#include <caffe/caffe.hpp>

#include <opencv2/core/core.hpp>

#include <opencv2/highgui/highgui.hpp>

#include <opencv2/imgproc/imgproc.hpp>

#include <iosfwd>

#include <memory>

#include <string>

#include <utility>

#include <vector>

using namespace caffe;  // NOLINT(build/namespaces)

using std::string;

/* Pair (label, confidence) representing a prediction. */

typedef std::pair<string, float> Prediction;

class Classifier {

public:

    Classifier(const string& model_file,

               const string& trained_file,

               const string& mean_file,

               const string& label_file);

    std::vector<Prediction> Classify(const cv::Mat& img, int N = 5);

private:

    void SetMean(const string& mean_file);

    std::vector<float> Predict(const cv::Mat& img);

    void WrapInputLayer(std::vector<cv::Mat>* input_channels);

    void Preprocess(const cv::Mat& img,

                    std::vector<cv::Mat>* input_channels);

private:

    shared_ptr<Net<float> > net_;

    cv::Size input_geometry_;

    int num_channels_;

    cv::Mat mean_;

    std::vector<string> labels_;

};

Classifier::Classifier(const string& model_file,

                       const string& trained_file,

                       const string& mean_file,

                       const string& label_file) {

#ifdef CPU_ONLY

    Caffe::set_mode(Caffe::CPU);

#else

    Caffe::set_mode(Caffe::GPU);

#endif

    /* Load the network. */

    net_.reset(new Net<float>(model_file, TEST));

    net_->CopyTrainedLayersFrom(trained_file);

    CHECK_EQ(net_->num_inputs(), 1) << "Network should have exactly one input.";

    CHECK_EQ(net_->num_outputs(), 1) << "Network should have exactly one output.";

    Blob<float>* input_layer = net_->input_blobs()[0];

    num_channels_ = input_layer->channels();

    CHECK(num_channels_ == 3 || num_channels_ == 1)

    << "Input layer should have 1 or 3 channels.";

    input_geometry_ = cv::Size(input_layer->width(), input_layer->height());

    /* Load the binaryproto mean file. */

    SetMean(mean_file);

    /* Load labels. */

    std::ifstream labels(label_file.c_str());

    CHECK(labels) << "Unable to open labels file " << label_file;

    string line;

    while (std::getline(labels, line))

        labels_.push_back(string(line));

    Blob<float>* output_layer = net_->output_blobs()[0];

    CHECK_EQ(labels_.size(), output_layer->channels())

    << "Number of labels is different from the output layer dimension.";

}

static bool PairCompare(const std::pair<float, int>& lhs,

                        const std::pair<float, int>& rhs) {

    return lhs.first > rhs.first;

}

/* Return the indices of the top N values of vector v. */

static std::vector<int> Argmax(const std::vector<float>& v, int N) {

    std::vector<std::pair<float, int> > pairs;

    for (size_t i = 0; i < v.size(); ++i)

        pairs.push_back(std::make_pair(v[i], i));

    std::partial_sort(pairs.begin(), pairs.begin() + N, pairs.end(), PairCompare);

    std::vector<int> result;

    for (int i = 0; i < N; ++i)

        result.push_back(pairs[i].second);

    return result;

}

/* Return the top N predictions. */

std::vector<Prediction> Classifier::Classify(const cv::Mat& img, int N) {

    std::vector<float> output = Predict(img);

    std::vector<int> maxN = Argmax(output, N);

    std::vector<Prediction> predictions;

    for (int i = 0; i < N; ++i) {

        int idx = maxN[i];

        predictions.push_back(std::make_pair(labels_[idx], output[idx]));

    }

    return predictions;

}

/* Load the mean file in binaryproto format. */

void Classifier::SetMean(const string& mean_file) {

    BlobProto blob_proto;

    ReadProtoFromBinaryFileOrDie(mean_file.c_str(), &blob_proto);

    /* Convert from BlobProto to Blob<float> */

    Blob<float> mean_blob;

    mean_blob.FromProto(blob_proto);

    CHECK_EQ(mean_blob.channels(), num_channels_)

    << "Number of channels of mean file doesn‘t match input layer.";

    /* The format of the mean file is planar 32-bit float BGR or grayscale. */

    std::vector<cv::Mat> channels;

    float* data = http://www.mamicode.com/mean_blob.mutable_cpu_data();

    for (int i = 0; i < num_channels_; ++i) {

        /* Extract an individual channel. */

        cv::Mat channel(mean_blob.height(), mean_blob.width(), CV_32FC1, data);

        channels.push_back(channel);

        data += mean_blob.height() * mean_blob.width();

