Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Serialized data format

Jump to bottom
Alex Taranov edited this page Aug 15, 2016 · 6 revisions
  • tiny-cnn saves networks weights in plain text format
  • all weights of layers are concatnated from input to output and printed as array of double

Opencv expansion (tiny-dnn version >= 0.1.1)

If you use tiny-cnn with Opencv, than another serialization way existed. To save and load network weights we can use cv::FileStorage. This approach makes possible to save network weights (and, if you add some code, all meta information about network architecture and layers) in *.xml or *.yml files.

To save weights and biases use:

template<typename T>
void saveCNNWeights(const char *filename, tiny_cnn::network<T> &_net) const
{
    cv::FileStorage fs(filename, cv::FileStorage::WRITE);
    if (fs.isOpened()) {        
    std::vector<tiny_cnn::float_t> _weights;
    std::vector<tiny_cnn::vec_t*> _w;
    for(size_t i = 0; i < m_net.depth(); i++) {
        // From tiny-cnn docs: number of elements differs by layer types and settings.
        // For example, in fully-connected layer with bias term,
        // weights[0] represents weight matrix and weights[1] represents bias vector.
        _w = m_net[i]->get_weights();
        tiny_cnn::vec_t *_v;
        for(size_t j = 0; j < _w.size(); j++) {
            _v = _w[j];
            _weights.insert(_weights.end(), _v->begin(), _v->end());
        }
    }
    fs << "weights" << _weights;
    fs.release();
    } else {
        CV_Error(Error::StsError, "File can't be opened for writing!");
    }
}

To load weights and biases:

template<typename T>
void CNNFaceRecognizer::load(const char *filename, tiny_cnn::network<T> &_net)
{
    cv::FileStorage fs(filename, cv::FileStorage::READ);
    if (fs.isOpened()) { 
    std::vector<tiny_cnn::float_t> _weights;
    fs["weights"] >> _weights;    
    int idx = 0;
    std::vector<tiny_cnn::vec_t*> _w;
    for(size_t i = 0; i < m_net.depth(); i++) {
        _w = m_net[i]->get_weights();
        tiny_cnn::vec_t *_v;
        for(size_t j = 0; j < _w.size(); j++) {
            _v = _w[j];
            for(size_t k = 0; k < _v->size(); k++)
                _v->at(k) = _weights[idx++];
        }
    }
    fs.release();
    } else {
        CV_Error(Error::StsError, "File can't be opened for reading!");
    }
}

Do not forget however, that before loading the network you should create it.

© Copyright 2018, tiny-dnn team

Clone this wiki locally