Skip to content

Commit

Permalink
Batched inference for cpp_classification
Browse files Browse the repository at this point in the history 
Example usage:
./build/examples/cpp_classification/classification.bin models/bvlc_reference_caffenet/deploy.prototxt \
models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel \
data/ilsvrc12/imagenet_mean.binaryproto \
data/ilsvrc12/synset_words.txt \
images/1.jpg images/2.jpg images/3.jpg images/4.jpg images/5.jpg
  • Loading branch information
@flx42
flx42 committed Feb 19, 2016
1 parent 0c89b38 commit be0bff1
Showing 1 changed file with 66 additions and 45 deletions.
111 changes: 66 additions & 45 deletions examples/cpp_classification/classification.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -25,14 +25,14 @@ class Classifier {
const string& mean_file,
const string& label_file);

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

private:
void SetMean(const string& mean_file);

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

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

void Preprocess(const cv::Mat& img,
std::vector<cv::Mat>* input_channels);
Expand Down Expand Up @@ -102,18 +102,24 @@ static std::vector<int> Argmax(const std::vector<float>& v, int N) {
}

/* Return the top N predictions. */
std::vector<Prediction> Classifier::Classify(const cv::Mat& img, int N) {
std::vector<float> output = Predict(img);

N = std::min<int>(labels_.size(), N);
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;
std::vector<std::vector<Prediction> > Classifier::Classify(const std::vector<cv::Mat>& imgs, int N) {
std::vector<std::vector<float> > outputs = Predict(imgs);

std::vector<std::vector<Prediction> > all_predictions;
for (int j = 0; j < outputs.size(); ++j) {
std::vector<float> output = outputs[j];

N = std::min<int>(labels_.size(), N);
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]));
}
all_predictions.push_back(predictions);
}

return all_predictions;
}

/* Load the mean file in binaryproto format. */
Expand Down Expand Up @@ -147,42 +153,48 @@ void Classifier::SetMean(const string& mean_file) {
mean_ = cv::Mat(input_geometry_, mean.type(), channel_mean);
}

std::vector<float> Classifier::Predict(const cv::Mat& img) {
std::vector<std::vector<float> > Classifier::Predict(const std::vector<cv::Mat>& imgs) {
Blob<float>* input_layer = net_->input_blobs()[0];
input_layer->Reshape(1, num_channels_,
input_layer->Reshape(imgs.size(), 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);

for (int i = 0; i < imgs.size(); ++i) {
std::vector<cv::Mat> input_channels;
WrapInputLayer(&input_channels, i);
Preprocess(imgs[i], &input_channels);
}
net_->ForwardPrefilled();

/* Copy the output layer to a std::vector */
std::vector<std::vector<float> > outputs;

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);
for (int i = 0; i < output_layer->num(); ++i) {
const float* begin = output_layer->cpu_data() + i * output_layer->channels();
const float* end = begin + output_layer->channels();
/* Copy the output layer to a std::vector */
outputs.push_back(std::vector<float>(begin, end));
}
return outputs;
}

/* 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) {
void Classifier::WrapInputLayer(std::vector<cv::Mat>* input_channels, int n) {
Blob<float>* input_layer = net_->input_blobs()[0];

int width = input_layer->width();
int height = input_layer->height();
float* input_data = 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;
int channels = input_layer->channels();
float* input_data = input_layer->mutable_cpu_data() + n * width * height * channels;
for (int i = 0; i < channels; ++i) {
cv::Mat channel(height, width, CV_32FC1, input_data);
input_channels->push_back(channel);
input_data += width * height;
}
}

Expand Down Expand Up @@ -221,13 +233,15 @@ void Classifier::Preprocess(const cv::Mat& img,
* 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) {
if (argc < 6) {
std::cerr << "Usage: " << argv[0]
<< " deploy.prototxt network.caffemodel"
<< " mean.binaryproto labels.txt img.jpg" << std::endl;
Expand All @@ -242,20 +256,27 @@ int main(int argc, char** argv) {
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);
std::vector<cv::Mat> imgs;
for (int i = 5; i < argc; ++i)
{
cv::Mat img = cv::imread(argv[i], -1);
CHECK(!img.empty()) << "Unable to decode image " << argv[i];
imgs.push_back(img);
}
std::vector<std::vector<Prediction> > all_predictions = classifier.Classify(imgs);

/* 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;
for (size_t i = 0; i < all_predictions.size(); ++i) {
std::cout << "---------- Prediction for "
<< argv[5 + i] << " ----------" << std::endl;

std::vector<Prediction>& predictions = all_predictions[i];
for (size_t j = 0; j < predictions.size(); ++j) {
Prediction p = predictions[j];
std::cout << std::fixed << std::setprecision(4) << p.second << " - \""
<< p.first << "\"" << std::endl;
}
std::cout << std::endl;
}
}
#else
Expand Down

2 comments on commit be0bff1

@pkrush
Copy link

@pkrush pkrush commented on be0bff1 Oct 1, 2016

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for posting this. I am going to be updating it a bit to handle LMDBs. I find this really strange this feature is not in Caffe already. Strange as I must be doing something wrong...

@mahaoyanghb
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks a lot. I implemented it, works well.
how do you find this method to speed up?
and what is main principle of this method? is it calculate all pictures on GPU simultaneously ?

Please sign in to comment.