How to give the input image which is to be convolved here ?

[debanjana-in-tech]

No description provided.

[davidstutz]

You have to convert your input image to vec_t, i.e. std::vector<float>. When your are running OpenCV, you can convert your image as follows (I follow the conversion as used in mnist_parser.h):

float denominator = 255;
float min = -1;
float max = 1;
int padding = 2;
vec_t sample(3*H + 2*padding)*(W + 2*padding), min);
for (int i = 0; i < H; i++) { // Go over all rows
    for (int j = 0; j < W; j++) { // Go over all columns
        for (int c = 0; c < 3; c++) { // Go through all channels
            sample[W*H*c + W*(i + padding) + (j + padding)] = image.at<cv::Vec3b>(i, j)[c]/denominator*(max - min) + min;
        }
    }
}

As you can see, padding is applied (I am using a kernel of size 5 x 5 in the first convolutional layer). If your kernel is of size 2_k + 1 x 2_k + 1, then you should padd the image by k on each side. Further, the image is transformed to lie in [-1,1].

[edgarriba]

@WithBestWishes Does this solve your problem?
Besides, check the provided example: #187

[xuanzhaopeng]

@davidstutz Could you kindly explain why should add padding k when I use 2k+1, 2k+1 kernel in first layer? I read the example cifar10, during the training, it use padding 0.
https://github.com/tiny-dnn/tiny-dnn/blob/master/examples/cifar10/train.cpp#L69

Thanks