Wrong results using trained caffe net from c++

Question

I tried to use my trained caffe net with my data from C++. I implemented standard caffe example classification.cpp for deploy. In train/test phase with python scripts the net achieved accuracy = 0.93, but now when I went to deploy I got some strange results. I have two classes:

environment
object

and I need to get the prob of object detection. I believed that the results will be presented in the form of two probs in Softmax output blob if the net have two outputs in FC-layer (prob1 + prob2 == 1.0f), but the result is puzzling. In output vector I get two identical values for every image. Here are input and output layers:

layer {
    name: "data"
    top:  "data"
    type: "Input"
    input_param { shape: { dim: 1 dim: 3 dim: 227 dim: 227 }}
}
layer {
    name: "fc6"
    top:  "fc6"
    type: "InnerProduct"
    bottom: "drop5"
    inner_product_param {
        num_output: 2
        weight_filler {
            type: "xavier"
            std: 0.1
        }
    }
}
layer {
    name: "prob"
    top:  "prob"
    type: "Softmax"
    bottom: "fc6"
}

My C++ code sample for the regular use:

Blob<float>* input_layer = m_net->input_blobs()[0];
input_layer->Reshape(1, m_numChannels, m_inputGeometry.height, m_inputGeometry.width);
m_net->Reshape();
std::vector<cv::Mat> input_channels;
Blob<float>* input_layer = m_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;
}

cv::split(image_float, *input_channels);
m_net->Forward();
Blob<float>* output_layer = m_net->output_blobs()[0];
const float* begin = output_layer->cpu_data();
const float* end = begin + output_layer->channels();
QVector<float> output = QVector<float>(end - begin, *begin);

In addition, the results are similar to random (and duplicated for each class), the smallest probability value is magic 0.443142. This value is often found in the output vector. What am I doing wrong?

Your input data is images right? Then why are you using only fully connected layers? Convolutional layers will work much better. — Dr. Snoopy
@Matias Here I describe only the input and output layers. The network includes convolutional layers too. I think my problem is not about accuracy. — esterlein

esterlein esterlein · Accepted Answer · 2016-07-29T14:18:24

So, the problem was beyond the scope of the topic. It's about difference between STL and Qt vectors. Original code

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

instead of

QVector<float> output(end - begin, *begin);

solves the issue.

Wrong results using trained caffe net from c++

1 Answers