    }

    /* Merge the separate channels into a single image. */

    cv::Mat mean;

    cv::merge(channels, mean);

    /* Compute the global mean pixel value and create a mean image

     * filled with this value. */

    cv::Scalar channel_mean = cv::mean(mean);

    mean_ = cv::Mat(input_geometry_, mean.type(), channel_mean);

}

std::vector<float> Classifier::Predict(const cv::Mat& img) {

    Blob<float>* input_layer = net_->input_blobs()[0];

    input_layer->Reshape(1, num_channels_,

                         input_geometry_.height, input_geometry_.width);

    /* Forward dimension change to all layers. */

    net_->Reshape();

    std::vector<cv::Mat> input_channels;

    WrapInputLayer(&input_channels);

    Preprocess(img, &input_channels);

    net_->ForwardPrefilled();

    /* Copy the output layer to a std::vector */

    Blob<float>* output_layer = net_->output_blobs()[0];

    const float* begin = output_layer->cpu_data();

    const float* end = begin + output_layer->channels();

    return std::vector<float>(begin, end);

}

/* Wrap the input layer of the network in separate cv::Mat objects

 * (one per channel). This way we save one memcpy operation and we

 * don‘t need to rely on cudaMemcpy2D. The last preprocessing

 * operation will write the separate channels directly to the input

 * layer. */

void Classifier::WrapInputLayer(std::vector<cv::Mat>* input_channels) {

    Blob<float>* input_layer = net_->input_blobs()[0];

    int width = input_layer->width();

    int height = input_layer->height();

    float* input_data = http://www.mamicode.com/input_layer->mutable_cpu_data();

    for (int i = 0; i < input_layer->channels(); ++i) {

        cv::Mat channel(height, width, CV_32FC1, input_data);

        input_channels->push_back(channel);

        input_data += width * height;

    }

}

void Classifier::Preprocess(const cv::Mat& img,

                            std::vector<cv::Mat>* input_channels) {

    /* Convert the input image to the input image format of the network. */

    cv::Mat sample;

    if (img.channels() == 3 && num_channels_ == 1)

        cv::cvtColor(img, sample, CV_BGR2GRAY);

    else if (img.channels() == 4 && num_channels_ == 1)

        cv::cvtColor(img, sample, CV_BGRA2GRAY);

    else if (img.channels() == 4 && num_channels_ == 3)

        cv::cvtColor(img, sample, CV_BGRA2BGR);

    else if (img.channels() == 1 && num_channels_ == 3)

        cv::cvtColor(img, sample, CV_GRAY2BGR);

    else

        sample = img;

    cv::Mat sample_resized;

    if (sample.size() != input_geometry_)

        cv::resize(sample, sample_resized, input_geometry_);

    else

        sample_resized = sample;

    cv::Mat sample_float;

    if (num_channels_ == 3)

        sample_resized.convertTo(sample_float, CV_32FC3);

    else

        sample_resized.convertTo(sample_float, CV_32FC1);

    cv::Mat sample_normalized;

    cv::subtract(sample_float, mean_, sample_normalized);

    /* This operation will write the separate BGR planes directly to the

     * input layer of the network because it is wrapped by the cv::Mat

     * objects in input_channels. */

    cv::split(sample_normalized, *input_channels);

    CHECK(reinterpret_cast<float*>(input_channels->at(0).data)

          == net_->input_blobs()[0]->cpu_data())

    << "Input channels are not wrapping the input layer of the network.";

}

int main(int argc, char** argv) {

    if (argc != 6) {

        std::cerr << "Usage: " << argv[0]

        << " deploy.prototxt network.caffemodel"

        << " mean.binaryproto labels.txt img.jpg" << std::endl;

        return 1;

    }

    ::google::InitGoogleLogging(argv[0]);

    string model_file   = argv[1];

    string trained_file = argv[2];

    string mean_file    = argv[3];

    string label_file   = argv[4];

    Classifier classifier(model_file, trained_file, mean_file, label_file);

    string file = argv[5];

    std::cout << "---------- Prediction for "

    << file << " ----------" << std::endl;

    cv::Mat img = cv::imread(file, -1);

    CHECK(!img.empty()) << "Unable to decode image " << file;

    std::vector<Prediction> predictions = classifier.Classify(img);

    /* Print the top N predictions. */

    for (size_t i = 0; i < predictions.size(); ++i) {

        Prediction p = predictions[i];

        std::cout << std::fixed << std::setprecision(4) << p.second << " - \""

        << p.first << "\"" << std::endl;

    }

}

深度学习caffe测试代码