3 -> 416 x 416 x 16 0.150 BFLOPs 1 max 2 x 2 / 2 416 x 416 x 16 -> 208 x 208 x 16 2 conv 32 3 x 3 / 1 208 x 208 x 16 -> 208 x 208 x 32 0.399 BFLOPs 3 max 2 x 2 / 2 208 x 208 x 32 -> 104 x 104 x 32 4 conv 64 3 x 3 / 1 104 x 104 x 32 -> 104 x 104 x 64 0.399 BFLOPs 5 max 2 x 2 / 2 104 x 104 x 64 -> 52 x 52 x 64 6 conv 128 3 x 3 / 1 52 x 52 x 64 -> 52 x 52 x 128 0.399 BFLOPs 7 max 2 x 2 / 2 52 x 52 x 128 -> 26 x 26 x 128 8 conv 256 3 x 3 / 1 26 x 26 x 128 -> 26 x 26 x 256 0.399 BFLOPs 9 max 2 x 2 / 2 26 x 26 x 256 -> 13 x 13 x 256 10 conv 512 3 x 3 / 1 13 x 13 x 256 -> 13 x 13 x 512 0.399 BFLOPs 11 max 2 x 2 / 1 13 x 13 x 512 -> 13 x 13 x 512 12 conv 1024 3 x 3 / 1 13 x 13 x 512 -> 13 x 13 x1024 1.595 BFLOPs 13 conv 256 1 x 1 / 1 13 x 13 x1024 -> 13 x 13 x 256 0.089 BFLOPs 14 conv 512 3 x 3 / 1 13 x 13 x 256 -> 13 x 13 x 512 0.399 BFLOPs 15 conv 255 1 x 1 / 1 13 x 13 x 512 -> 13 x 13 x 255 0.044 BFLOPs 16 yolo 17 route 13 18 conv 128 1 x 1 / 1 13 x 13 x 256 -> 13 x 13 x 128 0.011 BFLOPs 19 upsample 2x 13 x 13 x 128 -> 26 x 26 x 128 20 route 19 8 21 conv 256 3 x 3 / 1 26 x 26 x 384 -> 26 x 26 x 256 1.196 BFLOPs 22 conv 255 1 x 1 / 1 26 x 26 x 256 -> 26 x 26 x 255 0.088 BFLOPs 23 yolo Loading weights from ../yolov3-tiny.weights...Done! data/dog.jpg: Predicted in 0.163809 seconds. dog: 56% car: 52% truck: 56% car: 62% bicycle: 58%

Steps to train (Yolov2, tag v5.3)**

This section applies to git tag v5.3 and earlier. ie, 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 (train.txt) that will be needed in next step of training.

• 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), where 4 is '#of coords', and 5 is 'num' in the cfg file.
  • 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.

• Testing with Arapaho C++ API for detection

  Arapaho needs the darknet-cpp shared library (.so file on Linux, .dll on Windows). This can be built as below on Linux.

  make darknet-cpp-shared

  On Windows, the .dll is built by default. The windows build requires the .sln files from (https://github.com/prabindh/darknet-cpp-windows)

  Refer the file https://github.com/prabindh/darknet/blob/master/arapaho/arapaho_readme.txt for more details on running Arapaho.

How to file issues

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)

Darknet-cpp for Windows

Currently tested with VS2017, CUDA10.1 on Win10 upto RS5.

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

Darknet

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.