float CaffeFeatExtractor<Dtype>::extract_singleFeat(cv::Mat &image, Blob<Dtype> *features)
{
// Set the GPU/CPU mode for Caffe (here in order to be thread-safe)
if (gpu_mode)
{
Caffe::set_mode(Caffe::GPU);
Caffe::SetDevice(device_id);
}
else
{
Caffe::set_mode(Caffe::CPU);
}
cudaEvent_t start, stop;
if (timing)
{
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start, NULL);
}
// Initialize label to zero
int label = 0;
// Get pointer to data layer to set the input
caffe::shared_ptr<MemoryDataLayer<Dtype> > memory_data_layer = boost::dynamic_pointer_cast<caffe::MemoryDataLayer<Dtype> >(feature_extraction_net->layers()[0]);
// Set batch size to 1
if (memory_data_layer->batch_size()!=1)
{
memory_data_layer->set_batch_size(1);
cout << "BATCH SIZE = " << memory_data_layer->batch_size() << endl;
}
// Input preprocessing
// The image passed to AddMatVector must be same size as the mean image
// If not, it is resized:
// if it is downsampled, an anti-aliasing Gaussian Filter is applied
if (image.rows != mean_height || image.cols != mean_height)
{
if (image.rows > mean_height || image.cols > mean_height)
{
cv::resize(image, image, cv::Size(mean_height, mean_width), 0, 0, CV_INTER_LANCZOS4);
}
else
{
cv::resize(image, image, cv::Size(mean_height, mean_width), 0, 0, CV_INTER_LINEAR);
}
}
memory_data_layer->AddMatVector(vector<cv::Mat>(1, image),vector<int>(1,label));
size_t num_features = blob_names.size();
if(num_features!=1)
{
cout<< "Error! The list of features to be extracted has not size one!" << endl;
return -1;
}
// Run network and retrieve features!
// depending on your net's architecture, the blobs will hold accuracy and/or labels, etc
std::vector<Blob<Dtype>*> results = feature_extraction_net->Forward();
const caffe::shared_ptr<Blob<Dtype> > feature_blob = feature_extraction_net->blob_by_name(blob_names[0]);
int batch_size = feature_blob->num(); // should be 1
if (batch_size!=1)
{
cout << "Error! Retrieved more than one feature, exiting..." << endl;
return -1;
}
int channels = feature_blob->channels();
int width = feature_blob->width();
int height = feature_blob->height();
if (features==NULL)
{
features = new Blob<Dtype>(1, channels, height, width);
} else
{
features->Reshape(1, channels, height, width);
}
features->CopyFrom(*feature_blob);
if (timing)
{
// Record the stop event
cudaEventRecord(stop, NULL);
// Wait for the stop event to complete
cudaEventSynchronize(stop);
float msecTotal = 0.0f;
cudaEventElapsedTime(&msecTotal, start, stop);
return msecTotal;
}
else
{
return 0;
}
}
float CaffeFeatExtractor<Dtype>::extractBatch_singleFeat(vector<cv::Mat> &images, int new_batch_size, vector< Blob<Dtype>* > &features) {
// Set the GPU/CPU mode for Caffe (here in order to be thread-safe)
if (gpu_mode)
{
Caffe::set_mode(Caffe::GPU);
Caffe::SetDevice(device_id);
}
else
{
Caffe::set_mode(Caffe::CPU);
}
cudaEvent_t start, stop;
if (timing)
{
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start, NULL);
}
// Initialize the labels to zero
vector<int> labels(images.size(), 0);
// Get pointer to data layer to set the input
caffe::shared_ptr<MemoryDataLayer<Dtype> > memory_data_layer = boost::dynamic_pointer_cast<caffe::MemoryDataLayer<Dtype> >(feature_extraction_net->layers()[0]);
// Set batch size
if (memory_data_layer->batch_size()!=new_batch_size)
{
if (images.size()%new_batch_size==0)
{
memory_data_layer->set_batch_size(new_batch_size);
cout << "BATCH SIZE = " << memory_data_layer->batch_size() << endl;
}
else
{
if (images.size()%memory_data_layer->batch_size()==0)
{
cout << "WARNING: image number is not multiple of requested batch size, leaving the old one." << endl;
cout << "BATCH SIZE = " << memory_data_layer->batch_size() << endl;
} else
{
cout << "WARNING: image number is not multiple of batch size, setting it to 1 (performance issue)." << endl;
memory_data_layer->set_batch_size(1);
cout << "BATCH SIZE = " << memory_data_layer->batch_size() << endl;
}
}
} else
{
if (images.size()%memory_data_layer->batch_size()!=0)
{
cout << "WARNING: image number is not multiple of batch size, setting it to 1 (performance issue)." << endl;
memory_data_layer->set_batch_size(1);
cout << "BATCH SIZE = " << memory_data_layer->batch_size() << endl;
}
}
int num_batches = images.size()/new_batch_size;
// Input preprocessing
// The image passed to AddMatVector must be same size as the mean image
// If not, it is resized:
// if it is downsampled, an anti-aliasing Gaussian Filter is applied
for (int i=0; i<images.size(); i++)
{
if (images[i].rows != mean_height || images[i].cols != mean_height)
{
if (images[i].rows > mean_height || images[i].cols > mean_height)
{
cv::resize(images[i], images[i], cv::Size(mean_height, mean_width), 0, 0, CV_INTER_LANCZOS4);
}
else
{
cv::resize(images[i], images[i], cv::Size(mean_height, mean_width), 0, 0, CV_INTER_LINEAR);
}
}
}
memory_data_layer->AddMatVector(images,labels);
size_t num_features = blob_names.size();
if (num_features!=1)
{
cout<< "Error! The list of features to be extracted has not size one!" << endl;
return -1;
}
// Run network and retrieve features!
// depending on your net's architecture, the blobs will hold accuracy and/or labels, etc
std::vector<Blob<Dtype>*> results;
for (int b=0; b<num_batches; b++)
{
results = feature_extraction_net->Forward();
const caffe::shared_ptr<Blob<Dtype> > feature_blob = feature_extraction_net->blob_by_name(blob_names[0]);
int batch_size = feature_blob->num();
int channels = feature_blob->channels();
int width = feature_blob->width();
int height = feature_blob->height();
features.push_back(new Blob<Dtype>(batch_size, channels, height, width));
features.back()->CopyFrom(*feature_blob);
}
if (timing)
{
// Record the stop event
cudaEventRecord(stop, NULL);
// Wait for the stop event to complete
cudaEventSynchronize(stop);
float msecTotal = 0.0f;
cudaEventElapsedTime(&msecTotal, start, stop);
float msecPerImage = msecTotal/(float)images.size();
return msecPerImage;
}
else
{
return 0;
}
}