I implemented a modified version of the Caffe C++ example and while it works really well, it's incredibly slow because it only accepts images one by one. Ideally I'd like to pass Caffe a vector of 200 images and return the best prediction for each one. I received some great help from Fanglin Wang and implemented some of his recommendations, but am still having some trouble working out how to retrieve the best result from each image.
The Classify method is now passed a vector of cv::Mat
objects (variable input_channels
) which is a vector of grayscale floating point images. I've eliminated the preprocessing method in the code because I don't need to convert these images to floating point or subtract the mean image. I've also been trying to get rid of the N
variable because I only want to return the top prediction and probability for each image.
#include "Classifier.h"
using namespace caffe;
using std::string;
Classifier::Classifier(const string& model_file, const string& trained_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);
Blob<float>* input_layer = net_->input_blobs()[0];
num_channels_ = input_layer->channels();
input_geometry_ = cv::Size(input_layer->width(), input_layer->height());
/* 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 std::vector<cv::Mat> &input_channels) {
std::vector<float> output = Predict(input_channels);
std::vector<int> maxN = Argmax(output, 1);
int idx = maxN[0];
predictions.push_back(std::make_pair(labels_[idx], output[idx]));
return predictions;
}
std::vector<float> Classifier::Predict(const std::vector<cv::Mat> &input_channels, int num_images) {
Blob<float>* input_layer = net_->input_blobs()[0];
input_layer->Reshape(num_images, num_channels_,
input_geometry_.height, input_geometry_.width);
/* Forward dimension change to all layers. */
net_->Reshape();
WrapInputLayer(&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 + num_images * 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 = input_layer->mutable_cpu_data();
for (int i = 0; i < input_layer->channels() * num_images; ++i) {
cv::Mat channel(height, width, CV_32FC1, input_data);
input_channels->push_back(channel);
input_data += width * height;
}
}
UPDATE
Thank-you so much for your help Shai, I made the changes you recommended but seem to be getting some strange compilation issues I can't work out (I managed to sort out a few of the issues).
These are the changes I made:
Header File:
#ifndef __CLASSIFIER_H__
#define __CLASSIFIER_H__
#include <caffe/caffe.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <algorithm>
#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& label_file);
std::vector< std::pair<int,float> > Classify(const std::vector<cv::Mat>& img);
private:
std::vector< std::vector<float> > Predict(const std::vector<cv::Mat>& img, int nImages);
void WrapInputLayer(std::vector<cv::Mat>* input_channels, int nImages);
void Preprocess(const std::vector<cv::Mat>& img,
std::vector<cv::Mat>* input_channels, int nImages);
private:
shared_ptr<Net<float> > net_;
cv::Size input_geometry_;
int num_channels_;
std::vector<string> labels_;
};
#endif /* __CLASSIFIER_H__ */
Class File:
#define CPU_ONLY
#include "Classifier.h"
using namespace caffe; // NOLINT(build/namespaces)
using std::string;
Classifier::Classifier(const string& model_file,
const string& trained_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 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;
}
std::vector< std::pair<int,float> > Classifier::Classify(const std::vector<cv::Mat>& img) {
std::vector< std::vector<float> > output = Predict(img, img.size());
std::vector< std::pair<int,float> > predictions;
for ( int i = 0 ; i < output.size(); i++ ) {
std::vector<int> maxN = Argmax(output[i], 1);
int idx = maxN[0];
predictions.push_back(std::make_pair(labels_[idx], output[idx]));
}
return predictions;
}
std::vector< std::vector<float> > Classifier::Predict(const std::vector<cv::Mat>& img, int nImages) {
Blob<float>* input_layer = net_->input_blobs()[0];
input_layer->Reshape(nImages, 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, nImages);
Preprocess(img, &input_channels, nImages);
net_->ForwardPrefilled();
/* Copy the output layer to a std::vector */
Blob<float>* output_layer = net_->output_blobs()[0];
std::vector <std::vector<float> > ret;
for (int i = 0; i < nImages; i++) {
const float* begin = output_layer->cpu_data() + i*output_layer->channels();
const float* end = begin + output_layer->channels();
ret.push_back( std::vector<float>(begin, end) );
}
return ret;
}
/* 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, int nImages) {
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()* nImages; ++i) {
cv::Mat channel(height, width, CV_32FC1, input_data);
input_channels->push_back(channel);
input_data += width * height;
}
}
void Classifier::Preprocess(const std::vector<cv::Mat>& img,
std::vector<cv::Mat>* input_channels, int nImages) {
for (int i = 0; i < nImages; i++) {
vector<cv::Mat> channels;
cv::split(img[i], channels);
for (int j = 0; j < channels.size(); j++){
channels[j].copyTo((*input_channels)[i*num_channels_[0]+j]);
}
}
}
If I understand your problem correctly, you input n
images, expecting n
pairs of (label, prob)
, but getting only one such pair.
I believe these modifications should do the trick for you:
Classifier::Predict
should return a vector< vector<float> >
, that is a vector of probabilities per input image. That is a vector
of size n
of vectors of size output_layer->channels()
:
std::vector< std::vecot<float> >
Classifier::Predict(const std::vector<cv::Mat> &input_channels,
int num_images) {
// same code here...
/* changes here: Copy the output layer to a std::vector */
Blob<float>* output_layer = net_->output_blobs()[0];
std::vector< std::vector<float> > ret;
for ( int i = 0 ; i < num_images ; i++ ) {
const float* begin = output_layer->cpu_data() + i*output_layer->channels();
const float* end = begin + output_layer->channels();
ret.push_back( std::vector<float>(begin, end) );
}
return ret;
}
In Classifier::Classify
you need to process each vector<float>
through Argmax
independantly:
std::vector< std::pair<int,float> >
Classifier::Classify(const std::vector<cv::Mat> &input_channels) {
std::vector< std::vector<float> > output = Predict(input_channels);
std::vector< std::pair<int,float> > predictions;
for ( int i = 0 ; i < output.size(); i++ ) {
std::vector<int> maxN = Argmax(output[i], 1);
int idx = maxN[0];
predictions.push_back(std::make_pair(labels_[idx], output[idx]));
}
return predictions;
}
Unfortunately, I don't believe a parallelization of network Forward passes has been implemented. However, if you'd like you could simply implement your own wrapper to repeatedly run data through copies of your network, in parallel?
Have a look at How many images can you pass to Caffe at a time?
In the linked prototxt all you have to define is
input_shape {
dim: 64 // num of images
dim: 1
dim: 28 // height
dim: 28 // width
}
The existing implementation evaluates a batch of 64 images but not necessarily in parallel. However, if running on a GPU, processing a batch of 64 will be faster than 64 single-image batches.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With