OpenCV Neural Network Sigmoid Output - c++

Question

I have been using OpenCV for a quite time. I decided to check its power for Machine Learning lately. So I ended up with implementing a neural network for face recognition. To summarize my strategy for face recognition :
Read images from a csv of some face database.
Roll images to a Mat array row wise.
Apply PCA for dimensionality reduction.
Use projections of PCA to train the network.
Predict the test data using the trained network.
So everything was OK until the prediction stage. I was using the max responsed output unit to classify the face. So normally OpenCV's sigmoid implementation should give values in range of -1 to 1 which is stated at the docs. 1 is the max closure to class. After I got nearly 0 accuracy I checked the output responses for each class for each test data. I was suprised with the values : 14.53, -1.7 , #IND . If sigmoid was applied, how could i get these values ? Where am i doing wrong ?
To help you understand the matter and for the ones wondering how to apply PCA and use it with NN I m sharing my code :
Reading csv:
void read_csv(const string& filename, vector& images, vector& labels, char separator = ';')
{
std::ifstream file(filename.c_str(), ifstream::in);
if (!file)
{
string error_message = "No valid input file was given, please check the given filename.";
CV_Error(1, error_message);
}
string line, path, classlabel;
while (getline(file, line))
{
stringstream liness(line);
getline(liness, path, separator);
getline(liness, classlabel);
if(!path.empty() && !classlabel.empty())
{
Mat im = imread(path, 0);
images.push_back(im);
labels.push_back(atoi(classlabel.c_str()));
}
}
}
Rolling images row by row :
Mat rollVectortoMat(const vector<Mat> &data)
{
Mat dst(static_cast<int>(data.size()), data[0].rows*data[0].cols, CV_32FC1);
for(unsigned int i = 0; i < data.size(); i++)
{
Mat image_row = data[i].clone().reshape(1,1);
Mat row_i = dst.row(i);
image_row.convertTo(row_i,CV_32FC1, 1/255.);
}
return dst;
}
Converting vector of labels to Mat of labels
Mat getLabels(const vector<int> &data,int classes = 20)
{
Mat labels(data.size(),classes,CV_32FC1);
for(int i = 0; i <data.size() ; i++)
{
int cls = data[i] - 1;
labels.at<float>(i,cls) = 1.0;
}
return labels;
}
MAIN
int main()
{
PCA pca;
vector<Mat> images_train;
vector<Mat> images_test;
vector<int> labels_train;
vector<int> labels_test;
read_csv("train1k.txt",images_train,labels_train);
read_csv("test1k.txt",images_test,labels_test);
Mat rawTrainData = rollVectortoMat(images_train);
Mat rawTestData = rollVectortoMat(images_test);
Mat trainLabels = getLabels(labels_train);
Mat testLabels = getLabels(labels_test);
int pca_size = 500;
Mat trainData(rawTrainData.rows, pca_size,rawTrainData.type());
Mat testData(rawTestData.rows,pca_size,rawTestData.type());
pca(rawTrainData,Mat(),CV_PCA_DATA_AS_ROW,pca_size);
for(int i = 0; i < rawTrainData.rows ; i++)
pca.project(rawTrainData.row(i),trainData.row(i));
for(int i = 0; i < rawTestData.rows ; i++)
pca.project(rawTestData.row(i),testData.row(i));
Mat layers = Mat(3,1,CV_32SC1);
int sz = trainData.cols ;
layers.row(0) = Scalar(sz);
layers.row(1) = Scalar(1000);
layers.row(2) = Scalar(20);
CvANN_MLP mlp;
CvANN_MLP_TrainParams params;
CvTermCriteria criteria;
criteria.max_iter = 1000;
criteria.epsilon = 0.00001f;
criteria.type = CV_TERMCRIT_ITER | CV_TERMCRIT_EPS;
params.train_method = CvANN_MLP_TrainParams::BACKPROP;
params.bp_dw_scale = 0.1f;
params.bp_moment_scale = 0.1f;
params.term_crit = criteria;
mlp.create(layers,CvANN_MLP::SIGMOID_SYM);
int i = mlp.train(trainData,trainLabels,Mat(),Mat(),params);
int t = 0, f = 0;
for(int i = 0; i < testData.rows ; i++)
{
Mat response(1,20,CV_32FC1);
Mat sample = testData.row(i);
mlp.predict(sample,response);
float max = -1000000000000.0f;
int cls = -1;
for(int j = 0 ; j < 20 ; j++)
{
float value = response.at<float>(0,j);
if(value > max)
{
max = value;
cls = j + 1;
}
}
if(cls == labels_test[i])
t++;
else
f++;
}
return 0;
}
NOTE: I used AT&T 's first 20 class for my dataset.

Answer 1

Thanks to Canberk Baci's comment I managed to overcome sigmoid output discrepancy. Problem seems to be at default parameters of mlp 's create function which takes alpha and beta 0 as default. When they both are given as 1, sigmoid function works as it was stated in the docs and neural network can predict something but with errors of course.
And for the results of Neural Network :
By modifying some parameters like momentum etc, and without any illumunation correction algorithm, I got %72 accuracy on the dataset of (randomly sampled 936 train, 262 test images ) first 20 classes of CroppedYaleB from opencv tutorials. For the other factors to increase accuracy; when I applied PCA, I directly gave the reduced dimension size as 500. This may also reduce accuracy because retained variance may be below %95 or worse. So when I have free time I will apply these to increase accuracy :
Tan Triggs illumination correction
Train PCA with 0.95 as pca size to retain %95 variance.
Modify neural network parameters (I wish we had a less parametric NN in OpenCV library)
I shared these so that someone may wonder how to increase the classification accuracy of NN. I hope it helps.
By the way you can track the issue about this here:
http://code.opencv.org/issues/3583