search

schwittlick / ofxDarknet

darknet neural network addon for openFrameworks
MIT License
496 stars 89 forks source link
machine-learning-algorithms neural-networks

readme

ofxDarknet

ofxDarknet is a openFrameworks wrapper for 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. http://pjreddie.com/darknet/

Features

YOLO: Real-Time Object Detection (http://pjreddie.com/darknet/yolo/)

Darknet comes with two pre-trained models for this task. Additionally each has a smaller (and faster) but therefore less accurate version:

MS COCO dataset (80 different classes)

    std::string cfgfile = ofToDataPath( "cfg/tiny-yolo.cfg" );
    std::string weightfile = ofToDataPath( "tiny-yolo.weights" );
    std::string nameslist = ofToDataPath( "cfg/names.list" );
    darknet.init( cfgfile, weightfile, nameslist );

Pascal VOC dataset (20 different classes)

    std::string cfgfile = ofToDataPath( "cfg/tiny-yolo-voc.cfg" );
    std::string weightfile = ofToDataPath( "tiny-yolo-voc.weights" );
    std::string nameslist = ofToDataPath( "cfg/voc.names" );
    darknet.init( cfgfile, weightfile, nameslist );

YOLO2 with 9000 classes

    std::string datacfg = ofToDataPath( "cfg/combine9k.data" );
    std::string cfgfile = ofToDataPath( "cfg/yolo9000.cfg" );
    std::string weightfile = ofToDataPath( "yolo9000.weights" );
    darknet.init( cfgfile, weightfile, datacfg );
    float thresh = 0.25;
    std::vector< detected_object > detections = darknet.yolo( image.getPixelsRef(), thresh );

    for( detected_object d : detections )
    {
        ofSetColor( d.color );
        glLineWidth( ofMap( d.probability, 0, 1, 0, 8 ) );
        ofNoFill();
        ofDrawRectangle( d.rect );
        ofDrawBitmapStringHighlight( d.label + ": " + ofToString(d.probability), d.rect.x, d.rect.y + 20 );
    }

YOLO2

Imagenet Classification (http://pjreddie.com/darknet/imagenet/)

In order to classify an image with more classes, this is the spot. This classifies an image according to the 1000-class ImageNet Challenge.

    std::string cfgfile = ofToDataPath( "cfg/darknet.cfg" );
    std::string weightfile = ofToDataPath( "darknet.weights" );
    std::string nameslist = ofToDataPath( "cfg/imagenet.shortnames.list" );
    darknet.init( cfgfile, weightfile, nameslist );

    classifications = darknet.classify( image.getPixelsRef() );
    int offset = 20;
    for( classification c : classifications )
    {
        std::stringstream ss;
        ss << c.label << " : " << ofToString( c.probability );
        ofDrawBitmapStringHighlight( ss.str(), 20, offset );
        offset += 20;
    }

Classification

Deep Dream (http://pjreddie.com/darknet/nightmare/)

vgg-conv.cfg & vgg-conv.weights

    std::string cfgfile = ofToDataPath( "cfg/vgg-conv.cfg" );
    std::string weightfile = ofToDataPath( "vgg-conv.weights" );
    darknet.init( cfgfile, weightfile );

    int max_layer = 13;
    int range = 3;
    int norm = 1;
    int rounds = 4;
    int iters = 20;
    int octaves = 4;
    float rate = 0.01;
    float thresh = 1.0;
    nightmare = darknet.nightmate( image.getPixelsRef(), max_layer, range, norm, rounds, iters, octaves, rate, thresh );

DeepDream

Recurrent Neural Network (http://pjreddie.com/darknet/rnns-in-darknet/)

Darknet pre-trained weights files:

ofxDarknet custom pre-trained weight files (each trained for 20h on NVidia TitanX):

    std::string cfgfile = ofToDataPath( "cfg/rnn.cfg" );
    std::string weightfile = ofToDataPath( "shakespeare.weights" );
    darknet.init( cfgfile, weightfile );

    int character_count = 100;
    float temperature = 0.8;
    std::string seed_text = "openframeworks is ";
    std::string generated_text = darknet.rnn( character_count, seed_text, temperature );

RNN

You can train your own RNN models with darknet

    // no need to init
    darknet.train_rnn( ofToDataPath( "training_text.txt" ), "cfg/rnn.cfg" );

Go

Darknet has a policy network for Go. Read the original doc here.

In the example example-go is a 2-player game where darknet gives recommendations. To play, click on the square you wish to move a piece onto. More doc on this soon.

Go

Setup

Windows

Install the dependencies for building darknet on Windows 10:

There are some more necessary steps that don't work with the OF project generator:

OSX

First make sure to install CUDA 8.0 64bit (Driver & Toolkit). CUDA requires an NVIDIA graphics card and a reasonably recent Mac OS.

After that, projects should compile fine from the Project Generator. Make sure to download the necessary weights (links can be found here and include the required cfg files (found in the examples) in any app that opens them.

Building the library from source

If you want to make changes to the darknet lib, you can build it from source with cmake. cd into libs/darknet/cMake/ and then run:

cmake .
make

Note, you need to have CUDA and OpenCV installed on your system first.

Training your own models

YOLO

tcb

Credits

Reading